Scholarlink

<p>One of the important topics in remote sensing is land use land cover classification. This paper presents a framework that aims to correctly classify land use land cover into seven different landscape characteristics such as agricultural area, agricultural green area, greenery, land, mountain region, settlement, and water body using a convolutional neural network (CNN) model. Thousands of satellite images of different regions in India are obtained from the Google Maps platform using the Maps Static API. These images are pre-processed using image augmentation techniques, and the proposed CNN model is thoroughly trained with this dataset. Several existing deep CNN models such as VGG-16, ResNet-50, and InceptionV3 are also considered. Instead of training these CNN models from scratch, transfer learning is used and these pre-trained networks are fine-tuned by replacing the final layers with additional layers. All the models are trained rigorously using the satellite image dataset and their performances are compared. Results show that the proposed CNN model classifies the test dataset with 89.03% accuracy. Although ResNet-50 provides the most promising results with a classification accuracy of 90.94% on the test dataset, the proposed CNN model has much fewer parameters than ResNet-50, and therefore its training and prediction times are shorter. The model training time and the per-image prediction time of the proposed model are 77.06% and 30.55% less than those of ResNet-50, respectively. Some case studies are also conducted to show the application of the proposed model in classifying the land use land cover of different locations. Additionally, the efficacy of the model is validated by using it for land use land cover classification on the RGB version of the benchmark EuroSAT dataset. The results show that the proposed model can effectively perform land use land cover classification on the EuroSAT dataset with 94.02% accuracy.</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

A CNN-based framework for land use land cover classification of heterogeneous terrain using satellite images

  • 期刊论文

  • 发表日期：

    2025年6月2日

查看全文
  • Anurina Tarafdar,
  • Asif Iqbal Middya,
  • Sounak Banerjee,
  • Sunirmal Khatua,
  • Sarbani Roy

摘要

One of the important topics in remote sensing is land use land cover classification. This paper presents a framework that aims to correctly classify land use land cover into seven different landscape characteristics such as agricultural area, agricultural green area, greenery, land, mountain region, settlement, and water body using a convolutional neural network (CNN) model. Thousands of satellite images of different regions in India are obtained from the Google Maps platform using the Maps Static API. These images are pre-processed using image augmentation techniques, and the proposed CNN model is thoroughly trained with this dataset. Several existing deep CNN models such as VGG-16, ResNet-50, and InceptionV3 are also considered. Instead of training these CNN models from scratch, transfer learning is used and these pre-trained networks are fine-tuned by replacing the final layers with additional layers. All the models are trained rigorously using the satellite image dataset and their performances are compared. Results show that the proposed CNN model classifies the test dataset with 89.03% accuracy. Although ResNet-50 provides the most promising results with a classification accuracy of 90.94% on the test dataset, the proposed CNN model has much fewer parameters than ResNet-50, and therefore its training and prediction times are shorter. The model training time and the per-image prediction time of the proposed model are 77.06% and 30.55% less than those of ResNet-50, respectively. Some case studies are also conducted to show the application of the proposed model in classifying the land use land cover of different locations. Additionally, the efficacy of the model is validated by using it for land use land cover classification on the RGB version of the benchmark EuroSAT dataset. The results show that the proposed model can effectively perform land use land cover classification on the EuroSAT dataset with 94.02% accuracy.