Abstract. Illegal waste dumping poses growing environmental risks as rapid urbanization overwhelms traditional monitoring, prompting the need for scalable, data-driven anomaly detection using remote sensing and machine learning. This study presents a semi-supervised graph transformer neural network for detecting illegal dumping in CALABARZON, a region increasingly burdened by unregulated waste due to rapid urbanization. The model leverages multivariate time-series data from Sentinel-1 SAR and Sentinel-2 MSI imagery (2020–2021) to capture complex spatiotemporal patterns in environmental indicators. A harmonized graph-based representation integrates spectral indices, radar features, and spatial relationships across time, enabling semi-supervised training under extreme class imbalance and limited ground-truth labels. The model exhibited strong generalization, with PR AUC stabilizing between 0.80 and 0.85, and consistent reductions in Huber and total loss values, indicating reliable modeling of typical site dynamics. However, performance on the minority “waste” class was constrained by data sparsity, achieving high accuracy (0.88) and precision (0.98) for the “non-waste” class but limited recall (0.50) and low precision (0.17) for “waste,” resulting in a high false positive rate. Sensitivity analysis highlighted shortwave infrared bands (B11, B12), vegetation-sensitive bands (B6, B9), indices such as MNDWI, NDWI, NDI45, NDTI, and RVI, and radar-derived features (e.g., VH backscatter variability, terrain angle) as critical for detection. These findings underscore the value of integrating optical and radar data for monitoring heterogeneous dumping behavior. While the model shows strong potential for early detection of illegal dumping, further refinement is needed to improve class precision and ensure actionable deployment in environmental monitoring systems.
Broqueza et al. (Tue,) studied this question.