An operational framework to track individual farmland trees over time at national scales using PlanetScope imagery

An operational framework to track individual farmland trees over time at national scales using PlanetScope imagery | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article An operational framework to track individual farmland trees over time at national scales using PlanetScope imagery Florian Reiner, Dimitri Gominski, Rasmus Fensholt, Martin Brandt This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-4359628/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Trees outside forests, in particular on croplands, play a crucial role for food security and climate resilience in the Global South, but are threatened by increasing climatic change and human pressures. The dynamics of agroforestry systems and national cropland tree stocks are largely unknown, as currently no robust monitoring system exists to remotely detect single field trees and track changes at national scales. Here we present a framework to track cropland trees at the single tree level across multiple years, using a combination of satellite imagery, deep learning, and object-based change classification. The approach matches annual tree centre predictions to detect changes, such as individual tree losses from logging or tree mortality events. The slope between annual tree prediction confidence heatmaps is also used to detect areas of gains, with possible applications for monitoring plantation and restoration areas. The framework is designed for PlanetScope nano-satellite imagery, which offers unprecedented opportunities for detailed tree monitoring given the combined high spatial and temporal resolution. PlanetScope imagery, however, also come with a range of challenges, which are discussed and for which solutions are proposed. We demonstrate the framework by applying it to a national-scale case study of cropland trees in Tanzania from 2018 to 2022. Tree tracking Trees Outside Forests Agroforestry PlanetScope Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Figure 9 1. INTRODUCTION Agroforestry systems integrate trees in croplands and are a key component of smallholder agriculture across the Global South, providing wide-ranging benefits from both an ecological and socio-economic perspective [Jose 2009 , Bayala 2014]. Beyond yielding direct products such as fruit, fuelwood or animal fodder, the maintenance of tree cover on croplands has been shown to improve soil moisture, erosion resistance, and biodiversity, as well as contributing to mitigation of climate change in the form of direct carbon sequestration [Zomer 2016, Schnell 2015]. In dryland regions, cropland trees are also essential for providing climate resiliency, as local tree species both aid in preserving soil water content and continue to provide edible leaves and tree products during dry seasons, and long after crop failure due to drought [Quandt 2017]. Despite their numerous benefits, traditional agroforestry systems are impacted by changing climatic conditions and changing land management practices. While generally robust to water stress, increasing periods of prolonged drought can cause widespread tree die-offs due to hydraulic stress, fire and enhanced vulnerability to insect attack and fungal diseases [Choat 2018]. In addition to these direct mechanisms, cropland trees are also under significant threat of logging for fuelwood in times of drought, when other sources of combustible vegetation are scarce. Given that the intensity and frequency of climatic extremes are expected to continue rising in the next decades, there is an increasing risk of large-scale losses of cropland trees, with potential adverse effects on food security and local livelihoods in the Global South [Watts 2022]. However, the scale of such potential losses is currently not well understood, and while there have been some local studies indicating a decline in scattered tree cover [Plieninger 2009, Gibbons 2008], there is no clear picture of the long-term dynamics and stability of agroforestry systems at national and continental scales. In addition to existing agroforestry systems and community-led restoration efforts such as Farmer Managed Natural Regeneration, the increased awareness of tree-based carbon mitigation has also given rise to a variety of tree-planting initiatives aiming to increase tree cover in and along croplands or degraded woodland ecosystems. These schemes exist at multiple scales, from small-scale projects of non-governmental organisations (NGOs) working with local farmers [One Acre Fund 2022], to larger campaigns such as Rwanda’s and Ethiopia’s national agroforestry tree planting initiatives [New Times 2021 , UNEP 2019], up to trans-national initiatives such as the proposed Great Green Wall, which originally envisioned a vast tree-planting belt across the Sahel region [Goffner 2019]. While some projects are devoted primarily to improving agroforestry yields, biodiversity and ecosystem resilience, others are explicitly geared towards carbon credit schemes, with the continued growth of the global carbon market spawning a continuous rise in new tree-planting and reforestation schemes [Lefevbre 2021]. A major challenge for such projects is the lack of scalable monitoring tools to assess tree survival rates over multiple years, leading to a common criticism of ‘greenwashing’: that many of the promised carbon credits are not actually sequestered due to exaggeration of trees planted, lack of overview of tree survival, and losses of trees due to tree mortality factors [Duguma 2020]. There is thus an urgent need for developing a tool to monitor temporal changes in non-forest trees in general, and on croplands in particular, with potential users including local farmers, restoration practitioners, government ministries and the larger environmental research community. While the exact requirements of these actors will vary by use case, such that carbon mitigation schemes for example might also want to measure biomass and carbon, a first step is a simple monitoring of trees at large scale, that is able to reliably verify living trees and detect losses and potential gains. A further common requirement for nearly all users is low cost. Currently, for small projects monitoring is typically either done using normalised difference vegetation index (NDVI) trends based on coarse spatial resolution data as an integral measure of herbaceous and woody vegetation [Soares 2018, Bey 2021 ], or by repeated field visits, where the existence of single trees is manually verified over multiple years. However, at national scale the first is not reliable, and the latter is not cost-effective, and there remains a critical need for a scalable alternative, which currently does not exist. The major reason for this is the fundamental challenge in detecting cropland trees using remote sensing. Due to their scattered nature and lack of closed canopy, single agroforestry trees cannot be directly seen in medium-resolution satellite imagery. Instead, they have usually been mapped by their spectral characteristics in mixed pixels, using freely available medium resolution imagery such as Landsat [Hansen 2012, Higginbottom 2018, Venter 2018]. Additionally sub-pixel methods applied to Sentinel-1 and Sentinel-2 imagery have been used to map both forest and non-forest tree cover at near single-tree level [Zhang 2019], although detection of dynamics with these methods remains difficult [Brandt, J. 2023]. However, recent advances in machine learning methods applied to remote sensing have led to new breakthroughs in the mapping of non-forest tree cover. In particular, the combination of deep learning and high-resolution satellite imagery has allowed the mapping of individual trees at national and continental scales [Malkoç 2021, Brandt, M. 2020, Reiner 2023, Liu 2023]. This includes cropland trees, supporting the determination of accurate numbers of large cropland trees at national scales, and estimating their carbon stocks [Mugabowindekwe 2022]. Combined with an exploding amount of available high-resolution satellite data with new global coverage nano-satellite constellations such as PlanetScope, these methods now provide a unique opportunity to study the temporal dynamics of scattered cropland trees at national to global scales. Here we identify the opportunities and key challenges of monitoring cropland tree dynamics with PlanetScope data and propose several mitigation strategies to address the latter. We then integrate these methods into a framework to track individual trees over multiple years at national scale. It consists of four main components: A comprehensive satellite image download and compositing pipeline, a convolutional neural network to detect individual trees in an image, a tree matching algorithm to identify individual trees across years, and a change classification process to detect tree changes. The resulting end-to-end framework allows for accurate quantification of surviving and lost trees for a given area and temporal range and can be applied to any area due to global availability of PlanetScope satellite imagery. We demonstrate this application by a national-scale proof-of-concept case study of Tanzania, where we map individual cropland trees between 2018 and 2022. This framework may be useful as a tool for governments and environmental organisations to monitor the stability and resilience of agroforestry systems at national scale, and serve as the basis of an early-warning system for sustained losses in cropland tree cover. In addition, such a tool may be used to monitor landscape restoration projects and the planting of new cropland trees, as well as verify claims of tree numbers in large-scale plantations and carbon credit schemes. 2. MATERIALS AND METHODS 2.1 Overview We developed an end-to-end processing framework to detect changes of scattered individual trees across multiple years and large study areas, using satellite imagery. The method consists of a four-step processing flow (Fig. 1 ). First, raw satellite scenes are downloaded, quality-filtered and composited to produce yearly gap free mosaics covering the study area. Second, a deep learning framework is applied to detect individual tree crown locations with a given confidence per tree in each yearly mosaic. Thirdly, a matching algorithm is used to identify identical tree crowns in subsequent years, designed to be robust to small changes in detected location. Finally, a change classification scheme is applied, which takes into account yearly confidences and possible missed detections, to provide a final change class and change confidence. The overall framework is designed to be versatile in both temporal and spatial scope, able to provide output tree change maps for any study area and target input years. The detection of tree-level changes from PlanetScope satellite imagery faces a number of practical challenges, primarily related to inconsistencies in the way trees can be detected in the satellite images. Table 1 provides an overview of these challenges, and how we have addressed them, with further detail in the following sections. Table 1 Summary of challenges. A summary of the challenges faced in detecting individual tree changes using Planetscope data, along with the mitigation strategy that was used to address them. Challenge Mitigation Cost-effective availability of imagery at large spatial and high temporal scale Use of PlanetScope imagery. Alternatives include RapidEye, EarthDaily or Sentinel-1&2 for large trees or groups of trees Confusion of tree canopy infrared signal with high grasses and crops Scene selection from a narrow time window based on local vegetation phenology derived from ancillary satellite remote sensing data Incomplete scene coverage Progressive relaxation of filter criteria Inconsistent sharpness and scene quality, unreliable metadata Automated scene sharpness assessment based on blur kernel estimation [Anger 2019] Detection of young or very small crowns at single pixel size Selection of scenes with low solar altitude angle to detect the shadows of the small crowns High effort of hand-labelling a sufficient number of tree crown outlines for training Labelling and detection of tree crown centres, with trees modelled as Gaussian kernels Separation of clumped trees Extraction of local maxima from merged Gaussians in the predicted confidence heatmap Spatial shifts of detected trees between subsequent scene acquisitions Deploying a nearest-neighbour tree matching algorithm Variance in yearly prediction confidence Confidence-based weighted change detection Inconsistent detection of young trees Filtering by confidence thresholds Difficulty in reliably mapping gains Confidence slope maps to reveal local trends 2.2 Preparation of imagery Choice of imagery The primary component of a national-scale tree tracking framework is the underlying source of imagery, for which the main requirements are high spatial resolution, high temporal availability and feasible cost. Ideally, to detect single trees and small trees as individuals, imagery of the highest resolution possible should be chosen, but in practice this is not possible due to logistical and financial constraints. Amongst remote sensing techniques, UAV photography has the highest resolution (up to < 10 cm), but does not scale to national level. Sub-metre resolution aerial imagery has been successfully used to map trees at national scale [Malkoç 2021, Li 2022, Mugabowindekwe 2022], but this is still cost-prohibitive for most larger countries in the Global South, especially considering the need for multi-temporal acquisitions. Another option is globally available sub-metre satellite imagery such as Worldview-3 and Pleiades, however this is sold commercially with a high cost per km 2 , thus also rendering it prohibitively expensive to acquire for multiple years at national scale. In contrast, there is freely available public satellite imagery such as Landsat and Sentinel-1 & 2, yet at 30 m and 10 m respectively their resolution is insufficient to robustly detect single trees of various sizes. One imagery source that meets most of the requirements is the PlanetScope constellation, which provides near-daily 3 m resolution imagery globally at a lower cost compared to commercial sub-metre images, and is currently freely available in the tropics for non-commercial purposes through Norway’s International Climate and Forest Initiative [Planet 2021, NICFI 2021]. Compared to sub-metre imagery previously used in large-scale tree detection studies [Brandt, M. 2020], the 3 m resolution of PlanetScope is considerably lower, and will result in a limit to the minimum size of detected trees. Nonetheless, recent studies have demonstrated that mapping of single trees > 10–20 m 2 crown size is possible at national and continental scale with PlanetScope [Reiner 2023, Brandt, M. 2023]. Considering the combination of high temporal availability, global coverage and sufficient high spatial resolution, this is currently the most promising imagery source for a national tree tracking scheme, despite drawbacks such as the relatively lower spatial resolution that misses small trees and shrubs, and variance in image quality which are discussed below. We have developed the framework presented here to use primarily PlanetScope, however the core processing steps of tree detection, identification and change classification are agnostic of the underlying imagery, and it could thus easily be used with higher resolution imagery where available, or lower resolution imagery such as Sentinel-2, at the cost of an increase in minimum detected tree size. Acquisition of PlanetScope scenes The PlanetScope constellation consists of 180 ‘Dove’ nanosatellites providing daily coverage of the global land area at 3 m resolution [Planet 2021]. The constellation includes three generations of satellites: Dove Classic, Dove-R and SuperDove, where Dove and Dove-R provide 4-Band RGB/NIR imagery, and the newer SuperDoves provide 8 spectral bands. The Dove nanosatellites were launched in multiple rounds starting from 2014, with full constellation deployment achieved by 2018. As of 2023 this provides 5 years of near daily global coverage, resulting in a very large number of raw scenes that can be queried and downloaded from the Planet application programming interface (API). While there are images available at approximately daily rate, such a short revisit time is not meaningful for a tree tracking application. Instead, we use the archive of available scenes to create one high quality composite mosaic per year, and then track changes between these mosaics. For the processing framework, any number of years can be specified, and for the case study presented here we used the 5 years from 2018–2022. Planet Labs already provide annual and monthly base maps, which are composed from many raw scenes and harmonised to a smooth mosaic consistent in spectral distribution. However the use of these mosaics excludes the possibility to select optimal scenes, based on specific metadata criteria that have a large impact on the visibility of single tree crowns, such as the seasonal time of acquisition, the sun position at time of acquisition and the satellite view angle. The sun position in particular has a large effect, as a low solar elevation causes long tree shadows, which provide essential visual context to the machine learning model for detecting the tree, especially for young trees with a small crown size near or even below the native resolution. To create the annual mosaics, individual scenes are downloaded from the Planet API, using the PSScene item type and the ortho_analytic_4b_sr asset, which is an atmospherically corrected surface reflectance product. The Planet API provides functionality to query all available scenes for