Crop Agnostic Monitoring Driven by Deep Learning.

Area of Science:

• Agricultural Robotics
• Computer Vision
• Precision Agriculture

Background:

• Farmers need detailed crop management information, often gathered manually, which is time-consuming and costly.
• Current robotic platforms and AI algorithms are often limited to specific environments (e.g., farmland) or single platforms, creating a technology gap.

Purpose of the Study:

• To develop a novel, field-agnostic monitoring technique applicable to multiple robotic platforms and diverse agricultural settings (arable land and greenhouses).
• To enable accurate crop and weed identification, yield estimation, and ripeness assessment for enhanced agronomic decision-making.

Main Methods:

• Utilized instance segmentation as the core technique for object detection, classification, and area estimation.
• Developed a novel matching criterion for yield estimation, removing previous pixel-wise constraints.
• Validated the approach on two different robotic platforms in both arable farmland and glasshouse environments.

Main Results:

• Accurate estimation of sweet pepper count in greenhouses with 4.7% normalized absolute error and R² of 0.901.
• Improved performance by 50% in cluttered arable farmland scenes compared to prior methods.
• Successfully identified crop and weed species in arable farmland and sweet pepper ripeness in glasshouses.

Conclusions:

• The proposed field-agnostic monitoring technique offers a versatile solution for precision agriculture, adaptable to various robotic platforms and settings.
• The system provides farmers with actionable data for tasks like weed intervention, optimizing crop management.
• Instance segmentation and novel yield estimation methods are key to achieving high accuracy in diverse agricultural environments.