Advancements in Image Recognition for Urban Land Use: Multi-Scale CNN Extraction

This research introduces a specialized Convolutional Neural Network (CNN) architecture designed for effective feature extraction and classification of urban land use, with a specific emphasis on identifying buildings and roads using high-resolution Unmanned Aerial Vehicle (UAV) imagery. The CNN model consists of a meticulously crafted sequence of layers, including an input layer, four convolutional layers (which include a multi-scale convolutional layer), four pooling layers, and one fully connected layer. For the multi-scale convolutional layer, three sets of filters are employed: 128 filters with a size of 3x3 and a stride of 1, 128 filters with a size of 5x5 and a stride of 1 (with padding of 2), and 128 filters with a size of 7x7 and a stride of 1 (with padding of 3). Each of the convolutional layers is followed by a ReLU activation function. To enable feature extraction across various scales and to capture global dependencies, depth-multi-scale convolutional operations are initially applied to the input feature maps. This approach not only enhances contextual information modeling but also significantly reduces computational costs. In experimental evaluation, the proposed CNN model is assessed using UAV images from Fengshan District, Kaohsiung City. The results underscore the model's effectiveness in accurately detecting buildings and roads. The trained network achieves an impressive accuracy of 92% in distinguishing between these urban structures. This study narrows its focus to the detection of buildings and roads, contributing to the fields of urban area cartography, traffic planning, and land resource management.