a given area of interest, and returns a set of scene metadata such as ground sampling distance (GSD), sun angle, and view angle, as well as image quality indicators such as percent cloud, percent haze and percent shadow. To obtain images best suited for tree detection, the following initial query filters were used on the PSScene metadata: gsd < = 4.0, sun_elevation 95, clear_confidence > = 99, heavy_haze_percent < = 0, light_haze_percent < = 0, shadow_percent < = 0, snow_ice_percent < = 0, quality_category = ‘standard’. However, beyond these quality metrics and the sun angle, one of the main factors for successful tree detection is the phenological stage of local vegetation at the time of scene acquisition. Due to the differing phenology of grasses, crops, shrubs and trees over an annual growing cycle, there is an optimal time window in which the near-infrared signal of tree foliage is most clearly discernible from the background, and not confused with near-infrared from other vegetation such as crops or grasses (Fig. 2 a,b). This phenological time window with optimal pheno-spectral signatures is unique to local conditions, and also differs substantially between areas of deciduous and non-deciduous trees. Therefore, to select imagery for optimal tree detection, the framework divides the study areas into a grid of 1x1 degree tiles, and uses a different yearly target window for each tile, based on the local plant phenology. For each tile, the MODIS MOD09 surface reflectance product is used to determine the average time of key stages such as senescence, greendown, and dormancy [DiMiceli 2015]. For areas with evergreen woody plants dominating (> 90%), the target window is then taken as between mid-greendown and dormancy, when grasses have lost most of their near-infrared signal. For areas with deciduous woody plants dominating, an earlier window between senescence and mid-greendown is chosen, where grasses have passed their peak productivity, but trees still have full foliage (Fig. 2 c). The status of evergreen vs deciduous trees is computed for each tile, based on the average forest types from the ESA WorldCover land cover product [Zanaga 2020]. Initially, the framework queries the API for all scenes within this phenological window matching quality filters (Fig. 3 a). However, the target window length is different for each tile, and depending on local atmospheric conditions and cloud prevalence there are not always sufficient scenes available for creating a high quality cloud-free image mosaic. In this case a progressive retry algorithm is used, in which both sides of the temporal window are gradually extended to fill remaining gaps with additional scenes adjacent to the core target window, up to a maximum of 60 days. Where gaps remain, a second progressive loop is then used to gradually lower the minimum permissible metadata metric of ‘visible_confidence_percent’ down to a minimum of 60. Once all available scenes are found, the selected scenes are clipped to create a single seamless and non-overlapping coverage of the tile, stored as a footprint file and then ordered and downloaded in parallel (Fig. 3 b). The scenes are clipped before ordering to reduce quota use, as the Planet API quota is measured by actual area downloaded, not total area of scenes accessed. Automated quality filtering of imagery When downloading large numbers of raw scenes, it becomes clear that the PlanetScope scenes exhibit a certain variance in image quality and sharpness, even after applying the most restrictive metadata filters. In particular, there are occasional scenes suffering blurriness, graininess and band mis-alignment, which can be related to sensor problems, focus issues or unreliable metadata. In addition, the different generations of PlanetScope satellites have different sensor characteristics and were launched into different orbits, such that the older Dove Classic images actually have the lowest Ground Sample Distance (GSD) and thus sharpest images. The upgraded Dove-R satellites have a similar orbit, however both Dove Classic and Dove-R are being retired and are no longer available after April 2022. While the newer SuperDoves have substantially better radiometric performance, they are at a higher orbit and therefore have a higher GSD and lower resolution, which is the critical metric for detecting single trees. However, over time their orbit decays and GSD decreases again, such that recent SuperDove images are significantly sharper than the first batch from 2019. In a tree tracking context, this variance in sharpness is challenging, as a single blurry scene will directly result in missed tree detections for that year, which may cause false detection of disappeared trees. In other cases, very grainy images may also cause overestimation of trees, which can also lead to overestimation of disappeared trees as these falsely detected trees are then not detected in subsequent years. One strategy to reduce this variance would be to smooth merged mosaic to the lowest common sharpness, however this is not helpful in a tree tracking context where sharpness is critical. Instead, to mitigate these issues two approaches are proposed: Firstly, in the years where lower orbit Dove Classic and Dove-R scenes are available (2018–2021), only these scenes are selected, and then from 2022 onwards when they are unavailable only the lowest GSD SuperDove scenes are selected. Secondly, after downloading, an automated quality filtering is applied to detect blurry scenes using a kernel-based sharpness estimation, blacklist them, and redownload new replacement scenes until the entire mosaic is sharp. Sharpness is estimated using the blur kernel method developed by Anger et al. [Anger 2019], using a minimum threshold of 0.23 for the L2 norm sharpness score. To avoid misclassification, a mean score over 20 sub-windows of 512x512 pixels of the scene is used. Additionally scenes are whitelisted that are very small, or that are dominated (> 50%) by bare soil, water or wetlands according to the WorldCover land cover product, as they were found to produce falsely low sharpness scores [Zanaga 2020]. In a second step graininess is estimated as a scaled function of the mean standard deviation of all 4 bands aggregated across 20 sub samples, and a maximum threshold of 0.35 is used to exclude grainy scenes. Figure 4 shows examples of scenes blacklisted due to low sharpness or high graininess scores. As quality filtering and blacklisting only happens post-download, the entire tile filling process is done using an iterative approach where scenes are downloaded and checked in batches until the entire mosaic is filled. Each scene order requires about 10–30 minutes to be fulfilled by the API as stored raw scene assets are retrieved and prepared for download. Due to the iterative algorithm, which may require multiple rounds of ordering, it can therefore take more than 2 hours to complete a mosaic tile, but many tiles can be run in parallel to increase download speed. Image mosaicking into yearly tiles After downloading all scenes for a yearly tile, the scenes are composited into a mosaic. The raw scenes are kept in storage on disk in their original projection and lossless compression, but all mosaics are reprojected to a single configurable geographic projection, and can be compressed in a lossy format to reduce storage needs. By default EPSG:4326 is used as a global projection, and JP2OpenJPEG compression with quality 80 is applied as this was found to be near visually lossless and not to have any impact on prediction quality. This step also reduces the storage size per 1x1 degree tile from approximately 7 GB for the raw scenes to approximately 3 GB for the mosaic. Finally, to reduce seamlines between scenes, a histogram matching algorithm is applied using Landsat reference images for each grid tile [U.S. Geological Survey, 2022]. The Landsat reference images were produced as a temporal composite across 10 years, using Landsat images chosen from the same phenological time window per grid tile, to obtain a reference image that is spectrally stable across large areas and represents the typical histogram at that time of year. During histogram matching the surface reflectance of each PlanetScope scene is matched to the respective area in the Landsat image, resulting in a smooth final PlanetScope mosaic (Fig. 3 d). 2.3 Detection of individual tree crowns Crown centre based deep learning method To detect individual cropland trees, the framework uses a deep learning method with a convolutional neural network based on the UNet architecture [Ronneberger 2015]. Instead of direct tree crown segmentation as employed by Reiner et al, [Reiner 2023], the tree tracking framework uses a centre-based object detection method, where each tree crown is modelled by a Gaussian distribution around its centre during training [Luo 2021, Ventura 2022]. The CNN is first trained with PlanetScope images and the raster of crown centre Gaussians, and then applied to unseen PlanetScope images, where it predicts a heatmap that represents the confidence of a tree centre being present in each pixel. From this the local maxima are then extracted into a map of individual centres. This majorly reduces labelling time as trees can be labelled by their centre alone, and enables the tracking of specific tree instances over time. It also does not require any information about or delineation of the crown area during training, and improves the separation of individual trees when their crowns are connected. Training To train the model a large number of ground truth labels of tree centres are required. These can either be labelled by hand or extracted from LiDAR canopy height datasets. Compared to labelling of polygon tree canopy crowns, hand-labelling of centres is relatively quickly done by simple visual inspection on the source imagery, and utilising additional overlaying high-resolution imagery such as Google Earth. Alternatively, if model robustness across a wide range of ecosystems and local conditions requires a very large number of labels, the framework also supports extraction of tree centres from high-resolution canopy height models from aerial or UAV LiDAR. In a first step the raw LiDAR point clouds are converted to rasterised digital height maps, and then, after subtracting terrain, to canopy height maps. Subsequently, the local maxima of the height maps are used to separate individual tree crowns and extract their centres, based on a minimum distance parameter. The labelled tree centre points are then converted to a raster map of Gaussian distributions to be passed to the CNN during training (Fig. 5 c). Each tree is modelled as a Gaussian kernel around the centre with a fixed initial size. Ideally, the kernel size would match the real crown size. However, in the case of hand labelling of points this information is not available, and the sizes potentially extracted from LiDAR canopy height maps are also not reliable due to spatial and temporal mismatches between the LiDAR acquisition and the PlanetScope training image. Instead, to account for the variance of real crown sizes, an adaptive kernel method is used based on Luo et al, where the size of the Gaussian is continuously adapted during training with a dynamic scale map [Luo 2021]. We modify this method to draw exact Gaussian kernels on-the-fly, instead of their linear approximation, and furthermore only allow scale factors > 1. For PlanetScope, the standard deviation of the base Gaussian kernel size is set to 6.5 m, but this can be adapted to any size if other imagery sources are used. Prediction For prediction, each mosaic is split into patches of configurable size, typically 1024x1024 pixels, and predicted batchwise, with batch size determined by available video memory of the processing GPU. Reading, predicting and writing is done with queue-based multiprocessing to maximise GPU utilization, and the number of reader, predictors and writers is configurable to reduce bottlenecks due to disk or network speeds. The output of the prediction is a heatmap which represents the confidence of tree crown presence (Fig. 6 b). The local maxima of this heatmap are then extracted into a vector file of detected tree centres (Fig. 6 c), which includes as an attribute the confidence of the peak pixel, henceforth referred to as the tree’s confidence. An estimate of the tree crown area is stored, derived from the intensity and shape of the heatmap Gaussian surrounding the local maximum. Additionally, assuming that individual Gaussians tend to align with tree cover, we can threshold the heatmap with a fixed value (in practice 0.4 was found to give good results) to extract tree cover (Fig. 6 d). To improve robustness, the framework includes an ensemble prediction approach where multiple models can be trained and selected on different subsets of the training data. During prediction, images are then predicted with each individual model, and these output heatmaps are aggregated with a per-pixel mean into the final prediction. 2.4 Change detection frameworks The tree detection model provides two outputs for each year: A confidence heatmap raster layer, and a vector file with the extracted yearly tree centre positions, reflecting local confidence maxima. We then introduce two different methods for detecting tree changes based on these intermediate outputs. In the first method, the slope of the confidence map is analysed to produce a raster heatmap of confidence trend, which can be used to track growth for plantations or other areas where trees cannot be mapped as individuals, while the second aims to classify changes of individual trees, based on the yearly individual tree locations and confidences. Visualisation of confidence slope In effect, the heatmaps represent the CNN model’s confidence that a given pixel forms the centre of a tree crown. In addition to image quality, this confidence is directly affected by tree crown size, tree age and tree health. Therefore, by analysing changes in this confidence over time, trends in these underlying factors can be revealed. To this end, the framework merges the yearly heatmaps into a 3D spatio-temporal array, and then computes a linear regression across the time dimension, resulting in a map of the temporal slope of tree centre confidence per pixel. This map is useful in identifying areas where trees have increased in crown size, or new trees grown (positive slope), or areas where trees may have disappeared, been trimmed, or weakened due to fires or disease (negative slope). One advantage of this method is that the entire predicted area is considered, including all types of vegetation, not only single trees of a sufficient size to be detected with a high confidence. In general this method may be particularly suited to detecting gains, both of new individual trees, or of increasing canopy cover of shrublands (referred to as bush encroachment), woodlands or tree plantations. Figure 7 shows an example of an area with increasing tree cover from 2018–2022 in Tanzania, which is clearly identified from the confidence slope map. Individual tree change classification The individual tree change detection framework is based on predicting tree centre positions for each year, and then combining these yearly predictions to detect changes of individual trees such as disappeared trees. In a first step, tree detections are filtered by landcover and confidence. Next, a matching algorithm is applied to identify common trees across years, which is designed to be robust to slight changes in detected position. Finally, the presence, absence, and confidence of the yearly detections are supplied to a weighted classification algorithm to determine a final output change class per tree. After predicting each mosaic tile, the predictions are filtered by land cover, such that only trees on croplands are included in the further analysis. This is done by masking with an external land cover product, such as the ‘cropland’ class in ESA’s WorldCover [Zanaga 2020]. Significant uncertainty is introduced at this stage, as any misclassifications in the land cover product lead to inclusion of non-cropland areas for which the model was not trained, such as shrublands or dense woodlands. Another source of uncertainty stems from the detection of newly planted and small trees, which may be detected in some years but not others, due to their size being at the limit of detectability, given the spatial resolution of the data. When detected, these trees often have a very low confidence. To reduce uncertainty and false classifications from these small trees, another filtering step is therefore added to exclude them from the classification algorithm by first filtering with a minimum confidence threshold. For PlanetScope in Tanzania we used a threshold of 0.35. Identification of individual trees across years When overlaying multiple yearly predictions, it is clear that the predicted position of a given individual tree is not consistent across time, often exhibiting significant spatial shifts up to 10 m. This is due to multiple reasons both at the imagery and CNN level, including differences in scene conditions such as satellite view angle, sun angle and thus shadow length, phenological conditions, and scene sharpness, as well as uncertainty in the Gaussian kernel distribution and uncertainty in the extraction of local maxima. To identify specific trees, we therefore employ a nearest-neighbour approach, in which the closest tree centre in a subsequent year is considered the same tree, up to a certain buffer distance (15 m for PlanetScope). This is done with an iterative method. For each tree of the first year’s prediction, the nearest tree within the buffer of the second year is considered the same tree. The position of this tree is then updated to be the mean of these two years, before repeating the same procedure for the next year, and updating the position to the mean of all previous years (Fig. 8 f). The nearest neighbour is determined using a spatial index query which can be computationally heavy for tiles with millions of trees. Therefore, each tile is split into a configurable number of smaller pieces, which are processed in parallel and remerged. After processing all years sequentially, a tabular database is created including all trees detected in any of the years above a certain confidence threshold, their position for each year (or NULL), their confidence for each year (or NULL), and the mean position from all years. Classification of tree-level change The aim of the tree-level classification is to convert the tabular database of matched yearly tree detections into a change classification where each tree is allocated into one of four final classes (Table 2 ). Due to differences in image quality, it is not possible to reliably identify the exact year of change for the ‘disappeared’ or ‘gain’ classes, as some years without prediction may simply be caused by lower image sharpness. Table 2 Output tree change classes. Code Name Description 1 Remaining This is a remaining tree, detected in multiple years and with a high confidence. 