Multidimensional soil salinity data mining and evaluation from different satellites

Abstract

Soil salinization, a common land degradation mode, restricts the ecological environment and is a global issue due to climate change. Accurately, quickly and effectively monitoring soil salinity is critical for governmental institutions that develop hazard prevention and mitigation strategies. Remote sensing (RS) technology provides a viable alternative to traditional field work due to its large area coverage, abundant spectral information and nearly constant observations. Key issues in RS-based soil salinity monitoring include the lack of both data-mining techniques for obtaining spectral band information and comprehensive considerations of synergies among different spectra. The main objective of this study was to provide in-depth explorations of data mining and integration algorithms from different satellites to multidimensionally evaluate soil salinity models. The Ebinur Lake Wetland Reserve (Xinjiang Province, China) was selected as a case study. First, ground-measured visible and near infrared (VIS-NIR) spectral data were combined with the RS band to simulate Landsat 8 (L8) and Sentinel 2 (S2) and 3 (S3) data. Second, one-dimensional RS bands and 15 soil salinity and vegetation indices were selected, and 15 spectral data transformations (reciprocal, differential, absorbance, etc.) were obtained. Two- and three-dimensional spectral indices were constructed, and the response relationships between different spectral indices and soil electrical conductivity (EC) were comprehensively explored. Finally, an integrated multidimensional algorithm was used to estimate soil salinity in high-performance models for the three satellites. The results showed that all data-mining-based model combinations performed well for all satellites (R² > 0.80). However, with multidimensional model combinations, S3 presented the highest predictive capability (R² = 0.89, RMSE = 2.57 mS·cm^-1, RPD = 2.05), followed by S2 (R² = 0.86, RMSE = 2.71 mS·cm^-1, RPD = 1.90) and L8 (R² = 0.85, RMSE = 2.84 mS·cm^-1, RPD = 1.87). Therefore, data mining with integration algorithms in model combinations performs significantly better than previous models and could be considered a promising method for obtaining improved results from soil salinity susceptibility models in similar cases.

Keywords: Data mining; Integrated algorithm; Multidimensionality; Remote sensing; Soil salinity.