Microscopic detection of topsoil aggregate size based on LMTPI auto-DL framework in field environment

The topsoil aggregates diameter size (ADS) distribution reflects the micromorphological formation mechanisms of soil layers and serves as a quantitative indicator of pore structure, water retention, and soil mechanical properties. To further improve the efficiency, field safety, and parameter standardization of ADS analysis, this study proposes an auto-detection framework by integrating an improved LMTPI strategy, auto-DL, and model transfer in field. First, the vanishing maximum phenomenon was discovered in MTPI (Maximum Transfer Potential Index) auto-DL framework, so the improved LMTPI (Local MTPI) was used to extend its applicability to microscopic ADS datasets. The results demonstrate that the improved LMTPI reduces the required training dataset to 10.40% while decreasing the training time by 89.6%, which drives the subsequent auto-DL optimization process. Subsequently, an auto-DL framework based on LMTPI was employed to train an optimal detection model for ADS, and its results indicate that the automatically selected best Seg-ResNet50, with a mean Average Precision ( mAP ) of 98.43% and only 40.99M parameters. Finally, using a migration dataset including lighting changes, shooting angles, and foreign object interference, LMTPI-based transfer learning was further optimized for the pretrained model, where results show that model achieves 93.08% accuracy in indoor simulation environments and maintains 90.73% detection accuracy in actual field deployment of wheeled-legged robot. This research provides a technical reference for the auto-DL on-site soil detection and environmental assessments.