2 Misdetection Only detected in one year or with very low confidence. Hence, this is not a tree, or too young to be reliably classified. 3 Disappeared Detected with high confidence in at least one year, and then not predicted for all subsequent years. This tree has disappeared. 4 Gain Detected as a tree in multiple later years, but not predicted for all previous years. This may be a new tree. The tabular output of the tree-matching algorithm leads to a sequence of binary tree-no tree detections for each year. In principle, a simple change typology could be applied, where each case of possible tree-no tree sequences is mapped to an output change class. However, in practice this can easily lead to an overestimation of disappeared trees, as single yearly missing trees are often also due to image quality issues. Instead, a confidence-based method is used, in which the confidence of each annual prediction is taken into account. In a first step, the yearly confidences are multiplicatively combined to determine the confidence of a given binary sequence ‘scenario’ of yearly tree presence, such as ‘11111’ (tree detected each year) or ‘11100’ (tree detected three years, then not detected). Here each of these possible sequence scenarios is associated with the four final classes with a different weight factor, where the weights are manually chosen, such that for example sequence ‘11111’ corresponds to weights [1, 0, 0, 0] for the classes [remaining, misdetection, disappeared, gain]. In a second step, the confidences of each sequence scenario are then multiplied with the scenario-class weights, and these products are summed per class to obtain the final output class, and output confidence (Supplementary Table 1). The final output consists of the change class with the highest confidence, the value of which provides a ‘change confidence’ measure reflecting the certainty of the classification. After processing is completed, the framework saves all detected trees to disk as a geopackage file, with attributes including the position, change class, change confidence, confidence in each year and detected position in each year. The weights of how much each scenario contributes to the final change class can be manually tuned to adapt to the imagery and number of years used. For our PlanetScope scenario, it was found that detection of gains is not very reliable due to the low likelihood of suddenly detecting new young trees with 3 m resolution spatial imagery, and thus the weights of the gain class were set to 0 for nearly all change sequence scenarios. Instead, we used the confidence slope method described above to identify areas with gains (Fig. 7 ). Finally, the output classifications can be filtered by the final ‘change confidence’ described above, such that any uncertain changes below a minimum confidence threshold are excluded. For PlanetScope in Tanzania, we used a threshold of 0.8. Using such a high threshold implies that the observed changes are reliable but also that a larger number of actual changes may not be reported. To validate the output classifications, multitemporal ground truth data is required, which specifies the true change class for a subset of detections. As it can be quite challenging to acquire repeated yearly physical field data of single tree positions, this validation data can also be obtained by manually inspecting auxiliary imagery, such as historical Google Earth imagery. For each validation point, the state of remaining or disappeared tree is then visually confirmed for the first and last year of the sequence, resulting in the ground truth change class. Subsequently these are compared with the classified output change classes to determine the model’s overall accuracy per change class. 3. RESULTS We developed an end-to-end framework for tracking tree-level changes at national scale. The framework consists of a toolchain to download and preprocess satellite imagery, automatically identify single cropland tree crowns at scale, match detections from sequential annual detections to the same trees, and classify detected trees into remaining, gained or disappeared trees. The method is versatile and can be used for any number of years, any study area, and in principle any type of imagery. Tanzania case study To demonstrate the capabilities of this framework, we applied it to a proof-of-concept study case to track all cropland trees at national scale for a 5 year period from 2018–2022. We chose the country of Tanzania, with a land area of 945 000 km 2 , of which 188 000 km 2 is classified as cropland by the WorldCover land cover map [Zanaga 2020]. We employed a previously trained model developed for cropland trees in India [Brandt, M. 2023] to train a model for tree detection in Tanzania. We then used the framework to prepare PlanetScope imagery covering all Tanzania for the 5 target years, and then applied the model to produce annual tree centre predictions. The imagery pipeline downloaded 17 800 raw scenes with a total size of 4.1 TB, and generated 1.5 TB of mosaicked images across 125 tiles (Fig. 9 b). Image downloading, compositing and preprocessing for one year was completed in 48 hours with a 32-core CPU, and prediction of the entire country for one year was completed in approximately 10 hours on a single RTX3090 GPU. A total of 67,900,000 trees with a confidence > 0.35 were detected on croplands in at least one year, with an average tree density of 3.4 trees per ha across all cropland areas. After applying the change classification, we found 110,500 trees classified as ‘disappeared’ between 2018 and 2022 with a change confidence > 0.8, or approximately 0.16% of the total in 2018 if the prediction from 2018 is taken as a baseline. Aggregated output maps Beyond the prediction raster layers and tree databases per mosaic tile, the framework also generated several national-scale output maps, aggregated to a grid of 1 km. These include: Annual maps of percent tree crown cover, derived from the confidence heatmaps (Fig. 10e), annual density maps of average tree count per hectare (Fig. 10f), a density map of average count of trees classified as ‘disappeared’ between 2018 and 2022 with a change confidence > 0.8 (Fig. 10g), a slope map of the differences in prediction confidence between 2018 and 2022 (Fig. 10h), and a confidence map of the mean change confidence per 1 km pixel (Fig. 10i). The latter provides a measure of uncertainty for the change detection, reflecting the varying regional uncertainties of detected changes, which arise due to differences in image quality. Evaluation of change classification To evaluate the accuracy of reported changes, we manually inspected 500 randomly selected trees classified as ‘disappeared’ with a change confidence > 0.8, and used all imagery years plus auxiliary imagery such as Google Earth to create ground truth labels for each trees. We found a false detection of disappeared trees of 17.02%, for an accuracy of 82.98%. 4. DISCUSSION Agroforestry systems in the Global South are a cornerstone in providing food security for smallholder farmers, and are key to improving climate resiliency. However, these cropland trees are increasingly threatened by climatic changes and human pressures in the form of changing land management and unsustainable logging. Conversely, localised increases in cropland tree cover have been observed as the result of community-led restoration practices such as Farmer Managed Natural Regeneration, notably in Niger, Mali and Ethiopia [Haglund 2011, Reij 2016 ]. Additionally various ecosystem restoration and carbon mitigation schemes have been planting trees across the Global South, including some aiming to sell carbon credits, yet reliable data on survival rates and areas planted is often lacking [Duguma 2020, Reynolds 2012 ]. As a result, the overall extent of cropland tree dynamics is not clear, and there remains a critical need for a means to systematically assess agroforestry stability and cropland tree losses or gains at regional or national scales. The framework presented here is an end-to-end processing toolchain to provide such monitoring of cropland trees. It facilitates the monitoring of trees for any study area and any temporal range subject to image availability, and for a wide range of users including environmental NGOs, restoration practitioners, government ministries, carbon credit schemes and other tree-planting initiatives. Despite the support for any imagery source in principle, we currently see PlanetScope imagery as the best compromise for national scale studies, due to the combined requirements of large study areas, temporal availability, high spatial resolution, and low cost. Other possible imagery sources include RapidEye (5 m resolution), the upcoming Earth Daily constellation (5 m), or Sentinel-1 & 2 (10 m), although a loss of detail for smaller trees is associated with the latter. Yet the use of PlanetScope data comes with its own set of challenges. At a spatial resolution of 3 m, PlanetScope images are of sufficient resolution to detect single trees down to a crown size of approximately 5 m 2 , but only if the image is at its optimal quality. Generally, tree crowns > 30 m 2 can almost always be well detected [Reiner 2023]. However, the differences in observed image quality between scenes can result in different detectability thresholds for trees < 50 m 2 crown size, caused by reductions in the effective spatial resolution due to atmospheric conditions, sensor issues and varying satellite orbital height. Despite local quality checks implemented to remove the worst scenes, the issue of varying GSD and sharpness within the constellation remains, resulting in the detection of fewer trees in less sharp scenes. This is a major problem in a monitoring context as it contributes to false changes when trees are mistakenly classified as lost due to image quality issues. To alleviate this problem we developed the confidence-based method, which provides robustness against missing predictions in single years. However, this method currently requires manual weights for different classes, whereas ideally these weights should be determined automatically from a trained machine learning model. Additionally there are limits to this approach as the number of low confidence predictions increases, and in the case where the missing prediction is in the last year. These challenges can be mitigated to a certain degree by the inclusion of more temporal predictions to reduce the noise from missed predictions, either by the use of additional years (2017, 2023, 2024...), or by using multiple images per year. Another avenue would be to create yearly mosaics as temporal composites of many scenes, although there are challenges in stacking pixels at such high resolution due to apparent shifts in tree positions caused by differing georectification and the variance in view angles and sun angles in subsequent scenes. Ultimately, if perfect accuracy is required, a different image source such as sub-metre resolution aerial or UAV imagery should be chosen, although this would only be feasible for smaller study areas. Another limitation of the use of 3 m resolution data is the ability to detect tree gains. While the disappearance of a large tree due to logging produces an immediate and stark difference in the near-infrared signal, the image change due to a growing tree is gradual, starting from an undetectable sapling and moving to a single pixel crown without shadow before developing into a clearly recognizable crown with shadow. During this phase there are likely to be periods in which the young tree moves above and below the detection threshold in a PlanetScope image, due to differences in scene sharpness and satellite view angle. This would introduce considerable uncertainty to the change detection, and lead to confusion between ‘gains’ and ‘disappeared’ depending on the final year of the sequence. For mapping tree gains, it may therefore be better to choose only two years, with as large a time interval in-between as possible, e.g. change over a 10-year period. There is much scope for further work on the specific case of plantation monitoring with PlanetScope imagery, and it is likely that the geometric structure of plantations can aid machine learning models to detect saplings before their crowns become clearly visible. Furthermore, one additional component of this framework that is currently not considered, is the integration of biomass modelling at single tree level, which is key to quantifying the carbon impacts of removed trees at country scale. Allometric approaches have previously been used to map individual tree biomass and carbon at national level [Mugabowindekwe 2022], but these typically use imagery at < = 50cm resolution and exact segmentation of tree crowns. For this framework, while thresholding of the confidence heatmap is used to obtain tree cover, the resulting tree crown areas are not yet suitable for direct allometric analysis, due to a large uncertainty in crown area stemming from the variance in image quality, the Gaussian modelling of the centre, and the confidence threshold used. Despite these limitations, the presented framework offers a valuable new approach to the study of cropland tree dynamics and presents a pathway to long-term monitoring of agroforestry systems at single tree level. A recent application of the framework in India has led to the discovery of a large decline in agroforestry trees over the last 10 years [Brandt, M. 2023]. Such applications at national or continental scale allow policymakers across the public, private and NGO sectors to gain quantitative evidence on both the impacts of climate change and the results of land management policies on the condition and development of national agroforestry environments. At smaller scales, the low cost of PlanetScope imagery makes it feasible to acquire multi-year imagery for specific restoration projects by non-state users. Beyond agroforestry monitoring, the automated detection of tree survival and disappeared single trees may also contribute to the evaluation, verification and impact assessment of large-scale tree-planting projects, including carbon credit schemes and national ecosystem restoration projects. Declarations Acknowledgements FR and MB are supported by the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 947757 TOFDRY). MB also acknowledges funding from a DFF Sapere Aude grant (no. 9064–00049B). R.F. acknowledges support by the Villum Foundation through the project Deep Learning and Remote Sensing for Unlocking Global Ecosystem Resource Dynamics (DeReEco, grant no. 34306). Author contributions FR prepared the data, conducted the analyses, designed the figures, and wrote the manuscript. MB designed the study, sampled the training data and developed the previously trained model used. FR wrote the codes for the data preparation and data analyses, supported by XT. DG wrote the codes for the tree detection framework. MB, DG and RF reviewed the manuscript. Data accessibility Planetscope imagery was available through Norway's International Climate and Forest Initiative (NICFI) satellite data Level 2 programme. NICFI Planetscope imagery in tropical areas is available for non-commercial purposes from Planet Labs at https://www.planet.com/nicfi/. However, we did not use the basemaps provided in the frame of the NICFI programme but generated our own mosaics from the raw data. The derived tree cover maps produced in this study will be deposited in a Zenodo database, available at XX. Conflicts of interest The authors declare no conflicts of interest and no competing financial interests. References Anger, J., de Franchis, C. & Facciolo, G. Assessing the Sharpness of Satellite Images: Study of the Planetscope Constellation. 2019 IEEE International Geoscience and Remote Sensing Symposium 389–392 (2019). doi:10.1109/IGARSS.2019.8900526. Bayala, J., Sanou, J., Teklehaimanot, Z., Kalinganire, A. & Ouédraogo, S. J. Parklands for buffering climate risk and sustaining agricultural production in the Sahel of West Africa. Curr. Opin. Environ. Sustain. 6 , 28–34 (2014). Bey, A. & Meyfroidt, P. Improved land monitoring to assess large-scale tree plantation expansion and trajectories in Northern Mozambique. Environ. Res. Commun . 3 , 115009 (2021). Brandt, J., Ertel, J., Spore, J. & Stolle, F. Wall-to-wall mapping of tree extent in the tropics with Sentinel-1 and Sentinel-2. Remote Sens. Environ. 