CropCast: Context-aware AI framework for in-season crop growth forecasting

CropCast: Context-aware AI framework for in-season crop growth forecasting | 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 Article CropCast: Context-aware AI framework for in-season crop growth forecasting Ramana Heggadal Math, Ashutosh Tiwari, Lei Zhao, Juan Landivar, and 5 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9348900/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Open-field crop production is among the most weather-exposed economic activities on Earth. More than 40% of crops are lost before harvest, thanks to combined effects of abiotic and biotic stresses and flickering market economy, often exacerbated by delayed management interventions. Here we present CropCast, a context-aware continual crop growth forecasting framework for optimizing in-season management decisions using satellite-derived vegetation indices. CropCast integrates an encoder-decoder long short-term memory architecture with phenological context variables and a continual adaptation strategy that updates predictions as new satellite observations become available. Using multi-year PlanetScope imagery from cotton production regions in Texas, CropCast generates stable growth forecasts beginning as early as 50 days after planting and maintains predictive accuracy across unseen seasons and locations, reaching low mean absolute errors (2.5%) in forecasting. These forecasts further used as input for yield prediction provide high-resolution (9m) estimates of both yield and economic variability, enabling in-season dynamic intervention to optimize hyperlocal management strategies, and support risk-aware marketing decisions. Biological sciences/Plant sciences Biological sciences/Plant sciences/Plant stress responses Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Agricultural systems are under increasing pressure to produce more food and fiber while coping with climate variability, environmental degradation, and resource constraints. The global population is projected to reach 9.7 billion by 2050, driving a sharp increase in demand for agricultural production 1 , 2 . Agriculture remains one of the largest sources of employment worldwide, with approximately 26% of the global workforce (~ 916 million people) directly employed in agriculture, while nearly half of the global population depends on agrifood systems for their livelihoods 3 , . In many developing economies, agriculture contributes more than 25% of national gross domestic product, underscoring its importance for economic stability and rural development. Further, crop production remains highly vulnerable to environmental variability and management uncertainty. Global estimates suggest that up to 40% of potential crop production is lost annually to pests and plant diseases, representing economic losses exceeding USD 220 billion per year 4 . Climate-related disasters further exacerbate these impacts, with droughts, floods, and storms causing approximately USD 3.8 trillion in agricultural production losses globally over the past three decades 5 . Farmers and agronomists require timely and reliable information on how crops will grow, how they may respond to stress, and what yields can be expected for day-to-day management interventions 6 . At regional and national levels, such information supports policy decisions, stabilizes supply chains, and guides disaster response 7 . However, accurately modeling crop growth and development is challenging since these are dynamic processes shaped by interactions among genetics, environment, and management 8 . Since the 1960s, numerous process-based models have been developed to simulate crop growth using inputs such as weather, soil characteristics, and management practices 9 – 20 . These models have supported applications ranging from yield estimation and climate impact assessments to irrigation scheduling and regional production forecasting 19 , 21 , 22 . However, their operational deployment for large-scale crop monitoring is often constrained by extensive input data requirements, complex site-specific parameter calibration, and high computational costs 23 . As a result, producers, consultants, and extension agents continue to seek accessible tools that provide timely insight into future crop growth trajectories, yield potential, and management needs for their timely interventions 24 , 25 . Remote sensing is increasingly used to monitor crop growth and estimate yield using satellite, UAV, and terrestrial sensors 26 – 31 . Compared to in-situ measurements and airborne remote sensing, satellite observations enable large-scale monitoring of crop growth through repeated observations of vegetation dynamics, while UAV and ground systems offer higher spatial resolution but are often limited by operational cost, scalability, and processing requirements 32 – 36 . Multispectral satellite images provide spatially continuous measurements of canopy reflectance that can be used to derive indicators of plant vigor and physiological activity. Remote sensing studies commonly use vegetation indices derived from spectral bands to estimate crop properties such as biomass, leaf area index, and chlorophyll content 37 – 39 . Widely used indices include the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Green NDVI (GNDVI), which help reduce atmospheric effects, soil background influence, and saturation in dense vegetation canopies 38 – 46 . In addition to vegetation indices, several studies have used raw spectral bands from satellite sensors such as Landsat, MODIS, and Sentinel directly as input features for crop monitoring and yield prediction models 32 – 34 , 40 . Among these indicators, NDVI remains the most widely adopted vegetation metric because of its simplicity, strong relationship with canopy biomass and leaf area index, and long historical record across satellite platforms 37 – 39 . Time-series NDVI observations capture key phenological transitions including emergence, peak canopy development, and senescence, making them particularly useful for monitoring crop growth trajectories across seasons and regions 38 – 44 . The unprecedented rise in the availability of satellite observations has also stimulated the development of machine learning methods for crop growth analysis and yield prediction. Early studies applied statistical time-series models such as Autoregressive Integrated Moving Average (ARIMA) and Kalman filters to forecast vegetation dynamics from historical observations, performed well under stationary seasonal patterns 43 – 48 . Other, machine learning algorithms such as random forests, support vector regression, and gradient boosting have been used to model nonlinear relationships between satellite observations and crop productivity 49 – 53 . Deep learning approaches further improved the predictive capability by capturing spatial and temporal dependencies in remote sensing data. Convolutional neural networks have been applied to extract spatial features from multispectral imagery, while recurrent neural networks such as long short-term memory (LSTM), gated recurrent unit (GRU) and Transformer‑based architectures have further improved the modeling of temporal dynamics and spatial context architectures have demonstrated strong performance in modeling crop growth dynamics from satellite time series 54 – 58 . Hybrid frameworks that integrate satellite observations with climate data have also been proposed for regional crop yield prediction 59 – 67 . Despite their accurate performance, these models generally assume stationary conditions during inference and require additional fine‑tuning and retraining to infer accurately over new in‑season observations. Static models with fixed parameters degrade rapidly under weather extremes, novel management practices, or evolving cropping systems 47 – 49 . Further, retraining models with new data arrival is computationally expensive and operationally infeasible 50 – 53 . Moreover, model robustness across growth stages and environmental variability remains underexplored, leading to degraded forecasting stability 54 . Collectively, these limitations require a forecasting system that is adaptive, scalable, efficient, and resilient to the complexity of agricultural landscapes. To address the limitations of existing forecasting approaches, we introduce CropCast, a context-aware, continual autoregressive framework for adaptive crop growth forecasting using satellite remote sensing. To be operational and directly usable by farmers, CropCast uses satellite-derived NDVI measurements derived from high resolution (~ 3m) PlanetScope satellite imagery as input. Cotton provides an ideal case study for developing such forecasting systems and is used as a test crop in this study. As one of the most important fiber crops globally, cotton production requires multiple management decisions throughout the growing season, including irrigation scheduling, nutrient management, pest control, and harvest timing 55 – 59 . Despite these efforts, a significant portion of cotton acreage often goes unharvested, sometimes exceeding 20–40% during drought years, due to environmental and economic challenges 60 . For example, in 2022, Texas, USA experienced a 69% abandonment rate, primarily driven by drought, insect pressure, and low market prices. In-season crop growth forecasts and subsequent yield potential estimates can help producers optimize input use, reduce production costs, and manage market risk. Reliable early-season forecasts also enable growers to take advantage of pre-harvest marketing opportunities such as forward contracting or hedging strategies 61 – 63 . To develop CropCast, we develop an encoder-decoder LSTM architecture which models temporal dependencies and forecasts future time series. Figure 1 shows the overall CropCast framework comprising satellite image retrieval, preprocessing, modeling and forecast, and application (yield forecasting and marketing). The framework integrates an encoder-decoder long short-term memory architecture with phenological context variables including normalized days after planting and field-level vegetation statistics to capture both short-term growth dynamics and seasonal crop development patterns. Unlike conventional forecasting approaches that generate static predictions, CropCast produces autoregressive NDVI forecasts in incremental time windows and continually updates predictions as new satellite observations become available during the growing season, targeting predictions across diverse locations, seasons, and management regimes. The study aims to provide three contributions. First, we develop a continual forecasting framework that enables adaptive in-season crop growth prediction without retraining the full model. Second, we demonstrate that CropCast produces stable NDVI forecasts across seasons and locations beginning approximately 50 days after planting. Third, we show that predicted growth trajectories can be integrated into hyperlocal yield prediction and economic analysis, generating spatially explicit yield and profitability estimates at 9 m × 9 m resolution that support data-driven agricultural decision-making. Results Overview of time series data, modelling, and seasonal variations We performed data exploratory analysis and modeling with benchmark machine learning models to arrive at the CropCast architecture. Figure 2 a shows the location of the cotton fields used in this study. These fields named Sinton, Etter and Driscoll fields, lie in different cotton growing regions of the state, and therefore suitable for evaluating model transferability in space. Figure 2 B to D show the enlarged views for the Sinton, Etter and Driscoll fields respectively, obtained from PlanetScope satellite images, used for model training. For the Driscoll field, NDVI curves are shown in Fig. 2 D for four cropping year seasons (2021–2024), with shaded region showing the variance of the NDVI observations. Out of four years, data for the years 2021 to 2023 are used for training, while that of year 2024 is used for testing. While the general shape is bell‑like, the timing, slope, and amplitude of these phases vary with cultivar, planting date, management practices, and environmental stresses 63 – 65 . For example, delayed planting compresses the early growth phase, while drought can shorten the peak phase and accelerate decline. At early growth stage with minimal vegetation cover, NDVI has low (~ 0.2) values 66 . As plants emerge and leaf area expands, NDVI rises rapidly, reflecting increased chlorophyll content and canopy closure. This stage is often sensitive to planting date, seed quality, and early irrigation or fertilization decisions. Following the rapid growth phase, NDVI reaches a plateau or peak, typically corresponding to the crop’s maximum LAI and highest photosynthetic capacity 67 . This stage is associated with flowering and reproductive development. In well‑managed fields, NDVI values can remain at or near the peak for several weeks, depending on crop type and environmental conditions. As crops progress toward maturity and harvest, leaf senescence begins. Chlorophyll degradation and canopy thinning further lead to a gradual decline in NDVI 68 . Moreover, satellite observations are often affected by errors and offsets due to sensor geometry and atmospheric effects, radiometric calibration and noise, showing signal responses different than the theoretical and physical response of the crop growth curves. To reduce some of these effects, we aggregate native 3m x 3m pixels of the PlanetScope images into 9m x 9m grids by averaging reflectance values. This spatial aggregation reduces high-frequency noise caused by shadows, artifacts, and small-scale vegetation heterogeneity. We further propose using log-normal functions to fit the remote sensing observations and generate a daily NDVI growth curve resembling closeness to the physical crop growth behavior. To enhance contextual learning over time, we proposed the model to be trained on variable-length windows capturing local dynamics within the broader phenological cycle. Through the above, we create daily NDVI data for modeling and ensure that forecasts align with real physiological processes, improving interpretability and robustness. We begin our model design with a simple feedforward deep neural network trained with NDVI inputs from 1 to 60 days after planting (DAP) and train the model to regress NDVI for the remaining DAPs from fixed-length NDVI inputs of length 40. As expected, this model performs poorly across both temporal and spatial shifts (Supplementary Figures S2 A to C) when experimenting with different dropout ratios for combating overfitting. The model fails to capture the sequential nature of crop growth, making it unsuitable for modeling seasonal dynamics. To address this, we replace the feedforward architecture with a sequence-to-sequence long short-term memory (LSTM) network, which is better suited for learning temporal dependencies. Supplementary Figure S3 shows the forecasting outcomes with this model, showing significant improvement on validation data, i.e., unseen data points from the training years (2020–2022). It produces NDVI curves that capture the broad seasonal NDVI trajectory for held-out validation samples. However, when evaluated on the 2023 season, performance drops sharply (Supplementary Figure S3B). The predicted trajectories exhibit reduced spread relative to the observed NDVI, indicating that the model underestimates year-specific variability and does not generalize well to unseen seasonal conditions. This highlights a well-known challenge in cross-year time-series forecasting over the same cotton fields as year-specific environmental and agronomic signals reshape NDVI trajectories, producing distinct growth signatures that static models struggle to capture. To address the limitation of cross-year forecasting failure, we consider two approaches. The first approach is to significantly increase model complexity, e.g., adopting pre-trained Transformer architectures and hoping they might learn these shifts implicitly. The other is to make the model aware of the current season’s trend (context) and update its predictions accordingly. Initial Transformer experiments underperformed; despite their capacity, they generalized worse than simpler LSTMs (Supplementary Figure S2D). To enable context-aware learning, we find that using a few recent NDVI observations to adapt the model on-the-fly, we could forecast NDVI for the next 40–45 DAPs with high accuracy and maintain acceptable predictions. Motivated by this, we explored continual adaptive fine-tuning as a way to adapt the model to the current season’s trend. Our initial attempt involved fine-tuning the pretrained model using both recent and future NDVI observations. Since the base model was trained to predict the full remaining sequence, we assumed this setup will align the forecasts. While this approach did capture the overall trend, it produced unstable predictions, curves often began with a downward bump, exhibited unnatural dips near the end, or failed to progressively align as more data arrived (Supplementary Figure S2). We realized the issue stemming from reliance on the fine-tuning step to fix a base model that learned to stretch across the full time series, rather than generalizing