Identifying Waste Burning Plumes Using High-Resolution Satellite Imagery and Machine Learning: A Case Study in the Maldives

Abstract

A rapid increase in municipal solid waste generation has far outpaced resources to manage waste in many developing countries, resulting in the burning of trash in designated landfills or public places, the release of harmful air pollutants, and exposure of nearby populations. While some governments have recently banned open burning at municipal facilities, monitoring the success of mitigation strategies has been challenging due to the lack of adequate air pollution monitoring methodologies. To address this, we have developed a machine learning approach that utilizes high-resolution (3 m/pixel) satellite imagery and applied the methodology to detect plumes of smoke from waste burning on Thilafushi in the Maldives. We employed an image classification and semantic segmentation model based on a pretrained convolutional neural network to identify and locate plumes within images. Our approach achieved an average intersection over union (overlap) of 0.70 between visually identified plumes and the machine learning output as well as a pixel-level classification accuracy of 96.3% on our holdout testing data. Our results demonstrate the potential of machine learning models in detecting plumes from sources where measurements are not available, including wildfires, coal-fired power plants, and industrial plumes, as well as in tracking the progress of mitigation strategies.