Deep Learning Based Fossil-Fuel Power Plant Monitoring in High Resolution Remote Sensing Images: A Comparative Study

Haopeng Zhang,Qin Deng

doi:10.3390/rs11091117

Abstract

The frequent hazy weather with air pollution in North China has aroused wide attention in the past few years. One of the most important pollution resource is the anthropogenic emission by fossil-fuel power plants. To relieve the pollution and assist urban environment monitoring, it is necessary to continuously monitor the working status of power plants. Satellite or airborne remote sensing provides high quality data for such tasks. In this paper, we design a power plant monitoring framework based on deep learning to automatically detect the power plants and determine their working status in high resolution remote sensing images (RSIs). To this end, we collected a dataset named BUAA-FFPP60 containing RSIs of over 60 fossil-fuel power plants in the Beijing-Tianjin-Hebei region in North China, which covers about 123 km 2 of an urban area. We compared eight state-of-the-art deep learning models and comprehensively analyzed their performance on accuracy, speed, and hardware cost. Experimental results illustrate that our deep learning based framework can effectively detect the fossil-fuel power plants and determine their working status with mean average precision up to 0.8273, showing good potential for urban environment monitoring.

Highlights

Air pollution reaches a level pernicious to the health of the population and poses a threat to peoples’ daily life in the past few years in China
To ensure the fairness of comparison, we trained and tested the eight models, namely Faster Regions with Convolutional Neural Network (R-Convolutional Neural Network (CNN)) [18], Feature Pyramid Network (FPN) [22], Region-based Fully Convolutional Networks (R-fully convolutional networks (FCN)) [23], Deformable Convolutional Networks (DCN) [24], Single Shot MultiBox Detector (SSD) [26], Deconvolutional Single Shot Detector (DSSD) [28], YOLOv3 [30], and RetinaNet [31], on the same hardware, which is a server with a 2.5 GHz Central Processing Unit (CPU) and a Nvidia Tesla P4 Graphics Processing Unit (GPU)
Based on the data provided in [4], which contain 217 power plant images of 1-m spatial resolution with labeled locations and classes of chimney or condensing tower, we added 101 new images with 1-m spatial resolution collected from Google Earth and built a more comprehensive dataset named BUAA-FFPP60, meaning that it contains remote sensing images (RSIs) of over 60 fossil-fuel power plants (FFPP) collected by researchers from Beihang University (BUAA)

Summary

Introduction

Air pollution reaches a level pernicious to the health of the population and poses a threat to peoples’ daily life in the past few years in China. In the Beijing-Tianjin-Hebei region, serious haze pollution frequently appears in winter. The burning of fossil-fuel, e.g., coal, is one of the most important pollution sources in the Beijing-Tianjin-Hebei region. Without the difficulty of staffed daily monitoring, high resolution remote sensing provides an effective solution for such task by observing the running state of power plants from satellite or airborne imaging sensors. One is to find the power plants in high resolution remote sensing images (RSIs), and the other is to check the working status of them. It is better to solve this application problem automatically by recent artificial intelligence algorithms

Results

Discussion

Conclusion