Scholarlink

<p>There is a growing interest in the adoption of Google Earth Engine (GEE) for Land Use Land Cover (LULC) mapping across diverse geographical settings, yet guidance on selecting suitable classifiers and hyperparameter settings remains limited. This study presents a systematic comparison of all supervised machine learning classifiers within the GEE platform for a rapidly urbanizing urban setting in western India. Specifically, Support Vector Machine (SVM), Minimum Distance Classifier (MDC), Classification and Regression Trees (CART), Gradient Tree Boosting (GTB), K-Nearest Neighbours (KNN), Naive Bayes (NB), and Random Forest (RF) algorithms were evaluated under multiple hyperparameter variations. Sentinel-2 satellite imagery was pre-processed using cloud masking and image normalization, with 300 training points selected for each LULC category. The dataset was split, with 75% used for model training and 25% for testing. The classifiers were evaluated based on Kappa indices, which revealed significant variations in classification accuracy. SVM with a linear kernel delivered the highest accuracy (overall Kappa = 0.883), particularly for Built-Up (0.908) and Bareland (0.924) classes. GTB also performed strongly (Kappa = 0.870), followed by RF (Kappa = 0.862), both benefiting from ensemble-based learning. The MDC using the Mahalanobis distance achieved good accuracy (Kappa = 0.853), especially for Built-Up areas, while nonlinear SVM variants, Naïve Bayes, and several distance-based or shallow-tree configurations showed comparatively weaker results. Overall, the linear SVM emerged as the most effective classifier, with GTB, RF, and MDC also demonstrating strong performance, providing clear, data-driven guidance for selecting robust classifier–hyperparameter combinations for urban LULC mapping in GEE.</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

Machine Learning Classifiers for LULC Mapping - A Performance Comparison Using Google Earth Engine

  • 期刊论文

  • 发表日期：

    2025年12月5日

查看全文
  • Kratika Sharma,
  • Shobhit Chaturvedi

摘要

There is a growing interest in the adoption of Google Earth Engine (GEE) for Land Use Land Cover (LULC) mapping across diverse geographical settings, yet guidance on selecting suitable classifiers and hyperparameter settings remains limited. This study presents a systematic comparison of all supervised machine learning classifiers within the GEE platform for a rapidly urbanizing urban setting in western India. Specifically, Support Vector Machine (SVM), Minimum Distance Classifier (MDC), Classification and Regression Trees (CART), Gradient Tree Boosting (GTB), K-Nearest Neighbours (KNN), Naive Bayes (NB), and Random Forest (RF) algorithms were evaluated under multiple hyperparameter variations. Sentinel-2 satellite imagery was pre-processed using cloud masking and image normalization, with 300 training points selected for each LULC category. The dataset was split, with 75% used for model training and 25% for testing. The classifiers were evaluated based on Kappa indices, which revealed significant variations in classification accuracy. SVM with a linear kernel delivered the highest accuracy (overall Kappa = 0.883), particularly for Built-Up (0.908) and Bareland (0.924) classes. GTB also performed strongly (Kappa = 0.870), followed by RF (Kappa = 0.862), both benefiting from ensemble-based learning. The MDC using the Mahalanobis distance achieved good accuracy (Kappa = 0.853), especially for Built-Up areas, while nonlinear SVM variants, Naïve Bayes, and several distance-based or shallow-tree configurations showed comparatively weaker results. Overall, the linear SVM emerged as the most effective classifier, with GTB, RF, and MDC also demonstrating strong performance, providing clear, data-driven guidance for selecting robust classifier–hyperparameter combinations for urban LULC mapping in GEE.