DEVELOPMENT OF A METHOD OF ESTIMATION OF TOTAL FLOOR AREA USING DEEP LEARNING BASED ON NIGHTTIME-LIGHT DATA

Abstract

 In order to achieve the recycling society worldwide, it is important to analyze material stock and flow on a global scale and understand their accumulation trends and patterns. World-wide satellite images data is easily and quickly accessible, and it breaks the restriction of statistical data. Recently, Convolutional Neural Network (CNN) is received a lot of attentions as a technique in the analysis of satellite images. Our study made efforts to develop an advanced model for estimating the total floor area of buildings based on night light data using CNN. Three major metropolitans in Japan (Tokyo, Osaka, and Nagoya metropolitan area) were selected as the estimation areas and the learning areas, setting one metropolitan as learning area to estimate the buildings floor areas for the other two. And same procedure has been done three times as each metropolitan was changed to be learning area in turns. Additionally, we verified the accuracy of results to examine the effectiveness of our model. In the case of “Osaka learning for Nagoya estimation”, it was clear that the total floor area of buildings in Nagoya metropolitan area was 450 million m², which is close to the real value and indicates our model can be useful?