A New Deep Learning Method for Crop Yield Forecasting

Abstract:

The forecasting of crop yields is important for optimal nutrient management, crop market planning, crop insurance, and harvest management. Common yield forecasting approaches include conducting extensive manual surveys or using data from remote sensing. With the increasing amount of data provided by remote sensing imagery, there is a need for more sophisticated methods to extract the inherent spatiotemporal patterns of these data. Considerable progress has been made in this field by using deep Convolutional Neural Networks (CNN). However, no study before has investigated the use of Convolutional Long Short-Term Memory (ConvLSTM) for crop yield forecasting. In this study, ConvLSTM networks were developed for county-based soybean yield forecasting across the Contiguous United States (CONUS). We also proposed a new combined structure that integrates the ConvLSTM layers with the 3-Dimensional CNN (3DCNN) for more accurate and reliable spatiotemporal feature extraction. The models were trained by using historical yield data and MODIS satellite images over primary soybean growing counties in the CONUS. The forecasting performance of the developed models was compared against its traditional machine learning techniques and results indicate that the proposed combined structure significantly outperforms the other techniques and also performs better than both ConvLSTM and 3DCNN.