Hybrid ViT-UNet Framework for Accurate River Segmentation and Buffer Zone Mapping in High-Resolution Satellite Imagery

Journal of Engineering Issues and Solutions Land Cover Segmentation from Satellite Imagery Using U-Net with Custom Loss and Morphological Postprocessing

Enhancing Surface Water Mapping in the Mekong River Basin using Recursive Classification on Sentinel-2 Time-Series

Land Cover Segmentation from Satellite Imagery Using U-Net with Custom Loss and Morphological Postprocessing

Land cover classification is a fundamental task in remote sensing with direct applications in environmental monitoring, urban planning, and disaster management. In this project, we design and implement a semantic segmentation pipeline using a modified U-Net architecture from scratch for pixel-level classification of satellite images. The network was trained on the DeepGlobe Land Cover Classification dataset with significant attention to model generalization and class imbalance through the integration of a custom composite loss function that combines Focal Tversky Loss and Weighted Categorical Crossentropy. Data augmentation techniques and per-class weighting were used to address class imbalance, especially in underrepresented classes such as barren land and rangeland. Postprocessing techniques like morphological operations and class-wise median filtering further refined the segmentation outputs. The model achieved a mean Intersection-over-Union (IoU) of 0.67, with notable performance across dominant land types. A full Streamlit-based web interface was developed to enable real-time visualization and interaction with the model’s predictions. Additionally, we initiated a fine-tuning phase targeting Nepal- specific satellite imagery to adapt the model to the country's unique geography. A major ongoing challenge involves curating a clean and labeled dataset of Nepalese satellite patches, which is critical for domain adaptation and achieving optimal segmentation performance on local data. The project demonstrates a full end-to-end pipeline from model design to web deployment and sets a foundation for scalable geospatial AI applications tailored for Nepal.

AS-UNet: A Lightweight U-Net with Asymmetric Strip Attention and Joint Gating for RGB Optical Landslide Segmentation

Accurate delineation of landslides in RGB optical remote sensing imagery supports rapid disaster mapping and post-event assessment. This remains difficult because landslides are often small and irregular, resemble bare soil or disturbed vegetation, and acquire blurred boundaries when images are resized. We developed AS-UNet, a lightweight U-Net variant with three targeted modifications. The Asymmetric Strip Attention Module uses horizontal and vertical depthwise strip convolutions with parallel channel-spatial reweighting to capture anisotropic landslide morphology. The Channel-Spatial Joint Gate uses decoder semantics to filter selected skip connections while retaining channel-specific spatial responses. The Poly-Harmonized Gradient Dice Loss (PGD Loss) combines pixel-wise, region-overlap, gradient-density, and probability-regularization terms for imbalanced segmentation. At 128 × 128 input resolution, AS-UNet achieved a best-validation IoU of 80.46 ± 0.03% and an independent-test IoU of 77.82 ± 0.32% across three random seeds. AS-UNet contains 8.634 M parameters and processed 380.79 frames per second on the reported hardware. These results indicate a favorable balance between segmentation accuracy and computational efficiency for RGB optical landslide mapping.

Comparison of U-Net and Attention U-Net for Binary Flood Segmentation from UAV Imagery

Flood disasters consistently cause massive damage every year, making rapid mapping of affected areas crucial for coordinating emergency aid. The use of unmanned aerial vehicles (UAVs) offers a practical solution to obtain high-resolution aerial imagery, but manually identifying flood areas from hundreds of images remains time-consuming. This study analyzes and compares two deep learning segmentation architectures, U-Net and Attention U-Net, for automatic flood area detection from UAV RGB images. Both models were trained using 290 image-mask pairs from a public dataset, with a split of 70% for training, 10% for validation, and 20% for testing. Images were processed at a resolution of 256×256 pixels, normalized to the range [0,1], and augmented with horizontal flipping, brightness adjustment, and affine transformations. Attention U-Net enhances the standard U-Net structure by adding attention gates to all skip connections in the decoder to suppress irrelevant background features. Both models were evaluated across five independent training runs using different random seeds to assess result robustness. Across these runs, Attention U-Net achieved a marginally higher mean IoU (77.11% ± 0.76) and Dice/F1 (87.07% ± 0.49) compared to the U-Net baseline (IoU: 76.97% ± 0.54; Dice/F1: 86.98% ± 0.34), but a paired t-test revealed that these differences were not statistically significant (IoU: p = 0.77; Dice/F1: p = 0.77). These results suggest that, on this dataset, attention gates do not provide a measurable advantage over the standard U-Net architecture, establishing both as comparable practical baselines for future flood mapping research.

Landslide identification method based on enhanced DeepLabv3+ architecture: joint optimization of bottleneck attention and dual-module graph convolution