EXPLAINABLE DEEP LEARNING FRAMEWORK FOR SPATIO-TEMPORAL SOIL POLLUTION ASSESSMENT USING MULTI-SOURCE REMOTE SENSING DATA

Abstract

Soil pollution has become one of the more stubborn environmental problems we deal with — it creeps into food, water, and ecosystems without much fanfare until the damage is already done. The usual monitoring route  sending field teams to collect samples and running them through lab analysis  gives accurate numbers but only at a handful of points, far too sparse to really understand what's happening across a region or over time. This paper lays out a framework that pulls together multiple remote sensing sources and an explainable deep learning model to estimate soil heavy metal concentrations and pollution patterns across both space and time. Sentinel-2 optical imagery was cpmbined with  Sentinel-1 radar, MODIS land surface temperature, and a set of topographic and weather variables to build a feature space that captures the physical drivers behind metal accumulation and movement. The architecture pairs a convolutional encoder for spatial patterns with a recurrent module for temporal dynamics, and we wrap it in an integrated gradients layer so that every prediction can be traced back to the inputs that drove it. Tested across an industrial-agricultural belt in eastern India between 2019 and 2024, the framework hits an R² of 0.91 for lead and 0.88 for cadmium — comfortably above what regression, tree ensembles, or a plain convolutional network can manage. The explainability work shows that shortwave infrared bands and radar-derived moisture are doing most of the heavy lifting, with precipitation and temperature governing the seasonal swings. For agencies actually responsible for remediation planning, this offers something usable: transparent, defensible, and scalable.