Soil salinization, a major driver of global land degradation, threatens agricultural productivity, ecosystem stability, and sustainable land use. Although remote sensing provides valuable large-scale and high-frequency monitoring capabilities, its effectiveness is often limited by temporal noise and land-cover heterogeneity. To address the limitations of traditional single-model approaches in handling temporally noisy or mixed land-cover imagery, which often yield unstable results, this study develops a progressive inversion framework. This framework integrates temporal compositing with land-cover classification and is applied to Sentinel-2 imagery from the semiarid agropastoral ecotone of the Western Songnen Plain in China. Seasonal mean composites (May-October) were used to stabilize spectral responses, while separate inversion models for cropland and grassland captured distinct surface characteristics. Random forest and elastic net were employed for complementary feature selection, supported by log transformation and oversampling to alleviate data imbalance. Among five machine learning algorithms tested, the mean composite approach improved accuracy (from 0.40 to 0.51), with land-cover classification providing an additional 8% gain (R-2 = 0.60 for cropland and R-2 = 0.55 for grassland). Vegetation indices dominated grassland models, whereas salinity and topographic factors prevailed in cropland.