Interpretable Deep Learning Model for Crop Yield Prediction: A Case Study of Wheat Yield Prediction in Egypt.

Abstract Accurately predicting crop yield can be challenging due to the environmental, biological and technological factors that directly influence crops and farms. However, proper estimation of crop yield is of great importance to food security and crop harvest management. Recent studies have shown that machine learning (ML) or deep learning (DL) techniques can be used effectively for crop yield prediction due to their ability to capture patterns and make accurate predictions in time series data. In this study, we present an attention-based long short-term memory (LSTM) and extreme gradient boosting (XGB) hybrid regressor model for crop yield prediction. The LSTM block is trained to capture the temporal dependencies and also learn features from the time series input data, while the XGB is used to make predictions based on the trained extracted LSTM features. The dataset used in this research comprises average yield, climatic variables, soil and moderate-resolution imaging spectroradiometer (MODIS) data. Also, to deal with the black box nature associated with ML and DL models, we employ the explainable artificial intelligence (XAI) tool SHAP to interpret how our proposed LSTM-XGB method made predictions. Furthermore, we compared the performance of our proposed method to three other state-of-the-art (SOTA) models; LSTM, light gradient boosting machine regressor (LGBMR) and deep neural network (DNN). Results from our experiment show the superior performance of our proposed hybrid LSTM-XGB in comparison to other methods.