DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images (1805.06561v1)

Abstract: We present the DeepGlobe 2018 Satellite Image Understanding Challenge, which includes three public competitions for segmentation, detection, and classification tasks on satellite images. Similar to other challenges in computer vision domain such as DAVIS and COCO, DeepGlobe proposes three datasets and corresponding evaluation methodologies, coherently bundled in three competitions with a dedicated workshop co-located with CVPR 2018. We observed that satellite imagery is a rich and structured source of information, yet it is less investigated than everyday images by computer vision researchers. However, bridging modern computer vision with remote sensing data analysis could have critical impact to the way we understand our environment and lead to major breakthroughs in global urban planning or climate change research. Keeping such bridging objective in mind, DeepGlobe aims to bring together researchers from different domains to raise awareness of remote sensing in the computer vision community and vice-versa. We aim to improve and evaluate state-of-the-art satellite image understanding approaches, which can hopefully serve as reference benchmarks for future research in the same topic. In this paper, we analyze characteristics of each dataset, define the evaluation criteria of the competitions, and provide baselines for each task.

• The paper introduced a comprehensive challenge for satellite imagery parsing via segmentation, detection, and classification tasks.
• It provided detailed datasets and baseline evaluations, with metrics like a 0.545 IoU for road extraction and a 0.693 F1 score for building detection.
• The challenge underscores the potential of interdisciplinary collaboration to enhance remote sensing methodologies and practical applications in urban planning and environmental monitoring.

DeepGlobe 2018: A Challenge to Parse the Earth through Satellite Images

The DeepGlobe 2018 Satellite Image Understanding Challenge represents an ambitious effort to advance the state-of-the-art in remote sensing through targeted computer vision tasks. The challenge, co-located with CVPR 2018, introduces three specific competitions aimed at segmentation, detection, and classification tasks using satellite imagery datasets. This paper discusses the challenge's goals, the datasets, evaluation methodologies, and provides preliminary baselines against which future research can be compared.

Introduction

The DeepGlobe challenge is motivated by the rich yet underutilized potential of satellite imagery in computer vision. Unlike traditional datasets that focus on everyday images, satellite images provide structured and consistent information valuable for applications ranging from urban planning to climate research. The challenge is structured to not only evaluate state-of-the-art methods but also to foster collaboration between researchers from different fields, with a long-term objective of establishing meaningful benchmarks in satellite image analysis.

Datasets

The challenge comprises three datasets, each tailored to a specific problem: road extraction, building detection, and land cover classification.

1. Road Extraction Dataset: This dataset is sourced from DigitalGlobe +Vivid Images and includes images from Thailand, Indonesia, and India. The images are high-resolution (50 cm/pixel), and the dataset includes 8,570 images with pixel-wise road annotations.
2. Building Detection Dataset: Derived from the SpaceNet Building Detection Dataset, it includes annotated building footprints across urban landscapes in Las Vegas, Paris, Shanghai, and Khartoum. The dataset consists of 24,586 image tiles with corresponding building outlines.
3. Land Cover Classification Dataset: The dataset contains 1,146 images captured at 50 cm resolution and spans over 1,716.9 km². Each image is annotated with land cover types following the Anderson Classification.

Tasks and Evaluation

Each competition within the DeepGlobe challenge is defined by a distinct task and corresponding evaluation metrics.

Road Extraction: - Task: Binary classification of each pixel as road or non-road. - Metric: Intersection over Union (IoU), calculated as:

$IoU = \frac{TP}{TP + FP + FN}$

• Baseline: Modified DeepLab architecture with ResNet18 achieved an IoU score of 0.545.

Building Detection: - Task: Segmentation of building footprints from satellite images. - Metric: F1 Score, emphasizing precision and recall of detected buildings. True positives are identified by an IoU threshold of 0.5. - Baseline: Top algorithms from SpaceNet 2017 achieved an average F1 score of 0.693.

Land Cover Classification: - Task: Multi-class segmentation of land cover types. - Metric: IoU for each class, averaged across all land cover classes. - Baseline: A DeepLab-based model achieved a multi-class IoU score of 0.433.

Implications and Future Directions

The datasets and baselines provided by the DeepGlobe challenge are pivotal for advancing research in satellite image analysis. The challenge has demonstrated strong initial results, but also highlighted areas for improvement such as handling diverse terrain and improving classification granularity. Future research may benefit from integrating additional data sources (e.g., LiDAR), developing more robust data augmentation techniques, and employing sophisticated post-processing methods (e.g., Conditional Random Fields) to refine segmentation outcomes. By doing so, the benchmarks set by DeepGlobe can be pushed further, leading to more accurate and scalable solutions in remote sensing.

Conclusion

DeepGlobe 2018 marks a significant step in bridging the gap between computer vision and remote sensing. By providing well-structured challenges and datasets, it catalyzes research that can fundamentally enhance our understanding and monitoring of the Earth's surface. The collaborative effort and the benchmarks established through this challenge will no doubt inspire continued advancements in this interdisciplinary field.