MS-EfficientNet: Hierarchical Multi-Scale Feature Fusion for Land Cover Classification

Abstract

Land cover classification from remote sensing data is an essential task in Earth observation, playing a crucial role in various applications, including environmental monitoring, urban planning, agriculture, and natural resource management. This paper evaluates four advanced deep learning architectures: ResNet50, EfficientNetB3, ConvNeXt, and Vision Transformer (ViT) on EuroSAT dataset with consistent experimental setups. From comparative analysis, EfficientNetB3 yields the highest result compared to other baseline models; therefore, an advanced MS-EfficientNetB3 architecture is proposed to improve land cover classification performance. In particular, the proposed framework consists of two stages: multi-scale feature extraction and feature fusion stage that can extract not only the fine-grained local feature but also high-level semantic information in remote sensing images. Label smoothing and cosine annealing are used to improve model generalization and optimization. Experiments show that the proposed architecture yields a classification accuracy of 99.37% which surpasses the baseline EfficientNetB3 model. The ablation study proves the effectiveness of each part of the proposed framework, especially the role of multi-scale feature learning.

Mehr zum Titel