{"paper_id":"06520fcc-d08e-4f3e-8eaa-d39b212e057c","body_text":"A Deep Learning Approach to Predicting Pithomyces chartarum Sporulation for Livestock Protection | 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 A Deep Learning Approach to Predicting Pithomyces chartarum Sporulation for Livestock Protection Iúri Diogo, César Capinha, João Pinelo, Elizabeth Domingues, Mariana Ávila This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6453956/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Pithomycotoxicosis is a disease affecting grazing livestock, caused by ingestion of Pithomyces chartarum spores. These spores have been identified in various regions worldwide, including the Azores Archipelago (Portugal) since 1999. The severity of the disease is strongly linked to spore concentration, while spore abundance is known to depend on meteorological conditions. In this study, we develop and evaluate a deep learning-based framework to forecast the sporulation of P. chartarum on Terceira Island (Azores), using historical spore count data together with meteorological and topographic variables. Among 20 neural network architectures tested, a convolutional neural network (CNN) achieved the best performance in classifying high-risk conditions, with an area under the curve (AUC) of 0.81 on the validation set. Feature importance analysis identified mean daily temperature as the most influential variable for sporulation risk, consistent with known favorable conditions for fungal growth. Additionally, the results reveal a marked seasonal pattern in sporulation risk, shaped by short- to mid-term antecedent meteorological conditions. Our findings demonstrate that deep learning models can enhance predictive accuracy and deepen understanding of the environmental drivers of P. chartarum sporulation, thereby improving the performance of existing alert systems for livestock protection. Earth and environmental sciences/Ecology/Biogeography Biological sciences/Computational biology and bioinformatics/Machine learning Deep learning Time series prediction Pithomyces chartarum Pithomycotoxicosis Ecological prediction Alert system Figures Figure 1 Figure 2 Figure 3 Figure 4 1. Introduction Pithomyces chartarum is a cosmopolitan saprophytic fungus, found in warm temperate, sub-tropical, and tropical regions [ 1 ], that, like other fungi, uses sporulation as a means of surviving in harsh conditions and dispersing to new environments [ 2 , 3 ]. Its spores are usually found on several types of vegetation, including grass pastures and dead matter of grasses, living endophytically within them, which can then develop a wide range of symptoms such as brown necrotic leaf spots or lesions on live tissue, but can also occur as a saprophytic mould that can be found in plant debris at the base of pasture grasses [ 4 – 7 ]. Animals grazing on contaminated vegetation are likely to ingest the spores. The fungus produces the toxin sporidesmin, an hepatotoxin [ 8 ], and its ingestion by livestock causes pithomycotoxicosis, also known as facial eczema [ 9 , 10 ]. Affected animals are mainly sheep and cattle but also include, to varying degrees, guinea pigs, rabbits, rats, mice, red and fallow deer, goats and alpacas [ 10 – 12 ]. After infection, affected animals are known to develop hepatic photosensitivity resulting from liver damage, specifically injury to the biliary epithelial cells, which leads to acute bile duct obstruction and hepatic insufficiency [ 7 , 13 , 14 ]. Other commonly associated symptoms include reduced fertility and decreased production of milk, wool, and pelt, as well as transient diarrhoea, loss of appetite followed by weight loss, skin crusting, and, in severe cases, death [ 3 , 6 , 8 , 11 , 12 ]. In general, the severity of the intoxication is directly related to the spores concentration and the grazing period [ 3 , 12 , 15 ]. Clinical signs of infection do not appear until 10–14 days after the ingestion of the toxin [ 8 ] but could take up to 20 days to be visible [ 10 ]. According to Schöniger et al. [ 16 ], although rare, P. chartarum might cause onychomycosis in humans, and inhaling its spores may lead to asthma. One of the most common methods employed by farmers to prevent pithomycotoxicosis in livestock, is the administration of zinc feed supplements [ 11 ]. When ingested (mainly orally), zinc and sporidesmin combine, creating a stable mercaptide and eliminating the toxin during the autoxidation cycle [ 3 ]. However, this may lead to over or under administration, since the administration is based on the farmer’s past experience, as stated by Ávila et al. [ 3 ]. Additional prevention strategies involve restricting access to contaminated pastures and employing fungicide treatments [ 11 ]. Facial eczema was first reported in New Zealand and has since been identified in several Asian countries, including China and the Philippines, as well as in Australia, the United States, Turkey, South Africa, Uruguay, Argentina, and Brazil. In mainland Europe, the first reported case occurred in the Netherlands, with subsequent reports from France, Spain, Portugal, Hungary, and Italy [ 1 , 8 , 10 , 14 , 17 , 18 ]. The spread of facial eczema poses a serious threat not only to animal health but also to agricultural economies, due to animal mortality and reduced productivity and fertility [ 3 , 8 ]. For example, in New Zealand, the disease was estimated to cause annual losses of between $ 63 million and $ 126 million to sheep farmers alone between 1983 and 1988 [ 11 ], illustrating the substantial economic impact of P. chartarum . According to Pinto et al. [ 8 ], P. chartarum has been present on Azorean pastures since at least 1999, with 22 outbreaks recorded between 1999 and 2001 alone. Previous studies have shown a strong association between meteorological conditions and P. chartarum sporulation, suggesting that these variables may be useful for predicting its occurrence [ 3 ]. Sporulation is influenced by specific environmental conditions, particularly warm temperatures and high moisture levels, which stimulate fungal growth and spore production [ 4 , 19 ]. According to previous studies, the incubation period typically ranges from 3 to 5 days, depending on the season [ 4 ]. Sporulation typically occurs under conducive environmental conditions, including temperatures above 12°C during late summer and autumn, with an optimum around 24°C, and relative humidity exceeding 80% [ 4 , 12 , 15 , 18 ]. Despite moisture being required for sporulation, heavy or continuing rainfall reduces its occurrence and toxicity, since the spores can become saturated and sporidesmin is water soluble, therefore rainfall should range between 5 to 35 mm per week to enhance sporulation [ 3 , 4 , 11 , 19 ]. According to Dennis et al. [ 13 ] optimal conditions of warm and humid weather for at least three days provides enough time for a successful fungal sporulation and colonisation in pastures. However, in Azores, Portugal, fungal sporulation enhancement was recorded during periods of more than 10 consecutive days with minimum temperatures of at least 16°C and relative humidity exceeding 90% [ 8 ]. The presence of decaying plant material is another important factor promoting P. chartarum sporulation, particularly when it includes tissues rich in soluble organic compounds, such as freshly dead leaves [16, 19. Topographic features can influence P. chartarum sporulation by altering microclimatic conditions; for instance, higher sporulation levels have been observed in pastures sheltered by hedges and on north- or west-facing slopes, where heat accumulation is greater [ 20 ]. Natural light also appears to affect sporidesmin stability, with greater losses occurring on sunny days [ 9 ]. In 2024, Ávila et al. assessed the presence of P. chartarum in Azorean pastureland using IoT sensors and remote sensing. However, their approach lacked the predictive capability required for real-time risk assessment [ 3 ]. Given the significant health and economic consequences associated with P. chartarum , there is a clear need for early warning systems that support informed livestock management and improve milk production, while also enabling ongoing monitoring and scientific understanding of the fungus’s dynamics. Building on the work of Ávila et al. [ 3 ], this study aims to enhance predictive capabilities by applying deep learning techniques—specifically time-series neural networks—and by incorporating additional environmental variables to forecast P. chartarum sporulation. Deep learning is a subset of machine learning that relies on complex multilayered and interconnected nodes - Artificial Neural Networks (ANNs) - to process and recognize patterns from data [ 21 , 22 ]. They are composed of trainable parameters, applicable in various science disciplines, that enable a very high performance of tasks ranging from natural language processing to image and sound data classification, computer vision, clustering and prediction [ 21 , 23 , 24 ]. In ecology, deep learning has been increasingly used primarily for image, video and sound classification [ 22 – 24 ]. More recently, deep learning has shown potential for biodiversity monitoring and conservation planning, where, for example, Recurrent Neural Networks (RNNs) were applied to long-term arthropod monitoring data from Terceira Island (Azores), outperforming traditional models (e.g., Local Polynomial Regression) in goodness of fit and overall accuracy [ 25 ]. Compared to classical machine learning models, which rely on expert knowledge to identify predefined features, deep learning is considered powerful, by automatically extracting relevant patterns from raw data [ 22 , 23 , 26 ]. In the context of P. chartarum sporulation, this means, for example, using raw time series of weather variables as predictors in the model, rather than relying on predictors that summarize weather conditions over predefined time periods (e.g., weeks, months; [ 3 ]). In this study, we test the use of deep learning for early warning of P. chartarum sporulation levels. Ecologists often rely on deep learning models to accurately forecast future events, identifying the set of time series features that most accurately predict, for example, disease vector abundances, not needing to rely on human expertise [ 22 , 24 ]. Additionally, Rammer & Seidl [ 26 ] state that deep learning techniques are good at generalizing beyond test data and that is a key-factor in prediction exercises for ecological studies, performing better than other machine learning methods in terms of prediction accuracy. Therefore, we present and evaluate a deep learning-based approach for forecasting the sporulation of P. chartarum on Terceira Island (Azores), using spore abundance data together with meteorological and topographic variables as predictors. This approach is intended to be integrated into the workflow of the existing sporulation risk alert system. 2. Materials and Methods 2.1. Study Area The Azores archipelago (36 to 39° N, 25 to 31° W), situated in the Macaronesia region of the North Atlantic Ocean, is composed of nine islands [ 3 , 27 ]. The archipelago has a temperate oceanic climate, influenced by a semi-permanent subtropical Atlantic anticyclone (“Azores High”), with a low thermal amplitude (9–26°C), high humidity (annual average of 80%), and annual rainfall ranging from 900 to 3000 mm [ 3 , 8 , 27 ]. Terceira island has a surface area of 400 km2, making it the third largest of the archipelago; it has an elliptical shape, approximately 29 km from east to west, and 17.5 km from north to south, with its peak altitude in the Serra de Santa Bárbara (1021 m) [ 3 , 28 ]. According to the latest agricultural census in Portugal, in 2019 there were approximately 296196 grazing livestock in the Azores that could be affected by pithomycotoxicosis, more than 33% of which were dairy animals including cattle, sheep, and goats [ 29 ]. With such a large number of potentially endangered individuals, there could be a high negative impact on agriculture, a key economic sector of the archipelago [ 30 ]. 2.2. Data collection: P. chartarum spore count The spore count data was initially collected as part of the previous study led by [ 3 ]. Experts from Unicol collected multiple grass samples per week from various grazing locations in Terceira island, and quantified them at the Laboratório Regional de Veterinária using a washing method (spores/gram), from June 2021 to December 2023 [ 3 ]. Additionally, samples were collected during periods identified as both favorable and not favorable for sporulation [ 3 ], in order to evaluate the spores concentration over time. All grass samples were collected with permission, and sampling procedures adhered to established local agricultural and veterinary protocols. The sporulation and meteorological datasets, originally stored in databases, were downloaded and saved in Feather format to facilitate efficient processing; the entire analysis was then carried out programmatically. The data was read into Python (V3.12.3) using the Pandas and Pyarrow software libraries. After filtering the dataset to include only records from 2023, missing values were removed. Based on expert input from Unicol and spore counts per gram, two sporulation risk classes were defined: Low (≤ 10,000 spores/gram) and High (> 10,000 spores/gram). Although pastures are typically considered hazardous when spore concentrations exceed 30,000 spores per gram of grass, prolonged grazing under concentrations as low as 10,000 spores per gram can still pose a significant risk to livestock [ 31 ]. To enable binary classification, one-hot encoding was applied to convert the categorical classes into 0 and 1. 