Scholarlink

<p>Satellite imagery plays a crucial role in exploring land use inventories of urban areas. However, accurate land cover classification from satellite imagery remains a longstanding challenge. With recent advancements in artificial intelligence technology, Deep Learning algorithms have achieved success in understanding satellite images by means of Convolutional Neural Networks (CNNs). While there has been a notable emphasis on satellite image analysis to improve the accuracy of land cover classifications, it is imperative to emphasise the significance of data-driven optimisation techniques. This paper introduces a hybrid UNet-ResNet-50 architecture, which integrates the metaheuristic Particle Swarm Algorithm (PSA) in dynamic hyperparameter optimisation for multi-class semantic segmentation. The approach of this research leverages a UNet extractor with ResNet-50 backbone (UResNet-50) and augments it with a Particle Swarm Optimiser (PSO) to automate the hyperparameter tuning process for segmenting the DeepGlobe satellite dataset into seven meaningful classes, namely: urban, forest, rangeland, barren land, agriculture, water bodies and unknown. The PSO-UResNet-50 model demonstrated robust performance across four distinct locations, in terms of accuracy, precision, recall, F1-score and mIoU as follows: Location-1 (95.74%, 98.12%, 86.95%, 92.04%,88.17%); Location-2 (91.88%, 79.23%, 80.75%, 81.42%, 83.03%); Location-3 (99.44%, 93.97%, 87.42%, 88.68%, 90.77%); and Location-4 (96.20%, 94.03%, 89.75%, 92.16%, 88.97%). The proposed PSO-UResNet-50 model outperformed the conventional U-Net and hybrid UResNet-50, demonstrating the advantage of applying PSO in multi-class segmentation of satellite imagery. The principal contribution of this work lies in the development and validation of a novel, metaheuristic-optimised deep learning framework that addresses the land cover classification challenge inherent in satellite images.</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

Multi-class segmentation of land cover types for DigitalGlobe satellite imagery using deep hybrid UNet-ResNet-50 network optimised with metaheuristic particle swarm algorithm

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月9日

查看全文
  • Evans Annan Boah,
  • Yakubu Issaka,
  • Rebecca Arhinful,
  • Abraham Aidoo Borsah,
  • Wisdom Sena Aklamati,
  • Felix Tabase

摘要

Satellite imagery plays a crucial role in exploring land use inventories of urban areas. However, accurate land cover classification from satellite imagery remains a longstanding challenge. With recent advancements in artificial intelligence technology, Deep Learning algorithms have achieved success in understanding satellite images by means of Convolutional Neural Networks (CNNs). While there has been a notable emphasis on satellite image analysis to improve the accuracy of land cover classifications, it is imperative to emphasise the significance of data-driven optimisation techniques. This paper introduces a hybrid UNet-ResNet-50 architecture, which integrates the metaheuristic Particle Swarm Algorithm (PSA) in dynamic hyperparameter optimisation for multi-class semantic segmentation. The approach of this research leverages a UNet extractor with ResNet-50 backbone (UResNet-50) and augments it with a Particle Swarm Optimiser (PSO) to automate the hyperparameter tuning process for segmenting the DeepGlobe satellite dataset into seven meaningful classes, namely: urban, forest, rangeland, barren land, agriculture, water bodies and unknown. The PSO-UResNet-50 model demonstrated robust performance across four distinct locations, in terms of accuracy, precision, recall, F1-score and mIoU as follows: Location-1 (95.74%, 98.12%, 86.95%, 92.04%,88.17%); Location-2 (91.88%, 79.23%, 80.75%, 81.42%, 83.03%); Location-3 (99.44%, 93.97%, 87.42%, 88.68%, 90.77%); and Location-4 (96.20%, 94.03%, 89.75%, 92.16%, 88.97%). The proposed PSO-UResNet-50 model outperformed the conventional U-Net and hybrid UResNet-50, demonstrating the advantage of applying PSO in multi-class segmentation of satellite imagery. The principal contribution of this work lies in the development and validation of a novel, metaheuristic-optimised deep learning framework that addresses the land cover classification challenge inherent in satellite images.