292, 113574 (2023). Brandt, M. et al. An unexpectedly large count of trees in the West African Sahara and Sahel. Nature 587 , 78–82 (2020). Brandt, M., Gominski, D., Reiner, F. et al. Severe decline in large agroforestry trees in India over the past decade. Preprint available at: https://doi.org/10.21203/rs.3.rs-3340055/v1 (2023). DiMiceli, C., M. Carroll, R. Sohlberg, D. Kim, M. Kelly, J. Townshend. MOD44B MODIS/Terra Vegetation Continuous Fields Yearly L3 Global 250m SIN Grid V006. Duguma, L. A., Minang, P. A., Aynekulu, B. E., Carsan, S., Nzyoka, J., Bah, A., Jamnadass, R. H. From Tree Planting to Tree Growing: Rethinking Ecosystem Restoration Through Tree. ICRAF Working Paper No 304. doi: https://doi.org/10.5716/WP20001.PDF (2020). Goffner, D., Sinare, H. & Gordon, L. J. The Great Green Wall for the Sahara and the Sahel Initiative as an opportunity to enhance resilience in Sahelian landscapes and livelihoods. Reg. Environ. Chang. 19 , 1417–1428 (2019). Haglund, E., Ndjeunga, J., Snook, L. & Pasternak, D. Dry land tree management for improved household livelihoods: Farmer managed natural regeneration in Niger. J. Environ. Manage. 92 , 1696–1705 (2011). Higginbottom, T. P., Symeonakis, E., Meyer, H. & van der Linden, S. Mapping fractional woody cover in semi-arid savannahs using multi-seasonal composites from Landsat data. ISPRS J. Photogramm. Remote Sens. 139 , 88–102 (2018). Gibbons P, Lindenmayer DB, Fischer J, et al. The future of scattered trees in agricultural landscapes. Conserv Biol . 2008 , 22(5), 1309-1319 (2022). Jose, S. Agroforestry for ecosystem services and environmental benefits: an overview. Agroforest Syst 76 , 1–10 (2009). Lefebvre, D. et al. Assessing the carbon capture potential of a reforestation project. Sci. Reports 11 , 1–10 (2021). Liu, S. et al. The overlooked contribution of trees outside forests to tree cover and woody biomass across Europe. Science Advances 9, 37 (2023). Luo, Z. et al. Rethinking the Heatmap Regression for Bottom-up Human Pose Estimation. in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 13259–13268 (2021). doi:10.1109/CVPR46437.2021.01306. Malkoç, E., Rüetschi, M., Ginzler, C. & Waser, L. T. Countrywide mapping of trees outside forests based on remote sensing data in Switzerland. Int. J. Appl. Earth Obs. Geoinf. 100 , 102336 (2021). Mugabowindekwe, M. et al. Nation-wide mapping of tree-level aboveground carbon stocks in Rwanda. Nat. Clim. Chang. 2022 , 1–7 (2022). New Times, The; 2021. “Rwanda to plant over 43 million trees in new campaign” (accessed September 2023). https://www.newtimes.co.rw/article/189318/News/rwanda-to-plant-over-43-million-trees-in-new-campaign Norway’s International Climate And Forest Initiative (NICFI), 2021. (accessed May 2022). https://www.nicfi.no/. One Acre Fund, 2022. “Areas of Focus: Trees” (accessed September 2023). https://oneacrefund.org/what-we-do/areas-focus/trees Planet Labs PBC, 2022. PlanetScope Product Specifications (accessed May 2022). https://assets.planet.com/docs/Planet_PSScene_Imagery_Product_Spec_June_2021.pdf. Plieninger T, Levers C, Mantel M, Costa A, Schaich H, Kuemmerle T. Patterns and drivers of scattered tree loss in agricultural landscapes: orchard meadows in Germany (1968-2009). PLoS One . 10 (5): (2015) Quandt, A., Neufeldt, H., & McCabe, J. T. The role of agroforestry in building livelihood resilience to floods and drought in semiarid Kenya. Ecology and Society , 22 (3). (2017). Reij, C. & Garrity, D. Scaling up farmer-managed natural regeneration in Africa to restore degraded landscapes. Biotropica 48 , 834–843 (2016). Reiner, F., Brandt, M., Tong, X. et al. More than one quarter of Africa’s tree cover is found outside areas previously classified as forest. Nat Commun 14 , 2258 (2023). Reynolds, T. W. Institutional Determinants of Success Among Forestry-Based Carbon Sequestration Projects in Sub-Saharan Africa. World Dev. 40 , 542–554 (2012). Ronneberger, O., Fischer, P. & Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinform.) 9351 , 234–241 (2015). Schnell, S., Altrell, D., Ståhl, G. & Kleinn, C. The contribution of trees outside forests to national tree biomass and carbon stocks—a comparative study across three continents. Environ. Monit. Assess. 187 , 1–18 (2015). Soares, C. et al. Tracking tree canopy cover changes in space and time in High Nature Value Farmland to prioritize reforestation efforts. Int. J. Remote Sens . 39 , 4714–4726 (2018). United Nations Environment Programme, 2019. “ Spotlight on Ethiopia’s tree-planting programme”. (Accessed September 2023). https://www.unep.org/news-and-stories/story/spotlight-ethiopias-tree-planting-programme U.S. Geological Survey, Landsat 8 Level 2, Collection 2, Tier 1, available at: https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_L2 (accessed May 2022) USA National Phenology Network. 2019. Land Surface Phenology 2001-2017 for the United States. USA-NPN, Tucson, Arizona, USA, available at www.usanpn.org/data/land_surface_phenology. (accessed June 2022) Watts, M., Hutton, C., Mata Guel, E. O., Suckall, N. & Peh, K. S. H. Impacts of climate change on tropical agroforestry systems: A systematic review for identifying future research priorities. Front. For. Glob. Chang . 5 , 880621 (2022). Zanaga, D. et al. ESA WorldCover 10 m 2020 v100. (2021) doi:10.5281/ZENODO.5571936. Zhang, W., Brandt, M., Wang, Q., Prishchepov, A.V., Tucker, C.J., Li, Y., Lyu, H. and Fensholt, R. From woody cover to woody canopies: How Sentinel-1 and Sentinel-2 data advance the mapping of woody plants in savannas. Remote Sensing of Environment , 234 , p.111465. (2019). Zomer, R., Neufeldt, H., Xu, J. et al. Global Tree Cover and Biomass Carbon on Agricultural Land: The contribution of agroforestry to global and national carbon budgets. Sci Rep 6 , 29987 (2016). https://doi.org/10.1038/srep29987 Supplementary Tables Supplementary Table 1 is not available with this version. Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-4359628","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":298025231,"identity":"3e523f78-cf18-4e43-a40a-4f5d20c6b7c9","order_by":0,"name":"Florian Reiner","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAqElEQVRIiWNgGAWjYBAC9nYGxgcJQIYEEEs2EKOF5zADswHJWthAqknRwsz7rOLBrzsMkjMSGG/OIE4Lu9mNxL5nDNISCcyWG4jRYs/MxnYjsecwg5xEApvkA+JsYWMrIF0LQ8KPwyCHsUkS5TCgFmaJxIbDPJI9D5stifM+exvjxx9/DstJHE8+eLOHGC1gwNjGwAMkG4jWAAR/SFE8CkbBKBgFIw4AAI4JK8IjwcmpAAAAAElFTkSuQmCC","orcid":"https://orcid.org/0000-0003-1299-1983","institution":"University of Copenhagen","correspondingAuthor":true,"prefix":"","firstName":"Florian","middleName":"","lastName":"Reiner","suffix":""},{"id":298026713,"identity":"f5ca0b31-faab-47cf-a893-70b3174c5235","order_by":1,"name":"Dimitri Gominski","email":"","orcid":"https://orcid.org/0000-0002-8135-1341","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Dimitri","middleName":"","lastName":"Gominski","suffix":""},{"id":298027646,"identity":"f1c93fee-8372-45c1-ae95-3a35370398c7","order_by":2,"name":"Rasmus Fensholt","email":"","orcid":"https://orcid.org/0000-0003-3067-4527","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Rasmus","middleName":"","lastName":"Fensholt","suffix":""},{"id":298027647,"identity":"453977fb-5c95-4d83-a8ba-d3662b79ed39","order_by":3,"name":"Martin Brandt","email":"","orcid":"https://orcid.org/0000-0001-9531-1239","institution":"University of Copenhagen","correspondingAuthor":false,"prefix":"","firstName":"Martin","middleName":"","lastName":"Brandt","suffix":""}],"badges":[],"createdAt":"2024-05-02 14:13:57","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-4359628/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-4359628/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":55818099,"identity":"f422d2f2-475f-4cf9-830e-626602116147","added_by":"auto","created_at":"2024-05-03 20:53:14","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":67180,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSummary of main processing flow.\u003c/strong\u003e An overview of the main framework design, consisting of the four primary steps shown in green: downloading of images, detection of trees with a CNN neural network, identification of individual trees across years, and detection of tree-level changes. Intermediate products are shown in yellow, such as the annual composited mosaics, and the annual tree confidence heatmaps (see Methods). Output products are shown in blue, including the map of confidence slope, individual tree database, and the detected tree changes.\u0026nbsp;\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/2db863d33a83f91ac1c60684.png"},{"id":55818101,"identity":"e8346879-cfd9-4fed-8e0f-2cc081ad60fb","added_by":"auto","created_at":"2024-05-03 20:53:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":293737,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eEffect of phenology of grasses and crops on tree crown visibility. a,\u003c/strong\u003e A parkland landscape near phenological dormancy, with the tree canopies clearly visible against a lack of green grasses. \u003cstrong\u003eb, \u003c/strong\u003eThe same scene at near peak greenness, with the near-infrared signal from grasses merging with that of some tree canopies. Images are shown as false colour composites with near-infrared shown as red. \u003cstrong\u003ec, \u003c/strong\u003eThe timing of the phenological target windows for deciduous and evergreen trees, overlaid on a MODIS-based phenology diagram [USA-NPN 2022].\u0026nbsp;\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/ae71c97fceb9df6ca947164e.png"},{"id":55818104,"identity":"c21ac080-950a-4e85-abda-b56f245bf36e","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":1861201,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eGeneration of mosaics from PlanetScope raw scene imagery. a, \u003c/strong\u003eRetrieval of API metadata and cutlines for all scenes covering the target 1x1 degree tile. \u003cstrong\u003eb, \u003c/strong\u003eClipping of scenes to create a seamless and non-overlapping coverage of the tile. \u003cstrong\u003ec, \u003c/strong\u003eMosaic formed after merging downloaded partial scenes, with visible cutlines. \u003cstrong\u003ed,\u003c/strong\u003e Final mosaic after histogram matching each scene to a reference scene aggregated from Landsat imagery for the same area and phenological period [U.S. Geological Survey 2022].\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/d01f205ea3390c3a841abba9.png"},{"id":55818103,"identity":"eaea6acd-c5a0-4303-8b05-f1505eb5f279","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":254004,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFiltering of scenes by sharpness and graininess. a,\u003c/strong\u003e A very blurred PlanetScope scene that failed the automated quality filtering due to low a sharpness score. \u003cstrong\u003eb, \u003c/strong\u003eAnother PlanetScope scene which failed as ‘blurry’ with a sharpness score just below the minimum threshold of 0.23. \u003cstrong\u003ec, \u003c/strong\u003eA different scene of the same area as b, which failed quality filtering due to a graininess score above the maximum threshold of 0.35. \u003cstrong\u003ed, \u003c/strong\u003eA sharp PlanetScope scene of the same area, meeting both sharpness and graininess requirements. All scenes are shown in false colour with near-infrared as the red band.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/81734c95b2433f1d863b401c.png"},{"id":55818108,"identity":"4140bde4-1644-4f83-a436-0178dd149838","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":731638,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePoint-based training data. a,\u003c/strong\u003e An example of a training area with manually annotated point-based training data, with one point per tree centre. The background image is PlanetScope shown in false colour with near-infrared as red. \u003cstrong\u003eb, \u003c/strong\u003eThe same scene with Google Earth as background image. (Imagery © 2023 CNES / Airbus, Landsat / Copernicus, Maxar Technologies, Map data ©2023). \u003cstrong\u003ec,\u003c/strong\u003e The generated heatmap for this training area, with each tree crown modelled as a Gaussian kernel around its centre.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/c7d665c47dcf1ba6a2def9e8.png"},{"id":55818106,"identity":"8a072a6b-9710-4d6b-a7ad-296ed1d56ff1","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1047665,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrediction of individual tree crown centres. a,\u003c/strong\u003e A 2018 PlanetScope scene in near-infrared false colour for a field with trees in the Dodoma region, Tanzania. \u003cstrong\u003eb, \u003c/strong\u003eThe output prediction heatmap, representing the confidence of a tree centre being detected at each pixel. \u003cstrong\u003ec,\u003c/strong\u003e The tree centre locations (vector format) extracted from the peaks of the confidence map, overlaid on Google Earth imagery (Imagery © 2023 CNES / Airbus, Landsat / Copernicus, Maxar Technologies, Map data ©2023). \u003cstrong\u003ed, \u003c/strong\u003eThe tree cover derived from the confidence heatmap, overlaid on the PlanetScope scene in near-infrared false colour.\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/ec63df69ccc5ebb6897e22bf.png"},{"id":55818107,"identity":"0ffc9f65-884d-4116-9dca-e8438a82fe45","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1367206,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualisation of slope of confidence in crown centre detection. a,\u003c/strong\u003e A map of confidence slope in units of difference of percent confidence, representing the change in detected tree crown centre confidence from 2018 to 2022, for an area near Chenene village, Tanzania. \u003cstrong\u003eb,\u003c/strong\u003e detail of a, highlighting regions of possible tree crown gains (green) and disappearances (red). \u003cstrong\u003ec-d,\u003c/strong\u003e The corresponding PlanetScope scenes for 2018 and 2022 respectively, in false colour with near-infrared as red. \u003cstrong\u003ee-f, \u003c/strong\u003eClosest available historical imagery from Google Earth, providing visual confirmation of an increase in young tree crowns. (Imagery © 2015/2019 CNES / Airbus, Landsat / Copernicus, Maxar Technologies).\u003c/p\u003e","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/0f630b5437357fb5f120f658.png"},{"id":55818415,"identity":"1a9a6130-0132-474e-9301-edff46859017","added_by":"auto","created_at":"2024-05-03 21:01:14","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":959494,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTree matching and change detection across 5 years. a-e,\u003c/strong\u003e PlanetScope scenes and detected crown centres from 2018-2022 for a field near the village of Winza, southern Dodoma region, Tanzania. PlanetScope images are shown in false colour with near-infrared as red. \u003cstrong\u003ef, \u003c/strong\u003eThe combined yearly centre positions grouped into unique trees, and classified as ‘remaining or ‘disappeared’ change class. Centre predictions are grouped using a buffer, to account for the slight shifts between yearly centre locations. The background is a Google Earth image from 2022. \u003cstrong\u003eg-h, \u003c/strong\u003eHistorical Google Earth imagery from 2019 and 2022, providing visual confirmation of the disappearance of the two trees classified as ‘disappeared’. (Imagery © 2019/2023 CNES / Airbus, Landsat / Copernicus, Maxar Technologies).\u003c/p\u003e","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/457e8b7e24a63fc80ca6087a.png"},{"id":55818105,"identity":"47c423a5-e2ff-4740-afae-2be5bb60cc4a","added_by":"auto","created_at":"2024-05-03 20:53:15","extension":"png","order_by":9,"title":"Figure 9","display":"","copyAsset":false,"role":"figure","size":902979,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOutputs of national cropland tree tracking for Tanzania.\u003c/strong\u003e The different output products generated by the framework for a study area of Tanzanian croplands, shown in the regional context in \u003cstrong\u003ea.\u003c/strong\u003e The background map is from Natural Earth.\u003cstrong\u003e b, \u003c/strong\u003eThe downloaded scenes composited to 1x1 degree mosaics, shown with near-infrared as red band. \u003cstrong\u003ec,\u003c/strong\u003e The midpoint of the optimal temporal window used for each tile’s scene selection, based on the MODIS phenology aggregated per tile. \u003cstrong\u003ed, \u003c/strong\u003eThe average scene sharpness of all used scenes across 5 years, at 1 km resolution. No-data areas are due to skipping of the sharpness check for water and forest areas. \u0026nbsp;\u003cstrong\u003ee,\u003c/strong\u003e The percent tree crown cover in 2018 aggregated to 1 km, derived from the confidence heatmaps. \u003cstrong\u003ef, \u003c/strong\u003eThe total count of trees per hectare in 2018, aggregated to 1 km. \u003cstrong\u003eg,\u003c/strong\u003e The average number of trees classified as ‘disappeared’ between 2018-2022 with a change confidence \u0026gt; 0.8, aggregated to 1 km. \u003cstrong\u003eh, \u003c/strong\u003eIncrease and decrease in tree cover expressed as the slope of the confidence heatmaps from 2018-2022, in units of percent confidence, aggregated to 5 km. \u003cstrong\u003ei, \u003c/strong\u003eThe average change confidence aggregated to 1 km2, which represents a uncertainty map for the detected change per tree. No-data areas in grey in maps d-h represent non-cropland areas, where croplands are defined as all areas classified as ‘cropland’ by WorldCover [Zanaga 2020].\u003c/p\u003e","description":"","filename":"Figure9.png","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/8d0f8ec2741c9ea659e6ae65.png"},{"id":55818875,"identity":"fcb1c84a-235d-47df-9fda-1f74d757588e","added_by":"auto","created_at":"2024-05-03 21:17:22","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":8431181,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-4359628/v1/8a50c39c-ce2b-42b6-8675-d6cf2dfe2c11.