Satellite Image Segmentation to Identify Roads using U-Net Model
We test the accuracy and efficiency of Unet model on the dataset from Deep-Globe-Road-Extraction Challenge 2018. In disaster zones, especially in developing countries, maps and accessibility information are crucial for crisis response. Deep�Globe-Road-Extraction Challenge poses the challenge of automatically extracting roads and street networks from satellite images. Contains RGB colour code mappings for different classes in labels. The labels in this dataset have 2 classes: 'road' & 'background' • The training data for Road Challenge contains 6226 satellite imagery in RGB, size 1024x1024. • The imagery has 50cm pixel resolution, collected by DigitalGlobe's satellite. • The dataset contains 1243 validation and 1101 test images (but no masks). The proposed model was implemented using Keras framework and optimized by the Adam optimizer and a binary cross-entropy is used as the loss function.
In conclusion, this project successfully employed the UNet architecture for road segmentation from satellite imagery using the satellite-road-image�extraction dataset. With the proposed methodology, an impressive accuracy of 96% was achieved on the test set, while the validation accuracy of 95% further validated the robustness of the proposed method. The results obtained demonstrate the efficacy of the proposed UNet architecture for road segmentation from satellite imagery, which can be applied in various domains, including urban planning, transportation, and disaster management. Overall, this research provides valuable insights into the development of accurate and efficient road segmentation models for satellite imagery, which can contribute to the advancement of various fields of research and applications