Urban Land-Use Mapping and Classification using Deep Convolutional Neural Network with Multispectral Drone Imagery

Urban land use mapping is key for efficient urban planning, hazard management etc. Moreover, the growing sophistication of drones (UAV) helps to capture high resolution remote sensing data at low cost compared to conventional satellite and airborne systems. Nowadays, advanced artificial intelligence, especially use of deep convolutional neural networks (DCNN) has showcased notable capabilities in land use mapping because of its efficiency and cost effectiveness. Thus, in this paper, DCNN-based U-Net architecture is used for urban land use mapping on dataset of multispectral images captured from different areas of National Institute of Technology Kurukshetra, India using Drone. Further, we employed various pre-trained models like variants of ResNet, Squeeze-Excitation ResNet and VGG16 as the backbone of U-NET for efficient and faster training. Simulation evaluation highlight that SE-ResNet-34 consistently outperformed other models in complex scenarios by providing 89% accuracy. Further, SE-ResNet-18 and ResNet-18 are the most computationally efficient, offering a favorable trade-off between accuracy and processing time. Overall, results demonstrate the great potential of developed models in terms of accuracy, speed and memory usage, makes them suitable for land-use classification.