stable local dynamics. Hence, we decided to train the base model from scratch, making it retain the stable growth patterns across years that are transferable regardless of season-specific shifts. Instead of predicting the full NDVI curve in one go, CropCast learns to autoregress in smaller time chunks of incoming data (30–40 DAP time steps), progressively extending the curve. We structured training using chunked input sequences and variable-length inputs, allowing the model to specialize in immediate-term extension while deferring longer-term adjustment to future updates. Noticeably, we perform online fine-tuning only on the recent 5-day interval using ground truth information (NDVI) from that period without requiring backpropagation through the entire future sequence. Detailed explanation on the CropCast model architecture and other models used for comparison is provided in the Methods section. Transferability of CropCast across space and time With the above formulation, we trained CropCast on one cotton field (Driscoll) with three years of observation (2021–2023) for training and validation, and it required minor hyperparameter tuning to adapt to the differences in the linear, peak and senescence stages, demonstrating robustness to both temporal and regional variability. The test procedure involved three stages, (i) testing a nearby cotton field with similar weather and crop growth pattern, (ii) testing cross-year for a far cotton field with varied weather and climate characteristics, and (iii) testing cross-year performance over limited satellite data availability. We also experimented with varying input permutations, such as excluding group statistics or normalized DAP, to assess feature impact. Additionally, for the cross-region evaluation, we tested early forecasting for the Etter field starting from 40 DAPs onward to simulate in-season prediction scenarios. Figure 2 panels Bin, Cin and Din show the full season NDVI curves for these fields and B1 to B4, C1 to C4 and D1 to D4 show corresponding predictions obtained from CropCast for different times during the cropping season. For the first stage, we observe that the model adapted well to incoming data (Panels B1 to B4 in Fig. 2 ) for a nearby cotton field named in Sinton, Texas. We show the forecasts at four progressive dates in the cropping season (51, 81, 93 and 117 DAPs), demonstrating progressive in-season model predictions. CropCast forecasts align closely with ground truth NDVI data starting ~ 50 DAPs. However, forecasts with NDVI inputs from 81 DAP result in slight overestimation, improving incrementally with later observations. For the second stage, we test the Etter field, for which we notice slight overfitting in near future forecasts (which gets stable later) for NDVI inputs with DAP 51 (Fig. 2 C1). In this case, the forecasts improve with further in-season NDVI inputs (63, 81, 111) as shown in Figs. 2C2 to C4. The third stage test data is for the Driscoll field and year 2024, which had a long input data gap owing to cloud cover presence and rain during acquisitions, further challenging the forecasting ability. Figures 2D1 to D4 show the forecasts, where we again observe slight overestimation for input data corresponding to DAP 81, while getting improvements afterwards. In all the three testing stages, we observe slight overestimation of the NDVI forecasts, which improve as more in-season data arrives as input. Figure 3 shows the performance metrics for the three test stages. The mean absolute errors lie between the ranges 0.05 to 0.15 for Sinton (Fig. 3 A to D), 0.025 to 0.175 for Etter (Fig. 3 E to H), and 0.025 to 0.175 for Driscoll (Fig. 3 I to L). We observe that for the first and third stages, the MAE values rise after the second set of NDVI inputs arrive (~ 80 DAP) and reduces thereafter. For the second stage, we, however, see constant decline in MAE values as more in-season NDVI data arrives for prediction. Moreover, for all the test stages, the model effectively captures the trend for these forecasts at early DAPs, with predictions stabilizing rather than drifting over later season predictions. This is a key requirement for real-world deployment of CropCast, where it can be integrated with subsequent operational tasks like yield prediction. The forecasting mechanism begins around 50 DAP for the training and test datasets and continues to provide NDVI forecasts and its update using online finetuning, improving as more satellite image acquisitions arrive. Discussion Utilization of CropCast for in-season cotton yield forecasting Reflecting on the efficient performance of CropCast across different years and regions, we demonstrate its utilization for in-season yield forecasting to show how such a model tailors to real-world problem solving. Using the Driscoll field as a case study, we used cotton yield data from 2021–2023 for training, while keeping 2024 as an independent test year. Supplementary Figure S4 shows the soil properties, overall extent and gridded data for the Driscoll field. Here we use the forecasted NDVI coming from CropCast as input, and yield as the target variable for modeling. The advantage with CropCast forecasted data as input is the early availability of full season NDVI forecasts, which can help improve the yield model’s capability to learn from the full season growth pattern. This enables the yield model to learn from an anticipated complete growth trajectory rather than truncated observations, substantially improving early-stage prediction stability. Yield forecasts begin around ~ 50 DAP, aligning with the earliest reliable NDVI-driven crop growth signal. We use an LSTM model architecture shown in Fig. 1 with 3 LSTM layers, and RMSE as performance metric for evaluating model performance. Similar to the CropCast model forecasts, we predict yields at significant stages of crop growth, beginning well in time (~ 50 DAP) and continuing from thereon. Figure 4 illustrates how adaptive NDVI forecasting propagates into spatially explicit yield and profitability predictions. The predicted yields come from a LSTM model developed for predicting yields given CropCast forecasted NDVI. Figures 4 A provides a bar chart visualization of the average yield prediction for test year 2024, for DAPs 57, 81, 105, 117, 140 (defoliation day) and 163 (harvest day), corresponding to CropCast results for Driscoll field forecasts for year 2024. Early-stage predictions exhibit moderate uncertainty but converge toward harvest-time estimates as additional NDVI data becomes available. Further, we predict grid-level yield estimates at 9 m × 9 m resolution shown in Figs. 4A1 to A6. We observe strong spatial heterogeneity within the field for predictions throughout the season. Importantly, spatial yield structure emerges early in the season and stabilizes over time, suggesting that CropCast captures persistent growth patterns rather than transient noise. These estimates nevertheless highlight the capability of the CropCast model in contributing to high resolution yield prediction. We next carry out economic analysis of these yield estimates for crop management and marketing decisions. CropCast for crop management and marketing Beyond agronomic forecasting, the integration of yield predictions with economic metrics can provide actionable decision support to the farmers. Using forecasted yield \(\:\left({q}_{i,t}\right)\) and observed cotton prices \(\:{p}_{t}\) , we computed Breakeven Yield Margin (BYM) and profitability index (PI). Detailed descriptions on computing BYM and PI come in the Methodology section. Figures 4 B and 4 C show extension of the yield maps into economic interpretations. Figures 4B1 to B6 show spatial variability of BYM percentage over time. Early-season maps identify areas at risk of underperforming relative to cost thresholds. Figure 4 C quantifies expected profit per grid cell. Negative zones during mid-season (e.g., 105–117 DAP) indicate potential management intervention points, while recovery toward harvest reflects yield stabilization and favorable price interaction. The intervention points also get highlighted in the average BYM and PI plots shown in Figs. 4 B and 4 C respectively. These results demonstrate three critical operational advantages. The first advantage is grid-level management optimization. Grid-level forecasts allow variable-rate irrigation, fertilization, or pest control targeting zones with declining profitability signals. The second advantage can be risk-aware marketing decision. Knowing that cotton prices are often higher pre-harvest than post-harvest, early yield forecasts can enable growers and merchants to hedge production or engage in forward contracts with reduced uncertainty. The third advantage can be dynamic intervention planning. Since CropCast updates forecasts as new satellite data arrives, economic risk maps evolve during the season, supporting adaptive decision-making rather than static pre-season planning. Moreover, the model demonstrates that meaningful economic signals emerge well before harvest, highlighting the value of early, context-aware NDVI forecasting. Analysis and comparison of CropCast performance The coming generation of digital agriculture requires computationally efficient and scalable frameworks that can transform routinely available satellite data into reliable crop growth forecasts for real-time agricultural decision-making. We see CropCast as a solution to these requirements, addressing several persistent limitations in satellite-driven crop forecasting. Unlike many yield models that rely on peak-season or post-flowering observations, CropCast produces early forecasting capability (beginning around ~ 50 DAP), with improved stable forecasts as new observations are incorporated via continual fine-tuning and without the need of model retraining, enabling operability. Further, the framework relies primarily on NDVI, normalized DAP, and simple group statistics, rather than multi-source meteorological or soil datasets. Some state-of-the-art ML models integrate extensive climate and management inputs, increasing data requirements and limiting scalability 42 , 43 , 69 , 70 . CropCast demonstrates that structured temporal modeling combined with contextual encoding can achieve competitive accuracy with substantially fewer inputs. Furthermore, we observe that the proposed continual adaptation strategy freezing the encoder and fine-tuning only the decoder and output head preserved stable seasonal dynamics while adapting to year-specific deviations. This aligns with recent parameter-efficient fine-tuning paradigms but is tailored here for agricultural phenology. Compared to static sequence-to-sequence LSTMs which degrade sharply when evaluated on new seasons, CropCast updates itself performing well over cross year datasets 71 , 72 . Testing across Sinton, Etter, and Driscoll fields demonstrated robust cross-location generalization. Slight early overestimation was corrected as additional in-season data arrived, illustrating adaptive correction rather than divergence. Even under cloud-induced data gaps, forecasts remained stable and convergent, as evident for Driscoll prediction for the year 2024, where we miss 25 days of observation in the cropping season. The log-normal smoothing of NDVI sequences likely contributed to this robustness by anchoring predictions to biologically plausible growth shapes. Collectively, these characteristics distinguish CropCast from purely static deep learning models and computationally intensive process-based simulators. Rather than increasing architectural complexity (e.g., Transformers, which underperformed in our experiments), performance gains arise from contextual modeling and continual adaptation. Although demonstrated here for cotton, CropCast’s structure is crop-agnostic. The framework relies on a generalizable principle that vegetation indices follow phenology-dependent growth curves, and contextual timing (e.g., DAP) can normalize inter-season shifts. Further, the local fine-tuning by CropCast can correct season-specific deviations. Since many crops exhibit bell-shaped NDVI trajectories, the log-normal smoothing and chunked autoregressive forecasting can extend naturally to maize, wheat, soybean, and rice systems. Only crop-specific planting dates and yield model retraining would be required. Current limitations and scope for improvement Despite strong performance, there is scope for improvements in overall CropCast implementation and testing. One of those would be its validation, which is limited to Texas cotton systems. Further, the whole CropCast framework depends solely on NDVI-derived growth signals, which is convenient and computationally efficient, but will improve its generalization capability over cross-year and cross locations by including more input variables governing crop growth. The applications like yield modeling and defoliation date determination will also benefit. Furthermore, economic analysis presently assumes static production costs and observed prices. We can build scenarios based on dynamic production costs and prices, generating different solutions with different levels of confidence. Future work will incorporate multi-state datasets, additional spectral indices (e.g., EVI, different band combinations, modalities), scenario simulations, and dynamic price forecasting. Conclusion In this study, we introduce CropCast, a context-aware, in-season continual forecasting framework for adaptive crop growth prediction using satellite-derived NDVI. Unlike conventional static deep learning or process-based models, CropCast integrates phenological context, variable-length temporal modeling, and modular incremental fine-tuning to produce stable and transferable forecasts across years and locations. By training on multi-year cotton datasets and validating over independent cross-year and cross-region scenarios, we demonstrate that CropCast maintains forecasting stability beginning as early as ~ 50 days after planting. The framework adapts efficiently to new in-season observations by fine-tuning only the decoder and output layers, avoiding full retraining and enabling computationally efficient deployment. Beyond NDVI forecasting, we demonstrate probable application by showing high-resolution yield prediction and economic analysis. Forecasted growth trajectories translate into spatially explicit yield, breakeven margin, and profitability estimates at high (< 10 m) resolution. Importantly, meaningful economic signals emerge well before harvest, enabling hyperlocal management optimization, dynamic intervention planning, and risk-aware marketing decisions such as forward contracting or hedging. While demonstrated for Texas cotton systems, CropCast’s design is crop-agnostic and extensible to other phenology-driven crops. The framework’s reliance on vegetation dynamics, contextual timing, and lightweight continual adaptation makes it suitable for scalable, real-world agricultural monitoring, providing a practical bridge between satellite remote sensing data and operational farm decision-making. Methodology Data Collection and Preprocessing We acquired multispectral imagery through the PlanetScope application programming interface (API) for predefined Areas of Interest (AOIs) corresponding to three cotton fields located at two major cotton growing regions: South Plains and Coastal Bend in Texas. For each AOI and date, we filtered candidate scenes using a threshold on cloud cover, retaining only those with less than 20% overall cloud coverage. This threshold provides a reasonable guarantee of unobstructed visibility within the AOI while maintaining sufficient temporal density, by PlanetScope at high (~ 3m) spatial resolution. This process yielded georeferenced image subsets focused on the target fields. By automating the search query, filtering, and cropping (AOI extraction) pipeline through the API, we collected imagery across the entire growing season, ensuring continuous monitoring of crop development. Within each AOI, native 3m x 3m pixels are aggregated into 9m x 9m grid cells by averaging band reflectance values. This spatial aggregation reduces high-frequency noise caused by shadows, artifacts, and small-scale vegetation heterogeneity. For each grid cell, we compute the Normalized Difference Vegetation Index (NDVI) shown in Eq. (1), where NIR and Red denote the digital number obtained for the near-infrared and red bands, respectively. $$\:NDVI\:=\:\frac{NIR\:-\:Red}{NIR\:+\:Red}$$ (1) Although PlanetScope provides near-daily coverage, temporal gaps remain due to cloud cover and occasional observation dropouts. To obtain smooth and temporally aligned NDVI sequences, we fit a four-parameter log-normal growth curve to each grid’s NDVI time series, as shown in Eq. (2), where x is Days After Planting (DAP), \(\:{y}_{0}\) is the baseline value, a represents the NDVI amplitude, \(\:{x}_{0}\) denotes the central scale of the peak, and b represents the standard deviation. We select this log-normal function approximation since vegetation indices typically follow a bell-shaped trajectory over the season. We sample the fitted curve at uniform DAP intervals, converting irregular observations into smooth, consistent daily NDVI sequences for each grid. These preprocessed sequences for each grid over multiple cropping seasons and locations form the input time