Assessment of Soil Salinity in Arid Regions Using Remote Sensing and Machine Learning Models: Evidence from Karakalpakstan

Bekmurat Abdikairov (Institute of Agriculture and Agrotechnologies of Karakalpakstan, Nukus, 230109, Uzbekistan)
Mukhiddin Juliev (Institute of Fundamental and Applied Research at (TIIAME) National Research University, Tashkent, 100000, Uzbekistan Turin Polytechnic University in Tashkent, Tashkent, 100095, Uzbekistan Tashkent State Technical University, Tashkent, 100095, Uzbek)
Madinabonu Kholmurodova (Tashkent Institute of Irrigation and Agricultural Mechanization Engineers, National Research University, Tashkent, 100000, Uzbekistan.)
Zilolakhon Djanpulatova (Institute of Fundamental and Applied Research at (TIIAME) National Research University, Tashkent, 100000, Uzbekistan.)
Zuhra Khadjieva (Institute of Fundamental and Applied Research at (TIIAME) National Research University, Tashkent, 100000, Uzbekistan.)
Sojidabonu Turdalieva (Institute of Fundamental and Applied Research at (TIIAME) National Research University, Tashkent, 100000, Uzbekistan.)

Forum Geografi
Vol. 41 No. 1 (2027): April 2027

Soil salinization is a major environmental challenge in arid and semi-arid regions, reducing agricultural productivity and threatening sustainable land management. The spatiotemporal dynamics of soil salinity in the Shimbay district (Republic of Karakalpakstan, Uzbekistan) were investigated using multi-temporal Sentinel-2 imagery and the performance of Partial Least Squares Regression (PLSR), Random Forest (RF), and Multiple Linear Regression (MLR) models was compared. Soil salinity was assessed using the Normalized Difference Salinity Index (NDSI), Normalized Difference Vegetation Index (NDVI), Soil Moisture Index, and land surface temperature, together with field-measured soil electrical conductivity. The results revealed considerable interannual variability in soil salinity, with NDSI values ranging from 0.011 to 0.058 during 2018–2025 and an increasing salinity trend in 2024–2025. Among the evaluated models, PLSR achieved the highest predictive accuracy (R2 = 0.934, root mean square error (RMSE) = 0.398, mean absolute error (MAE) = 0.329), outperforming RF (R2 = 0.810) and MLR (R2 = 0.677). These findings demonstrate that integrating Sentinel-2-derived spectral indices with advanced modeling techniques provides an effective approach for assessing regional soil salinity. The proposed framework offers valuable support for sustainable land and water management in arid agricultural regions. However, further validation using larger datasets and broader environmental conditions is recommended.