Abstract: The accuracy of patch investigation and monitoring is closely linked to the scientific rigor and effectiveness of resource management decisions. Conventional quality inspection methods suffer from limited efficiency and reliability, failing to meet the demands for efficient and comprehensive quality inspections of patches. From the perspective of quality inspection applications, this study proposed a technical framework for patch anomaly detection based on prior category knowledge and convolutional neural networks. By integrating prior category knowledge into image data, the technical framework can effectively guide the anomaly detection model, enabling the efficient detection and precise localization of missed patches and the incorrect category attribution of patches. Additionally, a method for preparing quality problem samples was introduced, addressing the challenge of efficiently acquiring such samples. Experimental results obtained from multiple regions across China indicate that the proposed technical framework achieved a recall rate of up to 97.3% for omissions of changes in construction land (including dumping areas). Compared to the semantic segmentation algorithm, the technical framework improved the precision rate and the intersection over union by 11.6 percentage points and 5.9 percentage points, respectively while reducing the workload associated with manual investigations by 41.9%. Overall, the proposed technical framework significantly enhances patch anomaly detection capabilities, providing an innovative approach for patch anomaly detection in practical application scenarios.