Soil Salt Content Inversion Using Novel Salinity Indices Based on a Stacking Model Yufei Lan,Ruiyin Tang,Guohong Li,Xuqing Li,Yancang Wang,Haizhou Chen,Tingxuan Wang
North China Institute of Aerospace Engineering School of Remote Sensing and Information Engineering, Langfang 065000-
Hebei Province Collaborative Innovation Center for Space Remote Sensing Information Processing and Application, Langfang 065000-
Abstract
Soil salinization is a major threat to land and agriculture in coastal regions, particularly in China Coastal New Areas, including Tangshan, Cangzhou, and Qinhuangdao. Due to its coastal geography and climate, this region faces severe soil salinity issues, impacting agricultural productivity and environmental sustainability. Traditional field sampling is time-consuming and limited in coverage, inadequate for large-scale monitoring. While remote sensing is widely used, existing salinity indices, such as S1 and SI1, show low correlation in this region due to complex soil and environmental conditions, limiting prediction accuracy. This study develops new three-band salinity indices and integrates advanced stacking regression models to enhance soil salinity monitoring accuracy.Using 2025 Sentinel-2 multispectral imagery and 95 field-measured soil salinity samples via conductivity, six bands-Blue, Green, Red, NIR, SWIR1, SWIR2-were selected to construct novel three-band salinity indices through addition, subtraction, ratios, and power operations. The results confirm the superiority of the novel indices. The best-performing index achieved a maximum Pearson correlation coefficient of 0.60 with soil EC, a substantial improvement over the 0.41 from the most effective existing index. The Stacking model yielded an outstanding validation coefficient of determination of 0.870. This performance not only surpassed that of high-performing individual models like RF (R2=0.773) and XGBoost (R2=0.768), but also substantially exceeded the traditional Partial Least Squares Regression (PLSR) model (R2=0.526). The success of the Stacking model highlights its ability to effectively fuse the predictive strengths of its diverse base learners, enhancing overall robustness and accuracy. This research presents a robust methodology for regional soil salinity estimation, providing critical decision-making support for sustainable land management in the China Coastal Areas and other similar regions.
Keywords: Soil salinity, Sentinel-2, Indices constructions, Machine Learning, Stacking Regression
