DEEP LEARNING-BASED GEOSPATIAL INFRASTRUCTURE SEGMENTATION USING SATELLITE IMAGERY

Abstract

Satellite imagery has become an essential component of modern geospatial infrastructure monitoring systems, enabling large-scale analysis of urban environments for applications such as smart city planning, transportation management, disaster response, environmental monitoring, and infrastructure development.The main goal of remote sensing is to accurately identify the building and road segments from high resolution aerial images. Traditional image processing approaches and rule based segmentation methods generally do not perform well in extracting reliable information about complex urban areas since these are highly dependent on hand crafted features and manually defined heuristics. Deep learning techniques particularly encoder-decoder architectures, and transformer-based models have greatly enhanced semantic segmentation performance through providing an opportunity to automatically extract hierarchical features from the raw satellite imagery. This dissertation presents a deep learning architecture for extracting buildings and roads from satellite imagery with high accuracy. This proposed framework integrates both pipeline enhancements to enhance feature representation, segmentation accuracy, and computational efficiency. EfficientNet-V2 is used as the encoder backbone because of its ability to provide an optimal scale factor; also has efficient convolutional blocks and strong feature extraction capabilities. The U-Net decoder provides spatial detail through the use of skip connections and progressive upsampling operations. The proposed framework was tested using two datasets: the Massachusetts Buildings dataset and the Massachusetts Roadways dataset. These datasets contain high-resolution aerial imagery along with their respective ground truth segmentation masks. To further enhance the robustness and generality of the model, extensive pre-processing and data augmentation techniques were applied. Pre-processing included normalization, random cropping, flipping, rotating, scaling, and color transforming. Training the model utilized the previously identified hyperparameters and evaluation was conducted utilizing established segmentation metrics (Accuracy, Precision, Recall, F1-Score, IoU). Experimental results demonstrate that the proposed EfficientNet-V2-based U-Net model outperformed all other CNN architectures as well as transformer-based segmentation frameworks in terms of segmentation accuracy. Also, the proposed model provided better boundary preservation; greater structural continuity; and extracted more meaningful semantic features than previous architectures at a computationally efficient level required for large-scale geospatial applications. Results derived in this study demonstrate that deep-learning based segmentation frameworks v can greatly enhance automated extraction of infrastructure from satellite imagery contributing to developing intelligent geospatial monitoring systems for smart city and urban analytic purposes.