高分辨率遥感影像解译中的机器学习范式

High-resolution remote sensing image interpretation is a major topic in remote sensing information processing. It plays a vital role in the knowledge mining and intelligent analysis of remote sensing big data and has important application values in civil and military fields. The traditional methods of high-resolution remote sensing image interpretation generally use manual visual interpretation, which is time consuming and laborious and has low accuracy. Therefore, interpreting high-resolution remote sensing images automatically and efficiently is an urgent problem to be solved. The rapid development of artificial intelligence technology in recent years has made machine learning the mainstream research direction of high-resolution remote sensing image interpretation. In this study, we systematically review five kinds of representative machine learning paradigms on the basis of the typical tasks of high-resolution remote sensing image interpretation, such as object detection, scene classification, semantic segmentation, and hyperspectral image classification. Specifically, we introduce their definitions, typical methods, and applications. The representative machine learning paradigms include supervised learning (e.g., support vector machine, k-nearest neighbor, decision tree, random tree, and probabilistic graph model), semi-supervised learning (e.g., pure semi-supervised learning, transductive learning, and active learning), weakly supervised learning (e.g., multiple instance learning), unsupervised learning (e.g., clustering, principal component analysis, and sparse coding), and deep learning (e.g., stacked auto-encoder, deep belief network, convolutional neural network, and generative adversarial network). Then, we comprehensively analyze the strengths and limitations of the five kinds of machine learning paradigms and summarize their typical applications in remote sensing image interpretation. Finally, we summarize the development direction of high-resolution remote sensing image interpretation, such as few-shot learning, unsupervised deep learning, and reinforcement learning.