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eAn operational framework to track individual farmland trees over time at national scales using PlanetScope imagery\u003c/p\u003e","fulltext":[{"header":"1. INTRODUCTION","content":"\u003cp\u003eAgroforestry systems integrate trees in croplands and are a key component of smallholder agriculture across the Global South, providing wide-ranging benefits from both an ecological and socio-economic perspective [Jose \u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e2009\u003c/span\u003e, Bayala 2014]. Beyond yielding direct products such as fruit, fuelwood or animal fodder, the maintenance of tree cover on croplands has been shown to improve soil moisture, erosion resistance, and biodiversity, as well as contributing to mitigation of climate change in the form of direct carbon sequestration [Zomer 2016, Schnell 2015]. In dryland regions, cropland trees are also essential for providing climate resiliency, as local tree species both aid in preserving soil water content and continue to provide edible leaves and tree products during dry seasons, and long after crop failure due to drought [Quandt 2017].\u003c/p\u003e \u003cp\u003eDespite their numerous benefits, traditional agroforestry systems are impacted by changing climatic conditions and changing land management practices. While generally robust to water stress, increasing periods of prolonged drought can cause widespread tree die-offs due to hydraulic stress, fire and enhanced vulnerability to insect attack and fungal diseases [Choat 2018]. In addition to these direct mechanisms, cropland trees are also under significant threat of logging for fuelwood in times of drought, when other sources of combustible vegetation are scarce. Given that the intensity and frequency of climatic extremes are expected to continue rising in the next decades, there is an increasing risk of large-scale losses of cropland trees, with potential adverse effects on food security and local livelihoods in the Global South [Watts 2022]. However, the scale of such potential losses is currently not well understood, and while there have been some local studies indicating a decline in scattered tree cover [Plieninger 2009, Gibbons 2008], there is no clear picture of the long-term dynamics and stability of agroforestry systems at national and continental scales.\u003c/p\u003e \u003cp\u003eIn addition to existing agroforestry systems and community-led restoration efforts such as Farmer Managed Natural Regeneration, the increased awareness of tree-based carbon mitigation has also given rise to a variety of tree-planting initiatives aiming to increase tree cover in and along croplands or degraded woodland ecosystems. These schemes exist at multiple scales, from small-scale projects of non-governmental organisations (NGOs) working with local farmers [One Acre Fund 2022], to larger campaigns such as Rwanda\u0026rsquo;s and Ethiopia\u0026rsquo;s national agroforestry tree planting initiatives [New Times \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, UNEP 2019], up to trans-national initiatives such as the proposed Great Green Wall, which originally envisioned a vast tree-planting belt across the Sahel region [Goffner 2019]. While some projects are devoted primarily to improving agroforestry yields, biodiversity and ecosystem resilience, others are explicitly geared towards carbon credit schemes, with the continued growth of the global carbon market spawning a continuous rise in new tree-planting and reforestation schemes [Lefevbre 2021]. A major challenge for such projects is the lack of scalable monitoring tools to assess tree survival rates over multiple years, leading to a common criticism of \u0026lsquo;greenwashing\u0026rsquo;: that many of the promised carbon credits are not actually sequestered due to exaggeration of trees planted, lack of overview of tree survival, and losses of trees due to tree mortality factors [Duguma 2020].\u003c/p\u003e \u003cp\u003eThere is thus an urgent need for developing a tool to monitor temporal changes in non-forest trees in general, and on croplands in particular, with potential users including local farmers, restoration practitioners, government ministries and the larger environmental research community. While the exact requirements of these actors will vary by use case, such that carbon mitigation schemes for example might also want to measure biomass and carbon, a first step is a simple monitoring of trees at large scale, that is able to reliably verify living trees and detect losses and potential gains. A further common requirement for nearly all users is low cost. Currently, for small projects monitoring is typically either done using normalised difference vegetation index (NDVI) trends based on coarse spatial resolution data as an integral measure of herbaceous and woody vegetation [Soares 2018, Bey \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2021\u003c/span\u003e], or by repeated field visits, where the existence of single trees is manually verified over multiple years. However, at national scale the first is not reliable, and the latter is not cost-effective, and there remains a critical need for a scalable alternative, which currently does not exist.\u003c/p\u003e \u003cp\u003eThe major reason for this is the fundamental challenge in detecting cropland trees using remote sensing. Due to their scattered nature and lack of closed canopy, single agroforestry trees cannot be directly seen in medium-resolution satellite imagery. Instead, they have usually been mapped by their spectral characteristics in mixed pixels, using freely available medium resolution imagery such as Landsat [Hansen 2012, Higginbottom 2018, Venter 2018]. Additionally sub-pixel methods applied to Sentinel-1 and Sentinel-2 imagery have been used to map both forest and non-forest tree cover at near single-tree level [Zhang 2019], although detection of dynamics with these methods remains difficult [Brandt, J. 2023].\u003c/p\u003e \u003cp\u003eHowever, recent advances in machine learning methods applied to remote sensing have led to new breakthroughs in the mapping of non-forest tree cover. In particular, the combination of deep learning and high-resolution satellite imagery has allowed the mapping of individual trees at national and continental scales [Malko\u0026ccedil; 2021, Brandt, M. 2020, Reiner 2023, Liu 2023]. This includes cropland trees, supporting the determination of accurate numbers of large cropland trees at national scales, and estimating their carbon stocks [Mugabowindekwe 2022]. Combined with an exploding amount of available high-resolution satellite data with new global coverage nano-satellite constellations such as PlanetScope, these methods now provide a unique opportunity to study the temporal dynamics of scattered cropland trees at national to global scales.\u003c/p\u003e \u003cp\u003eHere we identify the opportunities and key challenges of monitoring cropland tree dynamics with PlanetScope data and propose several mitigation strategies to address the latter. We then integrate these methods into a framework to track individual trees over multiple years at national scale. It consists of four main components: A comprehensive satellite image download and compositing pipeline, a convolutional neural network to detect individual trees in an image, a tree matching algorithm to identify individual trees across years, and a change classification process to detect tree changes. The resulting end-to-end framework allows for accurate quantification of surviving and lost trees for a given area and temporal range and can be applied to any area due to global availability of PlanetScope satellite imagery. We demonstrate this application by a national-scale proof-of-concept case study of Tanzania, where we map individual cropland trees between 2018 and 2022.\u003c/p\u003e \u003cp\u003eThis framework may be useful as a tool for governments and environmental organisations to monitor the stability and resilience of agroforestry systems at national scale, and serve as the basis of an early-warning system for sustained losses in cropland tree cover. In addition, such a tool may be used to monitor landscape restoration projects and the planting of new cropland trees, as well as verify claims of tree numbers in large-scale plantations and carbon credit schemes.\u003c/p\u003e"},{"header":"2. MATERIALS AND METHODS","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Overview\u003c/h2\u003e \u003cp\u003eWe developed an end-to-end processing framework to detect changes of scattered individual trees across multiple years and large study areas, using satellite imagery. The method consists of a four-step processing flow (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). First, raw satellite scenes are downloaded, quality-filtered and composited to produce yearly gap free mosaics covering the study area. Second, a deep learning framework is applied to detect individual tree crown locations with a given confidence per tree in each yearly mosaic. Thirdly, a matching algorithm is used to identify identical tree crowns in subsequent years, designed to be robust to small changes in detected location. Finally, a change classification scheme is applied, which takes into account yearly confidences and possible missed detections, to provide a final change class and change confidence. The overall framework is designed to be versatile in both temporal and spatial scope, able to provide output tree change maps for any study area and target input years.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe detection of tree-level changes from PlanetScope satellite imagery faces a number of practical challenges, primarily related to inconsistencies in the way trees can be detected in the satellite images. Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e provides an overview of these challenges, and how we have addressed them, with further detail in the following sections.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003e\u003cb\u003eSummary of challenges.\u003c/b\u003e A summary of the challenges faced in detecting individual tree changes using Planetscope data, along with the mitigation strategy that was used to address them.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"2\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eChallenge\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMitigation\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCost-effective availability of imagery at large spatial and high temporal scale\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUse of PlanetScope imagery. Alternatives include RapidEye, EarthDaily or Sentinel-1\u0026amp;2 for large trees or groups of trees\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eConfusion of tree canopy infrared signal with high grasses and crops\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eScene selection from a narrow time window based on local vegetation phenology derived from ancillary satellite remote sensing data\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eIncomplete scene coverage\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eProgressive relaxation of filter criteria\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInconsistent sharpness and scene quality, unreliable metadata\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAutomated scene sharpness assessment based on blur kernel estimation [Anger 2019]\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDetection of young or very small crowns at single pixel size\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eSelection of scenes with low solar altitude angle to detect the shadows of the small crowns\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eHigh effort of hand-labelling a sufficient number of tree crown outlines for training\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eLabelling and detection of tree crown centres, with trees modelled as Gaussian kernels\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSeparation of clumped trees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eExtraction of local maxima from merged Gaussians in the predicted confidence heatmap\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpatial shifts of detected trees between subsequent scene acquisitions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDeploying a nearest-neighbour tree matching algorithm\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eVariance in yearly prediction confidence\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConfidence-based weighted change detection\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eInconsistent detection of young trees\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eFiltering by confidence thresholds\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDifficulty in reliably mapping gains\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eConfidence slope maps to reveal local trends\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Preparation of imagery\u003c/h2\u003e \u003cp\u003e \u003cem\u003eChoice of imagery\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe primary component of a national-scale tree tracking framework is the underlying source of imagery, for which the main requirements are high spatial resolution, high temporal availability and feasible cost. Ideally, to detect single trees and small trees as individuals, imagery of the highest resolution possible should be chosen, but in practice this is not possible due to logistical and financial constraints. Amongst remote sensing techniques, UAV photography has the highest resolution (up to \u0026lt;\u0026thinsp;10 cm), but does not scale to national level. Sub-metre resolution aerial imagery has been successfully used to map trees at national scale [Malko\u0026ccedil; 2021, Li 2022, Mugabowindekwe 2022], but this is still cost-prohibitive for most larger countries in the Global South, especially considering the need for multi-temporal acquisitions. Another option is globally available sub-metre satellite imagery such as Worldview-3 and Pleiades, however this is sold commercially with a high cost per km\u003csup\u003e2\u003c/sup\u003e, thus also rendering it prohibitively expensive to acquire for multiple years at national scale. In contrast, there is freely available public satellite imagery such as Landsat and Sentinel-1 \u0026amp; 2, yet at 30 m and 10 m respectively their resolution is insufficient to robustly detect single trees of various sizes.\u003c/p\u003e \u003cp\u003eOne imagery source that meets most of the requirements is the PlanetScope constellation, which provides near-daily 3 m resolution imagery globally at a lower cost compared to commercial sub-metre images, and is currently freely available in the tropics for non-commercial purposes through Norway\u0026rsquo;s International Climate and Forest Initiative [Planet 2021, NICFI 2021]. Compared to sub-metre imagery previously used in large-scale tree detection studies [Brandt, M. 2020], the 3 m resolution of PlanetScope is considerably lower, and will result in a limit to the minimum size of detected trees. Nonetheless, recent studies have demonstrated that mapping of single trees\u0026thinsp;\u0026gt;\u0026thinsp;10\u0026ndash;20 m\u003csup\u003e2\u003c/sup\u003e crown size is possible at national and continental scale with PlanetScope [Reiner 2023, Brandt, M. 2023]. Considering the combination of high temporal availability, global coverage and sufficient high spatial resolution, this is currently the most promising imagery source for a national tree tracking scheme, despite drawbacks such as the relatively lower spatial resolution that misses small trees and shrubs, and variance in image quality which are discussed below. We have developed the framework presented here to use primarily PlanetScope, however the core processing steps of tree detection, identification and change classification are agnostic of the underlying imagery, and it could thus easily be used with higher resolution imagery where available, or lower resolution imagery such as Sentinel-2, at the cost of an increase in minimum detected tree size.\u003c/p\u003e \u003cp\u003e \u003cem\u003eAcquisition of PlanetScope scenes\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe PlanetScope constellation consists of 180 \u0026lsquo;Dove\u0026rsquo; nanosatellites providing daily coverage of the global land area at 3 m resolution [Planet 2021]. The constellation includes three generations of satellites: Dove Classic, Dove-R and SuperDove, where Dove and Dove-R provide 4-Band RGB/NIR imagery, and the newer SuperDoves provide 8 spectral bands.\u003c/p\u003e \u003cp\u003eThe Dove nanosatellites were launched in multiple rounds starting from 2014, with full constellation deployment achieved by 2018. As of 2023 this provides 5 years of near daily global coverage, resulting in a very large number of raw scenes that can be queried and downloaded from the Planet application programming interface (API). While there are images available at approximately daily rate, such a short revisit time is not meaningful for a tree tracking application. Instead, we use the archive of available scenes to create one high quality composite mosaic per year, and then track changes between these mosaics. For the processing framework, any number of years can be specified, and for the case study presented here we used the 5 years from 2018\u0026ndash;2022.