2.3. Predictor variables Guided by the known ecological preferences of the fungus (see Introduction), we selected a set of predictive variables designed to capture the meteorological and topographic factors influencing sporulation. Previous studies found that the fungus sporulation is stimulated by warm temperatures and high levels of moisture [e.g. 4, 19]. Ávila et al. [ 3 ] explains how temperature and relative humidity data was collected in Terceira island specifically for this project: hourly data was collected through a network of Internet of Things (IoT) weather stations strategically positioned after an analysis performed by the AIR Centre team using data from three meteorological stations from the Instituto Português do Mar e da Atmosfera (IPMA). These measurements were then transmitted to the IoT server via the LoRaWAN protocol, and stored and accessed in a MySQL database of the AIR Centre. There are measurements available from 26 July 2021 onwards, from that network composed of 59 IoT weather stations [ 3 ]. Using Python 3.12.3, these hourly meteorological variables were used to first calculate the mean daily values, at a station level, because the existing literature suggests that the fungus is more influenced by general environmental conditions than by short-term fluctuations [e.g. 7, 14]. To account for a few implausible extreme values likely caused by equipment errors, temperature data were filtered to exclude measurements outside the 0°C to 32°C range. This threshold was based on IPMA’s monthly climate bulletins, which confirmed that the highest recorded temperature in the Azores between 2010 and 2023 was 31.9°C, observed in July 2015 [ 32 ]. Afterward, it was generated a complete set of date-station combinations and the data was then temporally interpolated to fill some data gaps originated by sensor downtime to ensure that there were measurements available for every day. To spatially match the temperature and relative humidity data with the sporulation data, we identified the nearest meteorological station for each sporulation record using the scipy.spatial library to build k-dimensional trees (cKDTree) and with a 0.5 distance margin. Temperature and relative humidity values ​​were then returned for the day of observation of each sporulation record and for the previous 360 days, given the fact that fungal predictive models can perform better when both short- and long-term environmental variation is represented [ 33 ]. Since it has been established that the topography of the terrain can also influence the amount of spores, affecting e.g. exposure and heat input [ 7 , 20 ], we also integrated into the model the following variables: elevation, slope, and aspect. Because variation in natural light and water can influence the number of spores, these topographic variables act as proxy variables of those relationships. Elevation data for Terceira Island was obtained in raster format and processed in R (V4.3.3) using the Terra and Raster libraries. First, the raster data was projected to the WGS84 coordinate system (EPSG:4326), and used to generate the terrain slope and aspect (both in degrees) with the terrain() function from the terra package. With Python 3.12.3 these variables were then interpolated as well, using an interpolation method based on the nearest neighbour. Lastly, for each sporulation record, we obtained the index of the closest topographic coordinate in order to retrieve the respective elevation, slope and aspect values. 2.4. Mcfly - A deep learning tool for time series classification and regression Modelling exercises that aim to classify time series data are generally regarded as Time Series Classification (TSC) tasks [ 23 ]. In this study, we use mcfly, a python library used for deep learning-based TSC, that automates most of the procedures of model architecture implementation and testing (‘AutoML’; [ 23 , 34 ]). This library makes use of TensorFlow, a machine learning library; allows classification of both univariate and multivariate time series data; and also the usage of hardware, such as GPUs. Additionally, it also serves as a wrapper for the Keras API framework, choosing the model’s architectures and hyperparameters based on a random search [ 34 ]. Mcfly AutoML workflow, generates and identifies suitable deep learning models with random architectures and hyperparameters from a predefined range of values, and entails the partition of the full dataset, including both dependent and predictor variables. These ‘candidate’ models are initially trained with a small subset of data and epochs, and their performance is evaluated through a validation dataset, so that the best model (i.e. with the highest performance) is selected and the trained with the full training data, and over optimal number of epochs [ 34 ]. There are four deep learning architectures that mcfly supports: Convolutional Neural Networks (CNN), Deep Convolutional Long Short-Term Memory networks (DeepConvLSTM), Residual Networks (ResNet), and InceptionTime. CNNs, despite being mostly used for object recognition, are also suitable for TSC given their ability to learn robust deep features [ 35 ]. According to van Kuppevelt et al. [ 34 ], the architecture has N convolutional layers with ReLU activation, a single hidden dense layer, and when generated by mcfly, does not use pooling layers, as reducing spatial size is often unnecessary. The DeepConvLSTM architecture combines convolutional layers with LSTM recurrent neural networks, where convolutions are applied per channel before connecting to the first LSTM layer and making it suitable for sequence-type input data like time series [ 24 , 34 ]. The ResNet architecture was also initially built for image recognition, but has since been successfully showing good results for TSC, by using multiple residual blocks, each containing three convolutions, and a skip connection [ 34 , 36 ]. Lastly, InceptionTime is a recent architecture that comprises a set of five deep learning models for TSC, each created by a stream of several inception modules and varying kernel sizes [ 34 , 37 ]. For this study, we adapted the methodology scheme used by Capinha et al. [ 23 ] and Ceia-Hasse et al. [ 24 ]. We randomly split our entire dataset into four partitions composed of: a training subset (50% of all instances; A), a ‘internal’validation subset (25%; B), a full training subset (75%; AB = A + B), and a last partition used as a final test to assess the performance of the chosen best model (25%; C). Using Python 3.7.12, we used mcfly to generate 20 candidate models. The modelgen function randomly chooses the model types, alternating between the architectures, in order to have approximately the same number of models per architecture [ 34 ]. These models were then trained during 5 epochs with the training (A), and for each epoch their performance was evaluated against the validation (B) data through the following metrics: loss values, area under the receiver operating characteristic curve (AUC) values, validation loss values, and validation AUC values. Through the AUC metric, the predictive ability of a model is considered “perfect” if the AUC = 1, “good” if AUC > 0.8 and that a model with a score of 0.5 does not discriminate better than randomly generated values [ 38 , 39 ]. The candidate model with the best performance in the validation subset is then chosen, according to measured AUC values. This model is then trained with the full training dataset (AB) during the number of epochs that maximized performance in the validation set. After this training, fungal sporulation is predicted and tested using the last data subset (C). To calculate the importance of each predictor variable, the permutation method was used. This method involves shuffling the values of each individual variable of the training data, and evaluating the decrease in the AUC. A higher reduction in the AUC translates to a greater importance of a variable [ 39 ]. The response curves for each predictive variable were also generated for the final test dataset partition using NumPy and Matplotlib, in order to help interpret the model and analyze how variations in their values influence the dependent variable. We generated them by creating copies of the test data, then modifying one feature at a time, replacing its values with a range from min to max, and passing the modified data to model.predict() . In order to further comprehend the relationship between the chosen predictive variables and the fungus sporulation, we generated, through the best performing model, a .csv file containing the class probability predictions for the test dataset, using the predict function. In this instance, being a binary classification, this function returned a matrix where each row has both the probability score of being classified as ‘low’ or ‘high’ risk of sporulation. Those predictions were used to plot a histogram of the distribution of predicted probabilities for the ‘low’ and ‘high’ risk classes, using the apply() function in Pandas, which compares row by row the scores in the \"low\" and \"high\" classes. 3. Results Out of the 20 models generated with mcfly, the one that demonstrated the best performance was a model with a CNN-type architecture, achieving a validation AUC of 0.83 (candidate model number 2, as referenced in Table 1 and Fig. 1 ). After further training for two additional epochs with the full training data (AB), the model demonstrated a steady increase in AUC throughout the epochs, achieving an internal AUC of 0.93 and an AUC of 0.81 in the test data. Table 1 Performance of Candidate Models. The best performing model achieved a validation AUC of 0.8258, having a CNN architecture. Candidate Model Architecture Train AUC Train Loss Val AUC Val Loss 0 InceptionTime 0.6819 0.6587 0.5645 33.0189 1 ResNet 0.5933 0.6735 0.5645 47.5007 2 CNN 0.9276 3.3416 0.8258 3.2641 3 DeepConvLSTM 0.8612 12.7140 0.7118 12.1372 4 InceptionTime 0.6290 0.6704 0.5602 1.6686 5 CNN 0.8192 0.9364 0.7375 1.0577 6 ResNet 0.5948 0.6759 0.4355 22.9626 7 DeepConvLSTM 0.6666 1.2840 0.4355 1.2616 8 DeepConvLSTM 0.8188 2.2299 0.5646 2.3523 9 ResNet 0.5543 0.7071 0.4363 13.4807 10 InceptionTime 0.6219 0.6803 0.5820 2.5696 11 CNN 0.7677 2.2624 0.7218 2.1406 12 InceptionTime 0.5941 0.6993 0.5645 25.0352 13 CNN 0.9613 1.4363 0.7136 1.7031 14 DeepConvLSTM 0.8487 14.9051 0.5470 14.2553 15 ResNet 0.7105 0.6296 0.5645 77.3314 16 ResNet 0.5984 0.6820 0.5543 5.4269 17 DeepConvLSTM 0.8417 12.8830 0.4839 13.0282 18 InceptionTime 0.5749 0.6857 0.5647 0.7139 19 CNN 0.7638 2.3186 0.6624 2.9733 The calculated permutation importance demonstrated that there was a higher decrease in the AUC when the mean daily temperature values were permuted, dropping by 0.24 (Fig. 2 ). The remaining variables caused much smaller performance drops, indicating minor contributions to the model. In second place in the hierarchy of importance comes elevation, with a loss of 0.02 followed by mean daily relative humidity with a AUC loss of just 0.01. The randomization of the remaining topographic variables (slope and aspect) showed negligeable losses in the AUC values. According to the response curves generated, in the case of the average daily temperature (Fig. 3 a), the temperature range where the probability of sporulation is higher is between approximately 10ºC to 20ºC, while the peak probability occurs when the temperature is around 15ºC. Extreme temperatures appear to be associated with lower sporulation capacity. In the case of the elevation, the second most important variable for the study, it is possible to observe that the probability peak is higher when the elevation values ​​are around 300 meters (Fig. 3 b). Very high elevation​​ translates into a sharp decrease in the probability of species’ sporulation. Regarding relative humidity, the third most important variable for the study, the peak probability occurs at around 75% relative humidity (Fig. 3 c). From this value onwards, the probability rapidly decreases. Our model also generated a .csv file containing class probability predictions for the test dataset, estimating the probability of each sample belonging to both ‘low’ and ‘high’ risk classes. Among the 287 predictions, 234 were classified as low risk of sporulation, while the remaining 53 were classified as high risk. As observed in the histogram from Fig. 4 , the distribution for both classes is distinct yet symmetric, with ‘low’ predictions concentrated near 1 and ‘high’ predictions between 0.05 e 0.4. Additionally and especially between 0.4 and 0.6, there appears to be a zone of ‘uncertainty’ where the classes appear in a more balanced way. 4. Discussion Using the Python programming language and the mcfly software library, we successfully implemented deep learning techniques to estimate the sporulation of Pithomyces chartarum on Terceira Island, Azores (Portugal). This effort addresses a critical need, as the fungus produces a toxin that, when ingested by grazing livestock, leads to facial eczema—a condition with severe implications for animal health and significant economic losses [e.g., 3, 9]. For this prediction task, we selected the candidate model with the highest validation AUC, which featured a convolutional neural network (CNN) architecture. This outcome is unsurprising, as CNNs are widely used in ecological modeling due to their strong performance in tasks such as species distribution prediction [ 21 , 22 ]. Notably, models employing a DeepConvLSTM architecture also showed promising results—particularly candidate model 3 (see Table 1 and Fig. 1 ). The selected CNN model was subsequently trained for two additional epochs, achieving an average AUC of 0.93 and a validation AUC of 0.81. These results reflect strong predictive performance, indicating the model’s capacity to effectively distinguish between high and low sporulation risk days. Our model generated 53 predictions identified as ‘high’ risk of sporulation. If used operationally, these would correspond to the areas on which preventive measures should be considered (e.g., avoid grazing and use preserved feed instead). When plotting the distribution of predicted probabilities for 'low' and 'high' classes (Fig. 4 ), the strong spikes around 0.1 and 0.9 suggest that the model is confident and has a clear ability to distinguish between ‘low’ and ‘high’ risk in many of its predictions. Being a binary classification, in most cases, when the model predicts a high probability for one of the classes, inevitably assigns a low probability score to the opposite class. Between the two meteorological variables used, our results show that there is a big emphasis on the mean daily temperature. This was the only variable that, when its values ​​were randomized, produced a significant drop in the model's performance and in its AUC score. The generated response curves help understand how the variation of predictive variables affects fungal sporulation. In the case of temperature, our results suggest that the temperature range where the probability of sporulation is higher is between approximately 10ºC to 20ºC, which is mostly within the temperature range normally identified in the literature as being most favorable for P. chartarum sporulation (i.e. 12ºC to 24ºC; [ 4 , 12 , 15 , 18 ]). However, even though 24ºC is usually identified as the optimal temperature for the fungus sporulation, we identified a different peak temperature, 15ºC. Despite that, in 2005, Pinto et al. analyzed pithomycotoxicosis in ruminants in the Azores, and with meteorological data from 1995 to 2001 stated that favourable conditions to fungal growth and sporulation were observed whenever there was more than 10 days with a minimum temperature of 16ºC [ 8 ]. According to Heras et al. [ 7 ], similar conditions were observed in Asturias, demonstrating the possible influence of oceanic climate on the fungus sporulation. Ávila et al. [ 3 ] identified average temperature over the 90 days preceding spore counts