기계학습 기반 수확상태 모니터링을 통한 트랙터 부착 감자수확기의 굴취 깊이 자동제어 시스템 개발 = Development of automatic digging depth control system for tractor-mounted potato harvester using machine learning-based harvest status monitoring

다국어 초록 (Multilingual Abstract)

Potato is one of the four most widely produced staple food crops worldwide, along with rice, maize, and wheat, due to its high yield and productivity per unit area. In Korea, the potato cultivation area has decreased from approximately 51,752 ha in 1975 to 23,573 ha in 2024, while potato imports have increased at an average annual rate of 5.6% over the past decade. Under the ongoing aging of the agricultural workforce, the development of agricultural machinery technology suited to domestic conditions is essential to improve potato self-sufficiency.
Smart agriculture is transforming traditional agricultural operations into automated, digital, and unmanned systems by integrating sensors and communication technologies into agricultural machinery and facilities. With the incorporation of artificial intelligence, smart agriculture aims to achieve management rationalization, high-quality production, and environmental protection. In open-field agriculture, various technologies such as autonomous driving of agricultural machinery, data-driven implement control, and tractor–implement management—have been developed, enabling farming practices that were traditionally dependent on operators’ skills to become more efficient, rapid, cost-effective, and less reliant on operator expertise.
The digging type potato harvester mounted on tractor which is widely accepted in Korea shows notable changes in potato yield, working speed and traction load depending on digging depth. In addition, the posture of an operator to control the digging depth while driving a tractor is burdensome to the musculoskeletal system, so considering the aging labor force, a smart automatic digging depth control system suitable for our conditions is required
Although several studies on the automatic digging depth control of harvester for underground crops have been reported at home and abroad, there has been no report on smart automatic digging depth control which integrates communication networks of the implement and that of the tractor for the tractor-mounted underground crop harvester. Also, a reliable method of estimating the digging depth on the irregularly deformed soil surface by the digging blade has not been yet developed, even though various types of non-contact distance sensor are available. Furthermore, there has been no report on an automatic control system that corrects the initially set digging depth by monitoring the harvesting condition in real time.
The objectives of this study are to develop a technology for estimating the digging depth based on the deposited soil profile in a working environment where direct measurement of the digging depth is difficult due to soil accumulation during the digging process, to develop a technology for real-time recognition of harvesting conditions, and to develop an advanced system that integrates the two technologies to overcome the limitations of existing studies, as well as to evaluate its performance. For estimating the digging depth, the depth map data were matched with the depth measured on the reference surface. Based on this method, a technology capable of estimating the digging depth even at digging points where soil accumulation occurs was developed. For real-time recognition of damaged potatoes, a deep learning-based monitoring technology that classifies normal and damaged potatoes and counts each category was developed.
For integration of the two technologies, an algorithm capable of correcting the initially set digging depth was developed by applying the previously developed real-time harvesting condition monitoring technology. In addition, by partially adopting the concept of the ISO-defined Tractor-implement management (TIM) system, a network system was designed and developed to enable the implementation of TIM even on tractors that do not comply with the ISO 11783 standard. Field experiments were conducted using a tractor without ISOBUS functionality, and field performance evaluations were carried out to verify the effectiveness of the proposed system.
To develop a method for estimating the actual digging depth at the working point, depth map data and reference surface–based depth measurements were simultaneously collected during actual harvesting operations. Three models—Linear Regression, Long Short-Term Memory (LSTM), and Multi-Layer Perceptron (MLP)—were selected as digging depth estimation algorithms, and their performances were compared using field experimental data. Based on the test dataset, the mean absolute error (MAE) was 2.06 cm for LSTM, 2.39 cm for MLP, and 3.54 cm for Linear Regression. In field validation experiments, the MAE values were 2.82 cm for MLP, 3.24 cm for LSTM, and 4.54 cm for Linear Regression. The MLP model, which demonstrated the best performance under field conditions reflecting actual operating environments, was selected as the final digging depth estimation model, confirming its suitability for real-time automatic digging depth control.
To assess the operational condition of potatoes during harvesting, a camera was installed on the potato harvester to collect images of potatoes conveyed on the conveyor during harvesting operations. In addition, 4K high-resolution images were acquired to analyze the effect of image resolution on object detection and classification performance. Two lightweight deep learning models, YOLOv8n and YOLOv10n, were selected for real-time potato detection and classification of detected potatoes into damaged and normal categories, and their performances were evaluated according to training dataset configuration and model architecture. Three datasets were constructed: a dataset consisting of application-resolution images only (Set 1), a dataset combining application-resolution and high-resolution images at an equal ratio (Set 2), and an expanded dataset in which high-resolution images were added to the application-resolution dataset (Set 3). YOLO-based models were trained under each condition. As a result, Model 5, which is based on YOLOv8n and trained on Set 3, exhibited the best performance, achieving a Precision of 0.954, Recall of 0.910, mAP@0.5 of 0.954, and mAP@0.5:0.95 of 0.770.The processing time per frame was 42.7 ms, confirming stable real-time detection of potato objects and reliable classification of damage status.
An integrated automatic digging depth control system was designed and implemented by partially adopting the concept of the ISO-defined Tractor–Implement Management (TIM) framework, allowing the digging depth estimation algorithm and the real-time harvesting condition monitoring algorithm to be reflected in tractor control. To ensure applicability to tractors without ISOBUS functionality, the system was configured such that the implement-side CCU performs digging depth estimation, harvesting condition evaluation, and depth correction decision-making, while the resulting commands are applied to the tractor three-point hitch lift control. Field experiments were conducted under four operating conditions: fixed digging depths of 18 cm, 21 cm, and 24 cm, and an automatic correction condition with an initial digging depth setting of 18 cm. The experimental results showed that the effective threshold exceedance ratio was 29.7% under the 18 cm fixed condition, 54.8% under the 21 cm fixed condition, and 50.1% under the 24 cm fixed condition. Under the automatic correction condition with an initial setting of 18 cm, the effective threshold exceedance ratio was 45.1%. In terms of harvesting performance, the harvest rates under the 18 cm and 21 cm fixed conditions were 63.9% and 66.2%, respectively. The harvest rate under the 24 cm fixed condition was 92.39%, while the automatic correction condition yielded a harvest rate of 90.81%. No significant difference was observed between the harvest rate of the 24 cm fixed condition and that of the automatic correction condition, indicating that when the initial digging depth is appropriately set, the contribution of the digging depth correction system is limited. These results indicate that the digging depth correction system primarily functions to compensate for deviations or errors in the initially set digging depth during operation, rather than providing additional performance improvements beyond those achieved through appropriate initial depth settings.
This study demonstrated the technical feasibility of automating digging depth adjustment in open-field agriculture by equipping a tractor-mounted potato harvester with cameras, a depth camera, an HMI for implement control configuration, and a CCU incorporating digging depth estimation algorithms, potato detection and classification algorithms, and control programs. Through the application of smart technologies, the study verified the potential to automate digging depth adjustment tasks that previously relied on operator experience and skill, thereby improving harvesting performance. Although the developed digging depth estimation algorithm and potato detection and classification algorithm can be applied to crops other than potatoes, achieving higher performance at increased operating speeds will require improvements in peripheral devices such as cameras and computing hardware. Furthermore, the field experiments conducted in this study did not evaluate the effects of soil type and soil condition on potato harvester performance, nor did they incorporate engine load or travel speed information into harvester control. Therefore, additional studies are required to address these limitations and to facilitate the commercialization of the proposed technologies.

