Explainable AI Based Wheat Crop Monitoring Using Vegetation Indices, Machine Learning and LSTM-Based Temporal Forecasting

Abstract

Accurate crop health monitoring is critical to increase the productivity in agriculture and for precision farming. This research introduces a machine learning and deep learning-based model explainable system for monitoring wheat crops. Twenty-five Sentinel-2A and Sentinel-2B images were acquired over three wheat fields, 8-acre, 2.5-acre and 2-acre in extent. Vegetation indices were extracted after processing using AOI clipping and cloud masking, that is NDVI, NDRE, GNDVI, EVI, MCARI and CIgreen. In the case where no labels exist, unsupervised K-Means clustering methods were used to create the crop health classes. The classes were then generated and classified using Random Forest, Support Vector Machine (SVM), K-Nearest Neighbor (KNN) and XGBoost Models. Fold cross validation method was used and the results of five-fold cross validation showed that SVM gave the best classification accuracy of 92.0%, and Random Forest and XGBoost gave 90.67%. SHAP analysis determined that the most influential features are the NDVI, NDRE, MCARI, and Chlorophyll. Moreover, the 100 epochs-trained LSTM model yielded an MAE of 0.1797 in its role of NDVI prediction. It was seen that the proposed combination of clustering, explainable machine learning, and temporal deep learning shows its effectiveness to assess health and predict performance of wheat crop.