as a key variable, showing a strong positive correlation with spore abundance—that is, higher average temperatures were associated with greater spore quantities. This finding underscores the importance of capturing seasonal patterns in sporulation forecasting. Our results extend this insight by suggesting that the influence of mean temperature may not be limited to the preceding 90 days, but may instead span up to 360 days. This highlights the possibility that P. chartarum sporulation is shaped by the broader environmental history of a given location, rather than by short-term conditions alone. Regarding elevation, this was the most important topographic variable for the model. According to the generated response curve (Fig. 3 b), the probability peak is higher when the elevation values ​​are around 300 meters. This situation is concerning, since the eastern part of Terceira Island is dominated by a huge expanse of pastureland, with an average altitude of around 390 meters, and an even more extensive plain to the north, with an average altitude of 200 meters [ 40 ]. Therefore, it is possible to confirm that there are several areas that may be within these sections with an elevation more susceptible to the fungus sporulation, and that represent pasture areas. These pasture areas also favour the availability of decaying plant material, which is pre-established as one of the ideal conditions for the fungus sporulation [ 16 , 19 ]. This peak in sporulation observed at elevations of 300 metres may be explained by the decrease in temperature associated with higher altitudes, since as previously noted, extreme temperatures (both too low or too high) are associated with a strong decrease in the fungus’ risk of sporulation. Regarding relative humidity, this variable occupies the third place in the permutation importance hierarchy (Fig. 3 c). Humidity is also one of the most frequently mentioned aspects in the literature when characterizing the fungus's suitability conditions, despite our results indicating a minor contribution to the model. It is usually a consensus that moist conditions, with relative humidity above 90% is more suitable for sporulation and fungal growth [ 13 , 18 ], however our results identified the range of 70–75% as the one with the greatest probability of risk. Unicol, an agricultural cooperative based on Terceira Island and with specialists responsible for the measurements of P. chartarum spores in grass samples used in this study, identifies the following necessary conditions for the formation of spores: temperature higher than 16ºC and humidity above 76% for 72 consecutive hours [ 31 ]. Nonetheless, despite the slight differences in identifying a range of relative humidity that translates into a higher risk of this species’ sporulation, it is proven that the fungus requires a moist environment in order for sporulation to occur and our results are aligned with this trend. Additionally, Ávila et al. [ 3 ] also reached a similar conclusion, stating that relative humidity seems to have a slight effect on the fungus sporulation and that this situation could be explained by the fact that Terceira island experiences a consistently high relative humidity. Overall, deep learning proved to be a powerful tool to analyze complex relationships between the chosen environmental variables and the sporulation data, further strengthening its position as a transformative tool regarding ecological studies and accuracy in prediction exercises. Our model ultimately allows for a better understanding of the ecological processes associated with P. chartarum sporulation, while enhancing the ability to forecast high-risk sporulation events. 5. Conclusions In this study, we successfully implemented a convolutional neural network (CNN) to forecast the sporulation of Pithomyces chartarum on Terceira Island, in the Azores archipelago (Portugal). The results demonstrate the potential of deep learning approaches to address ecological challenges, particularly in the context of fungal pathogens with significant animal health and economic impacts. P. chartarum produces a toxin that causes pithomycotoxicosis in grazing livestock, posing a serious threat to animal welfare and leading to substantial financial losses. Among the predictive variables tested, mean daily temperature emerged as the most influential driver of sporulation risk, in line with known ecological conditions favoring fungal growth. Our findings further suggest that temperature effects are not limited to short-term windows, but may extend across a full year, highlighting the importance of incorporating long-term environmental history into predictive models. The relatively limited contribution of other variables indicates that additional environmental or biological factors—such as precipitation, soil properties, or pasture type—may play a role and warrant inclusion in future modelling efforts to improve predictive accuracy. Importantly, the current deep learning-based model is not intended as a standalone solution, but as a component to be integrated into an existing early warning system already in use. By improving the reliability and accuracy of sporulation forecasts, this model enhances the system’s ability to provide timely alerts, enabling specialists to refine disease control strategies and supporting farmers in making informed management decisions. Furthermore, the framework developed here has the potential for adaptation to other regions affected by P. chartarum , highlighting its broader applicability in mitigating the impacts of pithomycotoxicosis. Declarations Additional Information The authors declare no competing interests. Funding This research was funded by the mobilizing agenda New Space Portugal, as part of Portugal's Recovery and Resilience Plan (RRP) - Project nº 02/C05-i01.01/2022.PC644936537-00000046; IAPMEI Projeto Nº11. Author Contribution I.D: Writing – original draft, Visualization, Code, Methodology, Investigation, Formal analysis, Data curation. C.C: Writing – review & editing, Validation, Code, Methodology, Formal analysis. J.P.: Writing – review & editing, Validation, Supervision, Project administration, Funding acquisition, Methodology, Formal analysis, Conceptualization, Resources. E.D: Methodology, Writing – review & editing. M.A: Methodology, Writing – review & editing. All authors have read and approved the final manuscript.​ Acknowledgement We would like to thank TERINOV for providing access to essential data from the IoT weather stations, and the Municipality of Angra do Heroísmo (Terceira) for their support in the initial development of the LoRaWAN Network. We would also like to acknowledge the Regional Veterinary Laboratory staff for their timely and consistent spore counts throughout the project, and UNICOL. Data Availability The data presented in this study is available upon request from the corresponding author. The dataset of the published report will be made publicly available in a repository. References Dingley, J. M. Pithomyces chartarum , its occurrence morphology, and taxonomy. N. Z. J. Agric. Res. 5 , 49–61 (1962). Huang, M. & Hull, C. M. Sporulation: how to survive on planet Earth (and beyond). Curr. Genet. 63 , 831–838 (2017). Ávila, M. et al. Assessing the presence of Pithomyces chartarum in pastureland using IoT sensors and remote sensing: the case study of Terceira Island (Azores, Portugal). Sensors 24 , 4485 (2024). Brook, P. J. Ecology of the fungus Pithomyces chartarum (Berk. & Curt.) M.B. Ellis in pasture in relation to facial eczema disease of sheep. N. Z. J. Agric. Res. 6 , 147–228 (1963). Wearn, J. Pithomyces chartarum – a fungus on the up? Field Mycol. 10 , 36–37 (2009). Cuttance, E. L., Laven, R. A. & Stevenson, M. A. Variability in measurement of Pithomyces chartarum spore counts. N. Z. Vet. J. 65 , 192–197 (2017). De las Heras, M. et al. Chronic pithomycotoxicosis associated with obstructive rhinopathy in sheep. Vet. Pathol. 59 , 950–959 (2022). Pinto, C. et al. Pithomycotoxicosis (facial eczema) in ruminants in the Azores, Portugal. Vet. Rec. 157 , 805–810 (2005). Marbrook, J. & Matthews, R. E. F. Loss of sporidesmin from spores of Pithomyces chartarum (Berk. & Curt.) M.B. Ellis. N. Z. J. Agric. Res. 5 , 223–236 (1962). Di Menna, M. E., Smith, B. L. & Miles, C. O. A history of facial eczema (pithomycotoxicosis) research. N. Z. J. Agric. Res. 52 , 345–376 (2009). Morris, C. A. et al. Inheritance of resistance to facial eczema: a review of research findings from sheep and cattle in New Zealand. N. Z. Vet. J. 52 , 205–215 (2004). Fernández, M. et al. Pathological study of facial eczema (pithomycotoxicosis) in sheep. Animals 11 , 1070 (2021). Dennis, N. A., Amer, P. R. & Meier, S. Brief communication: predicting the impact of climate change on the risk of facial eczema outbreaks throughout New Zealand. Proc. N. Z. Soc. Anim. Prod. 74 , 1–6 (2014). Liu, L. et al. Fungi, including Pithomyces chartarum , cause facial eczema and inflammation in grazing sheep in Western China. Microb. Pathog. 185 , 106451 (2023). Smith, J. D., Lees, F. T. & Crawley, W. E. Weather conditions, spore counts, and facial eczema in test sheep. N. Z. J. Agric. Res. 8 , 63–87 (1965). Schöniger, S. et al. Prototheca species and Pithomyces chartarum as causative agents of rhinitis and/or sinusitis in horses. J. Comp. Pathol. 155 , 121–125 (2016). Van Wuijckhuise, L. et al. First case of pithomycotoxicosis (facial eczema) in the Netherlands. Tijdschr. Diergeneeskd. 131 , 858–861 (2006). Dijkstra, E., Harkema, L. & Vellema, P. First case of pithomycotoxicosis in sheep in the Netherlands. Vet. Rec. Case Rep. 10 , e268 (2022). Mitchell, K. J., Thomas, R. G. & Clarke, R. T. J. Factors influencing the growth of Pithomyces chartarum in pasture. N. Z. J. Agric. Res. 4 , 566–577 (1961). Di Menna, M. E. & Bailey, J. R. Pithomyces chartarum spore counts in pasture. N. Z. J. Agric. Res. 16 , 343–351 (1973). Christin, S., Hervet, É. & Lecomte, N. Applications for deep learning in ecology. Methods Ecol. Evol. 10 , 1632–1644 (2019). Borowiec, M. L. et al. Deep learning as a tool for ecology and evolution. Methods Ecol. Evol. 13 , 1640–1660 (2022). Capinha, C., Ceia-Hasse, A., Kramer, A. M. & Meijer, C. Deep learning for supervised classification of temporal data in ecology. Ecol. Inform. 61 , 101252 (2021). Ceia-Hasse, A., Sousa, C. A., Gouveia, B. R. & Capinha, C. Forecasting the abundance of disease vectors with deep learning. Ecol. Inform. 78 , 102272 (2023). Lhoumeau, S., Pinelo, J. & Borges, P. A. Artificial intelligence for biodiversity: exploring the potential of recurrent neural networks in forecasting arthropod dynamics based on time series. Ecol. Indic. 171 , 113119 (2025). Rammer, W. & Seidl, R. Harnessing deep learning in ecology: an example predicting bark beetle outbreaks. Front. Plant Sci. 10 , 451705 (2019). Raposeiro, P. M. et al. Climate change facilitated the early colonization of the Azores Archipelago during medieval times. Proc. Natl. Acad. Sci. USA 118 , e2108236118 (2021). Portal do Ordenamento do Território dos Açores. Caracterização e identificação das paisagens dos Açores | Terceira. https://ot.azores.gov.pt/Unidades-Paisagem-Ficha.aspx?id=78 (2025). INE. Recenseamento Agrícola – Análise dos principais resultados – 2019. Inst. Nac. Estatística, Lisboa (2021). Almeida, A. M. de, Alvarenga, P. & Fangueiro, D. The dairy sector in the Azores Islands: possibilities and main constraints towards increased added value. Trop. Anim. Health Prod. 53 , 1–9 (2021). Unicol. Sistema de Alertas Unicol – SAU. https://www.unicol.pt/sistema-de-alertas-unicol-sau/ (2025). IPMA. Boletim climatológico mensal – Julho de 2015. Ministério da Agricultura e do Mar. http://www.ipma.pt (2015). Capinha, C. Predicting the timing of ecological phenomena using dates of species occurrence records: a methodological approach and test case with mushrooms. Int. J. Biometeorol. 63 , 1015–1024 (2019). Van Kuppevelt, D. et al. Mcfly: automated deep learning on time series. SoftwareX 12 , 100548 (2020). Zhao, B., Lu, H., Chen, S., Liu, J. & Wu, D. Convolutional neural networks for time series classification. J. Syst. Eng. Electron. 28 , 162–169 (2017). Fawaz, H. I. et al. Deep learning for time series classification: a review. Data Min. Knowl. Discov. 33 , 917–963 (2019). Fawaz, H. I. et al. InceptionTime: finding AlexNet for time series classification. Data Min. Knowl. Discov. 34 , 1936–1962 (2020). Phillips, S. J., Dudík, M. & Schapire, R. E. A maximum entropy approach to species distribution modeling. In Proc. 21st Int. Conf. Mach. Learn. 83–90 (ACM, 2004). Diogo, I., Sillero, N. & Capinha, C. Predicting the risk of invasion by broadleaf watermilfoil ( Myriophyllum heterophyllum ) in mainland Portugal. Heliyon 10 , e16744 (2024). SRAM. Livro das Paisagens dos Açores: Contributos para a Identificação e Caracterização das Paisagens dos Açores. https://ot.azores.gov.pt/Biblioteca-Publicacoes.aspx?id=84 (2005). Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {\"props\":{\"pageProps\":{\"initialData\":{\"identity\":\"rs-6453956\",\"acceptedTermsAndConditions\":true,\"allowDirectSubmit\":true,\"archivedVersions\":[],\"articleType\":\"Article\",\"associatedPublications\":[],\"authors\":[{\"id\":454450551,\"identity\":\"f01756b8-d188-4aae-83b5-b9cb97cc8c36\",\"order_by\":0,\"name\":\"Iúri Diogo\",\"email\":\"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA+0lEQVRIiWNgGAWjYBACNiA+wMAmAaSYDzAkkKiFLQGihY1ouxh4DJA4eACfRO7DAwxlFnLy7j0fPzzcYydnLt/78AFDxT27BlzmS6QbHGA4J2FseObsZomEZ8nGlm3sxgYMZ4qTcWtJYzjA2CaRuHFG7gaJhAMHEjccY2OTYGxLSMbpC4SWnMc/gFrqgVrYfxClZb5EDhvIlgQDoC0MQC12OLXwPGM4kAD0iwHPMTOLhAPJhhuOpTFLJJxJSMClRb49jfnDh7I6Ofn25sc3fxywkzc4fIzxw4eKBHtcWsAAZCAw3FBFEhvw6gFZh64Cvy2jYBSMglEwkgAAakRT7HqrXkoAAAAASUVORK5CYII=\",\"orcid\":\"\",\"institution\":\"Atlantic International Research Centre\",\"correspondingAuthor\":true,\"prefix\":\"\",\"firstName\":\"Iúri\",\"middleName\":\"\",\"lastName\":\"Diogo\",\"suffix\":\"\"},{\"id\":454450552,\"identity\":\"b0d3b402-d820-4f26-bf81-4111063df109\",\"order_by\":1,\"name\":\"César Capinha\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Centre of Geographical Studies, , Institute of Geography and Spatial Planning, University of Lisbon\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"César\",\"middleName\":\"\",\"lastName\":\"Capinha\",\"suffix\":\"\"},{\"id\":454450553,\"identity\":\"6de5aed1-3723-4a0f-9465-f441973fd4e6\",\"order_by\":2,\"name\":\"João Pinelo\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Atlantic International Research Centre\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"João\",\"middleName\":\"\",\"lastName\":\"Pinelo\",\"suffix\":\"\"},{\"id\":454450554,\"identity\":\"3df68e28-5070-4c5d-8470-d5070e2e03b9\",\"order_by\":3,\"name\":\"Elizabeth Domingues\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"UNICOL-Cooperativa Agrícola\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Elizabeth\",\"middleName\":\"\",\"lastName\":\"Domingues\",\"suffix\":\"\"},{\"id\":454450555,\"identity\":\"ca61875d-a971-40af-8d5f-945f9043f273\",\"order_by\":4,\"name\":\"Mariana Ávila\",\"email\":\"\",\"orcid\":\"\",\"institution\":\"Atlantic International Research Centre\",\"correspondingAuthor\":false,\"prefix\":\"\",\"firstName\":\"Mariana\",\"middleName\":\"\",\"lastName\":\"Ávila\",\"suffix\":\"\"}],\"badges\":[],\"createdAt\":\"2025-04-15 10:38:33\",\"currentVersionCode\":1,\"declarations\":\"\",\"doi\":\"10.21203/rs.3.rs-6453956/v1\",\"doiUrl\":\"https://doi.org/10.21203/rs.3.rs-6453956/v1\",\"draftVersion\":[],\"editorialEvents\":[],\"editorialNote\":\"\",\"failedWorkflow\":false,\"files\":[{\"id\":82605804,\"identity\":\"6d749db2-fef2-4e1a-bd9a-f70eddd683dc\",\"added_by\":\"auto\",\"created_at\":\"2025-05-13 10:03:36\",\"extension\":\"jpg\",\"order_by\":1,\"title\":\"Figure 1\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":82823,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eValidation AUC for candidate models. For each candidate model (0-19 on the vertical axis), each bar corresponds to one model with one of the four different architectures. The model with the highest validation AUC score is model 2, having a CNN architecture.