Zero-Shot Land Cover Segmentation with GroundingDINO-Segment Anything Model on WorldView-3 and DeepGlobe Data

Accurate and up-to-date land cover data is crucial for various applications, ranging from urban planning and climate change studies to environmental management and disaster response. This study evaluates the performance of the combined GroundingDINO and Segment Anything Model (SAM) approach for zero-shot land cover (LC) segmentation on very high-resolution (VHR) WorldView-3 imagery and the Deep-Globe dataset. The model was initially assessed on WorldView-3 imagery using a merged set of seven LC classes, yielding moderate performance with an overall Intersection over Union (IoU) of 0.4167 and an F1 score of 0.5883. Subsequent evaluation on the DeepGlobe dataset, without class-specific adjustments, showed an overall IoU of 0.5232 and an F1 score of 0.6869. Classes such as forest land and agriculture land achieved notably higher segmentation accuracy, while urban areas and rangelands presented significant challenges. These findings underscore the potential of prompt-driven segmentation in remote sensing, while also highlighting areas for improvement. Future work will focus on fine-tuning the model with remote sensing-specific LC data to enhance its ability to accurately capture complex land cover features.