Weakly Supervised Learning in Urban Remote Sensing: Concepts, methods, and applications

Urban remote sensing has become increasingly important, but the lack of large-scale high-quality annotated data has severely constrained the development of intelligent interpretation methods. This challenge is further aggravated by the complexity of urban scenes, including strong material heterogeneity, diverse 2D/3D spatial patterns, and high temporal dynamics. In this context, weakly supervised learning (WSL) has emerged as a promising paradigm for reducing annotation costs while enabling effective model training under limited supervision. This article presents a systematic review of WSL in urban remote sensing, organized from fundamental concepts to methodological paradigms and practical applications. Specifically, we first introduce the fundamental concepts of WSL and summarize three weak supervision scenarios: incomplete, inexact, and inaccurate supervision. We then review nine representative methodological categories and their major development branches, highlighting how different methods address the distinctive challenges of urban remote sensing. Furthermore, we examine the applications of WSL across seven major urban domains, thereby clarifying the relationship between methodological development and real-world urban tasks. Finally, we emphasize the need for benchmark datasets specifically designed for weakly supervised urban remote sensing to support fair and standardized evaluation, and we discuss several promising future directions, including foundation models (FMs), multisource and multimodal fusion, and continual learning.