Generating national very high-resolution land cover product of France without any labels: A comparative study

Abstract

Generating a national very high-resolution (VHR) land cover product is crucial for various applications, including environmental monitoring and urban planning. However, creating such a product often requires a large amount of labeled data over a target area, which can be expensive and challenging. In tackling these challenges, this work introduces a comparative analysis of three label-free techniques, including source-domain pretraining, pseudo-labels, and unsupervised domain adaptation (UDA), for developing the French national VHR land cover product. Three label-free techniques leverage the recent OpenEarthMap datasets and employ an advanced segmentation model, a fully Transformer-based network (FT-UNetFormer). The evaluation of these methods utilized the reference offered by the French datasets: FLAIR. Results indicated an overall product accuracy ranging from 82.1% to 85.5%, with a mean intersection over union (mIoU) fluctuating between 57% and 59%. Notably, the highest accuracy was achieved for buildings, while the lowest accuracy was obtained for bareland. Among the three methods, source-domain pretraining demonstrated adequacy but yielded lower accuracy. UDA exhibited very high accuracy; however, it came with considerable computational complexity. The pseudo-labels methods were identified as a viable trade-off between accuracy and computational efficiency. Ultimately, we will release the products derived from the three label-free techniques. The open availability of these products can contribute significantly to informed decision-making and sustainable development across various sectors.