Wheat yield predictions at a county and field scale with deep learning, machine learning, and google earth engine

Abstract

To meet the challenges of climate change, increasing population and food demand, a timely, accurate and reliable estimation of crop yield at a large scale is more imperative than ever for crop management, food security evaluation, food trade and policy-making. In this study, taking the major winter wheat production regions of China as an example, we compared a traditional machine learning method (random forest, RF) and three deep learning (DL) models, including DNN (deep neural networks), 1D-CNN (1D convolutional neural networks), and LSTM (long short-term memory networks) to predict crop yields by integrating publicly available data within the GEE (Google Earth Engine) platform, including climate, satellite, soil properties, and spatial information data. The results showed that all four models could capture winter wheat yield variations in all the county-years, with R2 of recorded and simulated yields ranging from 0.83 to 0.90 and RMSE ranging from 561.18 to 959.62 kg/ha. They all performed well for winter wheat yield prediction at a county level from 2011 to 2015, with mean R2≥0.85 and RMSE ≤ 768 kg/ha. At a field level, the spatial pattern of estimated winter wheat yield could capture the spatial heterogeneity and yield differences between individual fields across a county fairly well. However, only the DNN and RF models had relatively good performance at the field level, with mean R2 values of 0.71, 0.66 and RMSE values of 1127 kg/ha and 956 kg/ha, respectively. The model comparisons showed that the performance of RF was not always worse than DL at both the county and field levels. Our findings demonstrated a scalable, simple and inexpensive framework for estimating crop yields at various scales in a timely manner and with reliable accuracy, which has important implications for crop yield forecasting, agricultural disaster monitoring, food trade policy, and food security warning.