M.S. Defense: Aniruddh Saxena
Thursday, November 30, 2023
12:30 PM
Location: 347 Avery Hall
"SC-FUSE: A Feature Fusion Approach for Unpaved Road Detection from Remotely Sensed Images"
Road network extraction from remote sensing imagery is crucial for numerous applications, ranging from autonomous navigation to urban and rural planning. A particularly challenging aspect is the detection of unpaved roads, often underrepresented in research and data. These roads display variability in texture, width, shape, and surroundings, making their detection quite complex. This thesis addresses these challenges by creating a specialized dataset and introducing the SC-Fuse model.
Our custom dataset comprises high resolution remote sensing imagery which primarily targets unpaved roads of the American Midwest. To capture the diverse seasonal variation and their impact, the dataset includes images from different times of the year, capturing various weather conditions and offering a comprehensive view of these changing conditions.
To detect roads from our custom dataset we developed SC-Fuse model, a novel deep learning architecture designed to extract unpaved road networks from satellite imagery. This model leverages the strengths of dual feature extractors: the Swin Transformer and a Residual CNN. By combining features from these, SC-fuse captures the local as well as the global context of the images. The fusion of these features is done by a Feature Fusion Module which uses Linear Attention Mechanism, to optimize the computational efficiency. A LinkNet based decoder is used to ensure precise road network reconstruction.
The evaluation of SC-Fuse model is done using various metrics, including qualitative visual assessments, to test its effectiveness in unpaved road detection. We compare its performance against each of its constituent components, offering insights into the advantages of the fused approach. Additionally, we dive into an analysis of areas where the model has difficulty detecting the roads and explore potential reasons for inaccuracies and identify areas for future improvement.
Committee:
Ashok Samal, Advisor
Stephen Scott
Cody Stolle