Explainable transformer-based models for land use and land cover classification

Abstract

Accurate land use and land cover (LULC) mapping is crucial for environmental monitoring, sustainable resource management, and urban planning. Recent advancements in deep learning, particularly with transformerbased architectures have shown exceptional performance in image classification tasks, making them promising options for satellite-based LULC analysis. However, their practical implementation faces two significant challenges: (i) high computational costs during training and inference, and (ii) their inherently opaque decision-making processes, which limit their acceptance in high-stakes, policy-driven applications. To tackle these challenges, this study proposes a two-block framework that combines transformer-based LULC classification with model-agnostic explainability. The first block employs ViT and SwinT architectures, enhanced through transfer learning on ImageNet-21k and selective fine-tuning of the final layers, significantly reducing training costs without compromising accuracy. The second block uses Captum's Integrated Gradients to generate attribution maps, offering pixel-level insights into model predictions and enabling the detection of potential data biases. Extensive experiments conducted on the EuroSAT and PatternNet datasets demonstrate that the proposed framework achieves over 99% and 98% accuracy, respectively, while reducing computational time by up to 50 % compared to full fine-tuning. The results indicate that this approach enhances both efficiency and interpretability, paving the way for the practical, trustworthy, and ethically responsible deployment of transformer-based models in LULC mapping and broader remote sensing applications.