Foundation Models tillämpade på jordobservation

Satellite image analysis plays a crucial role in various fields, including environmental monitoring, disaster management, and urban planning. How- ever, training deep learning models for specific tasks such as land cover classification, semantic segmentation, and object detection typically requires substantial amounts of annotated data, which are costly and laborious to acquire in the remote sensing domain. Foundation models, pre-trained on massive datasets using self-supervised learning, offer a promising solution by learning general-purpose visual representations that can be efficiently adapted to diverse downstream tasks with reduced annotation requirements. This thesis evaluates the performance of vision foundation models in satellite imagery by comparing generalist models pre-trained on natural images (e.g., DINOv2) with domain-specific models specialized in satellite data (e.g., DOFA) and traditional task-specific architectures (e.g., ResNet, U-Net). The evaluation spans three key tasks: land classification, semantic segmentation, and object detection. For each task, performance was assessed using standard metrics such as accuracy and mean Intersection over Union (mIoU). The results show that generalist foundation models often match or surpass traditional models on RGB datasets, while domain-specific models excel on multispectral data thanks to specialized architectures. The findings highlights the potential of foundation models for satellite imagery analysis, while underscoring certain limitations, such as the detection of small objects and generalization across varied multispectral contexts. This work opens avenues for improving the adaptation of foundation models to the specific characteristics of satellite imagery, with potential applications in fields such as natural hazard management and environmental change detection.