Exploring urban land surface temperature with geospatial and regression modelling techniques in Uttarakhand using SVM, OLS and GWR models

Given the climate change challenges, Uttarakhand has become crucial for examining land dynamics and regional climate interactions. This study employed a Support Vector Machine (SVM) for land use and land cover mapping for 2024, achieving 94% accuracy and a Kappa coefficient of 0.90, indicating robust mapping. Key land indices such as NDVI, NDWI, NDBI, NDSI, and NDBaI were calculated, along with Land Surface Temperature (LST) from Landsat 8 imagery. These indices were selected for their relevance in representing vegetation health (NDVI), measuring water content (NDWI), assessing urban areas (NDBI), identifying snow cover (NDSI), and highlighting the barren land (NDBaI), which all influence LST. Hotspot analysis with Getis-Ord Gi∗ revealed spatial distribution patterns of LST. Regression analysis showed significant relationships: a strong positive correlation between LST and NDBI (R² = 0.78) and a substantial negative correlation between LST and NDSI (R² = −0.80). The strong positive correlation highlights how urbanization contributes to rising surface temperatures, while the substantial negative correlation underscores the cooling effect of snow cover, which is particularly relevant as reduced snow cover could lead to higher LST in the context of climate change. These correlations offer deeper insights into how land cover changes can exacerbate or mitigate climate impacts in Uttarakhand. Two regression models were used for statistical modeling and spatial analysis: Ordinary Least Squares (OLS) and Geographically Weighted Regression (GWR). In OLS, the results reveal non-stationarity (p = 0.000) with an R² value of 0.79 while GWR significantly enhanced performance, achieving an R² value of 0.94. The improved performance of GWR (R² = 0.94) compared to OLS (R² = 0.79) can be attributed to GWR’s ability to account for spatial non-stationarity. This method allows for variations in relationships between LST and explanatory variables across different locations, effectively capturing local patterns that OLS may overlook. Spatial autocorrelation analysis, utilizing Moran’s I, exhibited a decrease from 0.606 (OLS) to 0.02 (GWR), This reduction indicates that GWR effectively captures spatial non-stationarity, minimizing residual autocorrelation by modeling local relationships between LST and its predictors that often remain in global models like OLS, thereby demonstrating its advantages in heterogeneous regions. The findings underscore the importance of employing GWR to better elucidate the connection between LST and its predictors, specifically in regions characterized by spatial non-stationarity, thereby offering insights crucial for informed decision-making amidst changing climatic conditions.