목차 (Table of Contents)

• 1. 서론 1
• 1.1 연구 배경 1
• 1.2 연구 목적 7
• 2. 문헌 연구 9
• 2.1 노지 스마트 농업 9
• 2.2 감자 수확기계 12
• 2.3 굴취 깊이가 감자수확에 미치는 영향 16
• 2.4 굴취 깊이 자동제어 시스템 18
• 2.5 Tractor-implementmanagement 22
• 2.6 Depthmapdata활용 24
• 2.7 객체검출 기반 수확 중 작물 상태 및 손상 모니터링 27
• 3. Depth map 기반 굴취 깊이 추정 기술 개발 29
• 3.1 서론 29
• 3.2 재료 및 방법 33
• 3.3 결과 및 토론 49
• 3.4 결론 55
• 4. 인공지능 객체인식 기반 작업상태 모니터링 기술 개발 56
• 4.1 서론 56
• 4.2 재료 및 방법 59
• 4.3 결과 및 토론 71
• 4.4 결론 76
• 5. 통합 시스템 개발 77
• 5.1 서론 77
• 5.2 재료 및 방법 80
• 5.3 결과 및 결론 101
• 5.4 결론 107
• 6. 종합요약 및 결론 109
• Reference 112
• 부 록 116
• 感謝의 글 126