Applying deep neural networks and remote sensing to predict yields of major crops in the United States

Crop yield prediction has become increasingly important for early warning of food security, supply chain planning of the agriculture industry, and market prediction. Machine learning (ML) techniques such as decision trees, support vector machines, and random forests have become practical tools that can provide reasonable crop yield prediction. In recent years, with the increasing availability of remote sensing big data and rapid advances in ML, increasing opportunities for novel methods have emerged. Integrating remote sensing observations with deep learning methods such as Convolutional Neural Networks (CNNs) or Long Short-Term Memory (LSTM) has made great progress in improving the accuracy of crop yield prediction. Based on Moderate Resolution Imaging Spectroradiometer (MODIS) observations and county-level survey data, this study aims to evaluate the accuracy of different deep neural networks (CNN, LSTM, with and without Gaussian Process) to predict the three major crops in the US (corn, soybean and winter wheat) and explore how long before harvest these models can get reliable predictions. Our results indicate that deep neural networks show high accuracy in predicting the yields of different crop types and outperform traditional machine learning models. LSTM performs better in predicting soybean yield and CNN performs better in predicting corn yield; the Gaussian Process significantly improved the accuracy of LSTM but slightly reduced the accuracy of CNN. Deep learning models allow for near real-time forecasting of crop yield throughout the year and can usually provide reliable predictions 2-3 months before the harvest.