Scholarlink

Abstract <p>Using machine learning methods based on attribute data from satellite data products, the thermal anomalies associated with wildfires and gas flares in Eastern Siberia and the Russian Far East were classified. Based on a performance analysis of seven different machine learning models, the three models with the best quality metrics were selected (Random Forest, Extreme Gradient Boosting, and Categorical Boosting). Introducing additional features, such as meteorological data and spatiotemporal characteristics of thermal anomalies, improved the training quality metrics of selected models by 12–25% for MODIS satellite data and by 12–21% for VIIRS satellite data. Ensemble model construction approaches were used to improve classification efficiency. The best two ensemble models were tested on new data. Model for MODIS data using the Boosting approach correctly classified 88.6% of wildfires and 86.4% of gas flares. VIIRS data model, using the Stacking approach, correctly classified 97.6% of wildfires and 97.2% of gas flares. These results allow us to eliminate thermal anomalies caused by gas flares, which are falsely identified as wildfires, when analyzing fire activity and improve the accuracy of estimates of fire-related emissions of climate-active gases and aerosols.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Classification of Climate-Active Gas Emission Sources Based on Satellite Data Using Machine Learning Methods

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • V. G. Bondur,
  • N. V. Feoktistova,
  • O. S. Voronova

摘要

Abstract

Using machine learning methods based on attribute data from satellite data products, the thermal anomalies associated with wildfires and gas flares in Eastern Siberia and the Russian Far East were classified. Based on a performance analysis of seven different machine learning models, the three models with the best quality metrics were selected (Random Forest, Extreme Gradient Boosting, and Categorical Boosting). Introducing additional features, such as meteorological data and spatiotemporal characteristics of thermal anomalies, improved the training quality metrics of selected models by 12–25% for MODIS satellite data and by 12–21% for VIIRS satellite data. Ensemble model construction approaches were used to improve classification efficiency. The best two ensemble models were tested on new data. Model for MODIS data using the Boosting approach correctly classified 88.6% of wildfires and 86.4% of gas flares. VIIRS data model, using the Stacking approach, correctly classified 97.6% of wildfires and 97.2% of gas flares. These results allow us to eliminate thermal anomalies caused by gas flares, which are falsely identified as wildfires, when analyzing fire activity and improve the accuracy of estimates of fire-related emissions of climate-active gases and aerosols.