Identification of Illegal Urban Construction Based on VLMs and Remote Sensing Imagery

Abstract

Urbanization has led to a continuous increase in the demand for investigating illegal buildings, and manual inspection and drone patrol have shortcomings such as efficiency and airspace limitations. This paper researches urban illegal construction identification based on high-resolution remote sensing imagery and visual language models (Vision-Language Models, VLM). The LEVIR-CD public remote sensing dataset and the Qwen3.6-PLUS model were adopted. Multi-temporal remote sensing data and urban planning data are integrated for building a Visual Question Answering (VQA) judgment system. The experiment consisted of 10 training samples and 8 test samples. The model achieved a recognition accuracy of 62.5%, making it difficult to identify illegal buildings. Analysis revealed that insufficient sample size, poor model fit, and an unreasonable guided learning scheme were the contributing factors to insufficient accuracy. This study confirms the feasibility and long-term development prospects of this technical approach. Subsequent data expansion and algorithm optimization aim to improve the accuracy of urban building identification and illegal construction determination by training the Visual Language Model (VLM) to better adapt to high-resolution remote sensing images. At the same time, it aims to improve the stability of artificial intelligence technology in the field of urban illegal construction identification and processing, and reduce the proportion of manual identification.