Strategy for mapping soil salt contents during the bare soil period through a satellite image: Optimal calibration set combined with random forest

Abstract

Salt is a basic soil parameter that is especially sensitive to land degradation and global climate warming. Machine learning combined with satellite images offers the potential for effectively surveying soil salt contents (SSC) in the spatial and temporal dimensions. However, outlier spectral samples caused by environmental factors (e.g., particle size distribution, moisture, and vegetation) have interfered with the determination of the mapping relationship between ground-truth SSC and corresponding spectral features, resulting in a random forest (RF) mapping model with low generalization capability. Therefore, a spatial association module was introduced into the RF to remove the outlier spectral samples for building an optimal sample set in order to calibrate an improved random forest (IRF) model for SSC mapping. A total of 233 soil samples and a simultaneous Sentinel-2B multispectral image were acquired for the Weibei Plain of China. The spectral absorption features of SSC were characterized by three indices (band difference, band ratio, and band normalized). Then a spatial association function was developed to identify the outlier spectral samples for optimizing the calibration samples. Lastly, the SSC-responding bands and SSC-related indices were used as input variables to build the improved random forest (IRF) model for regional SSC mapping. Results showed that: (1) bands 3, 8, and 11 in the Sentinel-2B image were the SSC-responding bands, and band 11 had a significant correlation with SSC; (2) the band ratio index synthesized the multi-band spectral information, and effectively enhanced the spectral absorption signal of soil salt; the optimal SSC-related indices includedRI34 (reflectance ratio of band 3 and band 4),RI711,NDI611 (normalized reflectance values of band 6 and band 11), andDI45 (reflectance difference value of band 4 and band 5); (3) when the SSC-responding bands and optimal SSC-related indices were input into the IRF remote sensing estimation model, the accuracy parameters R2v and RPIQ were greater (0.89 and 3.52, respectively) than the comparable values with the RF model (R2v = 0.69; RPIQ = 2.90), with accuracy improved by 28.99 % and 21.38 %, respectively, indicating that satisfactory results for regional SSC mapping had been obtained; (4) SSC distribution showed that the values in the central part of the study area were slightly higher than in the southern and northern parts; the highest SSC area was mainly related to the distribution of saltern; additionally, agricultural activities and microtopography contributed significantly to the distribution of SSC at the farm scale. The IRF model based on the SSC-responding bands and optimal SSC-related indices provides a strategy for supporting regional mapping of SSC, and will be useful for environmental policy-making.