KG-MS-ResNet: A Knowledge-Guided Multi-Scale Attention Residual Network for Cultivated Land Change Monitoring

Abstract. Cultivated land conversion to built-up area is the core form of farmland non-agriculturalization and the main threat to farmland protection. China has strict requirements for safeguarding farmland security and holding the 1.8-billion-mu farmland red line. However, current remote sensing monitoring methods for cultivated land non-agriculturalization have two limitations: insufficient integration of domain prior knowledge, and the inability of purely data-driven models to achieve high Precision and Recall. To tackle these challenges, this study proposes a knowledge graph-enhanced method for cultivated land–built-up area change detection. First, a multi-scale knowledge analysis framework with feature, scene, and business knowledge layers is constructed, which integrates topographic factors, ecological indicators, and land surface indices, to form the semantic representation of domain knowledge. On this basis, KG-MS-ResNet, a knowledge fusion residual network is designed. It adopts ResNet-18 as the backbone and modifies the first convolutional layer to accommodate bi-temporal image inputs. TransE is employed to embed geographic indicators knowledge into multi-scale semantic vectors. A semantic–feature dual-path fusion strategy and a knowledge-guided attention mechanism are further proposed, enabling deep coupling and collaborative optimization between image features and domain prior knowledge. Experiments conducted in Pei County, Jiangsu Province, demonstrate that the proposed method outperforms the baseline ResNet across all evaluation metrics, with 4.84% higher Recall and 0.0752 higher F1-score. The proposed method effectively enhances the detection capability of true change regions. The results demonstrate that deeply integrating domain knowledge graphs with deep learning models can improve the performance of cultivated land–built-up area change detection.