series (~ 12,000 samples) for subsequent modeling and continual adaptation. $$\:y\:=\:{y}_{0}\:+\:\frac{a}{x}\:\text{exp}\left[-0.5{\left(\frac{\text{ln}\left(\frac{x}{{x}_{0}}\right)}{b}\right)}^{2}\right]\:$$ (2) To improve generalization and avoid overfitting a single temporal alignment, we designed the dataset to expose the model to multiple short, context-aware portions of each NDVI curve. Figure M1 illustrates this process. For each \(\:9m\:\) x \(\:9m\:\) grid cell, the full NDVI sequence is repeatedly sampled into variable-length input-target windows. A random start index is selected, and an input segment of length \(\:{l}_{1}\in\:\) \(\:\left[{i}_{1},\:{i}_{2}\right]\) is followed by a target segment of length \(\:{l}_{2}\:\in\:\:\left[{t}_{1},\:{t}_{2}\right]\) . This \(\:k\) -sampling strategy ensures that the model views the same input data curve from multiple points of view (early growth, peak, and decline) rather than memorizing a single fixed split. In addition to NDVI values, each input sequence is augmented with Days After Planting (DAP) indices and per-sample group statistics (mean and standard deviation), enabling the model to learn context-aware temporal dynamics across fields and seasons. To further enhance robustness, the target sequence is auto regressed using a mix of ground truth and previous prediction, as shown in Figure M2 . Here the green blocks ( g ) denote steps fed with ground truth (teacher forcing), while gray blocks ( p ) denote steps fed with the model’s own prior predictions. This combined strategy encourages the model to handle both ideal and self-generated contexts during inference. Because input and target windows vary in length, we implemented a custom batching pipeline. Before each training step, input sequences (NDVI and normalized DAP) and target sequences are dynamically padded to the maximum lengths in the batch, and the true sequence lengths are tracked. This allows sequence‑to‑sequence models such as LSTMs to correctly mask padded positions and prevent length mismatch issues. Group statistics for each sample are stacked alongside. This collator design preserves the variability in window lengths while ensuring computational efficiency. In this way, the dataset generation process (variable window sampling, context features, mixed teacher forcing, and dynamic batching) exposes the model to diverse temporal perspectives and improves its ability to generalize beyond a single input–target configuration. Model Architecture Instead of relying on deep, high-capacity architectures, we design a lightweight, flexible model architecture that can generalize across conditions and adapt with minimal fine-tuning, considering the crop growth dataset behavior as often sparse, noisy, and region-specific. Each input sample consists of a variable-length NDVI sequence, DAP, and group-level NDVI statistics (mean and standard deviation). This design encodes both local temporal dynamics and broader field-level context. The output is forecasted NDVI values for the next 40 DAPs. To mimic real-world variability, we use variable input lengths for NDVI, and DAP and group statistics as static input. The output is fixed at 40 days. To enrich training diversity, we slice multiple samples from each input data curve. The model we build is an encoder-decoder LSTM with 2 encoder LSTM layers, 2 decoder LSTM layers, and an output head (multi-layer perceptron layer). Figure M3 shows the modeling process. The encoder layers accept variable-length input (NDVI time series) with a padded sequence. The decoder LSTM layers perform autoregressive rollout, and the output head combines decoder state with group statistics and DAP. During training, we apply probabilistic teacher forcing ( p = 0.5) to stabilize decoding. To support real-time regional adaptation, we implement a continual fine-tuning strategy that incrementally updates the model as new NDVI sequences become available during the growing season. We implement modular adaptation by freezing the encoder to retain general spatiotemporal representations and fine-tune only the decoder and output head. This significantly reduces computational overhead and avoids overfitting, while still allowing the model to adapt to recent field-specific trends. We use a small batch of recently acquired NDVI sequences for fine-tuning and perform 1 to 2 epochs of gradient updates with a learning rate \(\:1{e}^{-4}\) . To stabilize training, we keep probabilistic teacher forcing. These steps enable fast, local adaptation without retraining the entire model, making it practical for dynamic environments where new observations trickle in gradually. Crop management and marketing methods To demonstrate the possible applications of the proposed model, we choose yield prediction as an example application. Building an LSTM model with three LSTM layers, we use the forecasted NDVI coming from the endoder-decoder LSTM as input to predict yields at multiple time steps of the growing season. For yield prediction, we use a separate model with three LSTM layers. For preparing model input, we initially compute the first derivative of the forecasted NDVI from the CropCast model for multiple DAPs (57,81,105, and 117) to mitigate the temporal shifts of the forecasts. From the first derivatives, we calculate a cumulative NDVI for these different DAPs, as feed this into the LSTM model shown in Fig. 1 . Subsequent to predicting yields at multiple time steps of the growing season, we develop a marketing method that can provide grid level calculations and decision-making capability to farmers. For each grid i at time t , the profit ( \(\:\pi\:\) ) is defined as the difference between total revenue and total production cost, as shown in Eq. (3), where \(\:{p}_{t}\) denotes the cotton price, \(\:{q}_{i,t}\) the yield, and \(\:{C}_{i}\) the grid cost. The breakeven point is obtained when \(\:{\pi\:}_{i,t}=0\) . Accordingly, the minimum yield (q) and price ( p ) required to cover total cost are shown in Eq. (4). Given the forecasted yield \(\:\left({q}_{i,t}\right)\) and the observed market prices \(\:{p}_{t}\) , we define the profitability index (PI) as the deviation of the market price and the breakeven price, shown in Eq. (5). $$\:{\pi\:}_{i,t}={p}_{t}\times\:{q}_{i,t}-{C}_{i}$$ (3) $$\:{q}_{i,t}^{BE}=\frac{{C}_{i}}{{p}_{i,t}},\:{p}_{i,t}^{BE}=\frac{{C}_{i}}{{q}_{i,t}}$$ (4) $$\:P{I}_{i,t}\:=\left(\:{p}_{t}-{p}_{i,t}^{BE}\right)\times\:{q}_{i,t}$$ (5) We further define the Breakeven Yield Margin (BYM) index as the proportional deviation of the estimated yield from its breakeven level defined in Eq. (6), where \(\:BY{M}_{i,t}\) denotes the corresponding BYM values for grid i at time t. Positive values of PI and BYM indicate economic viable production conditions within a grid. The cost of production is retrieved from the Texas A&M AgriLife, and the price of cotton is derived from United States Department of Agriculture (USDA)- Agricultural Market Service (AMS) (retrieved from: https://www.ams.usda.gov/market-news/cotton-tobacco (accessed 12/2025)) $$\:BY{M}_{i,t}=\frac{{q}_{i,t}-{q}_{i,t}^{BE}}{{q}_{i,t}^{BE}}$$ (6) Declarations Competing interests The authors declare no competing interests. Author contributions R.H.M. developed the methodology, implemented the forecasting framework, performed the formal analysis, and contributed to manuscript writing. A.T. contributed to study conceptualization, methodology development, model evaluation, interpretation of results, manuscript writing and revision. L.Z. contributed to methodology development, supervision, interpretation of results, and manuscript editing. J. Landivar contributed to agronomic interpretation, field study design, and manuscript revision. Y.C. contributed to the economic analysis framework, interpretation of profitability and marketing results, and manuscript revision. S.S. contributed to methodology development, interpretation of results, and manuscript writing and revision. J.L.S. and B.G. contributed to field and satellite image data collection and processing. M.B. conceived the study, contributed to project supervision, agronomic guidance, funding acquisition, and manuscript revision. All authors read and approved the final manuscript. Acknowledgements This research is part of the U.S. Department of Agriculture (USDA) Hatch Project number TEX0 9937. The authors acknowledge Cotton Incorporated (Texas State Support Committee, project number: 26-632TX) and Texas A&M AgriLife Research for funding this research. Authors also acknowledge Texas A&M Institute of Data Science (TAMIDS)- leadership members Dr. Nick Duffield and Dr. Drew Casey, the Data Science Capstone Program. Additionally, thanks to Mr. Jimmy Dodson and Mr. Jon Gwynn (DODSON FAMILY FARMS), Mr. Charles Ring, and the North Plains Ground Water Conservation District for enabling field experiments and data collection. Data and code availability Data is confidential as it belongs to farmers. Authors will be releasing a GitHub link to the code which is under finalization. Till then, code will be made available upon request. References Grafton RQ, Williams J, Jiang Q (2015) Food and water gaps to 2050: preliminary results from the global food and water system (GFWS) platform. Food Secur 7:209–220 World Population Prospects (2019) : Highlights. United Nations, Department of Economic and Social Affairs, Population Division (2019) Food and Agriculture Organization of the United Nations (2025) Employment indicators 2000–2023: global and regional trends. FAO Statistics Division Savary S, Willocquet L, Pethybridge SJ, Esker P, McRoberts N, Nelson A (2019) The global burden of pathogens and pests on major food crops. Nat Ecol Evol 3:430–439 FAO. The impact of disasters on agriculture and food security 2023. In The impact of disasters on agriculture and food security (2023) (2023) Getahun S, Kefale H, Gelaye Y (2024) Application of precision agriculture technologies for sustainable crop production and environmental sustainability: A systematic review. The Scientific World Journal 2126734 (2024) Lobell DB, Azzari G, Burke M, Gourlay S, Jin Z, Kilic T, Murray S (2020) Eyes in the sky, boots on the ground: Assessing satellite-and ground‐based approaches to crop yield measurement and analysis. Am J Agric Econ 102:202–219 Van Voorn GA, Boer MP, Truong SH, Friedenberg NA, Gugushvili S, McCormick R, Bustos Korts D, Messina CD, Van Eeuwijk FA (2023) A conceptual framework for the dynamic modeling of time-resolved phenotypes for sets of genotype-environment-management combinations: a model library. Front Plant Sci 14:1172359 Balboa GR, Archontoulis S, Salvagiotti F, Garcia FO, Stewart W, Francisco E, Prasad PV, Ciampitti IA (2019) A systems-level yield gap assessment of maize-soybean rotation under high-and low-management inputs in the Western US Corn Belt using APSIM. Agric Syst 174:145–154 Beegum S, Timlin D, Reddy KR, Reddy V, Sun W, Wang Z, Fleisher D, Ray C (2023) Improving the cotton simulation model, GOSSYM, for soil, photosynthesis, and transpiration processes. Sci Rep 13:7314 Chang Y, Latham J, Licht M, Wang L (2023) A data-driven crop model for maize yield prediction. Commun biology 6:439 Chapagain SK, Aryal A, Mohan G, Shrestha S, Mishra BK, Fukushi K (2022) Analysis of the climate change impact on water availability and the links between water pollution and economy for sustainable water resource management in Kaski District, Nepal. J Water Clim Change 13:3030–3045 Corbeels M, Chirat G, Messad S, Thierfelder C (2016) Performance and sensitivity of the DSSAT crop growth model in simulating maize yield under conservation agriculture. Eur J Agron 76:41–53 Ech-Chatir L, Er-Raki S, Rodriguez JC, Meddich A (2025) Advances in crop growth modeling: A review of perennial crop and beneficial soil microorganism approaches. Agric Water Manage 315:109548 Hearn A (1994) OZCOT: a simulation model for cotton crop management. Agric Syst 44:257–299 Jones JW, He J, Boote KJ, Wilkens P, Porter CH, Hu Z (2011) Estimating DSSAT cropping system cultivar-specific parameters using Bayesian techniques. Methods introducing Syst models into agricultural Res 2:365–393 Jones JW, Hoogenboom G, Porter CH, Boote KJ, Batchelor WD, Hunt LA, Wilkens PW, Singh U, Gijsman AJ, Ritchie JT (2003) The DSSAT cropping system model. Eur J Agron 18:235–265 Keating BA, Carberry PS, Hammer GL, Probert ME, Robertson MJ, Holzworth D, Huth NI, Hargreaves JN, Meinke H, Hochman Z (2003) An overview of APSIM, a model designed for farming systems simulation. Eur J Agron 18:267–288 Timlin D, Paff K, Han E (2024) The role of crop simulation modeling in assessing potential climate change impacts. Agrosystems Geosci Environ 7:e20453 Zhang N, Zhou X, Kang M, Hu B-G, Heuvelink E, Marcelis LF (2023) Machine learning versus crop growth models: an ally, not a rival. AoB Plants 15:plac061 Banerjee K, Dutta S, Das S, Sadhukhan R (2024) Crop simulation models as decision tools to enhance agricultural system productivity and sustainability–a critical review. Technol Agron 5 Abbas G, Fatima Z, Tariq M, Ahmed M, Rahman MHu, Nasim W, Rasul G, Ahmad S (2020) Applications of crop modeling in cotton production. Cotton Production and Uses: Agronomy, Crop Protection, and Postharvest Technologies. Springer Droutsas I, Challinor AJ, Deva CR, Wang E (2022) Integration of machine learning into process-based modelling to improve simulation of complex crop responses. silico Plants 4:diac017 Mba PC, Njoku JN, Uyeh DD (2025) Enhancing resilience in specialty crop production in a changing climate through smart systems adoption. Smart Agricultural Technol, 100897 De Alwis S, Hou Z, Zhang Y, Na MH, Ofoghi B, Sajjanhar A (2022) A survey on smart farming data, applications and techniques. Comput Ind 138:103624 Huete A, Didan K, Miura T, Rodriguez EP, Gao X, Ferreira LG (2002) Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sens Environ 83:195–213 Huete AR (1988) A soil-adjusted vegetation index (SAVI). Remote Sens Environ 25:295–309 Gitelson AA, Kaufman YJ, Merzlyak MN (1996) Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens Environ 58:289–298 Sarkar S, Cazenave AB, Oakes J, McCall D, Thomason W, Abbot L, Balota M (2020) High-throughput measurement of peanut canopy height using digital surface models. Plant Phenome J 3:e20003 Sarkar S, Cazenave AB, Oakes J, McCall D, Thomason W, Abbot L, Balota M (2021) Aerial high-throughput phenotyping of peanut leaf area index and lateral growth. Sci Rep 11:1–17 Chapu I, Okello DK, Okello RCO, Odong TL, Sarkar S, Balota M (2022) Exploration of Alternative Approaches to Phenotyping of Late Leaf Spot and Groundnut Rosette Virus Disease for Groundnut Breeding. Front Plant Sci 13:912332 Lobell DB, Asner GP (2003) Climate and management contributions to recent trends in US agricultural yields. Science 299:1032–1032 Lobell DB, Schlenker W, Costa-Roberts J (2011) Climate trends and global crop production since 1980. Science 333:616–620 Bolton DK, Friedl MA (2013) Forecasting crop yield using remotely sensed vegetation indices and crop phenology metrics. Agric For Meteorol 173:74–84 Sarkar S, Ramsey AF, Cazenave A-B, Balota M (2021) Peanut leaf wilting estimation from RGB color indices and logistic models. Front Plant Sci 12:658621 Sarkar S, Osorio Leyton J, Noa Yarasca E, Adhikari K, Hajda CB, Smith DR (2025) Integrating Remote Sensing and Soil Features for Enhanced Machine Learning-Based Corn Yield Prediction in the Southern US. Sensors Tucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens Environ 8:127–150 Zhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JC, Gao F, Reed BC, Huete A (2003) Monitoring vegetation phenology using MODIS. Remote Sens Environ 84:471–475 Rouse J, Haas R, Schell J, Deering D (1974) Monitoring vegetation systems in the great plains with ERTS proceeding. In: Third earth reserves technology satellite symposium, greenbelt: NASA SP-351 ) Justice CO, Townshend JR, Vermote EF, Masuoka E, Wolfe RE, Saleous N, Roy DP, Morisette JT (2002) An overview of MODIS Land data processing and product status. Remote Sens Environ 83:3–15 Huang J, Tian L, Liang S, Ma H, Becker-Reshef I, Huang Y, Su W, Zhang X, Zhu D, Wu W (2015) Improving winter wheat yield estimation by assimilation of the leaf area index from Landsat TM and MODIS data into the WOFOST model. Agric For Meteorol 204:106–121 Johnson DM (2014) An assessment of pre-and within-season remotely sensed variables for forecasting corn and soybean yields in the United States. Remote Sens Environ 141:116–128 Prasad AK, Chai L, Singh RP, Kafatos M (2006) Crop yield estimation model for Iowa using remote sensing and surface parameters. Int J Appl Earth Obs Geoinf 8:26–33 Ji Z, Pan Y, Zhu X, Wang J, Li Q (2021) Prediction of crop yield using phenological information extracted from remote sensing vegetation index. Sensors 21:1406 Sarkar S, Oakes J, Cazenave A-B, Burow MD, Bennett RS, Chamberlin KD, Wang N, White M, Payton P, Mahan J (2022) Evaluation of the US Peanut Germplasm Mini-Core Collection in the Virginia-Carolina Region Using Traditional and New High-Throughput Methods. Agronomy 12, 1945 Balota M, Sarkar S, Bennett RS, Burow MD (2024) Phenotyping Peanut Drought Stress with Aerial Remote-Sensing and Crop Index Data. Agriculture 14:565 Gamon JA, Field CB, Goulden ML, Griffin KL, Hartley AE, Joel G, Penuelas J, Valentini R (1995) Relationships between NDVI, canopy structure, and