Land Use and Land Cover Analysis for Solid Waste Management in NCT Delhi Using Fusion Aware Segformer

Abstract

The use of Land Use Land Cover (LULC) analysis is a fundamental requirement for urban solid waste management (SWM); however, conventional LULC analysis methods are not well suited to the spatio-temporal variability, multi-sensor heterogeneity, and seasonal variations of highly dynamic urban environments. To address this problem, in this work, we propose a novel Fusion Aware SegFormer, which is an upgraded SegFormer based architecture featuring a Spatio-Temporal Encoding framework with a Hierarchical Transformer module, and a Temporal Gated Cross Attention (TGCA) module to maintain the temporal consistency for multi-temporal inputs. In this context, the Adaptive Contextual Atrous Spatial Pyramid Pooling (AC-ASPP) is the last layer, which further improves spatial representation and contextual information under various scales. The model was developed in Python and trained using Sentinel-2 and Landsat-8 imagery to overcome the challenge of using different sensors and time to learn the features. Experimental results indicate that the proposed method outperforms existing baseline methods with 94.6% pixel accuracy, 83.5% mean IoU, 0.86 temporal consistency score, 0.87 F1-score, and 0.84 Kappa coefficient. In addition to being accurate, the model can be used for temporal change detection, spatial transition matrices, hotspot analysis, and integration with the SWM infrastructure in Delhi and can be applied for sustainable urban planning and smart waste management.