Combining CERES-Wheat model, Sentinel-2 data, and deep learning method for winter wheat yield estimation

Yi Xie

DOI: https://doi.org/10.1080/01431161.2022.2026521

IF: 3.531

2022-01-17

International Journal of Remote Sensing

Abstract:Timely and accurate regional wheat yield estimates is crucial to food security and sustainable development in the agricultural sector. In this study, a novel long short-term memory network (LSTM) model was developed to estimate wheat yield in Henan Province of China by combining field-measured yields, the time series of leaf area index (LAI) and yields simulated from CERES-Wheat model, and the LAI retrieved from Sentinel-2 imagery. The LSTM model was pretrained using the simulated LAI and yields at the agro-meteorological stations, and was retrained by the Sentinel-2 LAI and measured yields, hereafter referred to as retrained LSTM model. The yield estimation accuracy of the retrained LSTM model was compared with that of the LSTM model trained by the simulations alone (hereafter referred to as CGM-trained LSTM model) and that of the LSTM model trained by the Sentinel-2 LAI and measured yields alone (hereafter referred to as RS-trained LSTM model). The results showed that, at both the site and county levels, the retrained LSTM model provided improved accuracies compared with the CGM-trained and RS-trained LSTM models. It was because the accurate estimates of wheat yield at field scales were extended to regional scales through a combination of the advantages of both the crop growth models and satellite imagery. This study portrays the benefits of integrating crop growth models, satellite data and LSTM model in order to perform more reliable crop yield estimations over large areas.

imaging science & photographic technology,remote sensing