LSEVGG: An attention mechanism and lightweight-improved VGG network for remote sensing landscape image classification

Remote sensing landscape image classification is essential for environmental monitoring, land management, and ecological assessment, but presents critical challenges due to complex spatial distributions and high intra-class variability inherent in landscape scenes. Traditional deep convolutional neural networks, such as VGG16, though effective, are computationally intensive and unsuitable for deployment on resource-constrained platforms commonly used in landscape monitoring applications. In this paper, we propose LSEVGG, a novel and efficient CNN architecture that enhances the classic VGG structure through the integration of lightweight convolution techniques and channel attention mechanisms. Specifically, our approach incorporates depthwise separable convolutions and Squeeze-and-Excitation (SE) attention modules to create a model that is both computationally efficient and highly effective for landscape feature extraction. These modifications significantly reduce model complexity while enhancing the network’s ability to focus on key landscape feature regions and capture distinctive terrain characteristics. Experimental results on the NWPU-RESISC45 benchmark demonstrate that LSEVGG achieves a Top-1 classification accuracy of 82%, surpassing the original VGG16 by 17% and ResNet18 by 11%. The model exhibits particularly strong performance in identifying natural landscape categories, achieving 88%–92% accuracy for classes with uniform spatial patterns such as forests, beaches, and meadows, where the SE attention mechanism effectively captures distinctive textural and spatial features characteristic of different landscape types. Moreover, LSEVGG reduces the number of parameters by 99.8% and floating-point operations (FLOPs) by 88% compared to VGG16, highlighting its suitability for real-time and edge computing applications in landscape monitoring systems. These results confirm that our LSEVGG model offers a practical balance between classification accuracy and computational efficiency, making it well-suited for real-world remote sensing landscape analysis tasks.