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"1.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-6453956/v1/7d82bc61163a79115ab19013.jpg\"},{\"id\":82607076,\"identity\":\"187805bc-c156-47ff-ac3d-0f07652c4ffd\",\"added_by\":\"auto\",\"created_at\":\"2025-05-13 10:11:36\",\"extension\":\"jpg\",\"order_by\":2,\"title\":\"Figure 2\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":26080,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003ePermutance importance of each variable (vertical axis). The importance is obtained by calculating the AUC drop after shuffling the values of each individual variable of the training data. There is a higher decrease in the AUC when the mean daily temperature values are permuted.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"2.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-6453956/v1/854f74e51f032db8436abf0f.jpg\"},{\"id\":82607078,\"identity\":\"aafdd36c-4840-4c77-a34f-d1d072434d3a\",\"added_by\":\"auto\",\"created_at\":\"2025-05-13 10:11:36\",\"extension\":\"jpg\",\"order_by\":3,\"title\":\"Figure 3\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":68021,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eResponse curves for mean daily temperature (A), Elevation (B), and mean daily relative humidity (C). The curves are generated by creating copies of the test data, and modifying one feature at a time, replacing its values with a range from min to max, and passing the modified data to \\u003cem\\u003emodel.predict()\\u003c/em\\u003e. Optimal sporulation conditions include mean daily temperature between 10ºC to 20ºC, elevation of 300 meters, and 75% average relative humidity.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"3.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-6453956/v1/bee56163113fa84e866a49ab.jpg\"},{\"id\":82605811,\"identity\":\"6824b971-cd5d-4f4e-8878-4197790dd58e\",\"added_by\":\"auto\",\"created_at\":\"2025-05-13 10:03:36\",\"extension\":\"jpg\",\"order_by\":4,\"title\":\"Figure 4\",\"display\":\"\",\"copyAsset\":false,\"role\":\"figure\",\"size\":37139,\"visible\":true,\"origin\":\"\",\"legend\":\"\\u003cp\\u003eDistribution of predicted probabilities for 'low' and 'high' classes. The distribution for both classes is distinct yet symmetric, with ‘low’ predictions concentrated near 1 and ‘high’ predictions between 0.05 e 0.4.\\u003c/p\\u003e\",\"description\":\"\",\"filename\":\"4.jpg\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-6453956/v1/9f8f6efdc5c104c95aa97fd7.jpg\"},{\"id\":85730152,\"identity\":\"0b4ec1bb-17b9-4bc0-aaa4-cc558d6c3733\",\"added_by\":\"auto\",\"created_at\":\"2025-07-01 07:17:13\",\"extension\":\"pdf\",\"order_by\":0,\"title\":\"\",\"display\":\"\",\"copyAsset\":false,\"role\":\"manuscript-pdf\",\"size\":919189,\"visible\":true,\"origin\":\"\",\"legend\":\"\",\"description\":\"\",\"filename\":\"manuscript.pdf\",\"url\":\"https://assets-eu.researchsquare.com/files/rs-6453956/v1/07926cbd-2a6d-434a-b386-518b0120b107.pdf\"}],\"financialInterests\":\"No competing interests reported.\",\"formattedTitle\":\"A Deep Learning Approach to Predicting Pithomyces chartarum Sporulation for Livestock Protection\",\"fulltext\":[{\"header\":\"1. Introduction\",\"content\":\"\\u003cp\\u003e \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e is a cosmopolitan saprophytic fungus, found in warm temperate, sub-tropical, and tropical regions [\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e], that, like other fungi, uses sporulation as a means of surviving in harsh conditions and dispersing to new environments [\\u003cspan citationid=\\\"CR2\\\" class=\\\"CitationRef\\\"\\u003e2\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Its spores are usually found on several types of vegetation, including grass pastures and dead matter of grasses, living endophytically within them, which can then develop a wide range of symptoms such as brown necrotic leaf spots or lesions on live tissue, but can also occur as a saprophytic mould that can be found in plant debris at the base of pasture grasses [\\u003cspan additionalcitationids=\\\"CR5 CR6\\\" citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e]. Animals grazing on contaminated vegetation are likely to ingest the spores.\\u003c/p\\u003e \\u003cp\\u003eThe fungus produces the toxin sporidesmin, an hepatotoxin [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e], and its ingestion by livestock causes pithomycotoxicosis, also known as facial eczema [\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e]. Affected animals are mainly sheep and cattle but also include, to varying degrees, guinea pigs, rabbits, rats, mice, red and fallow deer, goats and alpacas [\\u003cspan additionalcitationids=\\\"CR11\\\" citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e]. After infection, affected animals are known to develop hepatic photosensitivity resulting from liver damage, specifically injury to the biliary epithelial cells, which leads to acute bile duct obstruction and hepatic insufficiency [\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e]. Other commonly associated symptoms include reduced fertility and decreased production of milk, wool, and pelt, as well as transient diarrhoea, loss of appetite followed by weight loss, skin crusting, and, in severe cases, death [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR6\\\" class=\\\"CitationRef\\\"\\u003e6\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e]. In general, the severity of the intoxication is directly related to the spores concentration and the grazing period [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e]. Clinical signs of infection do not appear until 10\\u0026ndash;14 days after the ingestion of the toxin [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e] but could take up to 20 days to be visible [\\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e]. According to Sch\\u0026ouml;niger et al. [\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e], although rare, \\u003cem\\u003eP. chartarum\\u003c/em\\u003e might cause onychomycosis in humans, and inhaling its spores may lead to asthma.\\u003c/p\\u003e \\u003cp\\u003eOne of the most common methods employed by farmers to prevent pithomycotoxicosis in livestock, is the administration of zinc feed supplements [\\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e]. When ingested (mainly orally), zinc and sporidesmin combine, creating a stable mercaptide and eliminating the toxin during the autoxidation cycle [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. However, this may lead to over or under administration, since the administration is based on the farmer\\u0026rsquo;s past experience, as stated by \\u0026Aacute;vila et al. [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Additional prevention strategies involve restricting access to contaminated pastures and employing fungicide treatments [\\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eFacial eczema was first reported in New Zealand and has since been identified in several Asian countries, including China and the Philippines, as well as in Australia, the United States, Turkey, South Africa, Uruguay, Argentina, and Brazil. In mainland Europe, the first reported case occurred in the Netherlands, with subsequent reports from France, Spain, Portugal, Hungary, and Italy [\\u003cspan citationid=\\\"CR1\\\" class=\\\"CitationRef\\\"\\u003e1\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR10\\\" class=\\\"CitationRef\\\"\\u003e10\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR14\\\" class=\\\"CitationRef\\\"\\u003e14\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR17\\\" class=\\\"CitationRef\\\"\\u003e17\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e]. The spread of facial eczema poses a serious threat not only to animal health but also to agricultural economies, due to animal mortality and reduced productivity and fertility [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e]. For example, in New Zealand, the disease was estimated to cause annual losses of between \\u003cspan\\u003e$\\u003c/span\\u003e63\\u0026nbsp;million and \\u003cspan\\u003e$\\u003c/span\\u003e126\\u0026nbsp;million to sheep farmers alone between 1983 and 1988 [\\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e], illustrating the substantial economic impact of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e.\\u003c/p\\u003e \\u003cp\\u003eAccording to Pinto et al. [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e], \\u003cem\\u003eP. chartarum\\u003c/em\\u003e has been present on Azorean pastures since at least 1999, with 22 outbreaks recorded between 1999 and 2001 alone. Previous studies have shown a strong association between meteorological conditions and \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation, suggesting that these variables may be useful for predicting its occurrence [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Sporulation is influenced by specific environmental conditions, particularly warm temperatures and high moisture levels, which stimulate fungal growth and spore production [\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e]. According to previous studies, the incubation period typically ranges from 3 to 5 days, depending on the season [\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e]. Sporulation typically occurs under conducive environmental conditions, including temperatures above 12\\u0026deg;C during late summer and autumn, with an optimum around 24\\u0026deg;C, and relative humidity exceeding 80% [\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e]. Despite moisture being required for sporulation, heavy or continuing rainfall reduces its occurrence and toxicity, since the spores can become saturated and sporidesmin is water soluble, therefore rainfall should range between 5 to 35 mm per week to enhance sporulation [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR11\\\" class=\\\"CitationRef\\\"\\u003e11\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e]. According to Dennis et al. [\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e] optimal conditions of warm and humid weather for at least three days provides enough time for a successful fungal sporulation and colonisation in pastures. However, in Azores, Portugal, fungal sporulation enhancement was recorded during periods of more than 10 consecutive days with minimum temperatures of at least 16\\u0026deg;C and relative humidity exceeding 90% [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eThe presence of decaying plant material is another important factor promoting \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation, particularly when it includes tissues rich in soluble organic compounds, such as freshly dead leaves [16, 19. Topographic features can influence \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation by altering microclimatic conditions; for instance, higher sporulation levels have been observed in pastures sheltered by hedges and on north- or west-facing slopes, where heat accumulation is greater [\\u003cspan citationid=\\\"CR20\\\" class=\\\"CitationRef\\\"\\u003e20\\u003c/span\\u003e]. Natural light also appears to affect sporidesmin stability, with greater losses occurring on sunny days [\\u003cspan citationid=\\\"CR9\\\" class=\\\"CitationRef\\\"\\u003e9\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eIn 2024, \\u0026Aacute;vila et al. assessed the presence of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e in Azorean pastureland using IoT sensors and remote sensing. However, their approach lacked the predictive capability required for real-time risk assessment [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Given the significant health and economic consequences associated with \\u003cem\\u003eP. chartarum\\u003c/em\\u003e, there is a clear need for early warning systems that support informed livestock management and improve milk production, while also enabling ongoing monitoring and scientific understanding of the fungus\\u0026rsquo;s dynamics. Building on the work of \\u0026Aacute;vila et al. [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e], this study aims to enhance predictive capabilities by applying deep learning techniques\\u0026mdash;specifically time-series neural networks\\u0026mdash;and by incorporating additional environmental variables to forecast \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation.