Scholarlink

<p>Spatial resolution matter in the detection of land cover along with the spectral reflectance received at satellite sensors. As we know in the era of machine learning, the spectral reflectance values in the form of digital numbers can be processed through ML algorithms to get meaningful output in the form of a classified image. In this study, the main objective was to find, whether spatial resolution affects the quantification of green infrastructure in semi-urban landscape? Especially, the green infrastructure like riparian vegetation, water body were measured in the Geographical Information System platform. While mapping the riparian vegetation, environmental researchers generally stuck at a point, where they are unable to decide the best ML algorithm for their research purpose. In this study, the six widely used ML algorithms were chosen in order to find various land covers area reflectance sensed by Linear Imaging and Self-scanning (LISS-4) and Sentinel-2 platform representing different spatial resolution. To assess this challenge, LISS-4 and Sentinel-2 imageries acquired on the same day were taken for processing through open-source SAGA GIS software by executing through six ML algorithms. Artificial Neural Network (ANN), K-Nearest Neighbours (KNN), Maximum Entropy (ME), Normal Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM) algorithms were applied on the training areas of stacked imageries of LISS-4 and Sentinel-2 bands. The output layer was validated using the F1-score, overall accuracy and Kappa coefficient. The results show that, SVM and ME algorithms were found to be closely match with the land cover condition present in the real-world coverage. Overall, the impact of spatial resolution on the variations in the computation of various land covers; it was revealed that the absolute difference between the average values of cumulative area under each land cover classified using six ML methods, using LISS-4 and Sentinel-2 dataset can be outlined as follows- agriculture 0.30&#xa0;km², barren land 2.01&#xa0;km², dense vegetation 0.49&#xa0;km², settlement 2.56&#xa0;km², and water body 0.23&#xa0;km². The ML-based outputs of final classified image using ME and SVM show nearly equal accuracy using LISS-4 data for F1-score and kappa coefficients. The research shows that, for semi-urban area to map blue and green infrastructure, SVM and ME algorithms-based outputs shows high reliability of acceptance as compared to other four method although the spatial resolution difference is 4.2&#xa0;m only. As this study deals with landscape morphology, land cover detection, for environmentalists it will guide to use thematic input for other ecosystem services-related research on riparian vegetation growth modelling, habitat suitability modelling, river bank erosion control, ecosystem restoration and future land cover forecasting.</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

Evaluation of Machine Learning Algorithms for Measuring the Green Infrastructure of Semi-urban Landscapes

  • 期刊论文

  • 发表日期：

    2026年4月16日

查看全文
  • Wasim Ayub Bagwan

摘要

Spatial resolution matter in the detection of land cover along with the spectral reflectance received at satellite sensors. As we know in the era of machine learning, the spectral reflectance values in the form of digital numbers can be processed through ML algorithms to get meaningful output in the form of a classified image. In this study, the main objective was to find, whether spatial resolution affects the quantification of green infrastructure in semi-urban landscape? Especially, the green infrastructure like riparian vegetation, water body were measured in the Geographical Information System platform. While mapping the riparian vegetation, environmental researchers generally stuck at a point, where they are unable to decide the best ML algorithm for their research purpose. In this study, the six widely used ML algorithms were chosen in order to find various land covers area reflectance sensed by Linear Imaging and Self-scanning (LISS-4) and Sentinel-2 platform representing different spatial resolution. To assess this challenge, LISS-4 and Sentinel-2 imageries acquired on the same day were taken for processing through open-source SAGA GIS software by executing through six ML algorithms. Artificial Neural Network (ANN), K-Nearest Neighbours (KNN), Maximum Entropy (ME), Normal Bayes (NB), Random Forest (RF), and Support Vector Machine (SVM) algorithms were applied on the training areas of stacked imageries of LISS-4 and Sentinel-2 bands. The output layer was validated using the F1-score, overall accuracy and Kappa coefficient. The results show that, SVM and ME algorithms were found to be closely match with the land cover condition present in the real-world coverage. Overall, the impact of spatial resolution on the variations in the computation of various land covers; it was revealed that the absolute difference between the average values of cumulative area under each land cover classified using six ML methods, using LISS-4 and Sentinel-2 dataset can be outlined as follows- agriculture 0.30 km², barren land 2.01 km², dense vegetation 0.49 km², settlement 2.56 km², and water body 0.23 km². The ML-based outputs of final classified image using ME and SVM show nearly equal accuracy using LISS-4 data for F1-score and kappa coefficients. The research shows that, for semi-urban area to map blue and green infrastructure, SVM and ME algorithms-based outputs shows high reliability of acceptance as compared to other four method although the spatial resolution difference is 4.2 m only. As this study deals with landscape morphology, land cover detection, for environmentalists it will guide to use thematic input for other ecosystem services-related research on riparian vegetation growth modelling, habitat suitability modelling, river bank erosion control, ecosystem restoration and future land cover forecasting.