\u003c/p\u003e \u003cp\u003ePlanet Labs already provide annual and monthly base maps, which are composed from many raw scenes and harmonised to a smooth mosaic consistent in spectral distribution. However the use of these mosaics excludes the possibility to select optimal scenes, based on specific metadata criteria that have a large impact on the visibility of single tree crowns, such as the seasonal time of acquisition, the sun position at time of acquisition and the satellite view angle. The sun position in particular has a large effect, as a low solar elevation causes long tree shadows, which provide essential visual context to the machine learning model for detecting the tree, especially for young trees with a small crown size near or even below the native resolution.\u003c/p\u003e \u003cp\u003eTo create the annual mosaics, individual scenes are downloaded from the Planet API, using the \u003cem\u003ePSScene\u003c/em\u003e item type and the \u003cem\u003eortho_analytic_4b_sr\u003c/em\u003e asset, which is an atmospherically corrected surface reflectance product. The Planet API provides functionality to query all available scenes for a given area of interest, and returns a set of scene metadata such as ground sampling distance (GSD), sun angle, and view angle, as well as image quality indicators such as percent cloud, percent haze and percent shadow. To obtain images best suited for tree detection, the following initial query filters were used on the PSScene metadata: gsd\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;4.0, sun_elevation\u0026thinsp;\u0026lt;\u0026thinsp;60, clear_confidence_percent\u0026thinsp;\u0026gt;\u0026thinsp;95, clear_confidence\u0026thinsp;\u0026gt;\u0026thinsp;=\u0026thinsp;99, heavy_haze_percent\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0, light_haze_percent\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0, shadow_percent\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0, snow_ice_percent\u0026thinsp;\u0026lt;\u0026thinsp;=\u0026thinsp;0, quality_category = \u0026lsquo;standard\u0026rsquo;.\u003c/p\u003e \u003cp\u003eHowever, beyond these quality metrics and the sun angle, one of the main factors for successful tree detection is the phenological stage of local vegetation at the time of scene acquisition. Due to the differing phenology of grasses, crops, shrubs and trees over an annual growing cycle, there is an optimal time window in which the near-infrared signal of tree foliage is most clearly discernible from the background, and not confused with near-infrared from other vegetation such as crops or grasses (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea,b). This phenological time window with optimal pheno-spectral signatures is unique to local conditions, and also differs substantially between areas of deciduous and non-deciduous trees. Therefore, to select imagery for optimal tree detection, the framework divides the study areas into a grid of 1x1 degree tiles, and uses a different yearly target window for each tile, based on the local plant phenology. For each tile, the MODIS MOD09 surface reflectance product is used to determine the average time of key stages such as senescence, greendown, and dormancy [DiMiceli 2015]. For areas with evergreen woody plants dominating (\u0026gt;\u0026thinsp;90%), the target window is then taken as between mid-greendown and dormancy, when grasses have lost most of their near-infrared signal. For areas with deciduous woody plants dominating, an earlier window between senescence and mid-greendown is chosen, where grasses have passed their peak productivity, but trees still have full foliage (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec). The status of evergreen vs deciduous trees is computed for each tile, based on the average forest types from the ESA WorldCover land cover product [Zanaga 2020].\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eInitially, the framework queries the API for all scenes within this phenological window matching quality filters (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). However, the target window length is different for each tile, and depending on local atmospheric conditions and cloud prevalence there are not always sufficient scenes available for creating a high quality cloud-free image mosaic. In this case a progressive retry algorithm is used, in which both sides of the temporal window are gradually extended to fill remaining gaps with additional scenes adjacent to the core target window, up to a maximum of 60 days. Where gaps remain, a second progressive loop is then used to gradually lower the minimum permissible metadata metric of \u0026lsquo;visible_confidence_percent\u0026rsquo; down to a minimum of 60. Once all available scenes are found, the selected scenes are clipped to create a single seamless and non-overlapping coverage of the tile, stored as a footprint file and then ordered and downloaded in parallel (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eb). The scenes are clipped before ordering to reduce quota use, as the Planet API quota is measured by actual area downloaded, not total area of scenes accessed.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eAutomated quality filtering of imagery\u003c/em\u003e \u003c/p\u003e \u003cp\u003eWhen downloading large numbers of raw scenes, it becomes clear that the PlanetScope scenes exhibit a certain variance in image quality and sharpness, even after applying the most restrictive metadata filters. In particular, there are occasional scenes suffering blurriness, graininess and band mis-alignment, which can be related to sensor problems, focus issues or unreliable metadata. In addition, the different generations of PlanetScope satellites have different sensor characteristics and were launched into different orbits, such that the older Dove Classic images actually have the lowest Ground Sample Distance (GSD) and thus sharpest images. The upgraded Dove-R satellites have a similar orbit, however both Dove Classic and Dove-R are being retired and are no longer available after April 2022. While the newer SuperDoves have substantially better radiometric performance, they are at a higher orbit and therefore have a higher GSD and lower resolution, which is the critical metric for detecting single trees. However, over time their orbit decays and GSD decreases again, such that recent SuperDove images are significantly sharper than the first batch from 2019. In a tree tracking context, this variance in sharpness is challenging, as a single blurry scene will directly result in missed tree detections for that year, which may cause false detection of disappeared trees. In other cases, very grainy images may also cause overestimation of trees, which can also lead to overestimation of disappeared trees as these falsely detected trees are then not detected in subsequent years. One strategy to reduce this variance would be to smooth merged mosaic to the lowest common sharpness, however this is not helpful in a tree tracking context where sharpness is critical.\u003c/p\u003e \u003cp\u003eInstead, to mitigate these issues two approaches are proposed: Firstly, in the years where lower orbit Dove Classic and Dove-R scenes are available (2018\u0026ndash;2021), only these scenes are selected, and then from 2022 onwards when they are unavailable only the lowest GSD SuperDove scenes are selected. Secondly, after downloading, an automated quality filtering is applied to detect blurry scenes using a kernel-based sharpness estimation, blacklist them, and redownload new replacement scenes until the entire mosaic is sharp. Sharpness is estimated using the blur kernel method developed by Anger et al. [Anger 2019], using a minimum threshold of 0.23 for the L2 norm sharpness score. To avoid misclassification, a mean score over 20 sub-windows of 512x512 pixels of the scene is used. Additionally scenes are whitelisted that are very small, or that are dominated (\u0026gt;\u0026thinsp;50%) by bare soil, water or wetlands according to the WorldCover land cover product, as they were found to produce falsely low sharpness scores [Zanaga 2020]. In a second step graininess is estimated as a scaled function of the mean standard deviation of all 4 bands aggregated across 20 sub samples, and a maximum threshold of 0.35 is used to exclude grainy scenes. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e shows examples of scenes blacklisted due to low sharpness or high graininess scores.\u003c/p\u003e \u003cp\u003eAs quality filtering and blacklisting only happens post-download, the entire tile filling process is done using an iterative approach where scenes are downloaded and checked in batches until the entire mosaic is filled. Each scene order requires about 10\u0026ndash;30 minutes to be fulfilled by the API as stored raw scene assets are retrieved and prepared for download. Due to the iterative algorithm, which may require multiple rounds of ordering, it can therefore take more than 2 hours to complete a mosaic tile, but many tiles can be run in parallel to increase download speed.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eImage mosaicking into yearly tiles\u003c/em\u003e \u003c/p\u003e \u003cp\u003eAfter downloading all scenes for a yearly tile, the scenes are composited into a mosaic. The raw scenes are kept in storage on disk in their original projection and lossless compression, but all mosaics are reprojected to a single configurable geographic projection, and can be compressed in a lossy format to reduce storage needs. By default EPSG:4326 is used as a global projection, and JP2OpenJPEG compression with quality 80 is applied as this was found to be near visually lossless and not to have any impact on prediction quality. This step also reduces the storage size per 1x1 degree tile from approximately 7 GB for the raw scenes to approximately 3 GB for the mosaic.\u003c/p\u003e \u003cp\u003eFinally, to reduce seamlines between scenes, a histogram matching algorithm is applied using Landsat reference images for each grid tile [U.S. Geological Survey, 2022]. The Landsat reference images were produced as a temporal composite across 10 years, using Landsat images chosen from the same phenological time window per grid tile, to obtain a reference image that is spectrally stable across large areas and represents the typical histogram at that time of year. During histogram matching the surface reflectance of each PlanetScope scene is matched to the respective area in the Landsat image, resulting in a smooth final PlanetScope mosaic (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ed).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Detection of individual tree crowns\u003c/h2\u003e \u003cp\u003e \u003cem\u003eCrown centre based deep learning method\u003c/em\u003e \u003c/p\u003e \u003cp\u003eTo detect individual cropland trees, the framework uses a deep learning method with a convolutional neural network based on the UNet architecture [Ronneberger 2015]. Instead of direct tree crown segmentation as employed by Reiner et al, [Reiner 2023], the tree tracking framework uses a centre-based object detection method, where each tree crown is modelled by a Gaussian distribution around its centre during training [Luo 2021, Ventura 2022]. The CNN is first trained with PlanetScope images and the raster of crown centre Gaussians, and then applied to unseen PlanetScope images, where it predicts a heatmap that represents the confidence of a tree centre being present in each pixel. From this the local maxima are then extracted into a map of individual centres. This majorly reduces labelling time as trees can be labelled by their centre alone, and enables the tracking of specific tree instances over time. It also does not require any information about or delineation of the crown area during training, and improves the separation of individual trees when their crowns are connected.\u003c/p\u003e \u003cp\u003e \u003cem\u003eTraining\u003c/em\u003e \u003c/p\u003e \u003cp\u003eTo train the model a large number of ground truth labels of tree centres are required. These can either be labelled by hand or extracted from LiDAR canopy height datasets. Compared to labelling of polygon tree canopy crowns, hand-labelling of centres is relatively quickly done by simple visual inspection on the source imagery, and utilising additional overlaying high-resolution imagery such as Google Earth. Alternatively, if model robustness across a wide range of ecosystems and local conditions requires a very large number of labels, the framework also supports extraction of tree centres from high-resolution canopy height models from aerial or UAV LiDAR. In a first step the raw LiDAR point clouds are converted to rasterised digital height maps, and then, after subtracting terrain, to canopy height maps. Subsequently, the local maxima of the height maps are used to separate individual tree crowns and extract their centres, based on a minimum distance parameter.\u003c/p\u003e \u003cp\u003eThe labelled tree centre points are then converted to a raster map of Gaussian distributions to be passed to the CNN during training (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec). Each tree is modelled as a Gaussian kernel around the centre with a fixed initial size. Ideally, the kernel size would match the real crown size. However, in the case of hand labelling of points this information is not available, and the sizes potentially extracted from LiDAR canopy height maps are also not reliable due to spatial and temporal mismatches between the LiDAR acquisition and the PlanetScope training image. Instead, to account for the variance of real crown sizes, an adaptive kernel method is used based on Luo et al, where the size of the Gaussian is continuously adapted during training with a dynamic scale map [Luo 2021]. We modify this method to draw exact Gaussian kernels on-the-fly, instead of their linear approximation, and furthermore only allow scale factors\u0026thinsp;\u0026gt;\u0026thinsp;1. For PlanetScope, the standard deviation of the base Gaussian kernel size is set to 6.5 m, but this can be adapted to any size if other imagery sources are used.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003ePrediction\u003c/em\u003e \u003c/p\u003e \u003cp\u003eFor prediction, each mosaic is split into patches of configurable size, typically 1024x1024 pixels, and predicted batchwise, with batch size determined by available video memory of the processing GPU. Reading, predicting and writing is done with queue-based multiprocessing to maximise GPU utilization, and the number of reader, predictors and writers is configurable to reduce bottlenecks due to disk or network speeds.\u003c/p\u003e \u003cp\u003eThe output of the prediction is a heatmap which represents the confidence of tree crown presence (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003eb). The local maxima of this heatmap are then extracted into a vector file of detected tree centres (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ec), which includes as an attribute the confidence of the peak pixel, henceforth referred to as the tree\u0026rsquo;s confidence. An estimate of the tree crown area is stored, derived from the intensity and shape of the heatmap Gaussian surrounding the local maximum. Additionally, assuming that individual Gaussians tend to align with tree cover, we can threshold the heatmap with a fixed value (in practice 0.4 was found to give good results) to extract tree cover (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e6\u003c/span\u003ed).\u003c/p\u003e \u003cp\u003eTo improve robustness, the framework includes an ensemble prediction approach where multiple models can be trained and selected on different subsets of the training data. During prediction, images are then predicted with each individual model, and these output heatmaps are aggregated with a per-pixel mean into the final prediction.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Change detection frameworks\u003c/h2\u003e \u003cp\u003eThe tree detection model provides two outputs for each year: A confidence heatmap raster layer, and a vector file with the extracted yearly tree centre positions, reflecting local confidence maxima. We then introduce two different methods for detecting tree changes based on these intermediate outputs. In the first method, the slope of the confidence map is analysed to produce a raster heatmap of confidence trend, which can be used to track growth for plantations or other areas where trees cannot be mapped as individuals, while the second aims to classify changes of individual trees, based on the yearly individual tree locations and confidences.\u003c/p\u003e \u003cp\u003e \u003cem\u003eVisualisation of confidence slope\u003c/em\u003e \u003c/p\u003e \u003cp\u003eIn effect, the heatmaps represent the CNN model\u0026rsquo;s confidence that a given pixel forms the centre of a tree crown. In addition to image quality, this confidence is directly affected by tree crown size, tree age and tree health. Therefore, by analysing changes in this confidence over time, trends in these underlying factors can be revealed. To this end, the framework merges the yearly heatmaps into a 3D spatio-temporal array, and then computes a linear regression across the time dimension, resulting in a map of the temporal slope of tree centre confidence per pixel. This map is useful in identifying areas where trees have increased in crown size, or new trees grown (positive slope), or areas where trees may have disappeared, been trimmed, or weakened due to fires or disease (negative slope). One advantage of this method is that the entire predicted area is considered, including all types of vegetation, not only single trees of a sufficient size to be detected with a high confidence. In general this method may be particularly suited to detecting gains, both of new individual trees, or of increasing canopy cover of shrublands (referred to as bush encroachment), woodlands or tree plantations. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e shows an example of an area with increasing tree cover from 2018\u0026ndash;2022 in Tanzania, which is clearly identified from the confidence slope map.