photosynthesis in three Californian vegetation types. Ecol Appl 5:28–41 Tan C-W, Zhang P-P, Zhou X-X, Wang Z-X, Xu Z-Q, Mao W, Li W-X, Huo Z-Y, Guo W-S, Yun F (2020) Quantitative monitoring of leaf area index in wheat of different plant types by integrating NDVI and Beer-Lambert law. Sci Rep 10:929 Rouse JW Jr., Haas RH, Schell JA, Deering DW (1974) Monitoring Vegetation Systems in the Great Plains with Erts. In: NASA Special Publication (eds Freden SC, Mercanti EP, Becker MA) Zheng H, Shen L, Tang A, Luo Y, Hu H, Du B, Wen Y, Tao D (2025) Learning from models beyond fine-tuning. Nat Mach Intell 7:6–17 Lei T, Bai J, Brahma S, Ainslie J, Lee K, Zhou Y, Du N, Zhao V, Wu Y, Li B (2023) Conditional adapters: Parameter-efficient transfer learning with fast inference. Adv Neural Inf Process Syst 36:8152–8172 Borzunov A, Ryabinin M, Chumachenko A, Baranchuk D, Dettmers T, Belkada Y, Samygin P, Raffel CA (2023) Distributed inference and fine-tuning of large language models over the internet. Adv Neural Inf Process Syst 36:12312–12331 Ding N, Qin Y, Yang G, Wei F, Yang Z, Su Y, Hu S, Chen Y, Chan C-M, Chen W (2023) Parameter-efficient fine-tuning of large-scale pre-trained language models. Nat Mach Intell 5:220–235 Camps-Valls G, Svendsen DH, Cortés-Andrés J, Mareno-Martínez Á, Pérez-Suay A, Adsuara J, Martín I, Piles M, Muñoz-Marí J, Martino L (2021) Physics-aware machine learning for geosciences and remote sensing. In: 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS ). IEEE Khan MA, Wahid A, Ahmad M, Tahir MT, Ahmed M, Ahmad S, Hasanuzzaman M (2020) World cotton production and consumption: An overview. Cotton production and uses: Agronomy, crop protection, and postharvest technologies , 1–7 Sun L, Wang Z, Xiong C, Gu J, Zheng Y, Ju F, Wang S, Hu W, Zhao W, Zhou Z (2024) Improving the soil K+/Na+ ratio under moderate salt stress synergistically increases the yield and quality of cotton fiber and cottonseed. Ind Crops Prod 213:118441 Shahzad AN, Rizwan M, Asghar MG, Qureshi MK, Bukhari SAH, Kiran A, Wakeel A (2019) Early maturing Bt cotton requires more potassium fertilizer under water deficiency to augment seed-cotton yield but not lint quality. Sci Rep 9:7378 Chen Z, Ma H, Xia J, Hou F, Shi X, Hao X, Hafeez A, Han H, Luo H (2017) Optimal pre-plant irrigation and fertilization can improve biomass accumulation by maintaining the root and leaf productive capacity of cotton crop. Sci Rep 7:17168 Ashraf S, Ch KM, Ashraf I, Akbar N (2024) A phenomenological inquiry into farmers’ experiences growing cotton in Punjab, Pakistan. Sci Rep 14:13846 Hightower M, USDA (2022) Drought drops U.S. cotton harvest to be lowest level since the 19th century. University of Arkansas, Division of Agriculture, Research and Extension Beegum S, Reddy V, Reddy KR (2023) Development of a cotton fiber quality simulation module and its incorporation into cotton crop growth and development model: GOSSYM. Comput Electron Agric 212:108080 Lin S, Deng M, Wei K, Wang Q, Su L (2023) A new regional cotton growth model based on reference crop evapotranspiration for predicting growth processes. Sci Rep 13:7368 Palla S, Sarkar S, Zhao L, Landivar J, Scott JL, Maeda M, McGinty J, Abugho S, Jung J, Rajan N (2025) Growth analysis of cotton using UAS derived multi temporal canopy features. Sci Rep Gwathmey CO, Tyler DD, Yin X (2010) Prospects for monitoring cotton crop maturity with normalized difference vegetation index. Agron J 102:1352–1360 Li H, Lascano R, Barnes E, Booker J, Wilson T, Bronson K, Segarra E (2001) Temporal patterns of cotton reflectance and NDVI-days lint yield modeling. In: Proc. Beltwide Cotton Conf ) Liu F, Xiao X, Qin Y, Liu L (2025) Identifying and mapping open-boll cotton fields in the southeastern United States with time series PhenoCam, Sentinel-2 and Sentinel-1 images. Int J Appl Earth Obs Geoinf 142:104723 Stone K, Billman ED, Bauer PJ, Sigua G (2022) Using NDVI for variable rate cotton irrigation prescriptions. Appl Eng Agric 38:787–795 Wang Y, Zhang Z, Xiao C, Zhang T, Yu K, Zhang C, Liao Q, Li F, Wan S, Chen G (2026) Characterizing Cotton Defoliation Progress via UAV-Based Multispectral-Derived Leaf Area Index and Analysis of Influencing Factors. Remote Sens 18:609 Van Klompenburg T, Kassahun A, Catal C (2020) Crop yield prediction using machine learning: A systematic literature review. Comput Electron Agric 177:105709 Cai Y, Guan K, Lobell D, Potgieter AB, Wang S, Peng J, Xu T, Asseng S, Zhang Y, You L (2019) Integrating satellite and climate data to predict wheat yield in Australia using machine learning approaches. Agric For Meteorol 274:144–159 Liu Z, Xu W, Feng J, Palaiahnakote S, Lu T Context-aware attention LSTM network for flood prediction. In: (2018) 24th international conference on pattern recognition (ICPR) ). IEEE (2018) Liu J, Wang G, Hu P, Duan L-Y, Kot AC (2017) Global context-aware attention lstm networks for 3d action recognition. In: Proceedings of the IEEE conference on computer vision and pattern recognition ) Additional Declarations There is NO Competing Interest. Supplementary Files CropCast.zip CropCast.zip Supplementary.docx Supplementary Information for CropCast: Context-aware AI framework for in-season crop growth forecasting Cite Share Download PDF Status: Under Review 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-9348900","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":632397346,"identity":"17e9a5ca-1a35-47a8-aaa0-8f607dcef60f","order_by":0,"name":"Ramana Heggadal Math","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA; Texas A\u0026M Institute of Data Science (TAMIDS), Texas A\u0026M University, College Station, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Ramana","middleName":"Heggadal","lastName":"Math","suffix":""},{"id":632397347,"identity":"bcb73585-c275-48ba-9978-cc7f4586f664","order_by":1,"name":"Ashutosh Tiwari","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Ashutosh","middleName":"","lastName":"Tiwari","suffix":""},{"id":632397348,"identity":"2c312b4f-257a-459c-8e8c-0305eb08c30e","order_by":2,"name":"Lei Zhao","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Lei","middleName":"","lastName":"Zhao","suffix":""},{"id":632397349,"identity":"a02e6894-cd39-49dc-b53a-e1810e4bc3f0","order_by":3,"name":"Juan Landivar","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Juan","middleName":"","lastName":"Landivar","suffix":""},{"id":632397350,"identity":"f2e8958a-765f-4e01-810c-29a5acf15c74","order_by":4,"name":"Yuri Calil","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Extension, Corpus Christi, Texas, USA; Department of Agriculture Economics, Texas A\u0026M University, College Station, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Yuri","middleName":"","lastName":"Calil","suffix":""},{"id":632397351,"identity":"e17a0da0-e5b3-4872-8b20-804f03578c7e","order_by":5,"name":"Sayantan Sarkar","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Sayantan","middleName":"","lastName":"Sarkar","suffix":""},{"id":632397352,"identity":"ec739477-52c8-4e37-b669-870252e19f15","order_by":6,"name":"Jose Landivar Scott","email":"","orcid":"https://orcid.org/0000-0002-6327-6093","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Jose","middleName":"Landivar","lastName":"Scott","suffix":""},{"id":632397353,"identity":"95375d63-f8e2-4d4d-b7f0-f08b4f77ebe4","order_by":7,"name":"Benjamin Ghansah","email":"","orcid":"","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA; Texas A\u0026M University-Corpus Christi, Corpus Christi, Texas, USA","correspondingAuthor":false,"prefix":"","firstName":"Benjamin","middleName":"","lastName":"Ghansah","suffix":""},{"id":632397345,"identity":"2697a33d-53aa-4561-8140-6e63ecf4b4cd","order_by":8,"name":"Mahendra Bhandari","email":"data:image/png;base64,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","orcid":"https://orcid.org/0000-0001-8450-2590","institution":"Texas A\u0026M AgriLife Research, Corpus Christi, Texas, USA; Department of Soil and Crop Sciences, Texas A\u0026M University, College Station, Texas","correspondingAuthor":true,"prefix":"","firstName":"Mahendra","middleName":"","lastName":"Bhandari","suffix":""}],"badges":[],"createdAt":"2026-04-07 19:35:52","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9348900/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9348900/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":108542362,"identity":"7720c365-b859-43d2-9843-bb6fd0e4efb0","added_by":"auto","created_at":"2026-05-05 19:09:23","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":892466,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 1. Proposed CropCast framework. Beginning with satellite image retrieval and pre-processing, the framework generates input datasets from in-season data and feeds the context aware encoder-decoder LSTM model which updates itself during the training, finally providing forecasted NDVI for the complete season. The forecasted estimates can be effectively used for several applications requiring crop growth behavior, such as yield prediction.\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/72bd28e00381600ffc945287.jpeg"},{"id":108542330,"identity":"ca8cd52a-4548-4579-b7be-81aab14ddc4c","added_by":"auto","created_at":"2026-05-05 19:09:12","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":924217,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 2. CropCast model testing across different test fields and across different years. Blue and red regions represent the input and forecasted data, respectively. The grey color zone represents the online fine-tuning period. A) location of the test cotton fields marked over Texas map, b), c) and d) show enlarged views of the three cotton fields. Panels Bin, Cin and Din show the interannual variability for available years of ground truth data used to prepare CropCast input. B1) to B4) shows NDVI forecasts for cotton field in Sinton, TX with input NDVI data corresponding to 51, 81, 93 and 117 days after planting. C1) to C4) shows NDVI forecasts for cotton field in Etter, TX with input NDVI data corresponding to 51, 63, 81 and 111 days after planting. D1) to D4) shows NDVI forecasts for cotton field in Driscoll, TX with input NDVI data corresponding to 57, 81, 105 and 117 days after planting.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/d59e9e79198b3ce2de33a94b.png"},{"id":108542361,"identity":"1620c501-c4d5-4f0c-9318-80286595e7a1","added_by":"auto","created_at":"2026-05-05 19:09:23","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3093442,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFigure 3. CropCast test performance scores. \u003c/strong\u003ea) to d) shows mean absolute errors (MAE) for forecasted NDVI for cotton field in Sinton, TX with input NDVI data corresponding to 51, 81, 93 and 117 days after planting. e) to h) shows corresponding MAE for cotton field in Etter, TX with input NDVI data corresponding to 51, 63, 81 and 111 days after planting, and i) to l) shows MAE for cotton field in Driscoll, TX with input NDVI data corresponding to 57, 81, 105 and 117 days after planting.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/c70f3f3c215e0526916466a0.png"},{"id":108804154,"identity":"43f64e1c-c5ad-4ce4-8c54-e252ab4fbbf9","added_by":"auto","created_at":"2026-05-08 15:16:44","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":16643674,"visible":true,"origin":"","legend":"\u003cp\u003eFigure 4: Application of CropCast for yield prediction and marketing. (A), (B) and (C) show the average predicted yield, average breakeven margin percentage, and profitability, respectively for the model predictions. Spatial maps highlighting grid level details on predictions at different stages of crop growth are shown from A1 to A6 for yield, B1 to B6 for breakeven margin, and C1 to C6 for profitability.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/78d68d6371f8dc34c494a402.png"},{"id":108805265,"identity":"bf469de3-2acf-431a-8aa7-ed25a9a1b96d","added_by":"auto","created_at":"2026-05-08 15:25:24","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":122827,"visible":true,"origin":"","legend":"\u003cp\u003eFigure M1. Processing input sequence in CropCast context-aware LSTM model. The upper part shows the processing of one input-target sequence, and the lower part shows multiple random windows sampled per input data curve.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/15f843131f7f40fcfd8abf84.png"},{"id":108542513,"identity":"87c2df7d-4e2e-4fd2-8f97-c57721f58660","added_by":"auto","created_at":"2026-05-05 19:10:15","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":66959,"visible":true,"origin":"","legend":"\u003cp\u003eFigure M2. Auto-regressed steps using ground truth (green) and previous predictions (gray).\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/74531a79e4be9c3a66e36230.png"},{"id":108542336,"identity":"47818343-0994-4505-912b-a8f2fd6853b6","added_by":"auto","created_at":"2026-05-05 19:09:18","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":158665,"visible":true,"origin":"","legend":"\u003cp\u003eFigure M3. Input data handling procedure for proposed context-aware continual finetuning encoder-decoder LSTM model in CropCast.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/e4fc3a7a34e1ca97d2b03b22.png"},{"id":109081484,"identity":"53bae882-fc9b-49a9-951d-1a841cd8cc8a","added_by":"auto","created_at":"2026-05-12 12:18:58","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":23819987,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/20d7dfbb-271b-45c6-90e7-8634361be9e5.pdf"},{"id":108542322,"identity":"bb09b45b-4221-4539-b6eb-5cd232976555","added_by":"auto","created_at":"2026-05-05 19:09:09","extension":"zip","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":4819480,"visible":true,"origin":"","legend":"CropCast.zip","description":"","filename":"CropCast.zip","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/40074677b3cfba1760204dbb.zip"},{"id":108542473,"identity":"f0e67363-71e8-426c-a647-a65bc0fa4439","added_by":"auto","created_at":"2026-05-05 19:09:54","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":13740757,"visible":true,"origin":"","legend":"Supplementary Information for CropCast: Context-aware AI framework for in-season crop growth forecasting","description":"","filename":"Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-9348900/v1/b0332ed48a39e6307c021868.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"CropCast: Context-aware AI framework for in-season crop growth forecasting","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAgricultural systems are under increasing pressure to produce more food and fiber while coping with climate variability, environmental degradation, and resource constraints. The global population is projected to reach 9.7\u0026nbsp;billion by 2050, driving a sharp increase in demand for agricultural production \u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e, \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Agriculture remains one of the largest sources of employment worldwide, with approximately 26% of the global workforce (~\u0026thinsp;916\u0026nbsp;million people) directly employed in agriculture, while nearly half of the global population depends on agrifood systems for their livelihoods\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003c/sup\u003e. In many developing economies, agriculture contributes more than 25% of national gross domestic product, underscoring its importance for economic stability and rural development. Further, crop production remains highly vulnerable to environmental variability and management uncertainty. Global estimates suggest that up to 40% of potential crop production is lost annually to pests and plant diseases, representing economic losses exceeding USD 220\u0026nbsp;billion per year\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e. Climate-related disasters further exacerbate these impacts, with droughts, floods, and storms causing approximately USD 3.8 trillion in agricultural production losses globally over the past three decades\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eFarmers and agronomists require timely and reliable information on how crops will grow, how they may respond to stress, and what yields can be expected for day-to-day management interventions \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. At regional and national levels, such information supports policy decisions, stabilizes supply chains, and guides disaster response \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. However, accurately modeling crop growth and development is challenging since these are dynamic processes shaped by interactions among genetics, environment, and management\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Since the 1960s, numerous process-based models have been developed to simulate crop growth using inputs such as weather, soil characteristics, and management practices\u003csup\u003e\u003cspan additionalcitationids=\"CR10 CR11 CR12 CR13 CR14 CR15 CR16 CR17 CR18 CR19\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. These models have supported applications ranging from yield estimation and climate impact assessments to irrigation scheduling and regional production forecasting \u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e, \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. However, their