Scholarlink

<p>Selecting an appropriate classification algorithm for satellite imagery is essential to ensure accurate and reliable land use and land cover (LULC) maps. Many previous studies compare machine learning and deep learning models using different input representations, making it difficult to isolate the effects of model architecture and input data. This study addresses this gap by systematically comparing three widely used machine learning algorithms (Random Forest, Support Vector Machine, and Decision Tree) and a deep neural network using identical feature-based inputs derived from Landsat imagery in a semi-arid region, while evaluating the contribution of spectral indices and topographic features. Rigorous validation, including cross-validation and external test region, was applied to assess spatial generalization. The results indicate that the deep neural network achieves the highest accuracy and stability, particularly for spectrally overlapping classes, while Random Forest and Support Vector Machine also perform strongly when enhanced with additional features. The Decision Tree generalizes better when using spectral bands only, whereas feature integration can lead to overfitting. Spatial generalization tests further show that performance in unseen regions ranges from moderate to good, highlighting the need for more geographically diverse training data. Overall, findings emphasize that optimal algorithm selection requires balancing accuracy, generalization capability, and computational cost.</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

Feature-Based Evaluation of ML and DNN Models for LULC Mapping in Semi-Arid Regions

  • 期刊论文

  • 发表日期：

    2026年5月28日

查看全文
  • Sadiq Tahmi,
  • Xiaoping Rui

摘要

Selecting an appropriate classification algorithm for satellite imagery is essential to ensure accurate and reliable land use and land cover (LULC) maps. Many previous studies compare machine learning and deep learning models using different input representations, making it difficult to isolate the effects of model architecture and input data. This study addresses this gap by systematically comparing three widely used machine learning algorithms (Random Forest, Support Vector Machine, and Decision Tree) and a deep neural network using identical feature-based inputs derived from Landsat imagery in a semi-arid region, while evaluating the contribution of spectral indices and topographic features. Rigorous validation, including cross-validation and external test region, was applied to assess spatial generalization. The results indicate that the deep neural network achieves the highest accuracy and stability, particularly for spectrally overlapping classes, while Random Forest and Support Vector Machine also perform strongly when enhanced with additional features. The Decision Tree generalizes better when using spectral bands only, whereas feature integration can lead to overfitting. Spatial generalization tests further show that performance in unseen regions ranges from moderate to good, highlighting the need for more geographically diverse training data. Overall, findings emphasize that optimal algorithm selection requires balancing accuracy, generalization capability, and computational cost.