Roof Material Classification from Aerial Imagery
- PDF / 1,356,003 Bytes
- 11 Pages / 612 x 792 pts (letter) Page_size
- 70 Downloads / 219 Views
oof Material Classification from Aerial Imagery R. A. Solovyev* Institute for Design Problems in Microelectronics of RAS (IPPM RAS), Moscow, 124681 Russia *e-mail: [email protected] Received April 23, 2020; revised June 29, 2020; accepted July 2, 2020
Abstract—this paper describes an algorithm for classification of roof materials using aerial photographs. Main advantages of the algorithm are proposed methods to improve prediction accuracy. Proposed methods includes: method of converting ImageNet weights of neural networks for using multichannel images; special set of features of second level models that are used in addition to specific predictions of neural networks; special set of image augmentations that improve training accuracy. In addition, complete flow for solving this problem is proposed. The following content is available in open access: solution code, weight sets and architecture of the used neural networks. The proposed solution achieved second place in the competition “Open AI Caribbean Challenge”. Keywords: aerial imagery, convolutional neural networks, aerial photographs, satellite images DOI: 10.3103/S1060992X20030133
INTRODUCTION Some areas of the world are under significant risk of natural hazards such as earthquakes, hurricanes and floods; these acts of nature can have devastating consequences. One such area is the Caribbean. Disaster risk is especially great when houses and buildings do not meet modern construction standards, which is not uncommon in poor and informal settlements. Buildings can be retrofit to better prepare them for disaster, but the traditional method for identifying high-risk buildings involves visual inspection while going door to door by foot. This process can take many weeks if not months and cost millions of dollars. To speed up this work, machine learning techniques can be helpful. So, the World Bank Global Program for Resilient Housing and WeRobotics teamed up to prepare aerial drone imagery of buildings across the Caribbean annotated with characteristics that matter to building inspectors [1]. A feature that is especially important is roof material. Actually, it is one of the main risk factors for earthquakes and hurricanes. To solve the problem of roof material classification, the competition—“Open AI Caribbean Challenge: Mapping Disaster Risk from Aerial Imagery” [2]—was held at DrivenData site. The goal of this challenge was to create new rooftop classifiers using aerial imagery in St. Lucia, Guatemala, and Colombia. As known, machine learning models, in particular, deep convolutional neural networks, have now become the standard in image classification [3–5] as well as in the processing of satellite images and aerial photographs [6–10]. Machine learning models developed by the contestants can most accurately show the risk of disasters using drone images, and thus will help speed up building inspections and reduce their costs. Thanks to this, additional resources will be allocated for disaster preparation where the most disruptive hazards are expected. 1. PROBLEM STATEMEN
Data Loading...