operational deployment for large-scale crop monitoring is often constrained by extensive input data requirements, complex site-specific parameter calibration, and high computational costs\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. As a result, producers, consultants, and extension agents continue to seek accessible tools that provide timely insight into future crop growth trajectories, yield potential, and management needs for their timely interventions\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRemote sensing is increasingly used to monitor crop growth and estimate yield using satellite, UAV, and terrestrial sensors\u003csup\u003e\u003cspan additionalcitationids=\"CR27 CR28 CR29 CR30\" citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. Compared to in-situ measurements and airborne remote sensing, satellite observations enable large-scale monitoring of crop growth through repeated observations of vegetation dynamics, while UAV and ground systems offer higher spatial resolution but are often limited by operational cost, scalability, and processing requirements\u003csup\u003e\u003cspan additionalcitationids=\"CR33 CR34 CR35\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. Multispectral satellite images provide spatially continuous measurements of canopy reflectance that can be used to derive indicators of plant vigor and physiological activity. Remote sensing studies commonly use vegetation indices derived from spectral bands to estimate crop properties such as biomass, leaf area index, and chlorophyll content\u003csup\u003e\u003cspan additionalcitationids=\"CR38\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. Widely used indices include the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil Adjusted Vegetation Index (SAVI), and Green NDVI (GNDVI), which help reduce atmospheric effects, soil background influence, and saturation in dense vegetation canopies\u003csup\u003e\u003cspan additionalcitationids=\"CR39 CR40 CR41 CR42 CR43 CR44 CR45\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. In addition to vegetation indices, several studies have used raw spectral bands from satellite sensors such as Landsat, MODIS, and Sentinel directly as input features for crop monitoring and yield prediction models\u003csup\u003e\u003cspan additionalcitationids=\"CR33\" citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e, \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. Among these indicators, NDVI remains the most widely adopted vegetation metric because of its simplicity, strong relationship with canopy biomass and leaf area index, and long historical record across satellite platforms\u003csup\u003e\u003cspan additionalcitationids=\"CR38\" citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e. Time-series NDVI observations capture key phenological transitions including emergence, peak canopy development, and senescence, making them particularly useful for monitoring crop growth trajectories across seasons and regions\u003csup\u003e\u003cspan additionalcitationids=\"CR39 CR40 CR41 CR42 CR43\" citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe unprecedented rise in the availability of satellite observations has also stimulated the development of machine learning methods for crop growth analysis and yield prediction. Early studies applied statistical time-series models such as Autoregressive Integrated Moving Average (ARIMA) and Kalman filters to forecast vegetation dynamics from historical observations, performed well under stationary seasonal patterns \u003csup\u003e\u003cspan additionalcitationids=\"CR44 CR45 CR46 CR47\" citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Other, machine learning algorithms such as random forests, support vector regression, and gradient boosting have been used to model nonlinear relationships between satellite observations and crop productivity\u003csup\u003e\u003cspan additionalcitationids=\"CR50 CR51 CR52\" citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. Deep learning approaches further improved the predictive capability by capturing spatial and temporal dependencies in remote sensing data. Convolutional neural networks have been applied to extract spatial features from multispectral imagery, while recurrent neural networks such as long short-term memory (LSTM), gated recurrent unit (GRU) and Transformer‑based architectures have further improved the modeling of temporal dynamics and spatial context architectures have demonstrated strong performance in modeling crop growth dynamics from satellite time series\u003csup\u003e\u003cspan additionalcitationids=\"CR55 CR56 CR57\" citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u003c/sup\u003e. Hybrid frameworks that integrate satellite observations with climate data have also been proposed for regional crop yield prediction\u003csup\u003e\u003cspan additionalcitationids=\"CR60 CR61 CR62 CR63 CR64 CR65 CR66\" citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e\u003c/sup\u003e. Despite their accurate performance, these models generally assume stationary conditions during inference and require additional fine‑tuning and retraining to infer accurately over new in‑season observations. Static models with fixed parameters degrade rapidly under weather extremes, novel management practices, or evolving cropping systems\u003csup\u003e\u003cspan additionalcitationids=\"CR48\" citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e. Further, retraining models with new data arrival is computationally expensive and operationally infeasible\u003csup\u003e\u003cspan additionalcitationids=\"CR51 CR52\" citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. Moreover, model robustness across growth stages and environmental variability remains underexplored, leading to degraded forecasting stability\u003csup\u003e\u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u003c/sup\u003e. Collectively, these limitations require a forecasting system that is adaptive, scalable, efficient, and resilient to the complexity of agricultural landscapes.\u003c/p\u003e \u003cp\u003eTo address the limitations of existing forecasting approaches, we introduce CropCast, a context-aware, continual autoregressive framework for adaptive crop growth forecasting using satellite remote sensing. To be operational and directly usable by farmers, CropCast uses satellite-derived NDVI measurements derived from high resolution (~\u0026thinsp;3m) PlanetScope satellite imagery as input. Cotton provides an ideal case study for developing such forecasting systems and is used as a test crop in this study. As one of the most important fiber crops globally, cotton production requires multiple management decisions throughout the growing season, including irrigation scheduling, nutrient management, pest control, and harvest timing\u003csup\u003e\u003cspan additionalcitationids=\"CR56 CR57 CR58\" citationid=\"CR55\" class=\"CitationRef\"\u003e55\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e59\u003c/span\u003e\u003c/sup\u003e. Despite these efforts, a significant portion of cotton acreage often goes unharvested, sometimes exceeding 20\u0026ndash;40% during drought years, due to environmental and economic challenges\u003csup\u003e\u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e60\u003c/span\u003e\u003c/sup\u003e. For example, in 2022, Texas, USA experienced a 69% abandonment rate, primarily driven by drought, insect pressure, and low market prices. In-season crop growth forecasts and subsequent yield potential estimates can help producers optimize input use, reduce production costs, and manage market risk. Reliable early-season forecasts also enable growers to take advantage of pre-harvest marketing opportunities such as forward contracting or hedging strategies\u003csup\u003e\u003cspan additionalcitationids=\"CR62\" citationid=\"CR61\" class=\"CitationRef\"\u003e61\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eTo develop CropCast, we develop an encoder-decoder LSTM architecture which models temporal dependencies and forecasts future time series. Figure\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e shows the overall CropCast framework comprising satellite image retrieval, preprocessing, modeling and forecast, and application (yield forecasting and marketing). The framework integrates an encoder-decoder long short-term memory architecture with phenological context variables including normalized days after planting and field-level vegetation statistics to capture both short-term growth dynamics and seasonal crop development patterns. Unlike conventional forecasting approaches that generate static predictions, CropCast produces autoregressive NDVI forecasts in incremental time windows and continually updates predictions as new satellite observations become available during the growing season, targeting predictions across diverse locations, seasons, and management regimes. The study aims to provide three contributions. First, we develop a continual forecasting framework that enables adaptive in-season crop growth prediction without retraining the full model. Second, we demonstrate that CropCast produces stable NDVI forecasts across seasons and locations beginning approximately 50 days after planting. Third, we show that predicted growth trajectories can be integrated into hyperlocal yield prediction and economic analysis, generating spatially explicit yield and profitability estimates at 9 m \u0026times; 9 m resolution that support data-driven agricultural decision-making.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eOverview of time series data, modelling, and seasonal variations\u003c/h2\u003e \u003cp\u003eWe performed data exploratory analysis and modeling with benchmark machine learning models to arrive at the CropCast architecture. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea shows the location of the cotton fields used in this study. These fields named Sinton, Etter and Driscoll fields, lie in different cotton growing regions of the state, and therefore suitable for evaluating model transferability in space. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB to D show the enlarged views for the Sinton, Etter and Driscoll fields respectively, obtained from PlanetScope satellite images, used for model training. For the Driscoll field, NDVI curves are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eD for four cropping year seasons (2021\u0026ndash;2024), with shaded region showing the variance of the NDVI observations. Out of four years, data for the years 2021 to 2023 are used for training, while that of year 2024 is used for testing. While the general shape is bell‑like, the timing, slope, and amplitude of these phases vary with cultivar, planting date, management practices, and environmental stresses\u003csup\u003e\u003cspan additionalcitationids=\"CR64\" citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e\u003c/sup\u003e. For example, delayed planting compresses the early growth phase, while drought can shorten the peak phase and accelerate decline. At early growth stage with minimal vegetation cover, NDVI has low (~\u0026thinsp;0.2) values\u003csup\u003e\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e\u003c/sup\u003e. As plants emerge and leaf area expands, NDVI rises rapidly, reflecting increased chlorophyll content and canopy closure. This stage is often sensitive to planting date, seed quality, and early irrigation or fertilization decisions. Following the rapid growth phase, NDVI reaches a plateau or peak, typically corresponding to the crop\u0026rsquo;s maximum LAI and highest photosynthetic capacity\u003csup\u003e\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e\u003c/sup\u003e. This stage is associated with flowering and reproductive development. In well‑managed fields, NDVI values can remain at or near the peak for several weeks, depending on crop type and environmental conditions. As crops progress toward maturity and harvest, leaf senescence begins. Chlorophyll degradation and canopy thinning further lead to a gradual decline in NDVI\u003csup\u003e\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e\u003c/sup\u003e. Moreover, satellite observations are often affected by errors and offsets due to sensor geometry and atmospheric effects, radiometric calibration and noise, showing signal responses different than the theoretical and physical response of the crop growth curves. To reduce some of these effects, we aggregate native 3m x 3m pixels of the PlanetScope images into 9m x 9m grids by averaging reflectance values. This spatial aggregation reduces high-frequency noise caused by shadows, artifacts, and small-scale vegetation heterogeneity. We further propose using log-normal functions to fit the remote sensing observations and generate a daily NDVI growth curve resembling closeness to the physical crop growth behavior. To enhance contextual learning over time, we proposed the model to be trained on variable-length windows capturing local dynamics within the broader phenological cycle. Through the above, we create daily NDVI data for modeling and ensure that forecasts align with real physiological processes, improving interpretability and robustness.\u003c/p\u003e \u003cp\u003eWe begin our model design with a simple feedforward deep neural network trained with NDVI inputs from 1 to 60 days after planting (DAP) and train the model to regress NDVI for the remaining DAPs from fixed-length NDVI inputs of length 40. As expected, this model performs poorly across both temporal and spatial shifts (Supplementary Figures S2 A to C) when experimenting with different dropout ratios for combating overfitting. The model fails to capture the sequential nature of crop growth, making it unsuitable for modeling seasonal dynamics. To address this, we replace the feedforward architecture with a sequence-to-sequence long short-term memory (LSTM) network, which is better suited for learning temporal dependencies. Supplementary Figure S3 shows the forecasting outcomes with this model, showing significant improvement on validation data, i.e., unseen data points from the training years (2020\u0026ndash;2022). It produces NDVI curves that capture the broad seasonal NDVI trajectory for held-out validation samples. However, when evaluated on the 2023 season, performance drops sharply (Supplementary Figure S3B). The predicted trajectories exhibit reduced spread relative to the observed NDVI, indicating that the model underestimates year-specific variability and does not generalize well to unseen seasonal conditions. This highlights a well-known challenge in cross-year time-series forecasting over the same cotton fields as year-specific environmental and agronomic signals reshape NDVI trajectories, producing distinct growth signatures that static models struggle to capture.\u003c/p\u003e \u003cp\u003eTo address the limitation of cross-year forecasting failure, we consider two approaches. The first approach is to significantly increase model complexity, e.g., adopting pre-trained Transformer architectures and hoping they might learn these shifts implicitly. The other is to make the model aware of the current season\u0026rsquo;s trend (context) and update its predictions accordingly. Initial Transformer experiments underperformed; despite their capacity, they generalized worse than simpler LSTMs (Supplementary Figure S2D). To enable context-aware learning, we find that using a few recent NDVI observations to adapt the model on-the-fly, we could forecast NDVI for the next 40\u0026ndash;45 DAPs with high accuracy and maintain acceptable predictions. Motivated by this, we explored continual adaptive fine-tuning as a way to adapt the model to the current season\u0026rsquo;s trend. Our initial attempt involved fine-tuning the pretrained model using both recent and future NDVI observations. Since the base model was trained to predict the full remaining sequence, we assumed this setup will align the forecasts. While this approach did capture the overall trend, it produced unstable predictions, curves often began with a downward bump, exhibited unnatural dips near the end, or failed to progressively align as more data arrived (Supplementary Figure S2). We realized the issue stemming from reliance on the fine-tuning step to fix a base model that learned to stretch across the full time series, rather than generalizing stable local dynamics. Hence, we decided to train the base model from scratch, making it retain the stable growth patterns across years that are transferable regardless of season-specific shifts. Instead of predicting the full NDVI curve in one go, CropCast learns to autoregress in smaller time chunks of incoming data (30\u0026ndash;40 DAP time steps), progressively extending the curve. We structured training using chunked input sequences and variable-length inputs, allowing the model to specialize in immediate-term extension while deferring longer-term adjustment to future updates. Noticeably, we perform online fine-tuning only on the recent 5-day interval using ground truth information (NDVI) from that period without requiring backpropagation through the entire future sequence. Detailed explanation on the CropCast model architecture and other models used for comparison is provided in the Methods section.