Accurate, automatic zero-shot wetland mapping from high resolution remote sensing imagery by prompting large foundation model (Segment Anything Model-SAM)

Wetlands are important landscape features that play a vital role in biological conservation and water resource management. Accurate and up-to-date monitoring of wetlands is essential for assessing and managing wetland resources at local and regional scales. However, refined mapping of wetlands using high resolution remote sensing data is still a challenge. This is mainly because traditional segmentation methods for high-resolution remote sensing images require manual interpretation, which is time and labour consuming. The recently introduced Segment Anything Model (SAM), a large foundation model for image segmentation, provides the ability to perform zero-shot segmentation on arbitrary images without training data. However, the default SAM model only generates object masks and does not assign labels. In this work, we introduce an accurate, automatic zero-shot refined wetland mapping method based on prompting SAM on high-resolution aerial imagery. We extract a number of negative and positive locations from existing low-resolution landcover products as prompt points, allowing SAM to achieve the ability to map wetlands from high resolution aerial images without manual annotation and training. Compared to the existing methods, the results show superior performance of the proposed method in refining wetland maps, which is close to manual annotation and matches well with ground truth data, making it a catalyst for driving the utility of SAM in land cover classification from high-resolution remote sensing imagery. To the best of our knowledge, this is the first attempt to adapt SAM to the remote sensing domain.