Optimizing deep neural networks for high-resolution land cover classification through data augmentation

Abstract

This study presents an innovative approach to high-resolution land cover classification using deep learning, tackling the challenge of working with an exceptionally small dataset. Manual annotation of land cover data is both time-consuming and labor-intensive, making data augmentation crucial for enhancing model performance. While data augmentation is a well-established technique, there has not been a comprehensive and comparative evaluation of a wide range of data augmentation methods specifically applied to land cover classification until now. Our work fills this gap by systematically testing eight different data augmentation techniques across four neural networks (U-Net, DeepLabv3?+, FCN, PSPNet) using 25 cm resolution images from Cantabria, Spain. In total, we generated 19 distinct training sets and trained and validated 72 models. The results show that data augmentation can boost model performance by up to 30%. The best model (DeepLabV3?+?with flip, contrast, and brightness adjustments) achieved an accuracy of 0.89 and an IoU of 0.78. Additionally, we utilized this optimized model to generate land cover maps for the years 2014, 2017, and 2019, validated at 580 samples selected based on a stratified sampling approach using CORINE Land Cover data, achieving an accuracy of 87.2%. This study not only provides a systematic ranking of data augmentation techniques for land cover classification but also offers a practical framework to help future researchers save time by identifying the most effective augmentation strategies for this specific task.