Research outputs from our ANRF-funded LHFIC-EARTH project
IEEE M2GARSS 2026 Published 🏆
Extracting region-of-interest (ROI) from very high-resolution remote sensing (RS) imagery is difficult due to irregular object boundaries, multi-scale spatial variations, and complex backgrounds—difficulties that are intensified by the extensive manual annotation required to generate high-quality training labels. To mitigate the annotation burden, a solution is to exploit the unlabeled images along with limited labeled images. In this paper, we propose a novel boundary-aware semi-supervised adversarial learning for region of interest extraction (BROIE) from RS images with limited number of labeled samples. To accomplish this, BROIE framework utilizes a single end-to-end UNet-based generator equipped with an EfficientNet encoder, UNet decoder, and a segmentation and edge refinement module (SAER) to perform both pseudo-label generation and ROI prediction. Dynamic tau-based confidence filtering selects reliable pseudo-labels from weakly annotated images, while adversarial training between the generator and a patch discriminator encourages reliable mask generation. Experiments on WHU building, EORSSD, and Vaihingen image benchmark datasets demonstrate improved ROI extraction in scarce-label scenarios, highlighting the robustness of the proposed model across diverse conditions.
ACRS 2026 Accepted 🏆
Extracting building footprints from satellite imagery is a challenging task due to ambiguous structural boundaries, high annotation costs, and the fact that buildings typically occupy only a small fraction of an image scene. Although transformer-based building extraction methods capture global context effectively, they often blur object boundaries—the very regions where precision is most critical for downstream urban analysis. To address this limitation, the proposed weakly supervised GeoRectNet model integrates an explicit edge branch alongside a standard vision transformer (ViT) encoder. The proposed convolutional neural network (CNN)-based edge branch preserves fine, high-frequency boundary details that transformers often smoothen out. A gated attention mechanism employed in the decoder controls how boundary signals merge into the extraction output. The loss function is designed to enhance geometric quality through region supervision, boundary refinement, distance-based regularization, and consistency constraints, with the goal of improving both semantic accuracy and structural fidelity. Training runs through a three-stage weakly supervised pipeline: pseudo-label generation, refinement and edge-safe fine-tuning designed for annotation budgets of 10% or less. Benchmarking on the WHU Building Dataset demonstrates a 65.19% mIoU and 71.61% Dice score, outperforming the ViT baseline by +3.02% mIoU and approaching fully supervised models (U-Net, DeepLabV3+) while using one-tenth of the annotations. Additional evaluations on the INRIA dataset confirm strong cross-domain generalizability, producing clean, vector-ready polygons for downstream GIS integration.