Scholarlink

Deep learning techniques have greatly advanced land-cover classification from remote sensing imagery, but their performance depends critically on choosing optimal hyperparameters. Manually tuning hyperparameters (e.g., learning rate, network depth, dropout rate) is time-consuming and often suboptimal. Metaheuristic algorithms offer an automated approach to this problem. In this work, we compare five metaheuristic optimizers—Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Differential Evolution (DE), African Vulture Optimization Algorithm (AVOA), and an Enhanced Dipper Throat Optimization Algorithm (EDTOA)—for hyperparameter tuning of convolutional neural networks (CNNs), a ResNet-50, and a U-Net. We evaluate these methods on two benchmark land-cover datasets: EuroSAT (patch-level multispectral image classification) and DeepGlobe (pixel-wise satellite image segmentation). Our data preprocessing includes normalization, data augmentation, and computing spectral indices (e.g., NDVI) to enrich the feature set. Each metaheuristic searches the hyperparameter space to maximize validation accuracy (for EuroSAT) or mean Intersection-over Union (mIoU) (for DeepGlobe). In addition to predictive performance, we analyze the computational cost (wall-clock time, epochs to convergence, GPU usage) of each optimizer to assess the trade-off between efficiency and accuracy. AVOA and EDTOA achieve the best results on both datasets (e.g., up to 98.5% accuracy on EuroSAT and 56% mIoU on DeepGlobe), outperforming the PSO, GA, and DE baselines while offering favorable cost-performance balance. These findings demonstrate that advanced metaheuristics can significantly improve deep model performance in land-cover classification. Our contributions include a comprehensive experimental comparison of five optimizers, a detailed methodology integrating spectral index features, a cost performance analysis, and reference results to guide future research.

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

Metaheuristic Optimization of Deep Learning Models for Land Cover Classification Using Remote Sensing Data

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Anil Antony,
  • R. Ganesh Kumar

摘要

Deep learning techniques have greatly advanced land-cover classification from remote sensing imagery, but their performance depends critically on choosing optimal hyperparameters. Manually tuning hyperparameters (e.g., learning rate, network depth, dropout rate) is time-consuming and often suboptimal. Metaheuristic algorithms offer an automated approach to this problem. In this work, we compare five metaheuristic optimizers—Particle Swarm Optimization (PSO), Genetic Algorithm (GA), Differential Evolution (DE), African Vulture Optimization Algorithm (AVOA), and an Enhanced Dipper Throat Optimization Algorithm (EDTOA)—for hyperparameter tuning of convolutional neural networks (CNNs), a ResNet-50, and a U-Net. We evaluate these methods on two benchmark land-cover datasets: EuroSAT (patch-level multispectral image classification) and DeepGlobe (pixel-wise satellite image segmentation). Our data preprocessing includes normalization, data augmentation, and computing spectral indices (e.g., NDVI) to enrich the feature set. Each metaheuristic searches the hyperparameter space to maximize validation accuracy (for EuroSAT) or mean Intersection-over Union (mIoU) (for DeepGlobe). In addition to predictive performance, we analyze the computational cost (wall-clock time, epochs to convergence, GPU usage) of each optimizer to assess the trade-off between efficiency and accuracy. AVOA and EDTOA achieve the best results on both datasets (e.g., up to 98.5% accuracy on EuroSAT and 56% mIoU on DeepGlobe), outperforming the PSO, GA, and DE baselines while offering favorable cost-performance balance. These findings demonstrate that advanced metaheuristics can significantly improve deep model performance in land-cover classification. Our contributions include a comprehensive experimental comparison of five optimizers, a detailed methodology integrating spectral index features, a cost performance analysis, and reference results to guide future research.