High-resolution remote sensing data enables the extraction of fine-detailed boundaries of open-pit minefields, which is crucial for various applications such as ecological restoration, environment impact assessment, mining field disaster minoring, etc. In the last decade, a variety of convolutional neural networks (CNNs) and vision transformers (ViT) approaches have been developed for extracting the boundary and coverage of open-pit minefields. However, these deep learning approaches are always computationally expensive in pretraining and fine-tuning the network parameters. In addition, the diverse land cover/land use of different open-pit minefields poses a big challenge in building a large-scale benchmark dataset. To conduct efficient open-pit minefield extraction with limited labelled data, this paper employs a large-scale foundation model called Segment Anything (SAM) to develop the few-shot learning strategy for extracting open-pit minefield with slightly fine-tuning SAM and without fine-tuning SAM, respectively. The experiment demonstrates that the proposed SAM-enhanced few-shot learning outperforms pretraining the-state-of-the-art semantic segmentation approaches in terms of extraction precision and time cost. We hope our work can provide a solution for complex open-pit minefield extraction with a small number of labelled datasets.

Coastal land use represents the combination of various land cover forms in a coastal area, which helps us understand the historical events, current conditions, and future progress of a coastal area. Currently, the emergence of high-resolution optical satellite images significantly extends the scope of coastal land cover recognition, and deep learning models provide a significant possibility of extracting high-level abstract features from an optical satellite image to characterize complicated coastal land covers. However, recognition systems for labeling are always defined differently for specific departments, organizations, and institutes. Moreover, considering the complexity of coastal land uses, it is impossible to create a benchmark dataset that fully covers all types of coastal land uses. To improve the transferability of high-level features generated by deep learning to reduce the burden of creating a massive amount of labeled data, this paper proposes an integrated framework to support semantically enriched coastal land use recognition, including foundation model-powered multi-label coastal land cover classification and conversion from coastal land cover mapping into coastal land use semantics with a vector space model (VSM). The experimental results prove that the proposed method outperformed the state-of-the-art deep learning approaches in complex coastal land use recognition.

Water resources are of critical importance across all ecological, social, and economic realms. Accurate extraction of water bodies is of significance to estimate the spatial coverage of water resources and to mitigate water-related disasters. Single-modal remote sensing images are often insufficient for accurate water body extraction due to limitations in spectral information, weather conditions, and speckle noises. Furthermore, state-of-the-art deep learning models may be constrained by data extensibility, feature transferability, model scalability, and task producibility. This manuscript presents an integrated GeoAI framework that enhances foundation models for efficient water body extraction with multi-modal remote sensing images. The proposed framework consists of a data augmentation module tailored for optical and synthetic aperture radar (SAR) remote sensing images, as well as extraction modules augmented by three popular foundation models, namely SAM, SAMRS, and CROMA. Specifically, optical and SAR images are preprocessed and augmented independently, encoded through foundation model backbones, and subsequently decoded to generate water body segmentation masks under single-modal and multi-modal settings.