Precision mapping of soil salinity in reclaiming salt-induced wasteland with UAV multispectral images and machine learning

• The SFE spectral indices apply to complex soil salinization environments. • Model robustness is enhanced by the integration of environmental covariates. • A cost-tailored soil-salinity inversion strategy was proposed for precision mapping. Identifying the spatial distribution and degree of soil salinity accurately and promptly is crucial for reclaiming salt-induced wasteland. It provides decision support for rapidly delineating the center of salt patches, zoning salinized areas, and rationally planning reclamation regions. This study investigated the potential of deriving Salinization-Feature Enhanced (SFE) spectral indices from multispectral (MS) band combinations for estimating field soil salinity content (SSC) in the Luohui Canal Irrigation District, Shaanxi China. The sensitivity of the SFE indices and conventional spectral indices (SIs) to SSC was analyzed using the Pearson correlation coefficient and the Variable Iterative Space Shrinkage Approach (VISSA). We assessed the accuracy of estimating surface SSC during the seedling stage under complex soil salinization environments, analyzing the effects of varying spatial resolutions and sampling densities. Extreme Learning Machine (ELM) and Random Forest Classifier (RFC) were employed to invert surface SSC and generate a digital soil salinity map. The results indicated that the SFE (|r| = 0.58) indices showed a stronger correlation with SSC compared with the SIs (|r| = 0.26), indicating superior performance in surface SSC inversion. The incorporation of environmental covariates enhanced the model’s inversion accuracy and stability. The sample dataset was divided into 70% training and 30% test subsets using stratified sampling. Surface SSC prediction models were developed using SFE indices and SIs with both the ELM and RFC algorithms. The ELM_SFE (accuracy = 0.68) and RFC_SIs (accuracy = 0.69) models achieved the best predictive accuracy. The ELM_SFE and RFC_SFE models demonstrated the best inversion performance in severely and moderately salt-induced wasteland, respectively. This study provides valuable insights for precision agricultural management in salt-induced wasteland reclamation.