Deep learning for predicting plant pathogen spread in a changing climate

Abstract

Addressing hunger and ensuring food security stands as a paramount goal within the United Nations' sustainable development goals. Food security, threatened by factors like climate change and resource limitations, demands innovative solutions like disease prediction models to safeguard crop yields. A significant threat to food security is the impact of crop diseases. Aligned with this goal, our study introduces a cutting-edge deep learning model designed to predict the spread of the Late Blight potato disease (Phytophthora infestans). This prediction is grounded in the assumption that diverse environmental data, gathered through sensors at various locations across a farm, exhibits variations. The study comprises three distinct models: a weather variation model, a synthetic plant growth model, and a disease spread simulation model governed by rules that capture the intricate relationship between environmental attributes and the disease. These models form the foundational input for three advanced deep learning models, each developed by harnessing the power of Neural Networks (NN) and Convolutional Neural Networks (CNN) respectively. The three models undertake the prediction of various aspects: Model 1 focuses on intra-plant transmission, while Model 2 is dedicated to predicting the inter-plant transmission within a one-block radius. Additionally, Model 3 engages in predicting multi-class neighbour identification, considering the specific plant affected. The first model (NN Model 1) excels in predicting intra-plant transmission, demonstrating high precision (88.664%) and accuracy (97.342%), albeit with a relatively lower recall (10.868%). The second model (NN Model 2) focuses on plant-to-plant infections, achieving a balanced performance with precision (57.884%) and recall (64.242%). The third model (NN Model 3) excels in multi-class tasks, boasting an exceptional AUC-ROC of 99.767, along with consistently high precision, recall, F1 score, and accuracy (84.475%). Transitioning to CNN models, the first model (CNN Model 1) exhibits commendable precision (77.404%) but faces challenges in recall (7.990%). The second model (CNN Model 2) strikes a balance, delivering high AUC-ROC (96.695), accuracy (96.933%), precision (57.700%), and recall (64.023%) for predicting infection status. The third model (CNN Model 3) excels in multi-class classification, achieving an AUC-ROC of 98.745, albeit with a slight dip in accuracy (83.375%). Overall, CNN models showcase superior AUC-ROC values, indicating heightened performance in binary classification tasks. Recommendations stemming from our findings include the utilization of NN for Model 1 due to its higher precision and F1 score. For infection status prediction, CNN Model 2 is recommended, boasting an impressive accuracy of 96.933%. Model 3 stands out as exceptional in both NN and CNN categories for multi-class tasks, registering accuracies of 84.475% and 83.375%, respectively. The findings in this study highlighted the significant influence of fluctuating weather conditions on the accuracy of the models.