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eTransferability of CropCast across space and time\u003c/h3\u003e\n\u003cp\u003eWith the above formulation, we trained CropCast on one cotton field (Driscoll) with three years of observation (2021\u0026ndash;2023) for training and validation, and it required minor hyperparameter tuning to adapt to the differences in the linear, peak and senescence stages, demonstrating robustness to both temporal and regional variability. The test procedure involved three stages, (i) testing a nearby cotton field with similar weather and crop growth pattern, (ii) testing cross-year for a far cotton field with varied weather and climate characteristics, and (iii) testing cross-year performance over limited satellite data availability. We also experimented with varying input permutations, such as excluding group statistics or normalized DAP, to assess feature impact. Additionally, for the cross-region evaluation, we tested early forecasting for the Etter field starting from 40 DAPs onward to simulate in-season prediction scenarios. Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e panels Bin, Cin and Din show the full season NDVI curves for these fields and B1 to B4, C1 to C4 and D1 to D4 show corresponding predictions obtained from CropCast for different times during the cropping season. For the first stage, we observe that the model adapted well to incoming data (Panels B1 to B4 in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) for a nearby cotton field named in Sinton, Texas. We show the forecasts at four progressive dates in the cropping season (51, 81, 93 and 117 DAPs), demonstrating progressive in-season model predictions. CropCast forecasts align closely with ground truth NDVI data starting\u0026thinsp;~\u0026thinsp;50 DAPs. However, forecasts with NDVI inputs from 81 DAP result in slight overestimation, improving incrementally with later observations. For the second stage, we test the Etter field, for which we notice slight overfitting in near future forecasts (which gets stable later) for NDVI inputs with DAP 51 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eC1). In this case, the forecasts improve with further in-season NDVI inputs (63, 81, 111) as shown in Figs.\u0026nbsp;2C2 to C4. The third stage test data is for the Driscoll field and year 2024, which had a long input data gap owing to cloud cover presence and rain during acquisitions, further challenging the forecasting ability. Figures\u0026nbsp;2D1 to D4 show the forecasts, where we again observe slight overestimation for input data corresponding to DAP 81, while getting improvements afterwards. In all the three testing stages, we observe slight overestimation of the NDVI forecasts, which improve as more in-season data arrives as input.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFigure \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e shows the performance metrics for the three test stages. The mean absolute errors lie between the ranges 0.05 to 0.15 for Sinton (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA to D), 0.025 to 0.175 for Etter (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eE to H), and 0.025 to 0.175 for Driscoll (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eI to L). We observe that for the first and third stages, the MAE values rise after the second set of NDVI inputs arrive (~\u0026thinsp;80 DAP) and reduces thereafter. For the second stage, we, however, see constant decline in MAE values as more in-season NDVI data arrives for prediction. Moreover, for all the test stages, the model effectively captures the trend for these forecasts at early DAPs, with predictions stabilizing rather than drifting over later season predictions. This is a key requirement for real-world deployment of CropCast, where it can be integrated with subsequent operational tasks like yield prediction. The forecasting mechanism begins around 50 DAP for the training and test datasets and continues to provide NDVI forecasts and its update using online finetuning, improving as more satellite image acquisitions arrive.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eUtilization of CropCast for in-season cotton yield forecasting\u003c/h2\u003e \u003cp\u003eReflecting on the efficient performance of CropCast across different years and regions, we demonstrate its utilization for in-season yield forecasting to show how such a model tailors to real-world problem solving. Using the Driscoll field as a case study, we used cotton yield data from 2021\u0026ndash;2023 for training, while keeping 2024 as an independent test year. Supplementary Figure S4 shows the soil properties, overall extent and gridded data for the Driscoll field. Here we use the forecasted NDVI coming from CropCast as input, and yield as the target variable for modeling. The advantage with CropCast forecasted data as input is the early availability of full season NDVI forecasts, which can help improve the yield model\u0026rsquo;s capability to learn from the full season growth pattern. This enables the yield model to learn from an anticipated complete growth trajectory rather than truncated observations, substantially improving early-stage prediction stability. Yield forecasts begin around ~\u0026thinsp;50 DAP, aligning with the earliest reliable NDVI-driven crop growth signal. We use an LSTM model architecture shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e with 3 LSTM layers, and RMSE as performance metric for evaluating model performance. Similar to the CropCast model forecasts, we predict yields at significant stages of crop growth, beginning well in time (~\u0026thinsp;50 DAP) and continuing from thereon. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e illustrates how adaptive NDVI forecasting propagates into spatially explicit yield and profitability predictions. The predicted yields come from a LSTM model developed for predicting yields given CropCast forecasted NDVI. Figures\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA provides a bar chart visualization of the average yield prediction for test year 2024, for DAPs 57, 81, 105, 117, 140 (defoliation day) and 163 (harvest day), corresponding to CropCast results for Driscoll field forecasts for year 2024. Early-stage predictions exhibit moderate uncertainty but converge toward harvest-time estimates as additional NDVI data becomes available. Further, we predict grid-level yield estimates at 9 m \u0026times; 9 m resolution shown in Figs.\u0026nbsp;4A1 to A6. We observe strong spatial heterogeneity within the field for predictions throughout the season. Importantly, spatial yield structure emerges early in the season and stabilizes over time, suggesting that CropCast captures persistent growth patterns rather than transient noise. These estimates nevertheless highlight the capability of the CropCast model in contributing to high resolution yield prediction. We next carry out economic analysis of these yield estimates for crop management and marketing decisions.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCropCast for crop management and marketing\u003c/h3\u003e\n\u003cp\u003eBeyond agronomic forecasting, the integration of yield predictions with economic metrics can provide actionable decision support to the farmers. Using forecasted yield \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left({q}_{i,t}\\right)\\)\u003c/span\u003e\u003c/span\u003e and observed cotton prices \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{t}\\)\u003c/span\u003e\u003c/span\u003e, we computed Breakeven Yield Margin (BYM) and profitability index (PI). Detailed descriptions on computing BYM and PI come in the Methodology section. Figures\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC show extension of the yield maps into economic interpretations. Figures\u0026nbsp;4B1 to B6 show spatial variability of BYM percentage over time. Early-season maps identify areas at risk of underperforming relative to cost thresholds. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC quantifies expected profit per grid cell. Negative zones during mid-season (e.g., 105\u0026ndash;117 DAP) indicate potential management intervention points, while recovery toward harvest reflects yield stabilization and favorable price interaction. The intervention points also get highlighted in the average BYM and PI plots shown in Figs.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB and \u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eC respectively. These results demonstrate three critical operational advantages. The first advantage is grid-level management optimization. Grid-level forecasts allow variable-rate irrigation, fertilization, or pest control targeting zones with declining profitability signals. The second advantage can be risk-aware marketing decision. Knowing that cotton prices are often higher pre-harvest than post-harvest, early yield forecasts can enable growers and merchants to hedge production or engage in forward contracts with reduced uncertainty. The third advantage can be dynamic intervention planning. Since CropCast updates forecasts as new satellite data arrives, economic risk maps evolve during the season, supporting adaptive decision-making rather than static pre-season planning. Moreover, the model demonstrates that meaningful economic signals emerge well before harvest, highlighting the value of early, context-aware NDVI forecasting.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eAnalysis and comparison of CropCast performance\u003c/h2\u003e \u003cp\u003eThe coming generation of digital agriculture requires computationally efficient and scalable frameworks that can transform routinely available satellite data into reliable crop growth forecasts for real-time agricultural decision-making. We see CropCast as a solution to these requirements, addressing several persistent limitations in satellite-driven crop forecasting. Unlike many yield models that rely on peak-season or post-flowering observations, CropCast produces early forecasting capability (beginning around ~\u0026thinsp;50 DAP), with improved stable forecasts as new observations are incorporated via continual fine-tuning and without the need of model retraining, enabling operability. Further, the framework relies primarily on NDVI, normalized DAP, and simple group statistics, rather than multi-source meteorological or soil datasets. Some state-of-the-art ML models integrate extensive climate and management inputs, increasing data requirements and limiting scalability\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e, \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e, \u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e\u003c/sup\u003e. CropCast demonstrates that structured temporal modeling combined with contextual encoding can achieve competitive accuracy with substantially fewer inputs. Furthermore, we observe that the proposed continual adaptation strategy freezing the encoder and fine-tuning only the decoder and output head preserved stable seasonal dynamics while adapting to year-specific deviations. This aligns with recent parameter-efficient fine-tuning paradigms but is tailored here for agricultural phenology. Compared to static sequence-to-sequence LSTMs which degrade sharply when evaluated on new seasons, CropCast updates itself performing well over cross year datasets\u003csup\u003e\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e, \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e\u003c/sup\u003e. Testing across Sinton, Etter, and Driscoll fields demonstrated robust cross-location generalization. Slight early overestimation was corrected as additional in-season data arrived, illustrating adaptive correction rather than divergence. Even under cloud-induced data gaps, forecasts remained stable and convergent, as evident for Driscoll prediction for the year 2024, where we miss 25 days of observation in the cropping season. The log-normal smoothing of NDVI sequences likely contributed to this robustness by anchoring predictions to biologically plausible growth shapes. Collectively, these characteristics distinguish CropCast from purely static deep learning models and computationally intensive process-based simulators. Rather than increasing architectural complexity (e.g., Transformers, which underperformed in our experiments), performance gains arise from contextual modeling and continual adaptation. Although demonstrated here for cotton, CropCast\u0026rsquo;s structure is crop-agnostic. The framework relies on a generalizable principle that vegetation indices follow phenology-dependent growth curves, and contextual timing (e.g., DAP) can normalize inter-season shifts. Further, the local fine-tuning by CropCast can correct season-specific deviations. Since many crops exhibit bell-shaped NDVI trajectories, the log-normal smoothing and chunked autoregressive forecasting can extend naturally to maize, wheat, soybean, and rice systems. Only crop-specific planting dates and yield model retraining would be required.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eCurrent limitations and scope for improvement\u003c/h3\u003e\n\u003cp\u003eDespite strong performance, there is scope for improvements in overall CropCast implementation and testing. One of those would be its validation, which is limited to Texas cotton systems. Further, the whole CropCast framework depends solely on NDVI-derived growth signals, which is convenient and computationally efficient, but will improve its generalization capability over cross-year and cross locations by including more input variables governing crop growth. The applications like yield modeling and defoliation date determination will also benefit. Furthermore, economic analysis presently assumes static production costs and observed prices. We can build scenarios based on dynamic production costs and prices, generating different solutions with different levels of confidence. Future work will incorporate multi-state datasets, additional spectral indices (e.g., EVI, different band combinations, modalities), scenario simulations, and dynamic price forecasting.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn this study, we introduce CropCast, a context-aware, in-season continual forecasting framework for adaptive crop growth prediction using satellite-derived NDVI. Unlike conventional static deep learning or process-based models, CropCast integrates phenological context, variable-length temporal modeling, and modular incremental fine-tuning to produce stable and transferable forecasts across years and locations. By training on multi-year cotton datasets and validating over independent cross-year and cross-region scenarios, we demonstrate that CropCast maintains forecasting stability beginning as early as ~\u0026thinsp;50 days after planting. The framework adapts efficiently to new in-season observations by fine-tuning only the decoder and output layers, avoiding full retraining and enabling computationally efficient deployment. Beyond NDVI forecasting, we demonstrate probable application by showing high-resolution yield prediction and economic analysis. Forecasted growth trajectories translate into spatially explicit yield, breakeven margin, and profitability estimates at high (\u0026lt;\u0026thinsp;10 m) resolution. Importantly, meaningful economic signals emerge well before harvest, enabling hyperlocal management optimization, dynamic intervention planning, and risk-aware marketing decisions such as forward contracting or hedging. While demonstrated for Texas cotton systems, CropCast\u0026rsquo;s design is crop-agnostic and extensible to other phenology-driven crops. The framework\u0026rsquo;s reliance on vegetation dynamics, contextual timing, and lightweight continual adaptation makes it suitable for scalable, real-world agricultural monitoring, providing a practical bridge between satellite remote sensing data and operational farm decision-making.\u003c/p\u003e"},{"header":"Methodology","content":" \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003cdiv id=\"Sec12\" class=\"Section3\"\u003e \u003ch2\u003eData Collection and Preprocessing\u003c/h2\u003e \u003cp\u003eWe acquired multispectral imagery through the PlanetScope application programming interface (API) for predefined Areas of Interest (AOIs) corresponding to three cotton fields located at two major cotton growing regions: South Plains and Coastal Bend in Texas. For each AOI and date, we filtered candidate scenes using a threshold on cloud cover, retaining only those with less than 20% overall cloud coverage. This threshold provides a reasonable guarantee of unobstructed visibility within the AOI while maintaining sufficient temporal density, by PlanetScope at high (~\u0026thinsp;3m) spatial resolution. This process yielded georeferenced image subsets focused on the target fields. By automating the search query, filtering, and cropping (AOI extraction) pipeline through the API, we collected imagery across the entire growing season, ensuring continuous monitoring of crop development. Within each AOI, native 3m x 3m pixels are aggregated into 9m x 9m grid cells by averaging band reflectance values. This spatial aggregation reduces high-frequency noise caused by shadows, artifacts, and small-scale vegetation heterogeneity. For each grid cell, we compute the Normalized Difference Vegetation Index (NDVI) shown in Eq.