Publicação
Development of a Deep Learning-Based Border Monitoring Tool Using Sentinel-2 Imagery
| Resumo: | Effective border monitoring in remote deserts is challenging because off-road vehicle tracks are faint and easily confused with natural terrain patterns. This thesis develops a deep-learning workflow for vehicle-track detection from Sentinel-2 multispectral imagery, demonstrated on a case study along the Chad-Libya border. A preprocessing pipeline was designed to create a 7-channel multispectral input from selected Sentinel-2 bands and track-enhancing indices. Two semantic segmentation models-U-Net and SegFormer were trained to predict binary track masks, using Focal Tversky loss to stabilize learning on thin, rare track pixels. Generalization was evaluated by training on AOI1 and testing on AOI2 across different dates, followed by targeted AOI2 fine-tuning to assess recoverability when transfer degrades. Results indicate that SegFormer generalizes more reliably to AOI2 than U-Net, suggesting that wider scene context improves detection of fragmented tracks. The experiments also show that strong domain shift can cause overly conservative, near-empty predictions; however, this failure is recoverable through lightweight AOI2 fine-tuning. After targeting AOI2 fine-tuning, the best SegFormer configuration achieved F1 = 0.6896 and IoU = 0.5263 on AOI2 (Tile 2). SegFormer achieves the best balance between detecting true tracks and limiting false alarms. This thesis contributes a practical multispectral workflow for scalable track-likelihood mapping, a controlled CNN-versus-transformer comparison under cross-AOI transfer, and an ablation-driven justification for an imbalance-aware objective suited to monitoring. Future work will explore Sentinel-1 SAR and SAR-optical fusion including incidence-angle effects, temporal consistency for change-driven monitoring, and broader evaluation across regions and seasons. |
|---|---|
| Autores principais: | Ambalavanan, Gomathy |
| Assunto: | Border monitoring Sentinel-2multispectral remote sensing off-road vehicle track detection semantic segmentation deep learning U-Net SegFormer class imbalance Focal Tversky loss domain shift transfer learning |
| Ano: | 2026 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso embargado |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | Effective border monitoring in remote deserts is challenging because off-road vehicle tracks are faint and easily confused with natural terrain patterns. This thesis develops a deep-learning workflow for vehicle-track detection from Sentinel-2 multispectral imagery, demonstrated on a case study along the Chad-Libya border. A preprocessing pipeline was designed to create a 7-channel multispectral input from selected Sentinel-2 bands and track-enhancing indices. Two semantic segmentation models-U-Net and SegFormer were trained to predict binary track masks, using Focal Tversky loss to stabilize learning on thin, rare track pixels. Generalization was evaluated by training on AOI1 and testing on AOI2 across different dates, followed by targeted AOI2 fine-tuning to assess recoverability when transfer degrades. Results indicate that SegFormer generalizes more reliably to AOI2 than U-Net, suggesting that wider scene context improves detection of fragmented tracks. The experiments also show that strong domain shift can cause overly conservative, near-empty predictions; however, this failure is recoverable through lightweight AOI2 fine-tuning. After targeting AOI2 fine-tuning, the best SegFormer configuration achieved F1 = 0.6896 and IoU = 0.5263 on AOI2 (Tile 2). SegFormer achieves the best balance between detecting true tracks and limiting false alarms. This thesis contributes a practical multispectral workflow for scalable track-likelihood mapping, a controlled CNN-versus-transformer comparison under cross-AOI transfer, and an ablation-driven justification for an imbalance-aware objective suited to monitoring. Future work will explore Sentinel-1 SAR and SAR-optical fusion including incidence-angle effects, temporal consistency for change-driven monitoring, and broader evaluation across regions and seasons. |
|---|