Simple is best: A single-CNN method for classifying remote sensing images

Abstract

Recently, researchers have proposed a lot of methods to boost the performance of convolutional neural networks (CNNs) for classifying remote sensing images (RSI). However, the methods' performance improvements were insignificant, while time and hardware costs increased dramatically due to re-modeling. To tackle this problem, this study sought a simple, lightweight, yet more accurate solution for RSI semantic classification (RSI-SC). At first, we proposed a set of mathematical derivations to analyze and identify the best way among different technical roadmaps. Afterward, we selected a simple route that can significantly boost a single CNN's performance while maintaining simplicity and reducing costs in time and hardware. The proposed method, called RE-EfficientNet, only consists of a lightweight EfficientNet-B3 and a concise training algorithm named RE-CNN. The novelty of RE-EfficientNet and RE-CNN includes the following: First, EfficientNet-B3 employs transfer learning from ImageNet-1K and excludes any complicated re-modeling. It can adequately utilize the easily accessible pre-trained weights for time savings and avoid the pre-training effect being weakened due to re-modeling. Second, RE-CNN includes an effective combination of data augmentation (DA) transformations and two modified training tricks (TTs). It can alleviate the data distribution shift from DA-processed training sets and make the TTs more effective through modification according to the inherent nature of RSI. Extensive experimental results on two RSI sets prove that RE-EfficientNet can surpass all 30 cutting-edge methods published before 2023. It gives a remarkable improvement of 0.50% to 0.75% in overall accuracy (OA) and a 75% or more reduction in parameters. The ablation experiment also reveals that RE-CNN can improve CNN OA by 0.55% to 1.10%. All the results indicate that RE-EfficientNet is a simple, lightweight and more accurate solution for RSI-SC. In addition, we argue that the ideas proposed in this work about how to choose an appropriate model and training algorithm can help us find more efficient approaches in the future.