Enhancing land cover classification with data augmentation in multispectral aerial imagery

Land cover classification is a fundamental task in remote sensing and environmental monitoring. Data augmentation is a widely used strategy to improve machine learning performance, but it is predominantly applied to RGB imagery. In this study, we evaluate the impact of data augmentation not only on RGB but also on full multispectral data, including RGB, NIR, and elevation bands from the FLAIR dataset. The study compares augmentation applied to the full dataset versus RGB-only data, using U-Net with ResNet50 and DeepLabV3+ with ResNeXt50 in both 19-class and 5-class aggregated classification scenarios. Results show that applying augmentation to all spectral bands significantly improves precision, F1-score, and IoU, outperforming RGB-only augmentation. In limited-data “Toy” scenarios, IoU increased by 13–15 percentage points with multispectral augmentation, compared to 11–12 points for RGB-only. DeepLabV3+ consistently surpasses U-Net, and class aggregation further reduces confusion between spectrally similar categories (F1-score up to 0.92, IoU up to 0.85). These findings highlight the importance of extending data augmentation beyond RGB to exploit full spectral richness for robust high-resolution land cover classification.