In view of rapid environmental degradation, the city of Delhi faces prolonged heat waves and rising temperatures. AI-based models have the potential to monitor and predict urban land surface temperature (LST) of Delhi during January 1 to December 31, 2023 in association with climatic (solar radiation, relative humidity), pollutant ( \(PM_{2.5}\) , \(PM_{10}\) , \(O_3\) , CO, NO, etc.), and other land cover factors (NDVI, NDBI). The present study aims to develop three STNN (Spatiotemporal Neural Network)-based weighted regression models (viz., STNNWR-v1, STNNWR-v2, and STNNWR-v3). Attention is paid to lowering the model complexity by reducing the number of parameters. The performance of STNNWR-v3 is significantly better in terms of evolution metrics and computation time when compared with baselines (For 30 independent trials, \(R^2\) score: \(0.934 \pm 0.002\) and MSE: \(2.743 \pm 0.102\) (mean ± standard deviation)). To support the generalizability of STNNWR-v3, the model is executed multiple times with different train-test settings. For each case, the entire dataset is partitioned so that six monitoring stations (randomly selected from 28 locations) are designated as the targeted locations, and the remaining locations are the training samples. The model demonstrates better predictive performance for such unknown (targeted) locations when trained on the remaining known locations. In several cases, the model achieves relatively high accuracy, and their R \(^2\) scores are greater than 0.70. In the case of the STNNWR framework, the optimized variant (STNNWR-v3) achieves a balance between model simplicity and predictive accuracy for utilizing a fewer number of model parameters. This reduced parameterization enhances both computational efficiency and the transparency of the model’s decision process as well.