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eIndividual tree change classification\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe individual tree change detection framework is based on predicting tree centre positions for each year, and then combining these yearly predictions to detect changes of individual trees such as disappeared trees. In a first step, tree detections are filtered by landcover and confidence. Next, a matching algorithm is applied to identify common trees across years, which is designed to be robust to slight changes in detected position. Finally, the presence, absence, and confidence of the yearly detections are supplied to a weighted classification algorithm to determine a final output change class per tree.\u003c/p\u003e \u003cp\u003eAfter predicting each mosaic tile, the predictions are filtered by land cover, such that only trees on croplands are included in the further analysis. This is done by masking with an external land cover product, such as the \u0026lsquo;cropland\u0026rsquo; class in ESA\u0026rsquo;s WorldCover [Zanaga 2020]. Significant uncertainty is introduced at this stage, as any misclassifications in the land cover product lead to inclusion of non-cropland areas for which the model was not trained, such as shrublands or dense woodlands.\u003c/p\u003e \u003cp\u003eAnother source of uncertainty stems from the detection of newly planted and small trees, which may be detected in some years but not others, due to their size being at the limit of detectability, given the spatial resolution of the data. When detected, these trees often have a very low confidence. To reduce uncertainty and false classifications from these small trees, another filtering step is therefore added to exclude them from the classification algorithm by first filtering with a minimum confidence threshold. For PlanetScope in Tanzania we used a threshold of 0.35.\u003c/p\u003e \u003cp\u003e \u003cem\u003eIdentification of individual trees across years\u003c/em\u003e \u003c/p\u003e \u003cp\u003eWhen overlaying multiple yearly predictions, it is clear that the predicted position of a given individual tree is not consistent across time, often exhibiting significant spatial shifts up to 10 m. This is due to multiple reasons both at the imagery and CNN level, including differences in scene conditions such as satellite view angle, sun angle and thus shadow length, phenological conditions, and scene sharpness, as well as uncertainty in the Gaussian kernel distribution and uncertainty in the extraction of local maxima. To identify specific trees, we therefore employ a nearest-neighbour approach, in which the closest tree centre in a subsequent year is considered the same tree, up to a certain buffer distance (15 m for PlanetScope). This is done with an iterative method. For each tree of the first year\u0026rsquo;s prediction, the nearest tree within the buffer of the second year is considered the same tree. The position of this tree is then updated to be the mean of these two years, before repeating the same procedure for the next year, and updating the position to the mean of all previous years (Fig.\u0026nbsp;\u003cspan refid=\"Fig8\" class=\"InternalRef\"\u003e8\u003c/span\u003ef). The nearest neighbour is determined using a spatial index query which can be computationally heavy for tiles with millions of trees. Therefore, each tile is split into a configurable number of smaller pieces, which are processed in parallel and remerged.\u003c/p\u003e \u003cp\u003eAfter processing all years sequentially, a tabular database is created including all trees detected in any of the years above a certain confidence threshold, their position for each year (or NULL), their confidence for each year (or NULL), and the mean position from all years.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eClassification of tree-level change\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe aim of the tree-level classification is to convert the tabular database of matched yearly tree detections into a change classification where each tree is allocated into one of four final classes (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Due to differences in image quality, it is not possible to reliably identify the exact year of change for the \u0026lsquo;disappeared\u0026rsquo; or \u0026lsquo;gain\u0026rsquo; classes, as some years without prediction may simply be caused by lower image sharpness.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOutput tree change classes.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"3\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCode\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eName\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDescription\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRemaining\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eThis is a remaining tree, detected in multiple years and with a high confidence.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eMisdetection\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eOnly detected in one year or with very low confidence. Hence, this is not a tree, or too young to be reliably classified.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e3\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eDisappeared\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDetected with high confidence in at least one year, and then not predicted for all subsequent years. This tree has disappeared.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eGain\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eDetected as a tree in multiple later years, but not predicted for all previous years. This may be a new tree.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eThe tabular output of the tree-matching algorithm leads to a sequence of binary tree-no tree detections for each year. In principle, a simple change typology could be applied, where each case of possible tree-no tree sequences is mapped to an output change class. However, in practice this can easily lead to an overestimation of disappeared trees, as single yearly missing trees are often also due to image quality issues. Instead, a confidence-based method is used, in which the confidence of each annual prediction is taken into account. In a first step, the yearly confidences are multiplicatively combined to determine the confidence of a given binary sequence \u0026lsquo;scenario\u0026rsquo; of yearly tree presence, such as \u0026lsquo;11111\u0026rsquo; (tree detected each year) or \u0026lsquo;11100\u0026rsquo; (tree detected three years, then not detected). Here each of these possible sequence scenarios is associated with the four final classes with a different weight factor, where the weights are manually chosen, such that for example sequence \u0026lsquo;11111\u0026rsquo; corresponds to weights [1, 0, 0, 0] for the classes [remaining, misdetection, disappeared, gain]. In a second step, the confidences of each sequence scenario are then multiplied with the scenario-class weights, and these products are summed per class to obtain the final output class, and output confidence (Supplementary Table\u0026nbsp;1).\u003c/p\u003e \u003cp\u003eThe final output consists of the change class with the highest confidence, the value of which provides a \u0026lsquo;change confidence\u0026rsquo; measure reflecting the certainty of the classification. After processing is completed, the framework saves all detected trees to disk as a geopackage file, with attributes including the position, change class, change confidence, confidence in each year and detected position in each year.\u003c/p\u003e \u003cp\u003eThe weights of how much each scenario contributes to the final change class can be manually tuned to adapt to the imagery and number of years used. For our PlanetScope scenario, it was found that detection of gains is not very reliable due to the low likelihood of suddenly detecting new young trees with 3 m resolution spatial imagery, and thus the weights of the gain class were set to 0 for nearly all change sequence scenarios. Instead, we used the confidence slope method described above to identify areas with gains (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e7\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eFinally, the output classifications can be filtered by the final \u0026lsquo;change confidence\u0026rsquo; described above, such that any uncertain changes below a minimum confidence threshold are excluded. For PlanetScope in Tanzania, we used a threshold of 0.8. Using such a high threshold implies that the observed changes are reliable but also that a larger number of actual changes may not be reported.\u003c/p\u003e \u003cp\u003eTo validate the output classifications, multitemporal ground truth data is required, which specifies the true change class for a subset of detections. As it can be quite challenging to acquire repeated yearly physical field data of single tree positions, this validation data can also be obtained by manually inspecting auxiliary imagery, such as historical Google Earth imagery. For each validation point, the state of remaining or disappeared tree is then visually confirmed for the first and last year of the sequence, resulting in the ground truth change class. Subsequently these are compared with the classified output change classes to determine the model\u0026rsquo;s overall accuracy per change class.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. RESULTS","content":"\u003cp\u003eWe developed an end-to-end framework for tracking tree-level changes at national scale. The framework consists of a toolchain to download and preprocess satellite imagery, automatically identify single cropland tree crowns at scale, match detections from sequential annual detections to the same trees, and classify detected trees into remaining, gained or disappeared trees. The method is versatile and can be used for any number of years, any study area, and in principle any type of imagery.\u003c/p\u003e \u003cp\u003e \u003cem\u003eTanzania case study\u003c/em\u003e \u003c/p\u003e \u003cp\u003eTo demonstrate the capabilities of this framework, we applied it to a proof-of-concept study case to track all cropland trees at national scale for a 5 year period from 2018\u0026ndash;2022. We chose the country of Tanzania, with a land area of 945 000 km\u003csup\u003e2\u003c/sup\u003e, of which 188 000 km\u003csup\u003e2\u003c/sup\u003e is classified as cropland by the WorldCover land cover map [Zanaga 2020].\u003c/p\u003e \u003cp\u003eWe employed a previously trained model developed for cropland trees in India [Brandt, M. 2023] to train a model for tree detection in Tanzania. We then used the framework to prepare PlanetScope imagery covering all Tanzania for the 5 target years, and then applied the model to produce annual tree centre predictions. The imagery pipeline downloaded 17 800 raw scenes with a total size of 4.1 TB, and generated 1.5 TB of mosaicked images across 125 tiles (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e9\u003c/span\u003eb). Image downloading, compositing and preprocessing for one year was completed in 48 hours with a 32-core CPU, and prediction of the entire country for one year was completed in approximately 10 hours on a single RTX3090 GPU.\u003c/p\u003e \u003cp\u003eA total of 67,900,000 trees with a confidence\u0026thinsp;\u0026gt;\u0026thinsp;0.35 were detected on croplands in at least one year, with an average tree density of 3.4 trees per ha across all cropland areas. After applying the change classification, we found 110,500 trees classified as \u0026lsquo;disappeared\u0026rsquo; between 2018 and 2022 with a change confidence\u0026thinsp;\u0026gt;\u0026thinsp;0.8, or approximately 0.16% of the total in 2018 if the prediction from 2018 is taken as a baseline.\u003c/p\u003e \u003cp\u003e \u003cem\u003eAggregated output maps\u003c/em\u003e \u003c/p\u003e \u003cp\u003eBeyond the prediction raster layers and tree databases per mosaic tile, the framework also generated several national-scale output maps, aggregated to a grid of 1 km. These include: Annual maps of percent tree crown cover, derived from the confidence heatmaps (Fig.\u0026nbsp;10e), annual density maps of average tree count per hectare (Fig.\u0026nbsp;10f), a density map of average count of trees classified as \u0026lsquo;disappeared\u0026rsquo; between 2018 and 2022 with a change confidence\u0026thinsp;\u0026gt;\u0026thinsp;0.8 (Fig.\u0026nbsp;10g), a slope map of the differences in prediction confidence between 2018 and 2022 (Fig.\u0026nbsp;10h), and a confidence map of the mean change confidence per 1 km pixel (Fig.\u0026nbsp;10i). The latter provides a measure of uncertainty for the change detection, reflecting the varying regional uncertainties of detected changes, which arise due to differences in image quality.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eEvaluation of change classification\u003c/em\u003e \u003c/p\u003e \u003cp\u003eTo evaluate the accuracy of reported changes, we manually inspected 500 randomly selected trees classified as \u0026lsquo;disappeared\u0026rsquo; with a change confidence\u0026thinsp;\u0026gt;\u0026thinsp;0.8, and used all imagery years plus auxiliary imagery such as Google Earth to create ground truth labels for each trees. We found a false detection of disappeared trees of 17.02%, for an accuracy of 82.98%.\u003c/p\u003e"},{"header":"4. DISCUSSION","content":"\u003cp\u003eAgroforestry systems in the Global South are a cornerstone in providing food security for smallholder farmers, and are key to improving climate resiliency. However, these cropland trees are increasingly threatened by climatic changes and human pressures in the form of changing land management and unsustainable logging. Conversely, localised increases in cropland tree cover have been observed as the result of community-led restoration practices such as Farmer Managed Natural Regeneration, notably in Niger, Mali and Ethiopia [Haglund 2011, Reij \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2016\u003c/span\u003e]. Additionally various ecosystem restoration and carbon mitigation schemes have been planting trees across the Global South, including some aiming to sell carbon credits, yet reliable data on survival rates and areas planted is often lacking [Duguma 2020, Reynolds \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2012\u003c/span\u003e]. As a result, the overall extent of cropland tree dynamics is not clear, and there remains a critical need for a means to systematically assess agroforestry stability and cropland tree losses or gains at regional or national scales.\u003c/p\u003e \u003cp\u003eThe framework presented here is an end-to-end processing toolchain to provide such monitoring of cropland trees. It facilitates the monitoring of trees for any study area and any temporal range subject to image availability, and for a wide range of users including environmental NGOs, restoration practitioners, government ministries, carbon credit schemes and other tree-planting initiatives. Despite the support for any imagery source in principle, we currently see PlanetScope imagery as the best compromise for national scale studies, due to the combined requirements of large study areas, temporal availability, high spatial resolution, and low cost. Other possible imagery sources include RapidEye (5 m resolution), the upcoming Earth Daily constellation (5 m), or Sentinel-1 \u0026amp; 2 (10 m), although a loss of detail for smaller trees is associated with the latter.\u003c/p\u003e \u003cp\u003eYet the use of PlanetScope data comes with its own set of challenges. At a spatial resolution of 3 m, PlanetScope images are of sufficient resolution to detect single trees down to a crown size of approximately 5 m\u003csup\u003e2\u003c/sup\u003e, but only if the image is at its optimal quality. Generally, tree crowns\u0026thinsp;\u0026gt;\u0026thinsp;30 m\u003csup\u003e2\u003c/sup\u003e can almost always be well detected [Reiner 2023]. However, the differences in observed image quality between scenes can result in different detectability thresholds for trees\u0026thinsp;\u0026lt;\u0026thinsp;50 m\u003csup\u003e2\u003c/sup\u003e crown size, caused by reductions in the effective spatial resolution due to atmospheric conditions, sensor issues and varying satellite orbital height. Despite local quality checks implemented to remove the worst scenes, the issue of varying GSD and sharpness within the constellation remains, resulting in the detection of fewer trees in less sharp scenes. This is a major problem in a monitoring context as it contributes to false changes when trees are mistakenly classified as lost due to image quality issues. To alleviate this problem we developed the confidence-based method, which provides robustness against missing predictions in single years. However, this method currently requires manual weights for different classes, whereas ideally these weights should be determined automatically from a trained machine learning model. Additionally there are limits to this approach as the number of low confidence predictions increases, and in the case where the missing prediction is in the last year. These challenges can be mitigated to a certain degree by the inclusion of more temporal predictions to reduce the noise from missed predictions, either by the use of additional years (2017, 2023, 2024...), or by using multiple images per year. Another avenue would be to create yearly mosaics as temporal composites of many scenes, although there are challenges in stacking pixels at such high resolution due to apparent shifts in tree positions caused by differing georectification and the variance in view angles and sun angles in subsequent scenes. Ultimately, if perfect accuracy is required, a different image source such as sub-metre resolution aerial or UAV imagery should be chosen, although this would only be feasible for smaller study areas.