Niloofar Azad | Searcharxiv

arXiv2026

Arctic-SAR: A Cross-Sensor RADARSAT Constellation Mission and Sentinel-1 Dataset and Fusion Benchmark for Multi-Task Sea Ice Mapping

Current sea ice mapping benchmarks are mostly built around single-sensor SAR products, limiting the evaluation of cross-sensor variability and the effectiveness of combining observations from different SAR platforms. Models trained on such datasets may also be sensitive to sensor-specific artifacts, radiometric characteristics, and acquisition conditions, which can reduce robustness and limit generalization. To address these limitations, this study introduces Arctic-SAR, a cross-sensor SAR dataset for Arctic sea ice mapping, constructed by matching RADARSAT Constellation Mission (RCM) dual-polarization imagery with Sentinel-1 scenes from the AI4Arctic Sea Ice Dataset. The dataset consists of 221 co-registered scenes and supports multi-task prediction of sea ice concentration (SIC), stage of development (SOD), and floe size (FLOE). In addition to the dataset, we establish a benchmark for cross-sensor SAR sea ice mapping by quantifying the contribution of RCM imagery and evaluating single-sensor baselines alongside representative input-, feature-, and decision-level fusion strategies within a consistent multi-task U-Net framework. Across three independent runs, the test results show that the effectiveness of cross-sensor fusion is strongly dependent on the fusion strategy and prediction task. Decision-level fusion achieves the highest mean combined score of $77.28\pm0.29\%$, essentially matching the Sentinel-1-only baseline at $77.26\pm0.13\%$, while achieving the highest SIC $R^2$ score. These findings show that Sentinel-1 and RCM provide task- and class-dependent information, but that simple joint cross-sensor fusion does not consistently improve generalization over a strong single-sensor baseline. Arctic-SAR therefore provides a benchmark for cross-sensor SAR fusion research in automated Arctic sea ice mapping.

eess.IV↗

arXiv2026

Sparse Data Tree Canopy Segmentation: Fine-Tuning Leading Pretrained Models on Only 150 Images

Tree canopy detection from aerial imagery is an important task for environmental monitoring, urban planning, and ecosystem analysis. Simulating real-life data annotation scarcity, the Solafune Tree Canopy Detection competition provides a small and imbalanced dataset of only 150 annotated images, posing significant challenges for training deep models without severe overfitting. In this work, we evaluate five representative architectures, YOLOv11, Mask R-CNN, DeepLabv3, Swin-UNet, and DINOv2, to assess their suitability for canopy segmentation under extreme data scarcity. Our experiments show that pretrained convolution-based models, particularly YOLOv11 and Mask R-CNN, generalize significantly better than pretrained transformer-based models. DeeplabV3, Swin-UNet and DINOv2 underperform likely due to differences between semantic and instance segmentation tasks, the high data requirements of Vision Transformers, and the lack of strong inductive biases. These findings confirm that transformer-based architectures struggle in low-data regimes without substantial pretraining or augmentation and that differences between semantic and instance segmentation further affect model performance. We provide a detailed analysis of training strategies, augmentation policies, and model behavior under the small-data constraint and demonstrate that lightweight CNN-based methods remain the most reliable for canopy detection on limited imagery.

cs.CV↗