Explainable AI for Remote Sensing
Quantitative evaluation of attribution methods on multispectral and SAR imagery (IGARSS 2023)
I built a PyTorch + Captum evaluation framework to compare five primary-attribution XAI methods (LIME, HiResCAM, GradSHAP, Saliency, Occlusion) across six metric categories including faithfulness, robustness, and localization. State-of-the-art backbones (ConvNeXt, Vision Transformers, Focal Networks) were fine-tuned for remote sensing, reaching 82% accuracy on SAR and 97% on RGB. A pipeline combining Otsu thresholding and superpixel segmentation enabled automated region-of-interest detection for systematic comparison of explainability across imaging modalities.