Man-made structures detection from space

Abstract

[en] With the development of affordable and recurrent remote sensing technology, we can now access frequent geospatial information in different levels of detail, ranging from 100m to 0.01m. The task of detecting various types of man-made structure and man-induced change has become a key problem in remote sensing image analysis. In this work we focus on providing an answer to the question: What is the optimal trade-off between resolution and cost when aiming at determining the existence of man-made structures in remote sensing images? Obtaining this value is important not only for designing optimal satellite sensors but also to use optimal data sources when developing data-based remote sensing products. At a global level, this knowledge contributes to understand the impact of our species on the planet. Our approach is based on developing a Deep Learning detector to classify human impact on aerial images. In particular, we exploit recent advances of Convolutional Neural Networks (CNN) that were successfully used for object detection and scene classification. We apply transfer learning by integrating a ResNet pre-trained on ImageNet to perform image classification on datasets of few thousand aerial images that we have manually collected and annotated. Using this classification pipeline we are able to determine the existence of man-made structure with an accuracy of 95% at the best resolution. We study the performance of our detector for resolutions ranging from 0.3m to 16m. We observe a linear decrease of the classification accuracy down to about 81% at the lowest resolution. Furthermore, we estimate the cost associated to build, launch, capture, and process satellite images to detect human impact. We estimate that monitoring the entire land surface of the earth at 1m resolution amounts for about $15 million. This cost increases by about two orders of magnitude at the best resolution studied here, and decreases by about one order of magnitude at a resolution of 10m per pixel. These results could be further improved by training a CNN on a labeled large scale remote sensing dataset. Nevertheless, our results suffice for studying the expansion of human kind using satellite imagery and provide valuable information for designing optimal satellite sensors.