TRANSFORMER-BASED ENVIRONMENTAL INTELLIGENCE FOR LAND-USE MONITORING AND INFORMED DECISION-MAKING

Abstract

Land-use monitoring plays a vital role in environmental management and informed decision-making, Land-use and land-cover (LULC) monitoring is critical for environmental management, urban planning, agricultural supervision, and climate-change assessment. However, reliable land-use classification from remote sensing imagery remains challenging due to high intra-class variability, strong inter-class spectral similarity (particularly among vegetation categories), complex urban spatial layouts, and domain shifts across sensors and spatial resolutions. To address these challenges, this study proposes a Swin-CBAM hybrid architecture that integrates the hierarchical, shifted-window self-attention of Swin Transformer-Tiny with Convolutional Block Attention Module (CBAM) refinement after each transformer stage. This design enables simultaneous modeling of global contextual dependencies and local discriminative cues through explicit channel-wise and spatial-wise feature recalibration. The proposed framework is evaluated on two widely used benchmark datasets EuroSAT (satellite imagery, 10 classes) and UC Merced Land Use (aerial imagery, 21 classes) under consistent preprocessing, augmentation, and training conditions. Experimental results demonstrate that Swin-CBAM consistently outperforms strong CNN and transformer baselines, achieving 97.76% accuracy and 97.62% macro-F1 on EuroSAT, and 95.71% accuracy and 95.48% macro-F1 on UC Merced. Notably, the model yields the largest gains on difficult and fine-grained categories (e.g., spectrally similar vegetation types and structurally complex residential scenes), reduces key confusion patterns, and improves cross-dataset generalization. Efficiency analysis shows that the proposed enhancement introduces only modest computational overhead relative to Swin-Tiny while delivering a favorable accuracy efficiency trade-off. Overall, the proposed Swin-CBAM framework provides an effective and deployable solution for transformer-based environmental intelligence, supporting reliable, interpretable, and scalable land-use monitoring for real-world decision-support systems.