Feature extraction of new-type urbanization from remote sensing images based on convolutional neural network

This study uses the Forest, Urban, and Suburban Land Use Dataset (FUSU) to improve the extraction accuracy of urbanization features from ultra-high-resolution remote sensing images. This study also intends to quantify the spatial correlation between these urbanization features and regional economic growth. Under standardized data preprocessing and unified training settings, this study compares four Convolutional Neural Network (CNN) models. The models include vanilla CNN, ResNet-18, DenseNet-121 and EfficientNet. Several urbanization indicators are extracted from remote sensing identification outputs. These indicators cover building density, road density, green space ratio and urban expansion rate. Combined with regional economic statistics, three spatial econometric models are adopted for subsequent analysis. The selected models are the spatial lag model, spatial error model and geographically weighted regression model. They are used to explore economic correlations and spatial heterogeneity. Experimental results verify that EfficientNet achieves optimal overall performance across all four data subsets. For the urban subset: accuracy = 0.92, precision = 0.89, recall = 0.91, F1-score = 0.90. For the forest subset: accuracy = 0.93, precision = 0.90, recall = 0.95, F1-score = 0.92. For the suburban subset: accuracy = 0.90, precision = 0.86, recall = 0.89, F1-score = 0.87. For the mixed subset: accuracy = 0.93, precision = 0.90, recall = 0.93, F1-score = 0.91. In regression evaluation, the mean square error (MSE) values of EfficientNet equal 0.030, 0.028, 0.040 and 0.033 for urban, forest, suburban and mixed subsets separately. Their corresponding coefficient of determination (R 2 ) values are 0.93, 0.95, 0.89 and 0.92 in sequence. Spatial econometric analysis shows that building density, road density, green space ratio, urban expansion rate and population density exert significant positive effects on regional gross domestic product (GDP). In conclusion, EfficientNet strikes a favorable trade-off between high classification accuracy and low computational overhead. Integrating this network with spatial econometric models can clarify intricate interactions between spatial urbanization signatures and regional economic development. The proposed framework supplies methodological references for urban layout design and coordinated regional development.