Evaluating spectral indices from MODIS to predict maize and soybean regional yields

Abstract

Abstract A regression model with spectral information and dummy variables was developed and evaluated for predicting regional maize and soybean yield in the agricultural rain-fed region of Córdoba, Argentina. The study area comprises eleven departments that currently harvest more than 80% of the provincial production of maize and soybean. In this study monthly Normalized Difference Vegetation Index (NDVI) product (MOD13C2) and daytime Land Surface Temperature (LST) product (MOD11C3) derived from the MODIS sensor on board of TERRA satellite were used as model input. From these data Temperature Vegetation Dryness Index (TVDI) was calculated and assessed also. In total, 19 summer crop seasons were analyzed between 2000/2001 and 2018/2019. There is a close and negative relationship between the NDVI, with both LST and TVDI. The best regression models with dummy variables were selected to estimate yield variation on a regional scale are integrated both with spectral information, as LST from January and NDVI from February, and factors linked to edaphic and management differences of each department, as well as the technological improvement in the model for soybean. By using an adaptation of the Leave One Out Cross-Validation (LOOCVad) technique, model accuracy was verified. The Residual Standard Error (RSE) obtained each year was, mostly, lower than that obtained for the entire record (general models). The mean RSE obtained for the set of years was 279.4 and 579.4 kg ha− 1 for soybean and maize, respectively, which are below those ​​obtained from the general models (354.7 and 788.6 kg ha− 1, respectively).