Scholarlink

<p>The lowlands of Lesotho are critical for socio-economic growth, hosting significant activities and infrastructure, yet they face pressures from population growth, urbanisation, and land degradation. This study assessed the performance of two leading machine learning algorithms, Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), in predicting land cover change from 1993 to 2039. The findings revealed a critical insight into model validation: while SVM consistently achieved higher quantitative accuracy (overall accuracy: 87–94%, Kappa coefficient: 0.83–0.92) compared to XGBoost (overall accuracy: 84–93%, Kappa coefficient: 0.79–0.90), qualitative assessment indicated that SVM produced an implausible spatiotemporal trend for built-up areas, exhibiting a pattern of increasing in 1993, decreasing in 2009, and increasing again in 2024, which contradicted known historical development in the Lowlands. This demonstrates that relying solely on accuracy metrics is insufficient for ensuring a robust workflow. By selecting the more plausible XGBoost results, the study provides a reliable analysis of historical land use/land cover (LULC) dynamics and a credible future prediction. Change detection results from XGBoost indicated an increase of 53.5% in built-up area between 1993 and 2024, a net gain of 2.3% in vegetation, a decrease of 9.6% in bare land, and a decline of 38.6% in water bodies. CA-Markov predicted the continuation of these observed trends between 2024 and 2039, with built-up area increasing by 5.2%, vegetation by 4.8%, bare land decreasing by 9.6%, and water declining by 3.8%. These insights are critical for supporting evidence-based and sustainable land use planning and management in Lesotho’s most vulnerable agroecological zone.</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

Improving land cover change modelling with machine learning: a comparative analysis of SVM and XGBoost in the Lesotho Lowlands

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月8日

查看全文
  • Lerata Lerata,
  • Paidamwoyo Mhangara,
  • Eskinder Gidey

摘要

The lowlands of Lesotho are critical for socio-economic growth, hosting significant activities and infrastructure, yet they face pressures from population growth, urbanisation, and land degradation. This study assessed the performance of two leading machine learning algorithms, Support Vector Machine (SVM) and Extreme Gradient Boosting (XGBoost), in predicting land cover change from 1993 to 2039. The findings revealed a critical insight into model validation: while SVM consistently achieved higher quantitative accuracy (overall accuracy: 87–94%, Kappa coefficient: 0.83–0.92) compared to XGBoost (overall accuracy: 84–93%, Kappa coefficient: 0.79–0.90), qualitative assessment indicated that SVM produced an implausible spatiotemporal trend for built-up areas, exhibiting a pattern of increasing in 1993, decreasing in 2009, and increasing again in 2024, which contradicted known historical development in the Lowlands. This demonstrates that relying solely on accuracy metrics is insufficient for ensuring a robust workflow. By selecting the more plausible XGBoost results, the study provides a reliable analysis of historical land use/land cover (LULC) dynamics and a credible future prediction. Change detection results from XGBoost indicated an increase of 53.5% in built-up area between 1993 and 2024, a net gain of 2.3% in vegetation, a decrease of 9.6% in bare land, and a decline of 38.6% in water bodies. CA-Markov predicted the continuation of these observed trends between 2024 and 2039, with built-up area increasing by 5.2%, vegetation by 4.8%, bare land decreasing by 9.6%, and water declining by 3.8%. These insights are critical for supporting evidence-based and sustainable land use planning and management in Lesotho’s most vulnerable agroecological zone.