Scholarlink

<p>The increasing frequency of forest fires, particularly in tropical deciduous forests, is causing severe damage to ecosystems. Currently, most available fire data is limited to active fire-point locations, and there is a lack of geospatial information of forest burn area and fire frequency in India. Timely and accurate mapping of fire scars is essential for planning appropriate forest management activities including enabling the assessment of fire frequency, risk zones, and identification of suitable areas for watch towers and fire closure areas etc. Current post-fire field surveys in India are inefficient, labour-intensive and lack objectivity. To address these challenges, this study proposes an automated method for forest burn scar extraction using spectral indices and machine learning algorithms with medium resolution Landsat-7 and 8 data. The study focuses on the tropical deciduous forests of Vidarbha region, Maharashtra state. Proposed approach involves two steps. First, generating the best Spectral Indices combination for burn scar delineation, which creates precise training samples for the segmentation model. Second, applying Deep Learning models to automatically map burn scars using the optimal outputs from the first step. This study presents an Automatic Weighted Average Ensembled learning U-Net (AWAE U-Net) model, where the learning of three individual backbone U-Net models i.e., VGG16, ResNet34 and Inception V3 were ensembled by applying the best weight calculated automatically. Experimental results show the model achieves 91.12% Intersection over Union and 93.51% F1 score for burn scar segmentation, demonstrating the effectiveness of the ensembled approach over single models.</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

A novel deep learning approach based on automatic weighted average ensemble for accurate forest burn scar extraction in Indian tropical deciduous forest

  • 期刊论文

  • 发表日期：

    2025年2月19日

查看全文
  • Amrita Singh,
  • A. O. Varghese,
  • Jugal Kishore Mani,
  • Ashish Kumar Sharma,
  • G. Sreenivasan,
  • Ashish Shrivastava

摘要

The increasing frequency of forest fires, particularly in tropical deciduous forests, is causing severe damage to ecosystems. Currently, most available fire data is limited to active fire-point locations, and there is a lack of geospatial information of forest burn area and fire frequency in India. Timely and accurate mapping of fire scars is essential for planning appropriate forest management activities including enabling the assessment of fire frequency, risk zones, and identification of suitable areas for watch towers and fire closure areas etc. Current post-fire field surveys in India are inefficient, labour-intensive and lack objectivity. To address these challenges, this study proposes an automated method for forest burn scar extraction using spectral indices and machine learning algorithms with medium resolution Landsat-7 and 8 data. The study focuses on the tropical deciduous forests of Vidarbha region, Maharashtra state. Proposed approach involves two steps. First, generating the best Spectral Indices combination for burn scar delineation, which creates precise training samples for the segmentation model. Second, applying Deep Learning models to automatically map burn scars using the optimal outputs from the first step. This study presents an Automatic Weighted Average Ensembled learning U-Net (AWAE U-Net) model, where the learning of three individual backbone U-Net models i.e., VGG16, ResNet34 and Inception V3 were ensembled by applying the best weight calculated automatically. Experimental results show the model achieves 91.12% Intersection over Union and 93.51% F1 score for burn scar segmentation, demonstrating the effectiveness of the ensembled approach over single models.