University of Twente Student Theses
Deep learning-based Digital Surface Model (DSM) generation using SAR image and building footprint data
Abdela, Nesredin (2023) Deep learning-based Digital Surface Model (DSM) generation using SAR image and building footprint data.
|
PDF
3MB |
| Abstract: | Digital Surface Models (DSMs) are crucial in urban planning, environmental monitoring, and disaster management. Although methods like stereo-photogrammetry, LiDAR, and InSAR have conventionally been used for DSM generation. Recently deep learning has been used to fill the gap between automating tasks and the need for accurate information for real-world applications. Synthetic Aperture Radar (SAR) imagery, with its unique attributes like side-looking geometry, all-weather operability, and day-and-night acquisition capabilities, present a unique opportunity for DSM reconstruction. This study has two main objectives: to assess the feasibility of deep learning for single SAR image-based DSM estimation and to explore enhancing DSM reconstruction accuracy in urban buildings by incorporating building footprint data. RADARSAT-2, TerraSAR-X SAR images, building footprint and ground truth DSM data were used. A fully convolutional neural network architecture with encoders and decoders sub-component is used for DSM estimation. The models are trained and evaluated on both single SAR images and combined SAR-image-building-footprint data using standard regression quality metrics and a structural similarity index (SISM). Results demonstrate deep learning's viability in DSM reconstruction using single SAR images, despite challenges in geometry, such as tilted building appearances and underestimation in shadow regions unseen by the sensor. Integration of building footprint data led to improved accuracy by addressing challenges faced in a single SAR model and produced well-defined building boundaries. Comparative analysis across RADARSAT-2 and TerraSAR-X datasets demonstrates competitive performance across varying spatial resolutions. Additionally, the performances of trained models were evaluated using cross-dataset and PAZ datasets promising inference was shown when using SAR and building footprint data. While accuracy from single SAR image-based predictions was limited, the trained model showed robustness when using both input data and SAR images acquired with similar frequency and looking direction. Future exploration involves adapting the model for diverse SAR datasets and data augmentation. This research demonstrates deep learning's capability in DSM reconstruction within urban settings and introduces integrating building footprint data to enhance estimation accuracy. The implications include urban planning, environmental assessment, and disaster management with a 2-3 meter RMSE accuracy level. Advanced co-registration techniques and diverse datasets are suggested for future work to enhance model performance across diverse land cover types and transferability to other SAR datasets, as a result broadening its applicability. Keywords: Digital Surface Model (DSM), Synthetic Aperture Radar (SAR), Deep Learning, Remote Sensing, Height Estimation. |
| Item Type: | Essay (Master) |
| Faculty: | ITC: Faculty of Geo-information Science and Earth Observation |
| Subject: | 54 computer science |
| Programme: | Geoinformation Science and Earth Observation MSc (75014) |
| Link to this item: | https://purl.utwente.nl/essays/97348 |
| Export this item as: | BibTeX EndNote HTML Citation Reference Manager |
Show download statistics for this publication
Show download statistics for this publicationRepository Staff Only: item control page
