Abstract:

Soil moisture is a significant factor influencing crop growth and a crucial aspect in environmental factors such as hydrology, ecology, and climate. It exerts profound impacts on natural environmental processes. The development and application of remote sensing technology have provided effective means for monitoring regional surface soil moisture. This study primarily utilized Sentinel data as the data source to extract characteristic parameters and construct an input parameter dataset. BP neural network optimized by genetic algorithms was then employed to reconstruct the soil moisture inversion model. The results indicated that 20 characteristic parameters extracted from Sentinel microwave and optical remote sensing images could be used to invert soil moisture content within the study area based on the BP neural network. However, redundant characteristic parameters result in low computational efficiency and longer processing time for the model. To address this, a feature selection algorithm is utilized to reduce the dimensionality of the feature subset. Feature screening was further conducted using the importance scores obtained from XGBoost. Eight optimal feature variables were ultimately determined, which retain the main information of the feature dataset while effectively reducing data redundancy. The inversion results demonstrated an R² value of 0.62 and an RMSE of 0.59%. The network runtime and memory usage were significantly improved compared to the full-feature GA-BP neural network, with an average reduction in runtime of 75 seconds and a decrease in memory usage by an average of 863.86MB. The inversion of soil moisture within the study area throughout the year revealed that July and September had the highest soil moisture content, with a maximum soil weight water content of 38.29% and an average of 14.52%. Conversely, January exhibit the lowest soil moisture content, with a maximum soil weight water content of 15.71% and an average of 12.52%. These patterns closely align with the precipitation trends observed during the year. The results of this study demonstrate that the proposed approach achieves rapid and accurate inversion of large-area soil moisture while maintaining high inversion precision. This study offers a novel approach for combining microwave and optical remote sensing data to invert soil moisture on farmland surfaces.

Key words: Soil moisture, Microwave remote sensing, Neural network, Genetic algorithm, Feature selection