Using hyperspectral data to estimate and map surface and subsurface soil salinity, pH, and calcium carbonates in arid region

Abstract Accurate prediction and mapping of soil properties are essential in environmentally sustainable land use. Traditional approaches are laborious and costly, however. This article demonstrates the use of Prismatic Imaging Sensor (PRISMA) with a cost-effective, rapid, and environmental-friendly approach of forecasting and mapping the soil pH, electrical conductivity (EC), and calcium carbonate (CaCO 3 ) of New Delta region in Egypt’s Western Desert based on hyperspectral data and machine learning (ML) techniques. The study integrates soil wet chemistry data, hyperspectral reflectance, multivariate regression, and ML models to improve soil property estimation. PRISMA hyperspectral imagery was acquired, processed, and classified to map land use and land cover (LULC). Seventy-four representative bare soil profiles were collected and their pH, EC, and CaCO 3 content were determined. Hyperspectral reflectance data for the samples were reaped, and noisy spectral bands were eliminated to improve data quality and prediction accuracy. Partial least squares regression (PLSR), random forest (RF), multivariate adaptive regression splines (MARS), and support vector regression (SVR) were the different ML models attempted. Spectral band selection was improved by the implementation of competitive adaptive reweighted sampling (CARS); and multiple linear regression (MLR) was used for the development of prediction equations to map soil properties over PRISMA image. Soil condition was found highly variable, ranging from slightly to extremely alkaline pH, non-calcareous to extremely calcareous in CaCO 3 content, and non-saline to highly saline soils. Of the models, PLSR provided the best fit estimates for surface soil pH (R² = 0.1186, RMSE = 3.115, RPD = 0.7902), surface soil EC (R² = 0.2281, RMSE = 1.196 dS/m, RPD = 1.132), surface CaCO 3 content (R² = 0.5984, RMSE = 2.73%, RPD = 1.817), subsurface soil EC (R² = 0.2557, RMSE = 1.7481 dS/m, RPD = 1.124), and subsurface CaCO 3 content (R² = 0.6092, RMSE = 2.32%, RPD = 1.779). RF yielded the highest performance for predicting subsurface soil pH (R² = 0.1517, RMSE = 2.876, RPD = 0.1777). Identification of relevant spectral bands, calibration of the prediction models, and their use in PRISMA imagery resulted in high-resolution maps of soil parameters. These findings are extremely helpful to improve land reclamation planning and the effectiveness of their application.