A Small-Sample Tillage Depth Recognition and Detection Method Integrating Multi-Scale Features and Physical Constraints

In China, tillage depth is a core performance indicator for certificating tillage machinery. The current manual measurement in field suffers from high subjectivity and poor traceability. This study proposed a tillage depth detection method called MSKe_PC_Transformer (Multi-Scale Kalman-enhanced Physical constraint Transformer). Multi-scale Kalman filtering extracts macroscopic trends, mesoscale fluctuations, and microscale details from soil penetration resistance sequences to construct a multi-scale feature representation. An attention-gating mechanism dynamically and adaptively fuses these features across scales. A physical constraint loss function based on prior knowledge of soil mechanics ensures that the model’s output conforms to the laws of soil mechanical behavior. Using custom-developed equipment, 99 sets of laboratory data and 300 sets of field data were collected for training and testing the MSKe_PC_Transformer model, which achieved an accuracy of 92.59% and a recall of 90.35%. Ablation experiments confirmed the contributions and necessity of each module. In field tests conducted in two regions, the accuracy rate for detection errors less than 1.5 cm was 93%, with the MAE and RMSE 1.03 cm and 1.19 cm, respectively. The results confirm the feasibility of deploying the proposed method as an objective and traceable alternative to manual inspection in tillage machinery certification. The established framework is extendable to other implements, such as subsoilers and moldboard plows, supporting the broader standardization of agricultural machinery certification in China.