用深度学习融合多分辨率遥感数据，实现高精度地物分割。
Semantic segmentation on multi-resolution optical and microwave data using deep learning
- 改进U-Net和VGG-UNet模型，处理光学与微波遥感图像
- 在1米分辨率影像上建筑/船舶识别准确率超95%，微波数据达96%以上
- 适用于复杂场景下的多类地物像素级分类，适合遥感分析研究者
当前深度学习与卷积神经网络（CNN）广泛应用于图像处理、分类和目标识别等领域。本文采用基于卷积神经网络的改进U-Net和VGG-UNet模型，自动从印度高分辨率遥感卫星影像中识别地物，并进行像素级分类。实验使用Cartosat 2S（约1米空间分辨率）数据集，模型对建筑物和船舶的检测准确率超过95%。另一组实验以RISAT-1的微波数据（不同分辨率）为输入，成功检测出船舶和树木，准确率高于96%。针对多类别分类任务，模型在多光谱Cartosat影像上训练，并通过真实样本验证。多标签分类结果的交并比（IoU）优于95%。共解决六类问题，根据复杂度不同，交并比在85%至98%之间。
原文摘要 · Abstract (English)
Presently, deep learning and convolutional neural networks (CNNs) are widely used in the fields of image processing, image classification, object identification and many more. In this work, we implemented convolutional neural network based modified U-Net model and VGG-UNet model to automatically identify objects from satellite imagery captured using high resolution Indian remote sensing satellites and then to pixel wise classify satellite data into various classes. In this paper, Cartosat 2S (~1m spatial resolution) datasets were used and deep learning models were implemented to detect building shapes and ships from the test datasets with an accuracy of more than 95%. In another experiment, microwave data (varied resolution) from RISAT-1 was taken as an input and ships and trees were detected with an accuracy of >96% from these datasets. For the classification of images into multiple-classes, deep learning model was trained on multispectral Cartosat images. Model generated results were then tested using ground truth. Multi-label classification results were obtained with an accuracy (IoU) of better than 95%. Total six different problems were attempted using deep learning models and IoU accuracies in the range of 85% to 98% were achieved depending on the degree of complexity.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。