Engineering Applications of Urban Green Space Planning in Mountainous Areas: An Improved Structure-based RS Land Class Information Extraction Method for U-Net Networks

Land type information is essential for urban planning, ecological monitoring and management. To improve ground class information extraction, the depth separable convolution operation and Dropout layer are added to the U-Net network structure. Additionally, the ReLU and TReLU functions are combined to create a ReLU-TReLU hybrid activation function, which is then used to create a method for extracting ground class information from remote sensing images based on IU-Net. The outcomes of the experiment demonstrated the ReLU-TReLU hybrid activation function’s quick convergence speed and robust learning capacity. Using accuracy, precision of producer, user accuracy, commission errors, omission errors, overall classification accuracy and Kappa coefficient as evaluation indicators, IU-Net achieved the highest test scores of 96.1%, 98.62%, 98.32%, 0.30, 0.90, 96.82% and 0.95, respectively. IU-Net had the lowest mistake rate and maximum accuracy of information extraction when compared to other approaches. Additionally, it often ran in less than 5 min, which is quicker and offers technical support for the land categorization project for the development of hilly urban green space.