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

Abstract

Figure 1: DeepGlobe Challenges: Example road extraction, building detection, and land cover classification training images superimposed on corresponding satellite images. AbstractWe present the DeepGlobe 2018 Satellite Image Understanding Challenge, which includes three public competitions for segmentation, detection, and classification tasks on satellite images ( Figure 1). Similar to other challenges in computer vision domain such as DAVIS[21] and COCO[33], 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.

Dense Dilated Convolutions’ Merging Network for Land Cover Classification

“…In this paper, we focus on two publicly used databases, namely the ISPRS 2D semantic labeling contest datasets [8], and the DeepGlobe land cover challenge dataset [9]. The ISPRS datasets are comprised of aerial images over two cities in Germany: Potsdam 1 and Vaihingen 2 , which have been labelled with six of the most common land cover classes: impervious surfaces, buildings, low vegetation, trees, cars and clutter.…”

“…The ISPRS datasets are comprised of aerial images over two cities in Germany: Potsdam 1 and Vaihingen 2 , which have been labelled with six of the most common land cover classes: impervious surfaces, buildings, low vegetation, trees, cars and clutter. The DeepGlobe land cover dataset consists of satellite data collected from the DigitalGlobe Vivid+ dataset [9], and focuses on rural areas. This includes seven types of land covers: urban (man-made, built up areas with human artifacts), agriculture (farms, cropland, orchards, vineyards, ornamental horticultural areas, and so on), rangeland (any non-forest, nonfarm, green land and grass), forest (any land with at least 20% tree crown density plus clear cuts), water (rivers, oceans, lakes, wetland, ponds), barren (mountain, rock, dessert, beach, land with no vegetation), and unknown (clounds and others).…”

“…1. Examples of land cover labels (right) and corresponding remote sensing images (left) from two different datasets (ISPRS Potsdam [8] and DeepGlobel [9]) separately. Semantic label colors are shown beside the ground truths.…”

“…Motivated by the recent success of depthwise separable convolutions [10], we also explore grouped convolutions [11] with strided operations adapted into our DDCM-Net, which is shown to further improve net speed and accuracy. We finally demonstrate the effectiveness of the adapted DDCM-Net on the DeepGlobe land cover challenge dataset [9]. In summary, our contributions are: 1) We propose a new computationally light-weight and scalable architecture based on dilated convolutions [12] that can be used as a simple, yet effective, encoder or decoder module for semantic segmentation tasks.…”

“…Our applied focus in this paper is land cover classification based on remote sensing. Lately, the FCNs and encoderdecoder architectures have been widely adapted and applied to the ISPRS [8] Semantic Labeling Contest [20], [21], [22], [23], [24], [25], [26], [27], [28], [29], and the DeepGlobe CVPR-2018 [9] challenge of automatic classification of land cover types [30], [31], [32], [33], [34], [35], [36]. Paisitkriangkrai et al [20] proposed a scheme for high-resolution land cover classification using a combination of a patchbased CNN and a random forest classification that is trained on hand-crafted features.…”