Researchers have developed LC-SLab, a novel deep learning framework designed for large-scale land cover classification using satellite imagery and sparse in-situ labels. This object-based approach assigns labels to coherent regions rather than individual pixels, aiming to produce more spatially coherent maps. The framework supports both input-level aggregation via graph neural networks and output-level aggregation through postprocessing semantic segmentation results, incorporating features from pre-trained networks to enhance performance on smaller datasets. Evaluations using Sentinel-2 imagery and LUCAS labels demonstrate that LC-SLab configurations can match or surpass the accuracy of pixel-wise models while significantly reducing map fragmentation. AI
IMPACT This framework could improve the accuracy and coherence of land cover maps derived from satellite data, benefiting various Earth science applications.
RANK_REASON The cluster describes a new research paper detailing a novel deep learning framework for a specific scientific application. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Earth science
- graph neural networks
- Johannes Leonhardt
- LC-SLab
- LUCAS
- semantic segmentation
- Sentinel-2
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
