Comparison of Deep Learning Models with Different Backbones for Building Footprints Extraction in Dense Residential Areas of Bhaktapur

Abstract

The rapid advancements of Artificial Intelligence, particularly deep learning, has enhanced features extraction mainly building footprints. However, models trained on developed countries struggle to perform better while testing on the datasets of high-density residential areas of developing countries like Nepal. This research aims to improve the performance of the models by developing and utilizing a region-specific dataset for Bhaktapur. Initially, models like Unet, PSPNet and LinkNet with different backbone architectures like resnet18, resnet34, resnet50, and vgg19 were trained on the Massachusetts dataset and the performance was poor when tested with the dense residential areas of Bhaktapur datasets. To address this issue, new datasets of Bhaktapur were introduced for high density residential area where houses are closely attached. The datasets were prepared by digitizing each house on the high resolution orthomosaics which was then converted to mask. Subsequently, the orthomosaic was patches into 300 x 300 with the corresponding mask. These patches were split into training, validation and testing datasets. Models with different backbones were trained with custom datasets applying data augmentation techniques, including random clipping of 256 x 256, flipping and rotation to prevent overfitting and make the model more robust to variations in real world data. The models with different backbones were validated and tested, and the best performance was with LinkNet model with vgg19 backbone with an IoU of 94.29% and F1 Score of 98.29%, demonstrating good results for dense residential areas. This study highlights the importance or need of region based custom datasets to improve the accuracy of deep learning segmentation models for building extractions, mainly on unique urban structure which can be useful for urban spatial planning, and disaster risk management and monitoring.