LC-SLab -- An object-based deep learning framework for large-scale land cover classification from satellite imagery and sparse in-situ labels

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Geospatial AI & Remote Sensing · Depth: Expert, extended

Summary

LC-SLab is a novel deep learning framework designed for large-scale land cover classification using satellite imagery and sparse in-situ labels. It addresses the issue of fragmented and noisy predictions from traditional pixel-wise methods by employing object-based classification, which assigns labels to semantically coherent image regions and enforces a minimum mapping unit (MMU) to control fragmentation. The framework supports both input-level aggregation via graph neural networks and output-level aggregation through postprocessing semantic segmentation models. Additionally, it integrates features from a pre-trained network to enhance performance on smaller datasets. Evaluations using annual Sentinel-2 composites and sparse LUCAS labels demonstrate that LC-SLab's object-based methods achieve comparable or superior accuracy to pixel-wise models while generating significantly more coherent land cover maps. Input-level aggregation proves more robust with limited data, whereas output-level aggregation excels with larger datasets. Notably, several LC-SLab configurations surpass the accuracy of established third-party products like ESA WorldCover and ESRI Land Cover.

Key takeaway

For Machine Learning Engineers developing large-scale land cover maps with sparse in-situ labels, you should adopt object-based deep learning frameworks like LC-SLab. This approach significantly reduces map fragmentation and can match or exceed pixel-wise accuracy. Evaluate input-level aggregation for smaller datasets and output-level for larger ones, and integrate pre-trained features to enhance label efficiency and model performance, especially with limited data.

Key insights

Object-based deep learning improves land cover map coherence and accuracy, especially with sparse labels, by enforcing a minimum mapping unit.

Principles

• Object-based classification reduces fragmentation and improves semantic consistency.
• MMU acts as regularization, improving accuracy for less powerful models.
• Pre-trained features boost accuracy and label efficiency with sparse data.

Method

LC-SLab defines objects via unsupervised oversegmentation, optionally extracts features from a pre-trained model, then classifies using either input-level aggregation (GNNs on object graphs) or output-level aggregation (postprocessing pixel-wise semantic segmentation logits). Optimization uses partial cross-entropy loss.

In practice

• For small datasets, use input-level aggregation (GNNs).
• For larger sparse datasets, use output-level aggregation.
• Integrate pre-trained features to boost small model accuracy.

Topics

• Land Cover Classification
• Object-Based Image Analysis
• Sparse Label Learning
• Graph Neural Networks
• Satellite Imagery
• Minimum Mapping Unit

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• MapSR: Prompt-Driven Land Cover Map Super-Resolution via Vision Foundation Models — MapSR enables high-resolution land-cover mapping with minimal annotation and compute by decoupling supervision from training.
• LDFE: Laplacian Decoupled Feature Enhancement Block for Dual-Stream CNN-based RGB-IR Object Detection — The LDFE block enhances RGB-IR object detection by decoupling features into global and local components for specialized fusion and denoising.
• SparseGF: A Height-Aware Sparse Segmentation Framework with Context Compression for Robust Ground Filtering Across Urban to Natural Scenes — SparseGF improves ground filtering by compressing context and using height-aware loss to prevent tall object misclassification.
• Global Offshore Wind Infrastructure: Deployment and Operational Dynamics from Dense Sentinel-1 Time Series — A new global Sentinel-1 SAR dataset enables high-resolution monitoring of offshore wind infrastructure deployment and operations.
• Self-Supervised Tree-level Biomass Estimation in Urban Environments From Airborne LiDAR and Optical Observations — Self-supervised deep learning combined with LiDAR and orthophotography enables accurate, scalable urban tree biomass estimation.
• A Spatial-Spectral-Frequency Interactive Network for Multimodal Remote Sensing Classification — Integrating spatial, spectral, and frequency domain fusion enhances multimodal remote sensing classification, especially with limited labels.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.