Multispectral Inversion of Starch Content in Rice Grains from Yingjiang County Based on Feature Band Selection Algorithms

The starch content in rice grains is a key factor in determining their quality. An optimal starch level not only ensures grain fullness, improving storage stability, but also enhances the stickiness and viscosity of cooked rice, thereby boosting its palatability and nutritional value. However, traditional methods for monitoring starch content are expensive and lack the capability to provide rapid spatial distribution information across large areas. To address this limitation, this study focuses on mature rice grains in the Yingjiang region, leveraging multispectral data from the Sentinel-2 satellite. First and second derivative transformations were applied to the multispectral reflectance data, followed by the use of three feature selection algorithms to identify key spectral bands. BP neural networks and ELM neural network regression models were then integrated to quantitatively estimate starch content across the study area. As a result, high-precision spatial distribution maps of starch content were generated, providing a novel and efficient method for large-scale rapid monitoring. The results demonstrate that, compared to full-band data, the use of SPA feature selection significantly improved the predictive accuracy of both BP and ELM models, despite a slight increase in the models’ MSE. Similarly, CARS feature selection also contributed substantially to enhancing the accuracy of the BP and ELM models. In contrast, UVE feature selection significantly reduced the MSE of the BP model, improving predictive precision, with the model achieving an R2 of 0.8061 and an MSE of 0.3896. This study highlights that the inversion method, which combines feature selection algorithms with machine learning models, can effectively enhance the predictive accuracy of starch content estimation. Among the tested approaches, the combination of UVE feature selection and BP neural networks delivered the best performance. These findings confirm the feasibility of utilizing Sentinel-2 satellite multispectral data for the quantitative inversion of agronomic parameters across large agricultural areas, providing robust technical support for precision agriculture.