A genetic algorithm-optimized backpropagation neural network model for predicting soil moisture content using spectral data

Accurate assessment of soil moisture content (SMC) is crucial for applications in climate science, hydrology, ecology, and agriculture. However, conventional SMC characterization and measurement are expensive, time-consuming, and have negative effects on soil. Recently, the application of multispectral technology provides a new idea for SMC accurate detection. The objective of this study was to develop and compare regression and machine learning algorithms to estimate SMC from multispectral images.

A multispectral sensor was used to collect spectral images of 125 soil samples from five distinct soil textures in Shanxi province at varying degrees of soil moisture, ranging from arid to fully saturated. A set of seven spectral parameters was derived from images, and predictive relationships were developed against laboratory-measured SMC. A linear regression (LR) model and a backpropagation neural network model based on genetic algorithm optimization (GA-BP) were compared in this study to predict SMC.

The results showed that (1) the spectral reflectance and SMC exhibit a clear negative correlation, and the lower the SMC, the larger the spectral reflectance is. (2) The GA-BP neural network model exhibits higher prediction accuracy and performance (R² = 0.978 ~ 0.990, RMSE = 0.366 ~ 0.799%, MAE = 0.360 ~ 0.890%). (3) The GA-BP model exhibits the excellent inversion precision for the fine sand soil (R² = 0.990, RMSE = 0.518%, MAE = 0.360%).

This study introduces an effective methodology for accurate estimation of SMC using multispectral remote sensing technology. It further underscores the significant effectiveness of employing backpropagation neural networks and genetic algorithms in SMC prediction, providing a rapid, precise, non-intrusive, and practical approach towards precision agriculture.