\\u003c/p\\u003e \\u003cp\\u003eDeep learning is a subset of machine learning that relies on complex multilayered and interconnected nodes - Artificial Neural Networks (ANNs) - to process and recognize patterns from data [\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e]. They are composed of trainable parameters, applicable in various science disciplines, that enable a very high performance of tasks ranging from natural language processing to image and sound data classification, computer vision, clustering and prediction [\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e]. In ecology, deep learning has been increasingly used primarily for image, video and sound classification [\\u003cspan additionalcitationids=\\\"CR23\\\" citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e\\u0026ndash;\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e]. More recently, deep learning has shown potential for biodiversity monitoring and conservation planning, where, for example, Recurrent Neural Networks (RNNs) were applied to long-term arthropod monitoring data from Terceira Island (Azores), outperforming traditional models (e.g., Local Polynomial Regression) in goodness of fit and overall accuracy [\\u003cspan citationid=\\\"CR25\\\" class=\\\"CitationRef\\\"\\u003e25\\u003c/span\\u003e]. Compared to classical machine learning models, which rely on expert knowledge to identify predefined features, deep learning is considered powerful, by automatically extracting relevant patterns from raw data [\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR26\\\" class=\\\"CitationRef\\\"\\u003e26\\u003c/span\\u003e]. In the context of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation, this means, for example, using raw time series of weather variables as predictors in the model, rather than relying on predictors that summarize weather conditions over predefined time periods (e.g., weeks, months; [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]).\\u003c/p\\u003e \\u003cp\\u003eIn this study, we test the use of deep learning for early warning of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation levels. Ecologists often rely on deep learning models to accurately forecast future events, identifying the set of time series features that most accurately predict, for example, disease vector abundances, not needing to rely on human expertise [\\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e]. Additionally, Rammer \\u0026amp; Seidl [\\u003cspan citationid=\\\"CR26\\\" class=\\\"CitationRef\\\"\\u003e26\\u003c/span\\u003e] state that deep learning techniques are good at generalizing beyond test data and that is a key-factor in prediction exercises for ecological studies, performing better than other machine learning methods in terms of prediction accuracy.\\u003c/p\\u003e \\u003cp\\u003eTherefore, we present and evaluate a deep learning-based approach for forecasting the sporulation of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e on Terceira Island (Azores), using spore abundance data together with meteorological and topographic variables as predictors. This approach is intended to be integrated into the workflow of the existing sporulation risk alert system.\\u003c/p\\u003e\"},{\"header\":\"2. Materials and Methods\",\"content\":\"\\u003cdiv id=\\\"Sec3\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.1. Study Area\\u003c/h2\\u003e \\u003cp\\u003eThe Azores archipelago (36 to 39\\u0026deg; N, 25 to 31\\u0026deg; W), situated in the Macaronesia region of the North Atlantic Ocean, is composed of nine islands [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e]. The archipelago has a temperate oceanic climate, influenced by a semi-permanent subtropical Atlantic anticyclone (\\u0026ldquo;Azores High\\u0026rdquo;), with a low thermal amplitude (9\\u0026ndash;26\\u0026deg;C), high humidity (annual average of 80%), and annual rainfall ranging from 900 to 3000 mm [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR27\\\" class=\\\"CitationRef\\\"\\u003e27\\u003c/span\\u003e]. Terceira island has a surface area of 400 km2, making it the third largest of the archipelago; it has an elliptical shape, approximately 29 km from east to west, and 17.5 km from north to south, with its peak altitude in the Serra de Santa B\\u0026aacute;rbara (1021 m) [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR28\\\" class=\\\"CitationRef\\\"\\u003e28\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eAccording to the latest agricultural census in Portugal, in 2019 there were approximately 296196 grazing livestock in the Azores that could be affected by pithomycotoxicosis, more than 33% of which were dairy animals including cattle, sheep, and goats [\\u003cspan citationid=\\\"CR29\\\" class=\\\"CitationRef\\\"\\u003e29\\u003c/span\\u003e]. With such a large number of potentially endangered individuals, there could be a high negative impact on agriculture, a key economic sector of the archipelago [\\u003cspan citationid=\\\"CR30\\\" class=\\\"CitationRef\\\"\\u003e30\\u003c/span\\u003e].\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec4\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.2. Data collection: \\u003cem\\u003eP. chartarum\\u003c/em\\u003e spore count\\u003c/h2\\u003e \\u003cp\\u003eThe spore count data was initially collected as part of the previous study led by [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Experts from Unicol collected multiple grass samples per week from various grazing locations in Terceira island, and quantified them at the Laborat\\u0026oacute;rio Regional de Veterin\\u0026aacute;ria using a washing method (spores/gram), from June 2021 to December 2023 [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e]. Additionally, samples were collected during periods identified as both favorable and not favorable for sporulation [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e], in order to evaluate the spores concentration over time. All grass samples were collected with permission, and sampling procedures adhered to established local agricultural and veterinary protocols.\\u003c/p\\u003e \\u003cp\\u003eThe sporulation and meteorological datasets, originally stored in databases, were downloaded and saved in Feather format to facilitate efficient processing; the entire analysis was then carried out programmatically. The data was read into Python (V3.12.3) using the Pandas and Pyarrow software libraries. After filtering the dataset to include only records from 2023, missing values were removed. Based on expert input from Unicol and spore counts per gram, two sporulation risk classes were defined: Low (\\u0026le;\\u0026thinsp;10,000 spores/gram) and High (\\u0026gt;\\u0026thinsp;10,000 spores/gram). Although pastures are typically considered hazardous when spore concentrations exceed 30,000 spores per gram of grass, prolonged grazing under concentrations as low as 10,000 spores per gram can still pose a significant risk to livestock [\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e]. To enable binary classification, one-hot encoding was applied to convert the categorical classes into 0 and 1.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec5\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.3. Predictor variables\\u003c/h2\\u003e \\u003cp\\u003eGuided by the known ecological preferences of the fungus (see Introduction), we selected a set of predictive variables designed to capture the meteorological and topographic factors influencing sporulation. Previous studies found that the fungus sporulation is stimulated by warm temperatures and high levels of moisture [e.g. 4, 19]. \\u0026Aacute;vila et al. [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e] explains how temperature and relative humidity data was collected in Terceira island specifically for this project: hourly data was collected through a network of Internet of Things (IoT) weather stations strategically positioned after an analysis performed by the AIR Centre team using data from three meteorological stations from the Instituto Portugu\\u0026ecirc;s do Mar e da Atmosfera (IPMA). These measurements were then transmitted to the IoT server via the LoRaWAN protocol, and stored and accessed in a MySQL database of the AIR Centre. There are measurements available from 26 July 2021 onwards, from that network composed of 59 IoT weather stations [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eUsing Python 3.12.3, these hourly meteorological variables were used to first calculate the mean daily values, at a station level, because the existing literature suggests that the fungus is more influenced by general environmental conditions than by short-term fluctuations [e.g. 7, 14]. To account for a few implausible extreme values likely caused by equipment errors, temperature data were filtered to exclude measurements outside the 0\\u0026deg;C to 32\\u0026deg;C range. This threshold was based on IPMA\\u0026rsquo;s monthly climate bulletins, which confirmed that the highest recorded temperature in the Azores between 2010 and 2023 was 31.9\\u0026deg;C, observed in July 2015 [\\u003cspan citationid=\\\"CR32\\\" class=\\\"CitationRef\\\"\\u003e32\\u003c/span\\u003e]. Afterward, it was generated a complete set of date-station combinations and the data was then temporally interpolated to fill some data gaps originated by sensor downtime to ensure that there were measurements available for every day. To spatially match the temperature and relative humidity data with the sporulation data, we identified the nearest meteorological station for each sporulation record using the scipy.spatial library to build k-dimensional trees (cKDTree) and with a 0.5 distance margin. Temperature and relative humidity values ​​were then returned for the day of observation of each sporulation record and for the previous 360 days, given the fact that fungal predictive models can perform better when both short- and long-term environmental variation is represented [\\u003cspan citationid=\\\"CR33\\\" class=\\\"CitationRef\\\"\\u003e33\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eSince it has been established that the topography of the terrain can also influence the amount of spores, affecting e.g. exposure and heat input [\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR20\\\" class=\\\"CitationRef\\\"\\u003e20\\u003c/span\\u003e], we also integrated into the model the following variables: elevation, slope, and aspect. Because variation in natural light and water can influence the number of spores, these topographic variables act as proxy variables of those relationships.\\u003c/p\\u003e \\u003cp\\u003eElevation data for Terceira Island was obtained in raster format and processed in R (V4.3.3) using the Terra and Raster libraries. First, the raster data was projected to the WGS84 coordinate system (EPSG:4326), and used to generate the terrain slope and aspect (both in degrees) with the \\u003cem\\u003eterrain()\\u003c/em\\u003e function from the terra package. With Python 3.12.3 these variables were then interpolated as well, using an interpolation method based on the nearest neighbour. Lastly, for each sporulation record, we obtained the index of the closest topographic coordinate in order to retrieve the respective elevation, slope and aspect values.\\u003c/p\\u003e \\u003c/div\\u003e \\u003cdiv id=\\\"Sec6\\\" class=\\\"Section2\\\"\\u003e \\u003ch2\\u003e2.4. Mcfly - A deep learning tool for time series classification and regression\\u003c/h2\\u003e \\u003cp\\u003eModelling exercises that aim to classify time series data are generally regarded as Time Series Classification (TSC) tasks [\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e]. In this study, we use mcfly, a python library used for deep learning-based TSC, that automates most of the procedures of model architecture implementation and testing (\\u0026lsquo;AutoML\\u0026rsquo;; [\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e]). This library makes use of TensorFlow, a machine learning library; allows classification of both univariate and multivariate time series data; and also the usage of hardware, such as GPUs. Additionally, it also serves as a wrapper for the Keras API framework, choosing the model\\u0026rsquo;s architectures and hyperparameters based on a random search [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eMcfly AutoML workflow, generates and identifies suitable deep learning models with random architectures and hyperparameters from a predefined range of values, and entails the partition of the full dataset, including both dependent and predictor variables. These \\u0026lsquo;candidate\\u0026rsquo; models are initially trained with a small subset of data and epochs, and their performance is evaluated through a validation dataset, so that the best model (i.e. with the highest performance) is selected and the trained with the full training data, and over optimal number of epochs [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eThere are four deep learning architectures that mcfly supports: Convolutional Neural Networks (CNN), Deep Convolutional Long Short-Term Memory networks (DeepConvLSTM), Residual Networks (ResNet), and InceptionTime. CNNs, despite being mostly used for object recognition, are also suitable for TSC given their ability to learn robust deep features [\\u003cspan citationid=\\\"CR35\\\" class=\\\"CitationRef\\\"\\u003e35\\u003c/span\\u003e]. According to van Kuppevelt et al. [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e], the architecture has N convolutional layers with ReLU activation, a single hidden dense layer, and when generated by mcfly, does not use pooling layers, as reducing spatial size is often unnecessary. The DeepConvLSTM architecture combines convolutional layers with LSTM recurrent neural networks, where convolutions are applied per channel before connecting to the first LSTM layer and making it suitable for sequence-type input data like time series [\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e]. The ResNet architecture was also initially built for image recognition, but has since been successfully showing good results for TSC, by using multiple residual blocks, each containing three convolutions, and a skip connection [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR36\\\" class=\\\"CitationRef\\\"\\u003e36\\u003c/span\\u003e]. Lastly, InceptionTime is a recent architecture that comprises a set of five deep learning models for TSC, each created by a stream of several inception modules and varying kernel sizes [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR37\\\" class=\\\"CitationRef\\\"\\u003e37\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eFor this study, we adapted the methodology scheme used by Capinha et al. [\\u003cspan citationid=\\\"CR23\\\" class=\\\"CitationRef\\\"\\u003e23\\u003c/span\\u003e] and Ceia-Hasse et al. [\\u003cspan citationid=\\\"CR24\\\" class=\\\"CitationRef\\\"\\u003e24\\u003c/span\\u003e]. We randomly split our entire dataset into four partitions composed of: a training subset (50% of all instances; A), a \\u0026lsquo;internal\\u0026rsquo;validation subset (25%; B), a full training subset (75%; AB\\u0026thinsp;=\\u0026thinsp;A\\u0026thinsp;+\\u0026thinsp;B), and a last partition used as a final test to assess the performance of the chosen best model (25%; C). Using Python 3.7.12, we used mcfly to generate 20 candidate models. The \\u003cem\\u003emodelgen\\u003c/em\\u003e function randomly chooses the model types, alternating between the architectures, in order to have approximately the same number of models per architecture [\\u003cspan citationid=\\\"CR34\\\" class=\\\"CitationRef\\\"\\u003e34\\u003c/span\\u003e]. These models were then trained during 5 epochs with the training (A), and for each epoch their performance was evaluated against the validation (B) data through the following metrics: loss values, area under the receiver operating characteristic curve (AUC) values, validation loss values, and validation AUC values. Through the AUC metric, the predictive ability of a model is considered \\u0026ldquo;perfect\\u0026rdquo; if the AUC\\u0026thinsp;=\\u0026thinsp;1, \\u0026ldquo;good\\u0026rdquo; if AUC\\u0026thinsp;\\u0026gt;\\u0026thinsp;0.8 and that a model with a score of 0.5 does not discriminate better than randomly generated values [\\u003cspan citationid=\\\"CR38\\\" class=\\\"CitationRef\\\"\\u003e38\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR39\\\" class=\\\"CitationRef\\\"\\u003e39\\u003c/span\\u003e].