Optimized Modified ResNet18: A Residual Neural Network for High Resolution
Songling Liu, Binxin Wang · 2024
High-resolution remote sensing images are invaluable across various fields, providing critical data for diverse applications. The effective analysis of these images hinges on accurate classification, a task for which convolutional neural networks (CNNs) have proven to be particularly adept, thanks to their efficiency and superior performance. However, challenges such as feature degradation in deeper network layers and insufficient training data can compromise classification accuracy. In this study, we introduce an optimized modified residual neural network model, OM ResNet18, which builds upon the established ResNet18 architecture. By leveraging transfer learning with the ImageNet dataset, we acquire pre-trained model weights that enhance the residual blocks within our model. We then train the optimized model using the AID dataset, subsequently comparing its performance against other pre-trained models, including AlexNet, VGG-16, and MobileNetV2, which also utilized transfer learning with ImageNet. Our experimental results highlight the superior performance of the proposed OM ResNet18 model. It achieved an accuracy (95.85%), precision (95.88%), recall (95.48%), F1-Score (95.60%), and kappa statistic (95.70%), surpassing the performance metrics of the other three neural networks. These findings underscore the effectiveness of our approach in improving the classification accuracy of high-resolution remote sensing images.