Scholarlink

<p>Urban landscapes pose significant challenges for land use/land cover (LULC) classification due to the inherent heterogeneity of built-up areas, vegetation, bare ground, and water bodies. This study aims to enhance the accuracy of urban LULC classification by combining multiple spectral indices within a&#xa0;machine-learning approach and statistically evaluating model performance. Using Kyiv, Ukraine, as a&#xa0;case study, we assess whether integrating vegetation, soil, and built-up indices with satellite-derived spectral reflectance improves the separation of urban LULC classes. Sentinel‑2 Level-2A images acquired during the spring–autumn periods of 2020–2022 were processed in Google Earth Engine to generate seasonally balanced mosaics. Nineteen spectral indices were grouped according to the physical features they highlight (water, built-up areas, bare soil, and vegetation) and combined with layers of terrain (slope and elevation), texture, and differential operators. Sixteen feature sets were then classified using three machine-learning algorithms: random forest (RF), classification and regression trees (CART), and support vector machines (SVM). Our results show that RF delivered the highest accuracy metrics. Random forest using feature set combination&#xa0;3 (targeted bare soils and built-up indices) achieved an overall accuracy (OA) of approximately 90.94%, with kappa ≈ 88.50% and F1 ≈ 91%. Classification and regression trees reached an OA of 87.76% using combination&#xa0;8, while SVM had an OA of 89.12% (best with combination&#xa0;5). McNemar tests indicated that RF’s improvements were statistically significant relative to SVM and CART when evaluated on the same feature sets. Kruskal–Wallis H tests further suggested that predictors most sensitive to the urban and other land classes (notably bare soil index, urban index, and short-wave infrared bands B11/B12) drive most model disagreement, highlighting confusion among the urban, sparse vegetation, and other land classes. Pairwise Z‑tests on kappa showed that the selected top combinations (combination&#xa0;3 for RF) significantly outperformed simpler sets (combination&#xa0;5 and&#xa0;8; |Z|&gt; 1.96). Across all models and feature sets, OA ranged from approximately 85% to 91% (with kappa up to ≈ 0.89). Class-wise results were strongest for water, dense vegetation, and urban land (user’s accuracies of roughly 95%–98%), whereas sparse vegetation and other lands remained the most challenging (approximately 76%–78% user accuracy). Notably, urban land exhibited high user’s accuracy but comparatively low producer’s accuracy&#xa0;(≈ 78%), indicating omission errors. Overall, these results show that targeted, multi-index feature sets materially improve urban LULC discrimination, with multiple statistical tests corroborating the superiority of the best-performing classifier feature set pairings.</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

Enhancing Urban Land Use/Land Cover Classification Through the Synergistic Use of Multiple Spectral Indices: a Case Study of Kyiv, Ukraine

  • 期刊论文

  • 发表日期：

    2026年1月28日

查看全文
  • Vadym Belenok,
  • Liliia Hebryn-Baidy,
  • Sergiy Kryachok,
  • Roman Tretiak,
  • Serhii Bodnar

摘要

Urban landscapes pose significant challenges for land use/land cover (LULC) classification due to the inherent heterogeneity of built-up areas, vegetation, bare ground, and water bodies. This study aims to enhance the accuracy of urban LULC classification by combining multiple spectral indices within a machine-learning approach and statistically evaluating model performance. Using Kyiv, Ukraine, as a case study, we assess whether integrating vegetation, soil, and built-up indices with satellite-derived spectral reflectance improves the separation of urban LULC classes. Sentinel‑2 Level-2A images acquired during the spring–autumn periods of 2020–2022 were processed in Google Earth Engine to generate seasonally balanced mosaics. Nineteen spectral indices were grouped according to the physical features they highlight (water, built-up areas, bare soil, and vegetation) and combined with layers of terrain (slope and elevation), texture, and differential operators. Sixteen feature sets were then classified using three machine-learning algorithms: random forest (RF), classification and regression trees (CART), and support vector machines (SVM). Our results show that RF delivered the highest accuracy metrics. Random forest using feature set combination 3 (targeted bare soils and built-up indices) achieved an overall accuracy (OA) of approximately 90.94%, with kappa ≈ 88.50% and F1 ≈ 91%. Classification and regression trees reached an OA of 87.76% using combination 8, while SVM had an OA of 89.12% (best with combination 5). McNemar tests indicated that RF’s improvements were statistically significant relative to SVM and CART when evaluated on the same feature sets. Kruskal–Wallis H tests further suggested that predictors most sensitive to the urban and other land classes (notably bare soil index, urban index, and short-wave infrared bands B11/B12) drive most model disagreement, highlighting confusion among the urban, sparse vegetation, and other land classes. Pairwise Z‑tests on kappa showed that the selected top combinations (combination 3 for RF) significantly outperformed simpler sets (combination 5 and 8; |Z|> 1.96). Across all models and feature sets, OA ranged from approximately 85% to 91% (with kappa up to ≈ 0.89). Class-wise results were strongest for water, dense vegetation, and urban land (user’s accuracies of roughly 95%–98%), whereas sparse vegetation and other lands remained the most challenging (approximately 76%–78% user accuracy). Notably, urban land exhibited high user’s accuracy but comparatively low producer’s accuracy (≈ 78%), indicating omission errors. Overall, these results show that targeted, multi-index feature sets materially improve urban LULC discrimination, with multiple statistical tests corroborating the superiority of the best-performing classifier feature set pairings.