Using Multi-Source Remote Sensing Data and Machine Learning Algorithms to Predict Maize Yield in China

Keywords: maize yield, machine learning, multisource data, optimal period
Session Type: Virtual Paper
Day: Wednesday
Session Start / End Time: 4/7/2021 08:00 AM (Pacific Time (US & Canada)) - 4/7/2021 10:20 AM (Pacific Time (US & Canada))
Room: Virtual 29

Abstract

Accurate and timely estimation of crop yield at a large scale is important for food security. However, few studies deeply investigated an adaptable method for maize yield estimation in whole China, furthermore, optimal estimation period has not been clearly discussed. In this study, multi-indicators (GPP, ET, Ts, LAI and soil properties), phenological information and machine learning algorithms (RFR and GBDT) were used for maize yield estimation, and aim to (1) evaluate the accuracy of yield estimation in China by multimodal data and machine learning, (2) explore the optimal period for the maize yield estimation with the accuracy at a high level, (3) analyze the adaptability of machine learning based maize yield estimation method over space. The results could be summarized as: (1) RFR algorithm outperformed GBDT in terms of maize yield estimation accuracy; (2) Ts showed the best performance of yield estimation in case of single indicator used. While multi indicators jointly used, GPP, Ts, ET and LAI combination yielded the greatest performance for yield estimation with R2 of 0.80 and rRMSE of 15.36% in RFR algorithm used; (3) The maize yield estimation accuracy decreases with the estimation period earlier, and the performance was still at a relative high level within at least 24 days before maize maturity (R2 > 0.80 and rRMSE < 15.39%); (4) Using machine learning and multi-indicators to estimate maize yield performed a strong adaptivity over space, which is able to cope with the spatial heterogeneity. This study provides a reliable reference for agricultural production manageme.