\\u003c/p\\u003e \\u003cp\\u003eThe candidate model with the best performance in the validation subset is then chosen, according to measured AUC values. This model is then trained with the full training dataset (AB) during the number of epochs that maximized performance in the validation set. After this training, fungal sporulation is predicted and tested using the last data subset (C).\\u003c/p\\u003e \\u003cp\\u003eTo calculate the importance of each predictor variable, the permutation method was used. This method involves shuffling the values of each individual variable of the training data, and evaluating the decrease in the AUC. A higher reduction in the AUC translates to a greater importance of a variable [\\u003cspan citationid=\\\"CR39\\\" class=\\\"CitationRef\\\"\\u003e39\\u003c/span\\u003e]. The response curves for each predictive variable were also generated for the final test dataset partition using NumPy and Matplotlib, in order to help interpret the model and analyze how variations in their values influence the dependent variable. We generated them by creating copies of the test data, then modifying one feature at a time, replacing its values with a range from min to max, and passing the modified data to \\u003cem\\u003emodel.predict()\\u003c/em\\u003e.\\u003c/p\\u003e \\u003cp\\u003eIn order to further comprehend the relationship between the chosen predictive variables and the fungus sporulation, we generated, through the best performing model, a .csv file containing the class probability predictions for the test dataset, using the predict function. In this instance, being a binary classification, this function returned a matrix where each row has both the probability score of being classified as \\u0026lsquo;low\\u0026rsquo; or \\u0026lsquo;high\\u0026rsquo; risk of sporulation. Those predictions were used to plot a histogram of the distribution of predicted probabilities for the \\u0026lsquo;low\\u0026rsquo; and \\u0026lsquo;high\\u0026rsquo; risk classes, using the \\u003cem\\u003eapply()\\u003c/em\\u003e function in Pandas, which compares row by row the scores in the \\\"low\\\" and \\\"high\\\" classes.\\u003c/p\\u003e \\u003c/div\\u003e\"},{\"header\":\"3. Results\",\"content\":\"\\u003cp\\u003eOut of the 20 models generated with mcfly, the one that demonstrated the best performance was a model with a CNN-type architecture, achieving a validation AUC of 0.83 (candidate model number 2, as referenced in Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e and Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e). After further training for two additional epochs with the full training data (AB), the model demonstrated a steady increase in AUC throughout the epochs, achieving an internal AUC of 0.93 and an AUC of 0.81 in the test data.\\u003c/p\\u003e \\u003cp\\u003e \\u003cdiv class=\\\"gridtable\\\"\\u003e\\u003ctable float=\\\"Yes\\\" id=\\\"Tab1\\\" border=\\\"1\\\"\\u003e \\u003ccaption language=\\\"En\\\"\\u003e \\u003cdiv class=\\\"CaptionNumber\\\"\\u003eTable 1\\u003c/div\\u003e \\u003cdiv class=\\\"CaptionContent\\\"\\u003e \\u003cp\\u003ePerformance of Candidate Models. The best performing model achieved a validation AUC of 0.8258, having a CNN architecture.\\u003c/p\\u003e \\u003c/div\\u003e \\u003c/caption\\u003e \\u003ccolgroup cols=\\\"6\\\"\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c1\\\" colnum=\\\"1\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"left\\\" class=\\\"colspec\\\" colname=\\\"c2\\\" colnum=\\\"2\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c3\\\" colnum=\\\"3\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c4\\\" colnum=\\\"4\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c5\\\" colnum=\\\"5\\\"\\u003e\\u003c/div\\u003e \\u003cdiv align=\\\"char\\\" char=\\\".\\\" class=\\\"colspec\\\" colname=\\\"c6\\\" colnum=\\\"6\\\"\\u003e\\u003c/div\\u003e \\u003cthead\\u003e \\u003ctr\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003eCandidate\\u003c/p\\u003e \\u003cp\\u003eModel\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eArchitecture\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003eTrain AUC\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003eTrain Loss\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003eVal AUC\\u003c/p\\u003e \\u003c/th\\u003e \\u003cth align=\\\"left\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003eVal Loss\\u003c/p\\u003e \\u003c/th\\u003e \\u003c/tr\\u003e \\u003c/thead\\u003e \\u003ctbody\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e0\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eInceptionTime\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.6819\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6587\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e33.0189\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e1\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eResNet\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5933\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6735\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e47.5007\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e2\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003eCNN\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e0.9276\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e3.3416\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e0.8258\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e\\u003cb\\u003e3.2641\\u003c/b\\u003e\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e3\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eDeepConvLSTM\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.8612\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e12.7140\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.7118\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e12.1372\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e4\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eInceptionTime\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.6290\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6704\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5602\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e1.6686\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e5\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eCNN\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.8192\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.9364\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.7375\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e1.0577\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e6\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eResNet\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5948\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6759\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.4355\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e22.9626\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e7\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eDeepConvLSTM\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.6666\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1.2840\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.4355\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e1.2616\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e8\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eDeepConvLSTM\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.8188\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2.2299\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5646\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e2.3523\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e9\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eResNet\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5543\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.7071\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.4363\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e13.4807\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e10\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eInceptionTime\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.6219\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6803\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5820\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e2.5696\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e11\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eCNN\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.7677\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2.2624\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.7218\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e2.1406\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e12\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eInceptionTime\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5941\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6993\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e25.0352\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e13\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eCNN\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.9613\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e1.4363\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.7136\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e1.7031\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e14\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eDeepConvLSTM\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.8487\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e14.9051\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5470\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e14.2553\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e15\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eResNet\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.7105\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6296\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5645\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e77.3314\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e16\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eResNet\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5984\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6820\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5543\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e5.4269\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e17\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eDeepConvLSTM\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.8417\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e12.8830\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.4839\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e13.0282\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e18\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eInceptionTime\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.5749\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e0.6857\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.5647\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e0.7139\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003ctr\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c1\\\"\\u003e \\u003cp\\u003e19\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"left\\\" colname=\\\"c2\\\"\\u003e \\u003cp\\u003eCNN\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c3\\\"\\u003e \\u003cp\\u003e0.7638\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c4\\\"\\u003e \\u003cp\\u003e2.3186\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c5\\\"\\u003e \\u003cp\\u003e0.6624\\u003c/p\\u003e \\u003c/td\\u003e \\u003ctd align=\\\"char\\\" char=\\\".\\\" colname=\\\"c6\\\"\\u003e \\u003cp\\u003e2.9733\\u003c/p\\u003e \\u003c/td\\u003e \\u003c/tr\\u003e \\u003c/tbody\\u003e \\u003c/colgroup\\u003e \\u003c/table\\u003e\\u003c/div\\u003e \\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eThe calculated permutation importance demonstrated that there was a higher decrease in the AUC when the mean daily temperature values were permuted, dropping by 0.24 (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig2\\\" class=\\\"InternalRef\\\"\\u003e2\\u003c/span\\u003e). The remaining variables caused much smaller performance drops, indicating minor contributions to the model. In second place in the hierarchy of importance comes elevation, with a loss of 0.02 followed by mean daily relative humidity with a AUC loss of just 0.01. The randomization of the remaining topographic variables (slope and aspect) showed negligeable losses in the AUC values.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eAccording to the response curves generated, in the case of the average daily temperature (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ea), the temperature range where the probability of sporulation is higher is between approximately 10\\u0026ordm;C to 20\\u0026ordm;C, while the peak probability occurs when the temperature is around 15\\u0026ordm;C. Extreme temperatures appear to be associated with lower sporulation capacity. In the case of the elevation, the second most important variable for the study, it is possible to observe that the probability peak is higher when the elevation values ​​are around 300 meters (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eb). Very high elevation​​ translates into a sharp decrease in the probability of species\\u0026rsquo; sporulation. Regarding relative humidity, the third most important variable for the study, the peak probability occurs at around 75% relative humidity (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ec). From this value onwards, the probability rapidly decreases.