Scholarlink

<p>Understanding tree community composition is central to monitoring biodiversity patterns, and recent advances in Earth observation (EO) offer powerful means to study these dynamics across large landscapes. In this study, we present an approach for classifying of tree communities by combining field-based information with machine learning that incorporates environmental factors and remote sensing data, including multi-temporal imagery and their derivatives. By leveraging EO data, five different scenarios with a combination of different EO variables were constructed and tested for their performance in accurately mapping 13 tree communities using a Random Forest classification approach. The evaluation results demonstrated that the utilisation of integrated EO data in scenario 5 yielded notably enhanced performance in accurately classifying tree communities with an overall accuracy of 83%, surpassing the other four scenarios. The F1-scores, representing the harmonic mean of precision (User Accuracy) and recall (Producer Accuracy) obtained range from 57.4% to 100% for the 13 tree communities, further affirming the efficacy of Scenario 5. The highest accuracy (F1 &gt; 83%) was achieved for <i>Pinus roxburghii</i>, mixed semi-evergreen, <i>Hevea brasiliensis</i>,<i> Xylia xylocarpa</i>,<i> Areca catechu</i>,<i> Acacia spp.</i> and <i>Tectona grandis</i> communities. The results of the variable importance analysis indicated that Sentinel-2 bands from December (peak growing season) and March (dry season), and bioclimatic variables, were the most important predictors and proved highly advantageous for discerning tree communities. The final map depicted composition, with <i>Tectona grandis</i> community occupying the largest area at 109.12 km<sup>2</sup>, followed by <i>Terminalia paniculata</i>-dominated community (98.56 km<sup>2</sup>) and the mixed semi-evergreen community (51.91 km<sup>2</sup>) in the region. The results demonstrate the robustness of the proposed methodology in discerning fine-scale tree community patterns and highlight its potential for application across diverse tropical forest regions. This approach provides valuable insights for advancing forest biodiversity assessments and supports informed conservation and landscape management in these ecologically rich environments.</p> Graphical Abstract <p></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

Fine-Scale Mapping of Tropical Forest Tree Communities with Open-Source Multisensor Earth Observation and Machine Learning

  • 期刊论文

  • 发表日期：

    2026年5月12日

查看全文
  • Naveen Babu Kanda,
  • Debabrata Behera,
  • Kurian Ayushi,
  • Ayyappan Narayanan,
  • Narayanaswamy Parthasarathy

摘要

Understanding tree community composition is central to monitoring biodiversity patterns, and recent advances in Earth observation (EO) offer powerful means to study these dynamics across large landscapes. In this study, we present an approach for classifying of tree communities by combining field-based information with machine learning that incorporates environmental factors and remote sensing data, including multi-temporal imagery and their derivatives. By leveraging EO data, five different scenarios with a combination of different EO variables were constructed and tested for their performance in accurately mapping 13 tree communities using a Random Forest classification approach. The evaluation results demonstrated that the utilisation of integrated EO data in scenario 5 yielded notably enhanced performance in accurately classifying tree communities with an overall accuracy of 83%, surpassing the other four scenarios. The F1-scores, representing the harmonic mean of precision (User Accuracy) and recall (Producer Accuracy) obtained range from 57.4% to 100% for the 13 tree communities, further affirming the efficacy of Scenario 5. The highest accuracy (F1 > 83%) was achieved for Pinus roxburghii, mixed semi-evergreen, Hevea brasiliensis, Xylia xylocarpa, Areca catechu, Acacia spp. and Tectona grandis communities. The results of the variable importance analysis indicated that Sentinel-2 bands from December (peak growing season) and March (dry season), and bioclimatic variables, were the most important predictors and proved highly advantageous for discerning tree communities. The final map depicted composition, with Tectona grandis community occupying the largest area at 109.12 km2, followed by Terminalia paniculata-dominated community (98.56 km2) and the mixed semi-evergreen community (51.91 km2) in the region. The results demonstrate the robustness of the proposed methodology in discerning fine-scale tree community patterns and highlight its potential for application across diverse tropical forest regions. This approach provides valuable insights for advancing forest biodiversity assessments and supports informed conservation and landscape management in these ecologically rich environments.

Graphical Abstract