Lightweight Transfer Learning for Water Body Segmentation Using Adaptor-Based Fine-Tuning

This paper presents a lightweight transfer learning approach for water body segmentation by applying Adaptor-based fine-tuning on general image datasets. Traditional deep learning models often require full-scale retraining for each new task, which is both computationally expensive and time-consuming. In contrast, Adaptor networks—lightweight modules that selectively fine-tune task-specific layers while retaining most pre-trained model parameters—offer an efficient alternative. Water bodies present unique challenges for segmentation, such as varying lighting, reflections, and seasonal fluctuations. These factors can confuse distinguishing water from land, particularly in cases where reflections resemble adjacent features. Adaptor-based fine-tuning helps to reduce computational costs while ensuring the model captures the fine distinctions between similar regions like shallow water and land. This paper evaluated the method on a dataset containing lakes, rivers, and wetlands under diverse environmental conditions to test the model's robustness. The results indicate that Adaptor-based fine-tuning achieves comparable performance to fully fine-tuned models, with a significant reduction in computational costs and training time. The method also demonstrated high precision in segmenting water bodies under challenging conditions, such as occlusions and reflections. This study highlights the potential of lightweight transfer learning in resource-constrained environments, with applications in environmental monitoring, hydrological modeling, and geographic information systems (GIS). By demonstrating the effectiveness of Adaptor networks, this work contributes to the broader field of efficient transfer learning, showcasing how minimal adjustments to pre-trained models can yield accurate task-specific performance.