Application of LAI and NDVI to model soybean yield in the regions of the Russian Far East

Abstract

Abstract Soybean yield modeling using remote sensing is an essential task in the south of the Russian Far East and makes it possible to plan sowing areas at the municipal level. This article presents a comparative assessment of the regression models’ accuracy, where the seasonal maxima of the LAI (Leaf Area Index) and NDVI (Normal Difference Vegetation Index), as well as the number of growing days (days with an average daily air temperature above 10°C), were considered as predictors. For four districts of the Amur Region and the Jewish Autonomous Region, MODIS (Moderate-resolution Imaging Spectroradiometer) data obtained from the arable land mask were used, using the Vega-Science web service, as well as soybean yield in 2010-2017. It was found that the maximum values of LAI and NDVI fall on weeks 31 to 33, which corresponds to the first half of August. In 2010-2017, the LAI-based models’ MAPE (Mean Absolute Percentage Error) was in the range 4.1 – 9.0%, and the RMSE (Root Mean Squared Error) was 0.06 to 0.13 t/ha. The corresponding errors of the regression model with NDVI were quite similar: MAPE 4.8 to 10.4%, RMSE 0.06 to 0.15 t/ha. This approach was evaluated with a ‘leave-one-year-out’ cross-validation procedure. There were no significant differences in the forecast error (APE) when using LAI and NDVI; at the same time, it was found that the quality of the regression model in the Tambovskiy and Oktyabrskiy districts is higher than in the Leninskiy and Mikhailovskiy districts. The median APE for Tambovskiy district was 7.2% for LAI and 8.8% for NDVI, for Oktyabrskiy the corresponding figures were 7.5% and 6.1%, for Leninskiy – 14.2% and 13.7%, and for Mikhailovskiy – 10.8% and 12.3%, respectively.