Transforming Low-Resolution CORINE Data into High-Resolution Landscape Maps with Semi-Supervised Deep Learning

Large-scale, high-resolution landscape mapping with precise classification is essential for understanding, managing, and protecting Earth's ecosystems. It provides granular spatial and thematic insights into land cover and land-use dynamics, allowing for a better representation of complex landscapes with multiple classes. By preserving fine-scale heterogeneity, such mapping enables the identification of subtle yet ecologically significant patterns, including habitat fragmentation, biodiversity hotspots, and land-use transitions. Despite the availability of several high-resolution global land cover products, there is a significant lack of detailed class information in these datasets. The existing classes are often too general and fail to accurately represent the inherent heterogeneity of landscapes. However, this task remains challenging due to intricate ground features, diverse landforms, and the limited availability of accurate training labels across extensive geographic regions.

In this study, we employed an efficient weakly supervised deep learning architecture to enable large-scale, high-resolution land cover mapping with detailed class distinctions. This was achieved by utilizing widely accessible and publicly available satellite products and global land cover (GLC) data, with a focus on Brandenburg, a federal state of Germany. We used the CORINE Land Cover (CLC) 2018 dataset as a low-resolution land cover label, alongside nine bands from Sentinel-2 MSI data and two bands (VV and VH) from Sentinel-1 SAR data, all at a 10-meter spatial resolution, organized into 256x256 pixel patches. While the CORINE dataset offers rich class information with 44 thematic classes (28 for Brandenburg), its coarse resolution (100 m) limits its utility for large-scale analyses. To address this, we enhanced the resolution of the dataset to 10 meters by integrating satellite data from hybrid sources. Additionally, we incorporated high-resolution global land cover databases, such as Dynamic World V1, into the model’s loss function to guide the generation of high-resolution data products while maintaining the same number of classes as CORINE. This framework addressed label noise resulting from the resolution mismatch between images and labels by combining a resolution-preserving CNN branch, a Transformer branch, a weakly supervised module, and a self-supervised loss function, enabling the automatic refinement of high-resolution land cover results without manual annotations.

Our results, obtained after running 30 epochs in the Google Colab Pro Python environment with a limited A100 GPU (~40 GB), show promising outcomes, with a gradual decrease in loss. The predicted validation data, aggregated into broader class categories, were compared with the Dynamic World dataset, yielding a match of 68%. Specific classes, such as cropland, vegetation, and grassland, demonstrated strong performance, with accuracy scores of 84%, 66%, and 55%, respectively. This framework generates high-resolution, detailed landscape maps with rich class information from accessible global land cover products, all without the need for manual annotation. It can also be applied across Europe, as the CORINE data covers the entire continent. While these results are encouraging, we are confident that further analyses, including additional training with more epochs and data, will improve performance even further.