Scholarlink

<p>Forest monitoring using remote sensing is effective due to its broad spatial and temporal coverage. In this study, we present an improved method for detecting forest cover change in the rapidly urbanizing Kathmandu Valley, Nepal. Random Forest classifiers with satellite imagery data from years 2013 and 2023, including both spectral bands and auxiliary variables such as elevation and indices derived from spectral bands were used. The proposed classification models achieved overall accuracies of 0.981 and 0.983 and Kappa coefficients of 0.962 and 0.968 for 2013 and 2023, respectively. The auxiliary variables reduced Out-of-Bag (OOB) error rates from ~ 0.025 to 0.015 in 2013 and from ~ 0.007 to 0.004 in 2023. Our analysis of the classified pixels shows that forest cover grew by 11.5% during the study period from 471 km<sup>2</sup> to 525 km<sup>2</sup>. The forest gain was most significant in the mid-elevation region (1400–1900&#xa0;m), that has optimal conditions for forest regeneration, while lower (&lt; 1400&#xa0;m) and higher elevation (&gt; 1900&#xa0;m) regions had minimal to no change. We performed a land cover transition analysis which showed forest gain occurred primarily on barren/unused land, with a gain-to-loss ratio of 4:1, whereas other land cover types had more balanced transitions. We conclude that including topographic and auxiliary data improves forest cover classification accuracy significantly.</p> Graphical Abstract <p></p> <p>This graphical abstract summarizes an assessment of forest cover change in Kathmandu Valley, Nepal, over a decade using techniques like remote sensing and machine learning. Landsat 8 (2013) and Landsat 8 &amp; 9 (2023) surface reflectance data were preprocessed for cloud masking and composited to derive spectral bands (B1–B7) and auxiliary indices (e.g., NDVI, EVI, NDBI, NDWI, NDBaI). Topographic data from the SRTM DEM was integrated as another auxiliary variable to enhance classification in the mountainous and spectrally complex urban region. Two Random Forest classification models were developed: Model A used spectral bands only, while Model B used spectral bands and available auxiliary variables. Model performance was optimized using Out-of-Bag (OOB) error rates, with Model B showing consistently lower OOB error and superior class-wise accuracy metrics. We performed a feature importance analysis in developed models and elevation as the most influential predictor in both years. The classified outputs were used to extract forest cover for 2013 and 2023 and a net forest gain of 11.5% (from 471 km<sup>2</sup> to 525 km<sup>2</sup>) was found. A transition analysis was performed that showed most forest gain occurred through conversion of barren or unused land, with the highest increase concentrated in mid-elevation zones (1400–1900&#xa0;m). The gain in this region is likely due to community forests and forest regeneration programs.</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

Multi-spectral Remote Sensing of Long-term Forest Cover Change in Kathmandu Valley Using an Improved Machine Learning Classifier

  • 期刊论文

  • 发表日期：

    2026年1月29日

查看全文
  • Dinesh Neupane,
  • Manisha Parajuli,
  • Arjun Rijal

摘要

Forest monitoring using remote sensing is effective due to its broad spatial and temporal coverage. In this study, we present an improved method for detecting forest cover change in the rapidly urbanizing Kathmandu Valley, Nepal. Random Forest classifiers with satellite imagery data from years 2013 and 2023, including both spectral bands and auxiliary variables such as elevation and indices derived from spectral bands were used. The proposed classification models achieved overall accuracies of 0.981 and 0.983 and Kappa coefficients of 0.962 and 0.968 for 2013 and 2023, respectively. The auxiliary variables reduced Out-of-Bag (OOB) error rates from ~ 0.025 to 0.015 in 2013 and from ~ 0.007 to 0.004 in 2023. Our analysis of the classified pixels shows that forest cover grew by 11.5% during the study period from 471 km2 to 525 km2. The forest gain was most significant in the mid-elevation region (1400–1900 m), that has optimal conditions for forest regeneration, while lower (< 1400 m) and higher elevation (> 1900 m) regions had minimal to no change. We performed a land cover transition analysis which showed forest gain occurred primarily on barren/unused land, with a gain-to-loss ratio of 4:1, whereas other land cover types had more balanced transitions. We conclude that including topographic and auxiliary data improves forest cover classification accuracy significantly.

Graphical Abstract

This graphical abstract summarizes an assessment of forest cover change in Kathmandu Valley, Nepal, over a decade using techniques like remote sensing and machine learning. Landsat 8 (2013) and Landsat 8 & 9 (2023) surface reflectance data were preprocessed for cloud masking and composited to derive spectral bands (B1–B7) and auxiliary indices (e.g., NDVI, EVI, NDBI, NDWI, NDBaI). Topographic data from the SRTM DEM was integrated as another auxiliary variable to enhance classification in the mountainous and spectrally complex urban region. Two Random Forest classification models were developed: Model A used spectral bands only, while Model B used spectral bands and available auxiliary variables. Model performance was optimized using Out-of-Bag (OOB) error rates, with Model B showing consistently lower OOB error and superior class-wise accuracy metrics. We performed a feature importance analysis in developed models and elevation as the most influential predictor in both years. The classified outputs were used to extract forest cover for 2013 and 2023 and a net forest gain of 11.5% (from 471 km2 to 525 km2) was found. A transition analysis was performed that showed most forest gain occurred through conversion of barren or unused land, with the highest increase concentrated in mid-elevation zones (1400–1900 m). The gain in this region is likely due to community forests and forest regeneration programs.