Study on the EC prediction of cracked soda saline-alkali soil based on texture analysis of high-resolution images from ground-based observation and machine learning methods

Abstract

The rapid and accurate measurement of soil salinity is essential for assessing the level of soil salinization. In natural conditions, the surface of cohesive soda saline-alkali soil experiences shrinkage and cracking phenomena during water evaporation. The propagation and development of these desiccation cracks can be effectively described using texture features due to their statistical randomness. This study aims to establish a relationship between electrical conductivity (EC) values and texture features derived from both gray-level co-occurrence matrix (GLCM) and wavelet decomposition, and to develop a prediction model for EC accordingly. To achieve these objectives, crack images on the surface of 200 soil samples with varying levels of salinity were obtained in the Songnen Plain field. GLCM was computed and wavelet decomposition was performed to extract texture features in different directions and scales. The results demonstrate a significant correlation between texture features and soil EC values. Among the various texture features, 12 GLCM texture features with a grayscale level of 2 and a step size of 1 pixel, and wavelet texture features derived from a 4th level orthogonal decomposition based on the coiflet-1 function were selected as the optimal parameters for establishing and comparing the predictive effects of two machine learning models on soil EC values. In comparison to the testing results of the BP neural network model (R2=0.83, RPD=1.64, RMSE=0.32 dS/m, MAE=0.18 dS/m), the random forest model exhibited higher accuracy and stability (R2=0.95, RPD=3.48, RMSE=0.21 dS/m, MAE=0.07 dS/m), indicating that although both machine learning models demonstrated rapid and nondestructive capabilities, the random forest method was better suited for EC prediction in soda saline-alkali soil due to its superior accuracy.