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e \\u003cp\\u003eOur model also generated a .csv file containing class probability predictions for the test dataset, estimating the probability of each sample belonging to both \\u0026lsquo;low\\u0026rsquo; and \\u0026lsquo;high\\u0026rsquo; risk classes. Among the 287 predictions, 234 were classified as low risk of sporulation, while the remaining 53 were classified as high risk. As observed in the histogram from Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e, the distribution for both classes is distinct yet symmetric, with \\u0026lsquo;low\\u0026rsquo; predictions concentrated near 1 and \\u0026lsquo;high\\u0026rsquo; predictions between 0.05 e 0.4. Additionally and especially between 0.4 and 0.6, there appears to be a zone of \\u0026lsquo;uncertainty\\u0026rsquo; where the classes appear in a more balanced way.\\u003c/p\\u003e \\u003cp\\u003e \\u003c/p\\u003e\"},{\"header\":\"4. Discussion\",\"content\":\"\\u003cp\\u003eUsing the Python programming language and the mcfly software library, we successfully implemented deep learning techniques to estimate the sporulation of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e on Terceira Island, Azores (Portugal). This effort addresses a critical need, as the fungus produces a toxin that, when ingested by grazing livestock, leads to facial eczema\\u0026mdash;a condition with severe implications for animal health and significant economic losses [e.g., 3, 9].\\u003c/p\\u003e \\u003cp\\u003eFor this prediction task, we selected the candidate model with the highest validation AUC, which featured a convolutional neural network (CNN) architecture. This outcome is unsurprising, as CNNs are widely used in ecological modeling due to their strong performance in tasks such as species distribution prediction [\\u003cspan citationid=\\\"CR21\\\" class=\\\"CitationRef\\\"\\u003e21\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR22\\\" class=\\\"CitationRef\\\"\\u003e22\\u003c/span\\u003e]. Notably, models employing a DeepConvLSTM architecture also showed promising results\\u0026mdash;particularly candidate model 3 (see Table\\u0026nbsp;\\u003cspan refid=\\\"Tab1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e and Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig1\\\" class=\\\"InternalRef\\\"\\u003e1\\u003c/span\\u003e). The selected CNN model was subsequently trained for two additional epochs, achieving an average AUC of 0.93 and a validation AUC of 0.81. These results reflect strong predictive performance, indicating the model\\u0026rsquo;s capacity to effectively distinguish between high and low sporulation risk days.\\u003c/p\\u003e \\u003cp\\u003eOur model generated 53 predictions identified as \\u0026lsquo;high\\u0026rsquo; risk of sporulation. If used operationally, these would correspond to the areas on which preventive measures should be considered (e.g., avoid grazing and use preserved feed instead). When plotting the distribution of predicted probabilities for 'low' and 'high' classes (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig4\\\" class=\\\"InternalRef\\\"\\u003e4\\u003c/span\\u003e), the strong spikes around 0.1 and 0.9 suggest that the model is confident and has a clear ability to distinguish between \\u0026lsquo;low\\u0026rsquo; and \\u0026lsquo;high\\u0026rsquo; risk in many of its predictions. Being a binary classification, in most cases, when the model predicts a high probability for one of the classes, inevitably assigns a low probability score to the opposite class.\\u003c/p\\u003e \\u003cp\\u003eBetween the two meteorological variables used, our results show that there is a big emphasis on the mean daily temperature. This was the only variable that, when its values ​​were randomized, produced a significant drop in the model's performance and in its AUC score. The generated response curves help understand how the variation of predictive variables affects fungal sporulation. In the case of temperature, our results suggest that the temperature range where the probability of sporulation is higher is between approximately 10\\u0026ordm;C to 20\\u0026ordm;C, which is mostly within the temperature range normally identified in the literature as being most favorable for \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation (i.e. 12\\u0026ordm;C to 24\\u0026ordm;C; [\\u003cspan citationid=\\\"CR4\\\" class=\\\"CitationRef\\\"\\u003e4\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR12\\\" class=\\\"CitationRef\\\"\\u003e12\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR15\\\" class=\\\"CitationRef\\\"\\u003e15\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e]). However, even though 24\\u0026ordm;C is usually identified as the optimal temperature for the fungus sporulation, we identified a different peak temperature, 15\\u0026ordm;C. Despite that, in 2005, Pinto et al. analyzed pithomycotoxicosis in ruminants in the Azores, and with meteorological data from 1995 to 2001 stated that favourable conditions to fungal growth and sporulation were observed whenever there was more than 10 days with a minimum temperature of 16\\u0026ordm;C [\\u003cspan citationid=\\\"CR8\\\" class=\\\"CitationRef\\\"\\u003e8\\u003c/span\\u003e]. According to Heras et al. [\\u003cspan citationid=\\\"CR7\\\" class=\\\"CitationRef\\\"\\u003e7\\u003c/span\\u003e], similar conditions were observed in Asturias, demonstrating the possible influence of oceanic climate on the fungus sporulation.\\u003c/p\\u003e \\u003cp\\u003e\\u0026Aacute;vila et al. [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e] identified average temperature over the 90 days preceding spore counts as a key variable, showing a strong positive correlation with spore abundance\\u0026mdash;that is, higher average temperatures were associated with greater spore quantities. This finding underscores the importance of capturing seasonal patterns in sporulation forecasting. Our results extend this insight by suggesting that the influence of mean temperature may not be limited to the preceding 90 days, but may instead span up to 360 days. This highlights the possibility that \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation is shaped by the broader environmental history of a given location, rather than by short-term conditions alone.\\u003c/p\\u003e \\u003cp\\u003eRegarding elevation, this was the most important topographic variable for the model. According to the generated response curve (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003eb), the probability peak is higher when the elevation values ​​are around 300 meters. This situation is concerning, since the eastern part of Terceira Island is dominated by a huge expanse of pastureland, with an average altitude of around 390 meters, and an even more extensive plain to the north, with an average altitude of 200 meters [\\u003cspan citationid=\\\"CR40\\\" class=\\\"CitationRef\\\"\\u003e40\\u003c/span\\u003e]. Therefore, it is possible to confirm that there are several areas that may be within these sections with an elevation more susceptible to the fungus sporulation, and that represent pasture areas. These pasture areas also favour the availability of decaying plant material, which is pre-established as one of the ideal conditions for the fungus sporulation [\\u003cspan citationid=\\\"CR16\\\" class=\\\"CitationRef\\\"\\u003e16\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR19\\\" class=\\\"CitationRef\\\"\\u003e19\\u003c/span\\u003e]. This peak in sporulation observed at elevations of 300 metres may be explained by the decrease in temperature associated with higher altitudes, since as previously noted, extreme temperatures (both too low or too high) are associated with a strong decrease in the fungus\\u0026rsquo; risk of sporulation.\\u003c/p\\u003e \\u003cp\\u003eRegarding relative humidity, this variable occupies the third place in the permutation importance hierarchy (Fig.\\u0026nbsp;\\u003cspan refid=\\\"Fig3\\\" class=\\\"InternalRef\\\"\\u003e3\\u003c/span\\u003ec). Humidity is also one of the most frequently mentioned aspects in the literature when characterizing the fungus's suitability conditions, despite our results indicating a minor contribution to the model. It is usually a consensus that moist conditions, with relative humidity above 90% is more suitable for sporulation and fungal growth [\\u003cspan citationid=\\\"CR13\\\" class=\\\"CitationRef\\\"\\u003e13\\u003c/span\\u003e, \\u003cspan citationid=\\\"CR18\\\" class=\\\"CitationRef\\\"\\u003e18\\u003c/span\\u003e], however our results identified the range of 70\\u0026ndash;75% as the one with the greatest probability of risk. Unicol, an agricultural cooperative based on Terceira Island and with specialists responsible for the measurements of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e spores in grass samples used in this study, identifies the following necessary conditions for the formation of spores: temperature higher than 16\\u0026ordm;C and humidity above 76% for 72 consecutive hours [\\u003cspan citationid=\\\"CR31\\\" class=\\\"CitationRef\\\"\\u003e31\\u003c/span\\u003e]. Nonetheless, despite the slight differences in identifying a range of relative humidity that translates into a higher risk of this species\\u0026rsquo; sporulation, it is proven that the fungus requires a moist environment in order for sporulation to occur and our results are aligned with this trend. Additionally, \\u0026Aacute;vila et al. [\\u003cspan citationid=\\\"CR3\\\" class=\\\"CitationRef\\\"\\u003e3\\u003c/span\\u003e] also reached a similar conclusion, stating that relative humidity seems to have a slight effect on the fungus sporulation and that this situation could be explained by the fact that Terceira island experiences a consistently high relative humidity.\\u003c/p\\u003e \\u003cp\\u003eOverall, deep learning proved to be a powerful tool to analyze complex relationships between the chosen environmental variables and the sporulation data, further strengthening its position as a transformative tool regarding ecological studies and accuracy in prediction exercises. Our model ultimately allows for a better understanding of the ecological processes associated with \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation, while enhancing the ability to forecast high-risk sporulation events.\\u003c/p\\u003e\"},{\"header\":\"5. Conclusions\",\"content\":\"\\u003cp\\u003eIn this study, we successfully implemented a convolutional neural network (CNN) to forecast the sporulation of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e on Terceira Island, in the Azores archipelago (Portugal). The results demonstrate the potential of deep learning approaches to address ecological challenges, particularly in the context of fungal pathogens with significant animal health and economic impacts. \\u003cem\\u003eP. chartarum\\u003c/em\\u003e produces a toxin that causes pithomycotoxicosis in grazing livestock, posing a serious threat to animal welfare and leading to substantial financial losses.\\u003c/p\\u003e \\u003cp\\u003eAmong the predictive variables tested, mean daily temperature emerged as the most influential driver of sporulation risk, in line with known ecological conditions favoring fungal growth. Our findings further suggest that temperature effects are not limited to short-term windows, but may extend across a full year, highlighting the importance of incorporating long-term environmental history into predictive models. The relatively limited contribution of other variables indicates that additional environmental or biological factors\\u0026mdash;such as precipitation, soil properties, or pasture type\\u0026mdash;may play a role and warrant inclusion in future modelling efforts to improve predictive accuracy.\\u003c/p\\u003e \\u003cp\\u003eImportantly, the current deep learning-based model is not intended as a standalone solution, but as a component to be integrated into an existing early warning system already in use. By improving the reliability and accuracy of sporulation forecasts, this model enhances the system\\u0026rsquo;s ability to provide timely alerts, enabling specialists to refine disease control strategies and supporting farmers in making informed management decisions. Furthermore, the framework developed here has the potential for adaptation to other regions affected by \\u003cem\\u003eP. chartarum\\u003c/em\\u003e, highlighting its broader applicability in mitigating the impacts of pithomycotoxicosis.\\u003c/p\\u003e\"},{\"header\":\"Declarations\",\"content\":\"\\u003cp\\u003e \\u003ch2\\u003eAdditional Information\\u003c/h2\\u003e \\u003cp\\u003eThe authors declare no competing interests.\\u003c/p\\u003e \\u003c/p\\u003e\\u003ch2\\u003eFunding\\u003c/h2\\u003e \\u003cp\\u003eThis research was funded by the mobilizing agenda New Space Portugal, as part of Portugal's Recovery and Resilience Plan (RRP) - Project n\\u0026ordm; 02/C05-i01.01/2022.PC644936537-00000046; IAPMEI Projeto N\\u0026ordm;11.\\u003c/p\\u003e\\u003ch2\\u003eAuthor Contribution\\u003c/h2\\u003e\\u003cp\\u003eI.D: Writing \\u0026ndash; original draft, Visualization, Code, Methodology, Investigation, Formal analysis, Data curation. C.C: Writing \\u0026ndash; review \\u0026amp; editing, Validation, Code, Methodology, Formal analysis. J.P.: Writing \\u0026ndash; review \\u0026amp; editing, Validation, Supervision, Project administration, Funding acquisition, Methodology, Formal analysis, Conceptualization, Resources. E.D: Methodology, Writing \\u0026ndash; review \\u0026amp; editing. M.A: Methodology, Writing \\u0026ndash; review \\u0026amp; editing. All authors have read and approved the final manuscript.​\\u003c/p\\u003e\\u003ch2\\u003eAcknowledgement\\u003c/h2\\u003e\\u003cp\\u003eWe would like to thank TERINOV for providing access to essential data from the IoT weather stations, and the Municipality of Angra do Hero\\u0026iacute;smo (Terceira) for their support in the initial development of the LoRaWAN Network. We would also like to acknowledge the Regional Veterinary Laboratory staff for their timely and consistent spore counts throughout the project, and UNICOL.\\u003c/p\\u003e\\u003ch2\\u003eData Availability\\u003c/h2\\u003e\\u003cp\\u003eThe data presented in this study is available upon request from the corresponding author. The dataset of the published report will be made publicly available in a repository.\\u003c/p\\u003e\"},{\"header\":\"References\",\"content\":\"\\u003col\\u003e\\n\\u003cli\\u003eDingley, J. M. \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e, its occurrence morphology, and taxonomy. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e5\\u003c/strong\\u003e, 49\\u0026ndash;61 (1962).\\u003c/li\\u003e\\n\\u003cli\\u003eHuang, M. \\u0026amp; Hull, C. M. Sporulation: how to survive on planet Earth (and beyond). \\u003cem\\u003eCurr. Genet.\\u003c/em\\u003e \\u003cstrong\\u003e63\\u003c/strong\\u003e, 831\\u0026ndash;838 (2017).\\u003c/li\\u003e\\n\\u003cli\\u003e\\u0026Aacute;vila, M. et al. Assessing the presence of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e in pastureland using IoT sensors and remote sensing: the case study of Terceira Island (Azores, Portugal). \\u003cem\\u003eSensors\\u003c/em\\u003e \\u003cstrong\\u003e24\\u003c/strong\\u003e, 4485 (2024).