Transformer-based land use and land cover classification with explainability using satellite imagery

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Geospatial Analysis

Background:

• Transformer models have advanced Land Use and Land Cover (LULC) analysis but suffer from high computational costs.
• Practical implementation of these powerful models is hindered by resource demands.

Purpose of the Study:

• To balance computational cost and accuracy in transformer-based LULC analysis.
• To enhance transparency and trust in AI-driven LULC models.
• To explore the application of Explainable AI (XAI) in this domain.

Main Methods:

• Utilized transfer learning and fine-tuning strategies to optimize transformer models.
• Employed Captum for bias detection and interpretation of AI decisions, ensuring transparency.
• Applied methods to satellite imagery for LULC classification.

Main Results:

• Transfer learning demonstrated potential for improving transformer-based satellite image classification.
• Strategic fine-tuning maintained model efficiency with minimal impact on accuracy.
• Captum facilitated the uncovering and mitigation of potential biases.

Conclusions:

• Transformer-based models, optimized with transfer learning and fine-tuning, offer efficient and accurate LULC analysis.
• Explainable AI (XAI) is crucial for transparent and trustworthy LULC modeling.
• This approach promotes fairness and broader adoption in fields like urban planning and environmental studies.