\u0026nbsp;(1), where NIR and Red denote the digital number obtained for the near-infrared and red bands, respectively.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:NDVI\\:=\\:\\frac{NIR\\:-\\:Red}{NIR\\:+\\:Red}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e(1)\u003c/h2\u003e \u003cp\u003eAlthough PlanetScope provides near-daily coverage, temporal gaps remain due to cloud cover and occasional observation dropouts. To obtain smooth and temporally aligned NDVI sequences, we fit a four-parameter log-normal growth curve to each grid\u0026rsquo;s NDVI time series, as shown in Eq.\u0026nbsp;(2), where \u003cem\u003ex\u003c/em\u003e is Days After Planting (DAP), \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{y}_{0}\\)\u003c/span\u003e\u003c/span\u003e is the baseline value, \u003cem\u003ea\u003c/em\u003e represents the NDVI amplitude, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{x}_{0}\\)\u003c/span\u003e\u003c/span\u003e denotes the central scale of the peak, and \u003cem\u003eb\u003c/em\u003e represents the standard deviation. We select this log-normal function approximation since vegetation indices typically follow a bell-shaped trajectory over the season. We sample the fitted curve at uniform DAP intervals, converting irregular observations into smooth, consistent daily NDVI sequences for each grid. These preprocessed sequences for each grid over multiple cropping seasons and locations form the input time series (~\u0026thinsp;12,000 samples) for subsequent modeling and continual adaptation.\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:y\\:=\\:{y}_{0}\\:+\\:\\frac{a}{x}\\:\\text{exp}\\left[-0.5{\\left(\\frac{\\text{ln}\\left(\\frac{x}{{x}_{0}}\\right)}{b}\\right)}^{2}\\right]\\:$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e(2)\u003c/h2\u003e \u003cp\u003eTo improve generalization and avoid overfitting a single temporal alignment, we designed the dataset to expose the model to multiple short, context-aware portions of each NDVI curve. Figure \u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003eM1\u003c/span\u003e illustrates this process. For each \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:9m\\:\\)\u003c/span\u003e\u003c/span\u003ex \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:9m\\:\\)\u003c/span\u003e\u003c/span\u003egrid cell, the full NDVI sequence is repeatedly sampled into variable-length input-target windows. A random start index is selected, and an input segment of length \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{l}_{1}\\in\\:\\)\u003c/span\u003e\u003c/span\u003e \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left[{i}_{1},\\:{i}_{2}\\right]\\)\u003c/span\u003e\u003c/span\u003e is followed by a target segment of length \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{l}_{2}\\:\\in\\:\\:\\left[{t}_{1},\\:{t}_{2}\\right]\\)\u003c/span\u003e\u003c/span\u003e. This \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:k\\)\u003c/span\u003e\u003c/span\u003e-sampling strategy ensures that the model views the same input data curve from multiple points of view (early growth, peak, and decline) rather than memorizing a single fixed split. In addition to NDVI values, each input sequence is augmented with Days After Planting (DAP) indices and per-sample group statistics (mean and standard deviation), enabling the model to learn context-aware temporal dynamics across fields and seasons. To further enhance robustness, the target sequence is auto regressed using a mix of ground truth and previous prediction, as shown in Figure \u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003eM2\u003c/span\u003e. Here the green blocks (\u003cem\u003eg\u003c/em\u003e) denote steps fed with ground truth (teacher forcing), while gray blocks (\u003cem\u003ep\u003c/em\u003e) denote steps fed with the model\u0026rsquo;s own prior predictions. This combined strategy encourages the model to handle both ideal and self-generated contexts during inference.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eBecause input and target windows vary in length, we implemented a custom batching pipeline. Before each training step, input sequences (NDVI and normalized DAP) and target sequences are dynamically padded to the maximum lengths in the batch, and the true sequence lengths are tracked. This allows sequence‑to‑sequence models such as LSTMs to correctly mask padded positions and prevent length mismatch issues. Group statistics for each sample are stacked alongside. This collator design preserves the variability in window lengths while ensuring computational efficiency. In this way, the dataset generation process (variable window sampling, context features, mixed teacher forcing, and dynamic batching) exposes the model to diverse temporal perspectives and improves its ability to generalize beyond a single input\u0026ndash;target configuration.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eModel Architecture\u003c/h2\u003e \u003cp\u003eInstead of relying on deep, high-capacity architectures, we design a lightweight, flexible model architecture that can generalize across conditions and adapt with minimal fine-tuning, considering the crop growth dataset behavior as often sparse, noisy, and region-specific. Each input sample consists of a variable-length NDVI sequence, DAP, and group-level NDVI statistics (mean and standard deviation). This design encodes both local temporal dynamics and broader field-level context. The output is forecasted NDVI values for the next 40 DAPs. To mimic real-world variability, we use variable input lengths for NDVI, and DAP and group statistics as static input. The output is fixed at 40 days. To enrich training diversity, we slice multiple samples from each input data curve. The model we build is an encoder-decoder LSTM with 2 encoder LSTM layers, 2 decoder LSTM layers, and an output head (multi-layer perceptron layer). Figure \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003eM3\u003c/span\u003e shows the modeling process. The encoder layers accept variable-length input (NDVI time series) with a padded sequence. The decoder LSTM layers perform autoregressive rollout, and the output head combines decoder state with group statistics and DAP. During training, we apply probabilistic teacher forcing (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.5) to stabilize decoding.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo support real-time regional adaptation, we implement a continual fine-tuning strategy that incrementally updates the model as new NDVI sequences become available during the growing season. We implement modular adaptation by freezing the encoder to retain general spatiotemporal representations and fine-tune only the decoder and output head. This significantly reduces computational overhead and avoids overfitting, while still allowing the model to adapt to recent field-specific trends. We use a small batch of recently acquired NDVI sequences for fine-tuning and perform 1 to 2 epochs of gradient updates with a learning rate \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:1{e}^{-4}\\)\u003c/span\u003e\u003c/span\u003e. To stabilize training, we keep probabilistic teacher forcing. These steps enable fast, local adaptation without retraining the entire model, making it practical for dynamic environments where new observations trickle in gradually.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eCrop management and marketing methods\u003c/h2\u003e \u003cp\u003eTo demonstrate the possible applications of the proposed model, we choose yield prediction as an example application. Building an LSTM model with three LSTM layers, we use the forecasted NDVI coming from the endoder-decoder LSTM as input to predict yields at multiple time steps of the growing season. For yield prediction, we use a separate model with three LSTM layers. For preparing model input, we initially compute the first derivative of the forecasted NDVI from the CropCast model for multiple DAPs (57,81,105, and 117) to mitigate the temporal shifts of the forecasts. From the first derivatives, we calculate a cumulative NDVI for these different DAPs, as feed this into the LSTM model shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Subsequent to predicting yields at multiple time steps of the growing season, we develop a marketing method that can provide grid level calculations and decision-making capability to farmers. For each grid \u003cem\u003ei\u003c/em\u003e at time \u003cem\u003et\u003c/em\u003e, the profit (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\pi\\:\\)\u003c/span\u003e\u003c/span\u003e) is defined as the difference between total revenue and total production cost, as shown in Eq.\u0026nbsp;(3), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{t}\\)\u003c/span\u003e\u003c/span\u003e denotes the cotton price, \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{q}_{i,t}\\)\u003c/span\u003e\u003c/span\u003e the yield, and \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{C}_{i}\\)\u003c/span\u003e\u003c/span\u003e the grid cost. The breakeven point is obtained when \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\pi\\:}_{i,t}=0\\)\u003c/span\u003e\u003c/span\u003e. Accordingly, the minimum yield (q) and price (\u003cem\u003ep\u003c/em\u003e) required to cover total cost are shown in Eq.\u0026nbsp;(4). Given the forecasted yield \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\left({q}_{i,t}\\right)\\)\u003c/span\u003e\u003c/span\u003e and the observed market prices \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{p}_{t}\\)\u003c/span\u003e\u003c/span\u003e, we define the profitability index (PI) as the deviation of the market price and the breakeven price, shown in Eq.\u0026nbsp;(5).\u003cdiv id=\"Equc\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equc\" name=\"EquationSource\"\u003e\n$$\\:{\\pi\\:}_{i,t}={p}_{t}\\times\\:{q}_{i,t}-{C}_{i}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e(3)\u003c/h2\u003e \u003cp\u003e \u003cdiv id=\"Equd\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equd\" name=\"EquationSource\"\u003e\n$$\\:{q}_{i,t}^{BE}=\\frac{{C}_{i}}{{p}_{i,t}},\\:{p}_{i,t}^{BE}=\\frac{{C}_{i}}{{q}_{i,t}}$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e(4)\u003c/h2\u003e \u003cp\u003e \u003cdiv id=\"Eque\" class=\"Equation\"\u003e \u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Eque\" name=\"EquationSource\"\u003e\n$$\\:P{I}_{i,t}\\:=\\left(\\:{p}_{t}-{p}_{i,t}^{BE}\\right)\\times\\:{q}_{i,t}$$\u003c/div\u003e \u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e(5)\u003c/h2\u003e \u003cp\u003eWe further define the Breakeven Yield Margin (BYM) index as the proportional deviation of the estimated yield from its breakeven level defined in Eq.\u0026nbsp;(6), where \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:BY{M}_{i,t}\\)\u003c/span\u003e\u003c/span\u003e denotes the corresponding BYM values for grid \u003cem\u003ei\u003c/em\u003e at time \u003cem\u003et.\u003c/em\u003e Positive values of PI and BYM indicate economic viable production conditions within a grid. The cost of production is retrieved from the Texas A\u0026amp;M AgriLife, and the price of cotton is derived from United States Department of Agriculture (USDA)- Agricultural Market Service (AMS) (retrieved from: \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.ams.usda.gov/market-news/cotton-tobacco\u003c/span\u003e\u003cspan address=\"https://www.ams.usda.gov/market-news/cotton-tobacco\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e (accessed 12/2025))\u003cdiv id=\"Equf\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equf\" name=\"EquationSource\"\u003e\n$$\\:BY{M}_{i,t}=\\frac{{q}_{i,t}-{q}_{i,t}^{BE}}{{q}_{i,t}^{BE}}$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec20\" class=\"Section2\"\u003e \u003ch2\u003e(6)\u003c/h2\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eCompeting interests\u003c/h2\u003e \u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eAuthor contributions\u003c/h2\u003e \u003cp\u003eR.H.M. developed the methodology, implemented the forecasting framework, performed the formal analysis, and contributed to manuscript writing. A.T. contributed to study conceptualization, methodology development, model evaluation, interpretation of results, manuscript writing and revision. L.Z. contributed to methodology development, supervision, interpretation of results, and manuscript editing. J. Landivar contributed to agronomic interpretation, field study design, and manuscript revision. Y.C. contributed to the economic analysis framework, interpretation of profitability and marketing results, and manuscript revision. S.S. contributed to methodology development, interpretation of results, and manuscript writing and revision. J.L.S. and B.G. contributed to field and satellite image data collection and processing. M.B. conceived the study, contributed to project supervision, agronomic guidance, funding acquisition, and manuscript revision. All authors read and approved the final manuscript.\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e \u003cp\u003eThis research is part of the U.S. Department of Agriculture (USDA) Hatch Project number TEX0 9937. The authors acknowledge Cotton Incorporated (Texas State Support Committee, project number: 26-632TX) and Texas A\u0026amp;M AgriLife Research for funding this research. Authors also acknowledge Texas A\u0026amp;M Institute of Data Science (TAMIDS)- leadership members Dr. Nick Duffield and Dr. Drew Casey, the Data Science Capstone Program. Additionally, thanks to Mr. Jimmy Dodson and Mr. Jon Gwynn (DODSON FAMILY FARMS), Mr. Charles Ring, and the North Plains Ground Water Conservation District for enabling field experiments and data collection.\u003c/p\u003e\u003ch2\u003eData and code availability\u003c/h2\u003e \u003cp\u003eData is confidential as it belongs to farmers. Authors will be releasing a GitHub link to the code which is under finalization. Till then, code will be made available upon request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGrafton RQ, Williams J, Jiang Q (2015) Food and water gaps to 2050: preliminary results from the global food and water system (GFWS) platform. Food Secur 7:209\u0026ndash;220\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Population Prospects (2019) : Highlights. United Nations, Department of Economic and Social Affairs, Population Division (2019)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFood and Agriculture Organization of the United Nations (2025) Employment indicators 2000\u0026ndash;2023: global and regional trends. FAO Statistics Division\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSavary S, Willocquet L, Pethybridge SJ, Esker P, McRoberts N, Nelson A (2019) The global burden of pathogens and pests on major food crops. Nat Ecol Evol 3:430\u0026ndash;439\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFAO. The impact of disasters on agriculture and food security 2023. In The impact of disasters on agriculture and food security (2023) (2023)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGetahun S, Kefale H, Gelaye Y (2024) Application of precision agriculture technologies for sustainable crop production and environmental sustainability: A systematic review. \u003cem\u003eThe Scientific World Journal\u003c/em\u003e 2126734 (2024)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLobell DB, Azzari G, Burke M, Gourlay S, Jin Z, Kilic T, Murray S (2020) Eyes in the sky, boots on the ground: Assessing satellite-and ground‐based approaches to crop yield measurement and analysis. Am J Agric Econ 102:202\u0026ndash;219\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Voorn GA, Boer MP, Truong SH, Friedenberg NA, Gugushvili S, McCormick R, Bustos Korts D, Messina CD, Van Eeuwijk FA (2023) A conceptual framework for the dynamic modeling of time-resolved phenotypes for sets of genotype-environment-management combinations: a model library. Front Plant Sci 14:1172359\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBalboa GR, Archontoulis S, Salvagiotti F, Garcia FO, Stewart W, Francisco E, Prasad PV, Ciampitti IA (2019) A systems-level yield gap assessment of maize-soybean rotation under high-and low-management inputs in the Western US Corn Belt using APSIM. Agric Syst 174:145\u0026ndash;154\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeegum S, Timlin D, Reddy KR, Reddy V, Sun W, Wang Z, Fleisher D, Ray C (2023) Improving the cotton simulation model, GOSSYM, for soil, photosynthesis, and transpiration processes. Sci Rep 13:7314\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChang Y, Latham J, Licht M, Wang L (2023) A data-driven crop model for maize yield prediction. Commun biology 6:439\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChapagain SK, Aryal A, Mohan G, Shrestha S, Mishra BK, Fukushi K (2022) Analysis of the climate change impact on water availability and the links between water pollution and economy for sustainable water resource management in Kaski District, Nepal. J Water Clim Change 13:3030\u0026ndash;3045\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCorbeels M, Chirat G, Messad S, Thierfelder C (2016) Performance and sensitivity of the DSSAT crop growth model in simulating maize yield under conservation