\u003c/p\u003e \u003cp\u003eAnother limitation of the use of 3 m resolution data is the ability to detect tree gains. While the disappearance of a large tree due to logging produces an immediate and stark difference in the near-infrared signal, the image change due to a growing tree is gradual, starting from an undetectable sapling and moving to a single pixel crown without shadow before developing into a clearly recognizable crown with shadow. During this phase there are likely to be periods in which the young tree moves above and below the detection threshold in a PlanetScope image, due to differences in scene sharpness and satellite view angle. This would introduce considerable uncertainty to the change detection, and lead to confusion between \u0026lsquo;gains\u0026rsquo; and \u0026lsquo;disappeared\u0026rsquo; depending on the final year of the sequence. For mapping tree gains, it may therefore be better to choose only two years, with as large a time interval in-between as possible, e.g. change over a 10-year period. There is much scope for further work on the specific case of plantation monitoring with PlanetScope imagery, and it is likely that the geometric structure of plantations can aid machine learning models to detect saplings before their crowns become clearly visible.\u003c/p\u003e \u003cp\u003eFurthermore, one additional component of this framework that is currently not considered, is the integration of biomass modelling at single tree level, which is key to quantifying the carbon impacts of removed trees at country scale. Allometric approaches have previously been used to map individual tree biomass and carbon at national level [Mugabowindekwe 2022], but these typically use imagery at \u0026lt;\u0026thinsp;=\u0026thinsp;50cm resolution and exact segmentation of tree crowns. For this framework, while thresholding of the confidence heatmap is used to obtain tree cover, the resulting tree crown areas are not yet suitable for direct allometric analysis, due to a large uncertainty in crown area stemming from the variance in image quality, the Gaussian modelling of the centre, and the confidence threshold used.\u003c/p\u003e \u003cp\u003eDespite these limitations, the presented framework offers a valuable new approach to the study of cropland tree dynamics and presents a pathway to long-term monitoring of agroforestry systems at single tree level. A recent application of the framework in India has led to the discovery of a large decline in agroforestry trees over the last 10 years [Brandt, M. 2023]. Such applications at national or continental scale allow policymakers across the public, private and NGO sectors to gain quantitative evidence on both the impacts of climate change and the results of land management policies on the condition and development of national agroforestry environments. At smaller scales, the low cost of PlanetScope imagery makes it feasible to acquire multi-year imagery for specific restoration projects by non-state users. Beyond agroforestry monitoring, the automated detection of tree survival and disappeared single trees may also contribute to the evaluation, verification and impact assessment of large-scale tree-planting projects, including carbon credit schemes and national ecosystem restoration projects.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFR and MB are supported by the European Research Council (ERC) under the European Union\u0026rsquo;s Horizon 2020 research and innovation programme (grant agreement no. 947757 TOFDRY). MB also acknowledges funding from a DFF Sapere Aude grant (no. 9064\u0026ndash;00049B). R.F. acknowledges support by the Villum Foundation through the project Deep Learning and Remote Sensing for Unlocking Global Ecosystem Resource Dynamics (DeReEco, grant no. 34306).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFR prepared the data, conducted the analyses, designed the figures, and wrote the manuscript. MB designed the study, sampled the training data and developed the previously trained model used. FR wrote the codes for the data preparation and data analyses, supported by XT. DG wrote the codes for the tree detection framework. MB, DG and RF reviewed the manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData accessibility\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePlanetscope imagery was available through Norway\u0026apos;s International Climate and Forest Initiative (NICFI) satellite data Level 2 programme. NICFI Planetscope imagery in tropical areas is available for non-commercial purposes from Planet Labs at https://www.planet.com/nicfi/. However, we did not use the basemaps provided in the frame of the NICFI programme but generated our own mosaics from the raw data.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe derived tree cover maps produced in this study will be deposited in a Zenodo database, available at XX.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConflicts of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare no conflicts of interest and no competing financial interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAnger, J., de Franchis, C. \u0026amp; Facciolo, G. Assessing the Sharpness of Satellite Images: Study of the Planetscope Constellation. \u003cem\u003e2019 IEEE International Geoscience and Remote Sensing Symposium\u003c/em\u003e 389\u0026ndash;392 (2019). doi:10.1109/IGARSS.2019.8900526.\u003c/li\u003e\n\u003cli\u003eBayala, J., Sanou, J., Teklehaimanot, Z., Kalinganire, A. \u0026amp; Ou\u0026eacute;draogo, S. J. Parklands for buffering climate risk and sustaining agricultural production in the Sahel of West Africa. \u003cem\u003eCurr. Opin. Environ. Sustain.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 28\u0026ndash;34 (2014).\u003c/li\u003e\n\u003cli\u003eBey, A. \u0026amp; Meyfroidt, P. Improved land monitoring to assess large-scale tree plantation expansion and trajectories in Northern Mozambique.\u003cem\u003e Environ. Res. Commun\u003c/em\u003e. \u003cstrong\u003e3\u003c/strong\u003e, 115009 (2021).\u003c/li\u003e\n\u003cli\u003eBrandt, J., Ertel, J., Spore, J. \u0026amp; Stolle, F. Wall-to-wall mapping of tree extent in the tropics with Sentinel-1 and Sentinel-2. Remote Sens. Environ. 292, 113574 (2023).\u003c/li\u003e\n\u003cli\u003eBrandt, M. et al. An unexpectedly large count of trees in the West African Sahara and Sahel. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e587\u003c/strong\u003e, 78\u0026ndash;82 (2020).\u003c/li\u003e\n\u003cli\u003eBrandt, M., Gominski, D., Reiner, F. et al. Severe decline in large agroforestry trees in India over the past decade. Preprint available at: https://doi.org/10.21203/rs.3.rs-3340055/v1 (2023).\u003c/li\u003e\n\u003cli\u003eDiMiceli, C., M. Carroll, R. Sohlberg, D. Kim, M. Kelly, J. Townshend. MOD44B MODIS/Terra Vegetation Continuous Fields Yearly L3 Global 250m SIN Grid V006.\u003c/li\u003e\n\u003cli\u003eDuguma, L. A., Minang, P. A., Aynekulu, B. E., Carsan, S., Nzyoka, J., Bah, A., Jamnadass, R. H. From Tree Planting to Tree Growing: Rethinking Ecosystem Restoration Through Tree. ICRAF Working Paper No 304. doi: https://doi.org/10.5716/WP20001.PDF (2020).\u003c/li\u003e\n\u003cli\u003eGoffner, D., Sinare, H. \u0026amp; Gordon, L. J. The Great Green Wall for the Sahara and the Sahel Initiative as an opportunity to enhance resilience in Sahelian landscapes and livelihoods. \u003cem\u003eReg. Environ. Chang.\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 1417\u0026ndash;1428 (2019).\u003c/li\u003e\n\u003cli\u003eHaglund, E., Ndjeunga, J., Snook, L. \u0026amp; Pasternak, D. Dry land tree management for improved household livelihoods: Farmer managed natural regeneration in Niger. \u003cem\u003eJ. Environ. Manage. \u003c/em\u003e\u003cstrong\u003e92\u003c/strong\u003e, 1696\u0026ndash;1705 (2011).\u003c/li\u003e\n\u003cli\u003eHigginbottom, T. P., Symeonakis, E., Meyer, H. \u0026amp; van der Linden, S. Mapping fractional woody cover in semi-arid savannahs using multi-seasonal composites from Landsat data. \u003cem\u003eISPRS J. Photogramm. \u003c/em\u003e\u003cem\u003eRemote Sens. \u003c/em\u003e\u003cstrong\u003e139\u003c/strong\u003e, 88\u0026ndash;102 (2018).\u003c/li\u003e\n\u003cli\u003eGibbons P, Lindenmayer DB, Fischer J, et al. The future of scattered trees in agricultural landscapes. \u003cem\u003eConserv Biol\u003c/em\u003e. \u003cstrong\u003e2008\u003c/strong\u003e, 22(5), 1309-1319 (2022).\u003c/li\u003e\n\u003cli\u003eJose, S. Agroforestry for ecosystem services and environmental benefits: an overview. \u003cem\u003eAgroforest Syst\u003c/em\u003e \u003cstrong\u003e76\u003c/strong\u003e, 1\u0026ndash;10 (2009).\u003c/li\u003e\n\u003cli\u003eLefebvre, D. et al. Assessing the carbon capture potential of a reforestation project. \u003cem\u003eSci. Reports\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1\u0026ndash;10 (2021).\u003c/li\u003e\n\u003cli\u003eLiu, S. et al. The overlooked contribution of trees outside forests to tree cover and woody biomass across Europe. \u003cem\u003eScience Advances\u003c/em\u003e \u003cstrong\u003e9, \u003c/strong\u003e37 (2023).\u003c/li\u003e\n\u003cli\u003eLuo, Z. et al. Rethinking the Heatmap Regression for Bottom-up Human Pose Estimation. in \u003cem\u003e2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)\u003c/em\u003e 13259\u0026ndash;13268 (2021). doi:10.1109/CVPR46437.2021.01306.\u003c/li\u003e\n\u003cli\u003eMalko\u0026ccedil;, E., R\u0026uuml;etschi, M., Ginzler, C. \u0026amp; Waser, L. T. Countrywide mapping of trees outside forests based on remote sensing data in Switzerland. \u003cem\u003eInt. J. Appl. Earth Obs. Geoinf.\u003c/em\u003e \u003cstrong\u003e100\u003c/strong\u003e, 102336 (2021).\u003c/li\u003e\n\u003cli\u003eMugabowindekwe, M. et al. Nation-wide mapping of tree-level aboveground carbon stocks in Rwanda. \u003cem\u003eNat. Clim. Chang.\u003c/em\u003e \u003cstrong\u003e2022\u003c/strong\u003e, 1\u0026ndash;7 (2022).\u003c/li\u003e\n\u003cli\u003eNew Times, The; 2021. \u003cem\u003e\u0026ldquo;Rwanda to plant over 43 million trees in new campaign\u0026rdquo;\u003c/em\u003e (accessed September 2023). https://www.newtimes.co.rw/article/189318/News/rwanda-to-plant-over-43-million-trees-in-new-campaign\u003c/li\u003e\n\u003cli\u003eNorway\u0026rsquo;s International Climate And Forest Initiative (NICFI), 2021. (accessed May 2022). https://www.nicfi.no/.\u003c/li\u003e\n\u003cli\u003eOne Acre Fund, 2022. \u0026ldquo;Areas of Focus: Trees\u0026rdquo; (accessed September 2023). https://oneacrefund.org/what-we-do/areas-focus/trees\u003c/li\u003e\n\u003cli\u003ePlanet Labs PBC, 2022. \u003cem\u003ePlanetScope Product Specifications\u003c/em\u003e (accessed May 2022). https://assets.planet.com/docs/Planet_PSScene_Imagery_Product_Spec_June_2021.pdf.\u003c/li\u003e\n\u003cli\u003ePlieninger T, Levers C, Mantel M, Costa A, Schaich H, Kuemmerle T. Patterns and drivers of scattered tree loss in agricultural landscapes: orchard meadows in Germany (1968-2009). \u003cem\u003ePLoS One\u003c/em\u003e. \u003cstrong\u003e10\u003c/strong\u003e(5): (2015)\u003c/li\u003e\n\u003cli\u003eQuandt, A., Neufeldt, H., \u0026amp; McCabe, J. T. The role of agroforestry in building livelihood resilience to floods and drought in semiarid Kenya. \u003cem\u003eEcology and Society\u003c/em\u003e, \u003cem\u003e22\u003c/em\u003e(3). (2017). \u003c/li\u003e\n\u003cli\u003eReij, C. \u0026amp; Garrity, D. Scaling up farmer-managed natural regeneration in Africa to restore degraded landscapes. \u003cem\u003eBiotropica \u003c/em\u003e\u003cstrong\u003e48\u003c/strong\u003e, 834\u0026ndash;843 (2016).\u003c/li\u003e\n\u003cli\u003eReiner, F., Brandt, M., Tong, X. \u003cem\u003eet al.\u003c/em\u003e More than one quarter of Africa\u0026rsquo;s tree cover is found outside areas previously classified as forest. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 2258 (2023). \u003c/li\u003e\n\u003cli\u003eReynolds, T. W. Institutional Determinants of Success Among Forestry-Based Carbon Sequestration Projects in Sub-Saharan Africa. \u003cem\u003eWorld Dev.\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 542\u0026ndash;554 (2012).\u003c/li\u003e\n\u003cli\u003eRonneberger, O., Fischer, P. \u0026amp; Brox, T. U-Net: Convolutional Networks for Biomedical Image Segmentation. \u003cem\u003eLect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinform.)\u003c/em\u003e \u003cstrong\u003e9351\u003c/strong\u003e, 234\u0026ndash;241 (2015).\u003c/li\u003e\n\u003cli\u003eSchnell, S., Altrell, D., St\u0026aring;hl, G. \u0026amp; Kleinn, C. The contribution of trees outside forests to national tree biomass and carbon stocks\u0026mdash;a comparative study across three continents. \u003cem\u003eEnviron. Monit. Assess.\u003c/em\u003e \u003cstrong\u003e187\u003c/strong\u003e, 1\u0026ndash;18 (2015).\u003c/li\u003e\n\u003cli\u003eSoares, C. et al. Tracking tree canopy cover changes in space and time in High Nature Value Farmland to prioritize reforestation efforts. \u003cem\u003eInt. J. Remote Sens\u003c/em\u003e. \u003cstrong\u003e39\u003c/strong\u003e, 4714\u0026ndash;4726 (2018).\u003c/li\u003e\n\u003cli\u003eUnited Nations Environment Programme, 2019. \u0026ldquo;\u003cem\u003eSpotlight on Ethiopia\u0026rsquo;s tree-planting programme\u0026rdquo;. \u003c/em\u003e(Accessed September 2023). https://www.unep.org/news-and-stories/story/spotlight-ethiopias-tree-planting-programme\u003c/li\u003e\n\u003cli\u003eU.S. Geological Survey, Landsat 8 Level 2, Collection 2, Tier 1, available at: https://developers.google.com/earth-engine/datasets/catalog/LANDSAT_LC08_C02_T1_L2 (accessed May 2022)\u003c/li\u003e\n\u003cli\u003eUSA National Phenology Network. 2019. Land Surface Phenology 2001-2017 for the United States. USA-NPN, Tucson, Arizona, USA, available at www.usanpn.org/data/land_surface_phenology. (accessed June 2022)\u003c/li\u003e\n\u003cli\u003eWatts, M., Hutton, C., Mata Guel, E. O., Suckall, N. \u0026amp; Peh, K. S. H. Impacts of climate change on tropical agroforestry systems: A systematic review for identifying future research priorities.\u003cem\u003e Front. For. Glob. Chang\u003c/em\u003e.\u003cstrong\u003e 5\u003c/strong\u003e, 880621 (2022).\u003c/li\u003e\n\u003cli\u003eZanaga, D. et al. ESA WorldCover 10 m 2020 v100. (2021) doi:10.5281/ZENODO.5571936.\u003c/li\u003e\n\u003cli\u003eZhang, W., Brandt, M., Wang, Q., Prishchepov, A.V., Tucker, C.J., Li, Y., Lyu, H. and Fensholt, R. From woody cover to woody canopies: How Sentinel-1 and Sentinel-2 data advance the mapping of woody plants in savannas. \u003cem\u003eRemote Sensing of Environment\u003c/em\u003e, \u003cstrong\u003e234\u003c/strong\u003e, p.111465. (2019).\u003c/li\u003e\n\u003cli\u003eZomer, R., Neufeldt, H., Xu, J. \u003cem\u003eet al.\u003c/em\u003e Global Tree Cover and Biomass Carbon on Agricultural Land: The contribution of agroforestry to global and national carbon budgets. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 29987 (2016). https://doi.org/10.1038/srep29987\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Supplementary Tables","content":"\u003cp\u003eSupplementary Table 1 is not available with this version.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"University of Copenhagen","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Tree tracking, Trees Outside Forests, Agroforestry, PlanetScope","lastPublishedDoi":"10.21203/rs.3.rs-4359628/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-4359628/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eTrees outside forests, in particular on croplands, play a crucial role for food security and climate resilience in the Global South, but are threatened by increasing climatic change and human pressures. The dynamics of agroforestry systems and national cropland tree stocks are largely unknown, as currently no robust monitoring system exists to remotely detect single field trees and track changes at national scales. Here we present a framework to track cropland trees at the single tree level across multiple years, using a combination of satellite imagery, deep learning, and object-based change classification. The approach matches annual tree centre predictions to detect changes, such as individual tree losses from logging or tree mortality events. The slope between annual tree prediction confidence heatmaps is also used to detect areas of gains, with possible applications for monitoring plantation and restoration areas. The framework is designed for PlanetScope nano-satellite imagery, which offers unprecedented opportunities for detailed tree monitoring given the combined high spatial and temporal resolution. PlanetScope imagery, however, also come with a range of challenges, which are discussed and for which solutions are proposed. We demonstrate the framework by applying it to a national-scale case study of cropland trees in Tanzania from 2018 to 2022.\u003c/p\u003e","manuscriptTitle":"An operational framework to track individual farmland trees over time at national scales using PlanetScope imagery","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-05-03 20:53:10","doi":"10.21203/rs.3.rs-4359628/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"573a1593-792a-406e-9b2a-2a68309c5cec","owner":[],"postedDate":"May 3rd, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2024-05-03T20:53:10+00:00","versionOfRecord":[],"versionCreatedAt":"2024-05-03 20:53:10","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-4359628","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-4359628","identity":"rs-4359628","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}