Attention based Peephole LSTM model for Soybean crop yield prediction

Abstract

In the agricultural domain, crop yield prediction is one of the most challenging tasks as the predicted yield depends on the input features such as the genomic type of crop, soil, and weather conditions that are taken into consideration. With the advent of precision agriculture, a lot of sensors have been used to measure different parameters related to the crop, which has brought in more challenges for feature selection. In this research work, an attention-based peephole LSTM model was designed to predict the yield of the Soybean crop in the US Corn Belt. The Attention block is designed to automatically learn the important features from the training data, while the Peephole LSTMs predict the yield of the crop using the weighted temporal features coming from the attention block. While comparing the generalized and site-specific modeling approaches in the nine states of corn belt with the same architecture, it was observed that on average the RMSE score of site-specific models were less than the generalized model by 1.31 bushels per hectare, making the site-specific approach a better option for yield prediction. Unmasking the black box property of the proposed architecture revealed that the model gave more weightage to precipitation and less weightage to vapor pressure.