[Paper Review] Building-Level Comparison of Microsoft and Google Open Building Footprints Datasets (Short Paper)

Wojciech Sirko, Sergii Kashubin+8 more|arXiv (Cornell University)|Jul 26, 2021

Remote-Sensing Image ClassificationEngineering17 references218 citations

TL;DR

This paper presents a scalable, U-Net-based deep learning pipeline for continent-wide building footprint detection across Africa using 50 cm satellite imagery. By integrating mixup data augmentation, self-training with soft KL loss, and optimized post-processing, the method achieves state-of-the-art performance, resulting in the Open Buildings dataset of 516 million footprints. - meta_description: A U-Net-based pipeline for continent-scale building detection in Africa using 50 cm satellite imagery, achieving 516M footprints via mixup and self-training. - objective: - Address large-scale building detection in developing regions with limited ground-truth data. - Improve instance segmentation performance on diverse African urban, rural, and desert contexts. - Develop a generalizable, scalable pipeline for continental-scale mapping using deep learning. - Generate a high-accuracy, open-access building footprint dataset for scientific and humanitarian use. - method: - Trained a U-Net model on 100k satellite images with 1.75M manually labeled buildings across Africa. - Applied mixup augmentation to improve mAP by +0.12. - Used self-training with soft KL loss to boost mAP by +0.06. - Employed distance-weighted loss with Gaussian convolutions to enhance boundary detection. - Implemented residual decoder blocks and morphological pre-processing to refine predictions. - Applied confidence-based deduplication and contouring to merge overlapping detections. - research_questions: - How do architectural choices and loss functions affect U-Net performance in building instance segmentation? - To what extent can mixup and self-training improve detection mAP on low-resource, high-variability satellite data? - How does pre-training on luminance, coarse/fine location data affect model generalization? - Can post-processing with confidence-based merging and contouring reduce false positives in large-scale outputs? - key_findings: - The final pipeline achieved a mean average precision (mAP) of 0.610 when pre-trained on ImageNet and fine-tuned with luminance data. - Mixup data augmentation improved mAP by +0.12 compared to baseline. - Self-training with soft KL loss yielded a +0.06 mAP gain. - The model maintained strong performance across diverse regions, including rural, urban, and desert areas. - The pipeline generated 516 million building footprints across Africa after deduplication and contouring. - Confidence score thresholds were derived to filter for 90% precision, enabling region-specific quality control.

ABSTRACT

Identifying the locations and footprints of buildings is vital for many practical and scientific purposes. Such information can be particularly useful in developing regions where alternative data sources may be scarce. In this work, we describe a model training pipeline for detecting buildings across the entire continent of Africa, using 50 cm satellite imagery. Starting with the U-Net model, widely used in satellite image analysis, we study variations in architecture, loss functions, regularization, pre-training, self-training and post-processing that increase instance segmentation performance. Experiments were carried out using a dataset of 100k satellite images across Africa containing 1.75M manually labelled building instances, and further datasets for pre-training and self-training. We report novel methods for improving performance of building detection with this type of model, including the use of mixup (mAP +0.12) and self-training with soft KL loss (mAP +0.06). The resulting pipeline obtains good results even on a wide variety of challenging rural and urban contexts, and was used to create the Open Buildings dataset of 516M Africa-wide detected footprints.

Motivation & Objective

• Address large-scale building detection in developing regions with limited ground-truth data.
• Improve instance segmentation performance on diverse African urban, rural, and desert contexts.
• Develop a generalizable, scalable pipeline for continental-scale mapping using deep learning.
• Generate a high-accuracy, open-access building footprint dataset for scientific and humanitarian use.

Proposed method

• Trained a U-Net model on 100k satellite images with 1.75M manually labeled buildings across Africa.
• Applied mixup data augmentation to improve mAP by +0.12.
• Used self-training with soft KL loss to boost mAP by +0.06.
• Employed distance-weighted loss with Gaussian convolutions to enhance boundary detection.
• Implemented residual decoder blocks and morphological pre-processing to refine predictions.
• Applied confidence-based deduplication and contouring to merge overlapping detections.

Experimental results

Research questions

• RQ1 How do architectural choices and loss functions affect U-Net performance in building instance segmentation?
• RQ2 To what extent can mixup and self-training improve detection mAP on low-resource, high-variability satellite data?
• RQ3 How does pre-training on luminance, coarse/fine location data affect model generalization?
• RQ4 Can post-processing with confidence-based merging and contouring reduce false positives in large-scale outputs?

Key findings

• The final pipeline achieved a mean average precision (mAP) of 0.610 when pre-trained on ImageNet and fine-tuned with luminance data.
• Mixup data augmentation improved mAP by +0.12 compared to baseline.
• Self-training with soft KL loss yielded a +0.06 mAP gain.
• The model maintained strong performance across diverse regions, including rural, urban, and desert areas.
• The pipeline generated 516 million building footprints across Africa after deduplication and contouring.
• Confidence score thresholds were derived to filter for 90% precision, enabling region-specific quality control.