Development of remote sensing-based yield prediction models at the maturity stage of boro rice using parametric and nonparametric approaches

Abstract

This research aims to develop rice yield prediction models using satellite remote sensing-based vegetation indices at the optimum harvesting time before flash flooding. Five relevant vegetation indices, the normalized difference vegetation index (NDVI), normalized difference water index (NDWI), rice growth vegetation index (RGVI), moisture stress index (MSI), and leaf area index (LAI), were used to develop several empirical yield prediction models for rice production. For this research, Sentinel-2 images with 10 m spatial resolution was used for the haor area of Bangladesh. To calibrate and validate the remote sensing images at such large spatial and temporal scales, ground reference data of the vegetation indices were used. The generated models were validated using both parametric (simple and multiple) and nonparametric (artificial neural network, ANN) regression analyses. The crop yield models that were developed using regression analyses showed very significant agreement with the ground reference yield information. The best estimated performances for the RGVI (R2 = 0.44), NDVI (R2=0.63), NDVI (R2 = 0.55), and NDVI (R2 = 0.67) in the simple regression analyses were observed for 2017, 2018, and 2019 and the average of seasons from 2017 to 2019. On the other hand, the composite NDVI-RGVI (R2 = 0.65), NDVI-NDWI (R2 = 0.56), and NDVI-MSI (R2 = 0.69) indices were the best-performing vegetation indices for developing boro rice yield prediction models using multiple regression. Nevertheless, in the ANN-based machine-learning results, NDVI exhibited higher accuracy for the average boro rice season (2017–2019) by using a simple regression approach (R2 = 0.84) and multiple regression analysis (R2 = 0.91) of the average NDVI-MSI composite index. Validation between the actual and predicted yields showed that more than 70% of the study area can be accurately predicted using vegetation indices with ground reference mean yield data. Moreover, in 2018, the predicted yields by using simple and multiple linear regression were 4.25 and 4.23 MT/ha, respectively. The developed models are applicable 118–132 days after planting (DAT) in any similar environment for agricultural practices. Therefore, the yield prediction models of boro rice at the maturity stage can be useful for farm risk management, insurance premium determinations, and relevant stakeholder decision-making to mitigate the effects of extreme flash flood events.