Design and field validation of a vision-based graded compaction intensity control system for wheat seeding

Spatial heterogeneity in surface soil structure makes it difficult for conventional seeder press-wheel devices to match compaction intensity to local seedbed conditions. We developed a vision-perception-driven system that classifies soil surfaces into small-clod regions, medium-clod regions, and large-clod regions and maps these classes to light, moderate, and heavy compaction commands. Surface images acquired ahead of the press wheel were processed using an improved YOLOv11n model incorporating C3k2_CoT, DyHead, and Focaler-DIoU. A regional evaluation index combining clod area and number proportions was then used to generate graded displacement commands for an electric linear actuator. The improved detector achieved a precision of 86.3%, recall of 90.8%, mAP50 of 95.4%, and inference speed of 41.02 FPS. Relative to baseline YOLOv11n, recall and mAP50 increased by 4.9 and 1.1 percentage points, respectively. Across three forward speeds and three soil-region classes, actuator displacement tracking accuracy ranged from 86.11% to 94.44%, with mean absolute errors (MAE) of 0.69–1.14 mm and root mean square errors (RMSE) of 1.04–1.36 mm. Separate fixed-extension field trials showed a nonlinear effect of actuator extension on seeding quality. Within the tested conditions, a forward speed of 1.5 m·s −1 and an actuator extension of 25 mm produced a 100% qualified sowing-depth rate and a 92.73% sowing-depth stability coefficient. These results demonstrate the feasibility of integrating pre-compaction soil-structure perception with graded actuator control under the tested conditions, providing experimental evidence for adaptive compaction regulation in wheat seeding equipment.