\\u003c/li\\u003e\\n\\u003cli\\u003eBrook, P. J. Ecology of the fungus \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e (Berk. \\u0026amp; Curt.) M.B. Ellis in pasture in relation to facial eczema disease of sheep. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e6\\u003c/strong\\u003e, 147\\u0026ndash;228 (1963).\\u003c/li\\u003e\\n\\u003cli\\u003eWearn, J. \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e \\u0026ndash; a fungus on the up? \\u003cem\\u003eField Mycol.\\u003c/em\\u003e \\u003cstrong\\u003e10\\u003c/strong\\u003e, 36\\u0026ndash;37 (2009).\\u003c/li\\u003e\\n\\u003cli\\u003eCuttance, E. L., Laven, R. A. \\u0026amp; Stevenson, M. A. Variability in measurement of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e spore counts. \\u003cem\\u003eN. Z. Vet. J.\\u003c/em\\u003e \\u003cstrong\\u003e65\\u003c/strong\\u003e, 192\\u0026ndash;197 (2017).\\u003c/li\\u003e\\n\\u003cli\\u003eDe las Heras, M. et al. Chronic pithomycotoxicosis associated with obstructive rhinopathy in sheep. \\u003cem\\u003eVet. Pathol.\\u003c/em\\u003e \\u003cstrong\\u003e59\\u003c/strong\\u003e, 950\\u0026ndash;959 (2022).\\u003c/li\\u003e\\n\\u003cli\\u003ePinto, C. et al. Pithomycotoxicosis (facial eczema) in ruminants in the Azores, Portugal. \\u003cem\\u003eVet. Rec.\\u003c/em\\u003e \\u003cstrong\\u003e157\\u003c/strong\\u003e, 805\\u0026ndash;810 (2005).\\u003c/li\\u003e\\n\\u003cli\\u003eMarbrook, J. \\u0026amp; Matthews, R. E. F. Loss of sporidesmin from spores of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e (Berk. \\u0026amp; Curt.) M.B. Ellis. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e5\\u003c/strong\\u003e, 223\\u0026ndash;236 (1962).\\u003c/li\\u003e\\n\\u003cli\\u003eDi Menna, M. E., Smith, B. L. \\u0026amp; Miles, C. O. A history of facial eczema (pithomycotoxicosis) research. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e52\\u003c/strong\\u003e, 345\\u0026ndash;376 (2009).\\u003c/li\\u003e\\n\\u003cli\\u003eMorris, C. A. et al. Inheritance of resistance to facial eczema: a review of research findings from sheep and cattle in New Zealand. \\u003cem\\u003eN. Z. Vet. J.\\u003c/em\\u003e \\u003cstrong\\u003e52\\u003c/strong\\u003e, 205\\u0026ndash;215 (2004).\\u003c/li\\u003e\\n\\u003cli\\u003eFern\\u0026aacute;ndez, M. et al. Pathological study of facial eczema (pithomycotoxicosis) in sheep. \\u003cem\\u003eAnimals\\u003c/em\\u003e \\u003cstrong\\u003e11\\u003c/strong\\u003e, 1070 (2021).\\u003c/li\\u003e\\n\\u003cli\\u003eDennis, N. A., Amer, P. R. \\u0026amp; Meier, S. Brief communication: predicting the impact of climate change on the risk of facial eczema outbreaks throughout New Zealand. \\u003cem\\u003eProc. N. Z. Soc. Anim. Prod.\\u003c/em\\u003e \\u003cstrong\\u003e74\\u003c/strong\\u003e, 1\\u0026ndash;6 (2014).\\u003c/li\\u003e\\n\\u003cli\\u003eLiu, L. et al. Fungi, including \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e, cause facial eczema and inflammation in grazing sheep in Western China. \\u003cem\\u003eMicrob. Pathog.\\u003c/em\\u003e \\u003cstrong\\u003e185\\u003c/strong\\u003e, 106451 (2023).\\u003c/li\\u003e\\n\\u003cli\\u003eSmith, J. D., Lees, F. T. \\u0026amp; Crawley, W. E. Weather conditions, spore counts, and facial eczema in test sheep. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e8\\u003c/strong\\u003e, 63\\u0026ndash;87 (1965).\\u003c/li\\u003e\\n\\u003cli\\u003eSch\\u0026ouml;niger, S. et al. \\u003cem\\u003ePrototheca\\u003c/em\\u003e species and \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e as causative agents of rhinitis and/or sinusitis in horses. \\u003cem\\u003eJ. Comp. Pathol.\\u003c/em\\u003e \\u003cstrong\\u003e155\\u003c/strong\\u003e, 121\\u0026ndash;125 (2016).\\u003c/li\\u003e\\n\\u003cli\\u003eVan Wuijckhuise, L. et al. First case of pithomycotoxicosis (facial eczema) in the Netherlands. \\u003cem\\u003eTijdschr. Diergeneeskd.\\u003c/em\\u003e \\u003cstrong\\u003e131\\u003c/strong\\u003e, 858\\u0026ndash;861 (2006).\\u003c/li\\u003e\\n\\u003cli\\u003eDijkstra, E., Harkema, L. \\u0026amp; Vellema, P. First case of pithomycotoxicosis in sheep in the Netherlands. \\u003cem\\u003eVet. Rec. Case Rep.\\u003c/em\\u003e \\u003cstrong\\u003e10\\u003c/strong\\u003e, e268 (2022).\\u003c/li\\u003e\\n\\u003cli\\u003eMitchell, K. J., Thomas, R. G. \\u0026amp; Clarke, R. T. J. Factors influencing the growth of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e in pasture. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e4\\u003c/strong\\u003e, 566\\u0026ndash;577 (1961).\\u003c/li\\u003e\\n\\u003cli\\u003eDi Menna, M. E. \\u0026amp; Bailey, J. R. \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e spore counts in pasture. \\u003cem\\u003eN. Z. J. Agric. Res.\\u003c/em\\u003e \\u003cstrong\\u003e16\\u003c/strong\\u003e, 343\\u0026ndash;351 (1973).\\u003c/li\\u003e\\n\\u003cli\\u003eChristin, S., Hervet, \\u0026Eacute;. \\u0026amp; Lecomte, N. Applications for deep learning in ecology. \\u003cem\\u003eMethods Ecol. Evol.\\u003c/em\\u003e \\u003cstrong\\u003e10\\u003c/strong\\u003e, 1632\\u0026ndash;1644 (2019).\\u003c/li\\u003e\\n\\u003cli\\u003eBorowiec, M. L. et al. Deep learning as a tool for ecology and evolution. \\u003cem\\u003eMethods Ecol. Evol.\\u003c/em\\u003e \\u003cstrong\\u003e13\\u003c/strong\\u003e, 1640\\u0026ndash;1660 (2022).\\u003c/li\\u003e\\n\\u003cli\\u003eCapinha, C., Ceia-Hasse, A., Kramer, A. M. \\u0026amp; Meijer, C. Deep learning for supervised classification of temporal data in ecology. \\u003cem\\u003eEcol. Inform.\\u003c/em\\u003e \\u003cstrong\\u003e61\\u003c/strong\\u003e, 101252 (2021).\\u003c/li\\u003e\\n\\u003cli\\u003eCeia-Hasse, A., Sousa, C. A., Gouveia, B. R. \\u0026amp; Capinha, C. Forecasting the abundance of disease vectors with deep learning. \\u003cem\\u003eEcol. Inform.\\u003c/em\\u003e \\u003cstrong\\u003e78\\u003c/strong\\u003e, 102272 (2023).\\u003c/li\\u003e\\n\\u003cli\\u003eLhoumeau, S., Pinelo, J. \\u0026amp; Borges, P. A. Artificial intelligence for biodiversity: exploring the potential of recurrent neural networks in forecasting arthropod dynamics based on time series. \\u003cem\\u003eEcol. Indic.\\u003c/em\\u003e \\u003cstrong\\u003e171\\u003c/strong\\u003e, 113119 (2025).\\u003c/li\\u003e\\n\\u003cli\\u003eRammer, W. \\u0026amp; Seidl, R. Harnessing deep learning in ecology: an example predicting bark beetle outbreaks. \\u003cem\\u003eFront. Plant Sci.\\u003c/em\\u003e \\u003cstrong\\u003e10\\u003c/strong\\u003e, 451705 (2019).\\u003c/li\\u003e\\n\\u003cli\\u003eRaposeiro, P. M. et al. Climate change facilitated the early colonization of the Azores Archipelago during medieval times. \\u003cem\\u003eProc. Natl. Acad. Sci. USA\\u003c/em\\u003e \\u003cstrong\\u003e118\\u003c/strong\\u003e, e2108236118 (2021).\\u003c/li\\u003e\\n\\u003cli\\u003ePortal do Ordenamento do Territ\\u0026oacute;rio dos A\\u0026ccedil;ores. Caracteriza\\u0026ccedil;\\u0026atilde;o e identifica\\u0026ccedil;\\u0026atilde;o das paisagens dos A\\u0026ccedil;ores | Terceira. https://ot.azores.gov.pt/Unidades-Paisagem-Ficha.aspx?id=78 (2025).\\u003c/li\\u003e\\n\\u003cli\\u003eINE. \\u003cem\\u003eRecenseamento Agr\\u0026iacute;cola \\u0026ndash; An\\u0026aacute;lise dos principais resultados \\u0026ndash; 2019.\\u003c/em\\u003e Inst. Nac. Estat\\u0026iacute;stica, Lisboa (2021).\\u003c/li\\u003e\\n\\u003cli\\u003eAlmeida, A. M. de, Alvarenga, P. \\u0026amp; Fangueiro, D. The dairy sector in the Azores Islands: possibilities and main constraints towards increased added value. \\u003cem\\u003eTrop. Anim. Health Prod.\\u003c/em\\u003e \\u003cstrong\\u003e53\\u003c/strong\\u003e, 1\\u0026ndash;9 (2021).\\u003c/li\\u003e\\n\\u003cli\\u003eUnicol. Sistema de Alertas Unicol \\u0026ndash; SAU. https://www.unicol.pt/sistema-de-alertas-unicol-sau/ (2025).\\u003c/li\\u003e\\n\\u003cli\\u003eIPMA. \\u003cem\\u003eBoletim climatol\\u0026oacute;gico mensal \\u0026ndash; Julho de 2015.\\u003c/em\\u003e Minist\\u0026eacute;rio da Agricultura e do Mar. http://www.ipma.pt (2015).\\u003c/li\\u003e\\n\\u003cli\\u003eCapinha, C. Predicting the timing of ecological phenomena using dates of species occurrence records: a methodological approach and test case with mushrooms. \\u003cem\\u003eInt. J. Biometeorol.\\u003c/em\\u003e \\u003cstrong\\u003e63\\u003c/strong\\u003e, 1015\\u0026ndash;1024 (2019).\\u003c/li\\u003e\\n\\u003cli\\u003eVan Kuppevelt, D. et al. Mcfly: automated deep learning on time series. \\u003cem\\u003eSoftwareX\\u003c/em\\u003e \\u003cstrong\\u003e12\\u003c/strong\\u003e, 100548 (2020).\\u003c/li\\u003e\\n\\u003cli\\u003eZhao, B., Lu, H., Chen, S., Liu, J. \\u0026amp; Wu, D. Convolutional neural networks for time series classification. \\u003cem\\u003eJ. Syst. Eng. Electron.\\u003c/em\\u003e \\u003cstrong\\u003e28\\u003c/strong\\u003e, 162\\u0026ndash;169 (2017).\\u003c/li\\u003e\\n\\u003cli\\u003eFawaz, H. I. et al. Deep learning for time series classification: a review. \\u003cem\\u003eData Min. Knowl. Discov.\\u003c/em\\u003e \\u003cstrong\\u003e33\\u003c/strong\\u003e, 917\\u0026ndash;963 (2019).\\u003c/li\\u003e\\n\\u003cli\\u003eFawaz, H. I. et al. InceptionTime: finding AlexNet for time series classification. \\u003cem\\u003eData Min. Knowl. Discov.\\u003c/em\\u003e \\u003cstrong\\u003e34\\u003c/strong\\u003e, 1936\\u0026ndash;1962 (2020).\\u003c/li\\u003e\\n\\u003cli\\u003ePhillips, S. J., Dud\\u0026iacute;k, M. \\u0026amp; Schapire, R. E. A maximum entropy approach to species distribution modeling. In \\u003cem\\u003eProc. 21st Int. Conf. Mach. Learn.\\u003c/em\\u003e 83\\u0026ndash;90 (ACM, 2004).\\u003c/li\\u003e\\n\\u003cli\\u003eDiogo, I., Sillero, N. \\u0026amp; Capinha, C. Predicting the risk of invasion by broadleaf watermilfoil (\\u003cem\\u003eMyriophyllum heterophyllum\\u003c/em\\u003e) in mainland Portugal. \\u003cem\\u003eHeliyon\\u003c/em\\u003e \\u003cstrong\\u003e10\\u003c/strong\\u003e, e16744 (2024).\\u003c/li\\u003e\\n\\u003cli\\u003eSRAM. \\u003cem\\u003eLivro das Paisagens dos A\\u0026ccedil;ores: Contributos para a Identifica\\u0026ccedil;\\u0026atilde;o e Caracteriza\\u0026ccedil;\\u0026atilde;o das Paisagens dos A\\u0026ccedil;ores.\\u003c/em\\u003e https://ot.azores.gov.pt/Biblioteca-Publicacoes.aspx?id=84 (2005).\\u003c/li\\u003e\\n\\u003c/ol\\u003e\"}],\"fulltextSource\":\"\",\"fullText\":\"\",\"funders\":[],\"hasAdminPriorityOnWorkflow\":false,\"hasManuscriptDocX\":true,\"hasOptedInToPreprint\":true,\"hasPassedJournalQc\":\"\",\"hasAnyPriority\":false,\"hideJournal\":true,\"highlight\":\"\",\"institution\":\"\",\"isAcceptedByJournal\":false,\"isAuthorSuppliedPdf\":false,\"isDeskRejected\":\"\",\"isHiddenFromSearch\":false,\"isInQc\":false,\"isInWorkflow\":false,\"isPdf\":false,\"isPdfUpToDate\":true,\"isWithdrawnOrRetracted\":false,\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true},\"keywords\":\"Deep learning, Time series prediction, Pithomyces chartarum, Pithomycotoxicosis, Ecological prediction,; Alert system\",\"lastPublishedDoi\":\"10.21203/rs.3.rs-6453956/v1\",\"lastPublishedDoiUrl\":\"https://doi.org/10.21203/rs.3.rs-6453956/v1\",\"license\":{\"name\":\"CC BY 4.0\",\"url\":\"https://creativecommons.org/licenses/by/4.0/\"},\"manuscriptAbstract\":\"\\u003cp\\u003ePithomycotoxicosis is a disease affecting grazing livestock, caused by ingestion of \\u003cem\\u003ePithomyces chartarum\\u003c/em\\u003e spores. These spores have been identified in various regions worldwide, including the Azores Archipelago (Portugal) since 1999. The severity of the disease is strongly linked to spore concentration, while spore abundance is known to depend on meteorological conditions. In this study, we develop and evaluate a deep learning-based framework to forecast the sporulation of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e on Terceira Island (Azores), using historical spore count data together with meteorological and topographic variables. Among 20 neural network architectures tested, a convolutional neural network (CNN) achieved the best performance in classifying high-risk conditions, with an area under the curve (AUC) of 0.81 on the validation set. Feature importance analysis identified mean daily temperature as the most influential variable for sporulation risk, consistent with known favorable conditions for fungal growth. Additionally, the results reveal a marked seasonal pattern in sporulation risk, shaped by short- to mid-term antecedent meteorological conditions. Our findings demonstrate that deep learning models can enhance predictive accuracy and deepen understanding of the environmental drivers of \\u003cem\\u003eP. chartarum\\u003c/em\\u003e sporulation, thereby improving the performance of existing alert systems for livestock protection.\\u003c/p\\u003e\",\"manuscriptTitle\":\"A Deep Learning Approach to Predicting Pithomyces chartarum Sporulation for Livestock Protection\",\"msid\":\"\",\"msnumber\":\"\",\"nonDraftVersions\":[{\"code\":1,\"date\":\"2025-05-13 10:03:30\",\"doi\":\"10.21203/rs.3.rs-6453956/v1\",\"editorialEvents\":[{\"type\":\"communityComments\",\"content\":0}],\"status\":\"published\",\"journal\":{\"display\":true,\"email\":\"info@researchsquare.com\",\"identity\":\"researchsquare\",\"isNatureJournal\":false,\"hasQc\":true,\"allowDirectSubmit\":true,\"externalIdentity\":\"\",\"sideBox\":\"\",\"snPcode\":\"\",\"submissionUrl\":\"/submission\",\"title\":\"Research Square\",\"twitterHandle\":\"researchsquare\",\"acdcEnabled\":true,\"dfaEnabled\":false,\"editorialSystem\":\"\",\"reportingPortfolio\":\"\",\"inReviewEnabled\":false,\"inReviewRevisionsEnabled\":true}}],\"origin\":\"\",\"ownerIdentity\":\"4d7b86de-5726-4dcf-a1bf-2d67ec10ad12\",\"owner\":[],\"postedDate\":\"May 13th, 2025\",\"published\":true,\"recentEditorialEvents\":[],\"rejectedJournal\":[],\"revision\":\"\",\"amendment\":\"\",\"status\":\"posted\",\"subjectAreas\":[{\"id\":48325905,\"name\":\"Earth and environmental sciences/Ecology/Biogeography\"},{\"id\":48325906,\"name\":\"Biological sciences/Computational biology and bioinformatics/Machine learning\"}],\"tags\":[],\"updatedAt\":\"2025-07-01T07:09:03+00:00\",\"versionOfRecord\":[],\"versionCreatedAt\":\"2025-05-13 10:03:30\",\"video\":\"\",\"vorDoi\":\"\",\"vorDoiUrl\":\"\",\"workflowStages\":[]},\"version\":\"v1\",\"identity\":\"rs-6453956\",\"journalConfig\":\"researchsquare\"},\"__N_SSP\":true},\"page\":\"/article/[identity]/[[...version]]\",\"query\":{\"redirect\":\"/article/rs-6453956\",\"identity\":\"rs-6453956\",\"version\":[\"v1\"]},\"buildId\":\"8U1c8b4HqxoKbykW_rLl7\",\"isFallback\":false,\"isExperimentalCompile\":false,\"dynamicIds\":[84888],\"gssp\":true,\"scriptLoader\":[]}","source_license":"CC-BY-4.0","license_restricted":false}