Beyond Parallel Trends: A Confounder-Adjusted, Spatially Explicit Test of Climate–Land-Use Associations in a Western Ghats-Adjacent Semi-Arid District, South India

Abstract

Land-use/land-cover (LULC) change and climate variability are often reported together in district-level studies, yet their statistical association is rarely tested after accounting for topographic and anthropogenic influences. This study examines Erode District, Tamil Nadu, a semi-arid district adjacent to the Western Ghats, using a confounder-adjusted and spatially explicit framework. Sentinel-2 imagery was classified for 2016 and 2024 using a Random Forest model, achieving test-set overall accuracies of 75.7% and 74.8% and Kappa values of 0.68 and 0.66, respectively. Agricultural land declined by 8.0%, built-up area increased by 3.2%, and forest and tree cover increased by 14.3%. Climate trends were evaluated across 156 gridded nodes for 1990–2025 using autocorrelation-aware Mann-Kendall testing and Sen's slope estimation. Minimum temperature increased significantly at all nodes (+0.0237 °C/year, p=0.002), while wind speed declined significantly (-0.0059 m/s/year, p<0.001). Logistic regression on a 1 km grid showed that the agriculture-to-built-up transition remained significantly associated with the local minimum-temperature trend after adjustment for elevation, slope, distance to road, and population. Confounder-adjusted models consistently improved model fit, with AIC reductions ranging from 7.9 to 108.3 points. Residual spatial autocorrelation remained significant, indicating unresolved spatial structure. The findings support a replicable framework for testing climate–land-use associations while avoiding causal interpretation.