Lightweight Convolutional Neural Network Based on Transfer Learning for Remote Sensing Image Classification

Remote sensing image classification plays a crucial role in environmental monitoring and land cover and land use studies. This paper proposes a lightweight convolutional neural network method based on transfer learning, utilizing MobileNetV2 as the base model and conducting experiments on the UC Merced remote sensing image dataset. By utilizing depthwise separable convolutions and inverted residual structures, MobileNetV2 achieves a significant reduction in the number of model parameters (2.25 M) and computational cost (0.33 G FLOPs) while retaining strong feature extraction capabilities. Experimental results demonstrate that the proposed method achieves a high classification accuracy (0.8714 ± 0.0100) on the UC Merced dataset, along with a notable decrease in computational resource usage and training time (4906.84 seconds). When compared to DenseNet121, ResNet18, ConvNeXt-Tiny, and GoogLeNet, MobileNetV2 demonstrates better performance in precision, recall, F1-score, and Kappa coefficient. This study provides an efficient and lightweight solution for remote sensing image classification tasks, with practical application potential.