agriculture. Eur J Agron 76:41\u0026ndash;53\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEch-Chatir L, Er-Raki S, Rodriguez JC, Meddich A (2025) Advances in crop growth modeling: A review of perennial crop and beneficial soil microorganism approaches. Agric Water Manage 315:109548\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHearn A (1994) OZCOT: a simulation model for cotton crop management. Agric Syst 44:257\u0026ndash;299\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones JW, He J, Boote KJ, Wilkens P, Porter CH, Hu Z (2011) Estimating DSSAT cropping system cultivar-specific parameters using Bayesian techniques. Methods introducing Syst models into agricultural Res 2:365\u0026ndash;393\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJones JW, Hoogenboom G, Porter CH, Boote KJ, Batchelor WD, Hunt LA, Wilkens PW, Singh U, Gijsman AJ, Ritchie JT (2003) The DSSAT cropping system model. Eur J Agron 18:235\u0026ndash;265\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKeating BA, Carberry PS, Hammer GL, Probert ME, Robertson MJ, Holzworth D, Huth NI, Hargreaves JN, Meinke H, Hochman Z (2003) An overview of APSIM, a model designed for farming systems simulation. Eur J Agron 18:267\u0026ndash;288\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTimlin D, Paff K, Han E (2024) The role of crop simulation modeling in assessing potential climate change impacts. Agrosystems Geosci Environ 7:e20453\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang N, Zhou X, Kang M, Hu B-G, Heuvelink E, Marcelis LF (2023) Machine learning versus crop growth models: an ally, not a rival. AoB Plants 15:plac061\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanerjee K, Dutta S, Das S, Sadhukhan R (2024) Crop simulation models as decision tools to enhance agricultural system productivity and sustainability\u0026ndash;a critical review. Technol Agron 5\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbbas G, Fatima Z, Tariq M, Ahmed M, Rahman MHu, Nasim W, Rasul G, Ahmad S (2020) Applications of crop modeling in cotton production. Cotton Production and Uses: Agronomy, Crop Protection, and Postharvest Technologies. Springer\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDroutsas I, Challinor AJ, Deva CR, Wang E (2022) Integration of machine learning into process-based modelling to improve simulation of complex crop responses. silico Plants 4:diac017\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMba PC, Njoku JN, Uyeh DD (2025) Enhancing resilience in specialty crop production in a changing climate through smart systems adoption. Smart Agricultural Technol, 100897\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDe Alwis S, Hou Z, Zhang Y, Na MH, Ofoghi B, Sajjanhar A (2022) A survey on smart farming data, applications and techniques. Comput Ind 138:103624\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuete A, Didan K, Miura T, Rodriguez EP, Gao X, Ferreira LG (2002) Overview of the radiometric and biophysical performance of the MODIS vegetation indices. Remote Sens Environ 83:195\u0026ndash;213\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuete AR (1988) A soil-adjusted vegetation index (SAVI). Remote Sens Environ 25:295\u0026ndash;309\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGitelson AA, Kaufman YJ, Merzlyak MN (1996) Use of a green channel in remote sensing of global vegetation from EOS-MODIS. Remote Sens Environ 58:289\u0026ndash;298\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar S, Cazenave AB, Oakes J, McCall D, Thomason W, Abbot L, Balota M (2020) High-throughput measurement of peanut canopy height using digital surface models. Plant Phenome J 3:e20003\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar S, Cazenave AB, Oakes J, McCall D, Thomason W, Abbot L, Balota M (2021) Aerial high-throughput phenotyping of peanut leaf area index and lateral growth. Sci Rep 11:1\u0026ndash;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChapu I, Okello DK, Okello RCO, Odong TL, Sarkar S, Balota M (2022) Exploration of Alternative Approaches to Phenotyping of Late Leaf Spot and Groundnut Rosette Virus Disease for Groundnut Breeding. Front Plant Sci 13:912332\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLobell DB, Asner GP (2003) Climate and management contributions to recent trends in US agricultural yields. Science 299:1032\u0026ndash;1032\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLobell DB, Schlenker W, Costa-Roberts J (2011) Climate trends and global crop production since 1980. Science 333:616\u0026ndash;620\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBolton DK, Friedl MA (2013) Forecasting crop yield using remotely sensed vegetation indices and crop phenology metrics. Agric For Meteorol 173:74\u0026ndash;84\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar S, Ramsey AF, Cazenave A-B, Balota M (2021) Peanut leaf wilting estimation from RGB color indices and logistic models. Front Plant Sci 12:658621\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar S, Osorio Leyton J, Noa Yarasca E, Adhikari K, Hajda CB, Smith DR (2025) Integrating Remote Sensing and Soil Features for Enhanced Machine Learning-Based Corn Yield Prediction in the Southern US. Sensors\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTucker CJ (1979) Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens Environ 8:127\u0026ndash;150\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, Friedl MA, Schaaf CB, Strahler AH, Hodges JC, Gao F, Reed BC, Huete A (2003) Monitoring vegetation phenology using MODIS. Remote Sens Environ 84:471\u0026ndash;475\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouse J, Haas R, Schell J, Deering D (1974) Monitoring vegetation systems in the great plains with ERTS proceeding. In: \u003cem\u003eThird earth reserves technology satellite symposium, greenbelt: NASA SP-351\u003c/em\u003e)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJustice CO, Townshend JR, Vermote EF, Masuoka E, Wolfe RE, Saleous N, Roy DP, Morisette JT (2002) An overview of MODIS Land data processing and product status. Remote Sens Environ 83:3\u0026ndash;15\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHuang J, Tian L, Liang S, Ma H, Becker-Reshef I, Huang Y, Su W, Zhang X, Zhu D, Wu W (2015) Improving winter wheat yield estimation by assimilation of the leaf area index from Landsat TM and MODIS data into the WOFOST model. Agric For Meteorol 204:106\u0026ndash;121\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJohnson DM (2014) An assessment of pre-and within-season remotely sensed variables for forecasting corn and soybean yields in the United States. Remote Sens Environ 141:116\u0026ndash;128\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePrasad AK, Chai L, Singh RP, Kafatos M (2006) Crop yield estimation model for Iowa using remote sensing and surface parameters. Int J Appl Earth Obs Geoinf 8:26\u0026ndash;33\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJi Z, Pan Y, Zhu X, Wang J, Li Q (2021) Prediction of crop yield using phenological information extracted from remote sensing vegetation index. Sensors 21:1406\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSarkar S, Oakes J, Cazenave A-B, Burow MD, Bennett RS, Chamberlin KD, Wang N, White M, Payton P, Mahan J (2022) Evaluation of the US Peanut Germplasm Mini-Core Collection in the Virginia-Carolina Region Using Traditional and New High-Throughput Methods. \u003cem\u003eAgronomy\u003c/em\u003e 12, 1945\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBalota M, Sarkar S, Bennett RS, Burow MD (2024) Phenotyping Peanut Drought Stress with Aerial Remote-Sensing and Crop Index Data. Agriculture 14:565\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGamon JA, Field CB, Goulden ML, Griffin KL, Hartley AE, Joel G, Penuelas J, Valentini R (1995) Relationships between NDVI, canopy structure, and photosynthesis in three Californian vegetation types. Ecol Appl 5:28\u0026ndash;41\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTan C-W, Zhang P-P, Zhou X-X, Wang Z-X, Xu Z-Q, Mao W, Li W-X, Huo Z-Y, Guo W-S, Yun F (2020) Quantitative monitoring of leaf area index in wheat of different plant types by integrating NDVI and Beer-Lambert law. Sci Rep 10:929\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRouse JW Jr., Haas RH, Schell JA, Deering DW (1974) Monitoring Vegetation Systems in the Great Plains with Erts. In: \u003cem\u003eNASA Special Publication\u003c/em\u003e (eds Freden SC, Mercanti EP, Becker MA)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZheng H, Shen L, Tang A, Luo Y, Hu H, Du B, Wen Y, Tao D (2025) Learning from models beyond fine-tuning. Nat Mach Intell 7:6\u0026ndash;17\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLei T, Bai J, Brahma S, Ainslie J, Lee K, Zhou Y, Du N, Zhao V, Wu Y, Li B (2023) Conditional adapters: Parameter-efficient transfer learning with fast inference. Adv Neural Inf Process Syst 36:8152\u0026ndash;8172\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBorzunov A, Ryabinin M, Chumachenko A, Baranchuk D, Dettmers T, Belkada Y, Samygin P, Raffel CA (2023) Distributed inference and fine-tuning of large language models over the internet. Adv Neural Inf Process Syst 36:12312\u0026ndash;12331\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDing N, Qin Y, Yang G, Wei F, Yang Z, Su Y, Hu S, Chen Y, Chan C-M, Chen W (2023) Parameter-efficient fine-tuning of large-scale pre-trained language models. Nat Mach Intell 5:220\u0026ndash;235\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCamps-Valls G, Svendsen DH, Cort\u0026eacute;s-Andr\u0026eacute;s J, Mareno-Mart\u0026iacute;nez \u0026Aacute;, P\u0026eacute;rez-Suay A, Adsuara J, Mart\u0026iacute;n I, Piles M, Mu\u0026ntilde;oz-Mar\u0026iacute; J, Martino L (2021) Physics-aware machine learning for geosciences and remote sensing. In: \u003cem\u003e2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS\u003c/em\u003e). IEEE\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKhan MA, Wahid A, Ahmad M, Tahir MT, Ahmed M, Ahmad S, Hasanuzzaman M (2020) World cotton production and consumption: An overview. \u003cem\u003eCotton production and uses: Agronomy, crop protection, and postharvest technologies\u003c/em\u003e, 1\u0026ndash;7\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSun L, Wang Z, Xiong C, Gu J, Zheng Y, Ju F, Wang S, Hu W, Zhao W, Zhou Z (2024) Improving the soil K+/Na+ ratio under moderate salt stress synergistically increases the yield and quality of cotton fiber and cottonseed. Ind Crops Prod 213:118441\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eShahzad AN, Rizwan M, Asghar MG, Qureshi MK, Bukhari SAH, Kiran A, Wakeel A (2019) Early maturing Bt cotton requires more potassium fertilizer under water deficiency to augment seed-cotton yield but not lint quality. Sci Rep 9:7378\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Z, Ma H, Xia J, Hou F, Shi X, Hao X, Hafeez A, Han H, Luo H (2017) Optimal pre-plant irrigation and fertilization can improve biomass accumulation by maintaining the root and leaf productive capacity of cotton crop. Sci Rep 7:17168\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAshraf S, Ch KM, Ashraf I, Akbar N (2024) A phenomenological inquiry into farmers\u0026rsquo; experiences growing cotton in Punjab, Pakistan. Sci Rep 14:13846\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHightower M, USDA (2022) Drought drops U.S. cotton harvest to be lowest level since the 19th century. University of Arkansas, Division of Agriculture, Research and Extension\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBeegum S, Reddy V, Reddy KR (2023) Development of a cotton fiber quality simulation module and its incorporation into cotton crop growth and development model: GOSSYM. Comput Electron Agric 212:108080\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin S, Deng M, Wei K, Wang Q, Su L (2023) A new regional cotton growth model based on reference crop evapotranspiration for predicting growth processes. Sci Rep 13:7368\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePalla S, Sarkar S, Zhao L, Landivar J, Scott JL, Maeda M, McGinty J, Abugho S, Jung J, Rajan N (2025) Growth analysis of cotton using UAS derived multi temporal canopy features. Sci Rep\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eGwathmey CO, Tyler DD, Yin X (2010) Prospects for monitoring cotton crop maturity with normalized difference vegetation index. Agron J 102:1352\u0026ndash;1360\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLi H, Lascano R, Barnes E, Booker J, Wilson T, Bronson K, Segarra E (2001) Temporal patterns of cotton reflectance and NDVI-days lint yield modeling. In: \u003cem\u003eProc. Beltwide Cotton Conf\u003c/em\u003e)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu F, Xiao X, Qin Y, Liu L (2025) Identifying and mapping open-boll cotton fields in the southeastern United States with time series PhenoCam, Sentinel-2 and Sentinel-1 images. Int J Appl Earth Obs Geoinf 142:104723\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eStone K, Billman ED, Bauer PJ, Sigua G (2022) Using NDVI for variable rate cotton irrigation prescriptions. Appl Eng Agric 38:787\u0026ndash;795\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang Y, Zhang Z, Xiao C, Zhang T, Yu K, Zhang C, Liao Q, Li F, Wan S, Chen G (2026) Characterizing Cotton Defoliation Progress via UAV-Based Multispectral-Derived Leaf Area Index and Analysis of Influencing Factors. Remote Sens 18:609\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eVan Klompenburg T, Kassahun A, Catal C (2020) Crop yield prediction using machine learning: A systematic literature review. Comput Electron Agric 177:105709\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai Y, Guan K, Lobell D, Potgieter AB, Wang S, Peng J, Xu T, Asseng S, Zhang Y, You L (2019) Integrating satellite and climate data to predict wheat yield in Australia using machine learning approaches. Agric For Meteorol 274:144\u0026ndash;159\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu Z, Xu W, Feng J, Palaiahnakote S, Lu T Context-aware attention LSTM network for flood prediction. In: (2018) \u003cem\u003e24th international conference on pattern recognition (ICPR)\u003c/em\u003e). IEEE (2018)\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu J, Wang G, Hu P, Duan L-Y, Kot AC (2017) Global context-aware attention lstm networks for 3d action recognition. In: \u003cem\u003eProceedings of the IEEE conference on computer vision and pattern recognition\u003c/em\u003e)\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-9348900/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9348900/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eOpen-field crop production is among the most weather-exposed economic activities on Earth. More than 40% of crops are lost before harvest, thanks to combined effects of abiotic and biotic stresses and flickering market economy, often exacerbated by delayed management interventions. Here we present CropCast, a context-aware continual crop growth forecasting framework for optimizing in-season management decisions using satellite-derived vegetation indices. CropCast integrates an encoder-decoder long short-term memory architecture with phenological context variables and a continual adaptation strategy that updates predictions as new satellite observations become available. Using multi-year PlanetScope imagery from cotton production regions in Texas, CropCast generates stable growth forecasts beginning as early as 50 days after planting and maintains predictive accuracy across unseen seasons and locations, reaching low mean absolute errors (2.5%) in forecasting. These forecasts further used as input for yield prediction provide high-resolution (9m) estimates of both yield and economic variability, enabling in-season dynamic intervention to optimize hyperlocal management strategies, and support risk-aware marketing decisions.\u003c/p\u003e","manuscriptTitle":"CropCast: Context-aware AI framework for in-season crop growth forecasting","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-05-05 19:08:07","doi":"10.21203/rs.3.rs-9348900/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"communications-earth-and-environment","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"commsenv","sideBox":"Learn more about [Communications Earth and Environment](https://www.nature.com/commsenv/)","snPcode":"","submissionUrl":"","title":"Communications Earth \u0026 Environment","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Communications Series","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"7351e280-fcf4-424c-8b0d-103242474f98","owner":[],"postedDate":"May 5th, 2026","published":true,"recentEditorialEvents":[{"type":"editorAssigned","content":"","date":"2026-04-30T11:14:04+00:00","index":"","fulltext":""}],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":67313439,"name":"Biological sciences/Plant sciences"},{"id":67313440,"name":"Biological sciences/Plant sciences/Plant stress responses"}],"tags":[],"updatedAt":"2026-05-05T19:08:07+00:00","versionOfRecord":[],"versionCreatedAt":"2026-05-05 19:08:07","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9348900","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9348900","identity":"rs-9348900","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}