Toward Seasonal Guidelines for Robust Deep-Learning Sentinel-2 Building Detection in Different Area Types

2024-01-29 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Advanced, extended

Summary

This study introduces a Sentinel-2 building-detection framework designed to quantify the effects of seasonality and built-up environment heterogeneity on deep learning model performance. Utilizing a multi-temporal Sentinel-2 dataset from the Warsaw region (2019-2025) and official Polish topographic database (BDOT10k) ground truth, the research systematically evaluated U-Net and DeepLabV3+ convolutional segmentation backbones. U-Net consistently outperformed DeepLabV3+, achieving higher IoU, F1-score, and Balanced Accuracy. While L2A atmospherically corrected products showed a marginal advantage over L1C, the study found peak classification accuracy from May to September, with significant degradation in winter months (December-January) due to low sun angles and snow cover. Performance also varied substantially across settlement types, with industrial and dense urban areas yielding higher accuracy than suburban and dispersed rural regions. The framework also produces 2.8m resolution output from 10m input.

Key takeaway

For Machine Learning Engineers developing building detection systems with Sentinel-2 imagery, you should prioritize using L2A products and imagery acquired between May and August. This period offers optimal conditions for robust classification, especially when employing a Collective Summer Model to minimize phenological differences. Be aware that performance significantly degrades in winter months and for dispersed rural areas, necessitating careful model selection or temporal composites to avoid overfitting to transient conditions like snow cover.

Key insights

Seasonal and built-up environment variations significantly impact deep learning model performance for Sentinel-2 building detection.

Principles

• U-Net architectures generally outperform DeepLabV3+ for urban building classification.
• Building detection accuracy is highest in summer (May-September), lowest in winter.
• Dense urban areas yield higher detection accuracy than low-density or rural settlements.

Method

The study used scene-specific fine-tuning of U-Net and DeepLabV3+ on a multi-temporal Sentinel-2 dataset, followed by cross-temporal inference to assess seasonal and typological transferability.

In practice

• Use U-Net for Sentinel-2 building detection.
• Classify L2A data for better stability across dates.
• Implement a Collective Summer Model (May-August) for robust annual mapping.

Topics

• Sentinel-2 Imagery
• Building Detection
• Semantic Segmentation
• Deep Learning
• Seasonal Variability
• Urban Mapping

Code references

• microsoft/GlobalMLBuildingFootprints
• sentinel-hub/sentinel2-cloud-detector

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

Related on AIssential

• Toward Seasonal Guidelines for Robust Deep-Learning Sentinel-2 Building Detection in Different Area Types — A framework quantifies seasonal and environmental impacts on Sentinel-2 building detection, guiding robust model selection.
• Monocular Building Height Estimation from PhiSat-2 Imagery: Dataset and Method — PhiSat-2 imagery, combined with a novel two-stream ordinal network, significantly improves monocular building height estimation.
• Building Extraction from Remote Sensing Imagery under Hazy and Low-light Conditions: Benchmark and Baseline — A new benchmark and network address building extraction challenges in hazy and low-light remote sensing imagery.
• Context-Aware Slum Mapping in Sub-Saharan Africa Using Sentinel-1 Texture and Local Climate Zones — Integrating Sentinel-1 SAR textures with Sentinel-2 optical data significantly enhances informal settlement mapping accuracy and seasonal stability in Sub-Saharan Africa.
• How Does Urban Context Relate to Residential Building Health? A Vision-POI Fusion Framework for Building-Level Housing Inspection — Fusing multi-view visual data with urban POI context enhances building health assessment accuracy.
• Uncertainty Quantification for EO Regression Tasks: Building Height, Tree Canopy Height and Above-ground Biomass Estimation — Deep learning models can provide reliable per-pixel uncertainty estimates for Earth Observation regression tasks using Gaussian or Quantile approaches.

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.