Satellite Imagery and Deep Learning Combined for Wheat Yield Forecasting

In the context of effective resource management and ensuring nutritional stability, precise forecasting of crop yields becomes essential. The development of artificial intelligence methodologies, coupled with satellite imagery, has emerged as a powerful strategy for predicting crop yields in modern times. In This study, deep learning algorithms based on LSTM (Long Short-Term Memory) were developed to efficiently optimize and extract from Sentinel-2 data spatiotemporal information of wheat yield. To estimate accurately wheat yield in Morocco, several machine learning and deep learning techniques such as Random Forest, LSTM, Bi-LSTM (Bidirectional LSTM), stacked LSTM, etc. were used and compared. The optimized Bi-LSTM model accurately estimates wheat yield based on NDVI (normalized difference vegetation index) data and weather data (temperature, precipitation). Three datasets gathered from satellite imagery were used which are temperature data, precipitation data and NDVI data combined for training and testing the proposed model. After data processing, different machine learning and deep learning algorithms were compared, and the result showed that Bi-LSTM estimates wheat yield accurately. The proposed and optimized Bi-LSTM model reached a satisfactory accuracy at the sizable regional scale. The obtained result demonstrates that the RMSE (Root Mean Square Error) score was 6.22 and the loss was 6.61‧10–4 after 20 epochs of training the proposed model, which overcomes most of the existing methods.