[Paper Review] Towards practical object detection for weed spraying in precision agriculture

Adrian Salazar-Gomez, Madeleine Darbyshire+3 more|arXiv (Cornell University)|Sep 22, 2021

Smart Agriculture and AI12 citations

TL;DR

This paper proposes a practical framework for object detection in precision agriculture weed spraying by introducing three new metrics—weed coverage rate (WCR), area sprayed, and inference speed—to evaluate real-world feasibility beyond traditional mAP. Using YOLOv5m and Faster R-CNN models on real farm data, it demonstrates that models with lower mAP can outperform in practical spraying efficiency, with Faster R-CNN achieving 85.78% WCR while spraying only 17.08% of the area, significantly reducing herbicide use.

ABSTRACT

The evolution of smaller, faster processors and cheaper digital storage mechanisms across the last 4-5 decades has vastly increased the opportunity to integrate intelligent technologies in a wide range of practical environments to address a broad spectrum of tasks. One exciting application domain for such technologies is precision agriculture, where the ability to integrate on-board machine vision with data-driven actuation means that farmers can make decisions about crop care and harvesting at the level of the individual plant rather than the whole field. This makes sense both economically and environmentally. However, the key driver for this capability is fast and robust machine vision -- typically driven by machine learning (ML) solutions and dependent on accurate modelling. One critical challenge is that the bulk of ML-based vision research considers only metrics that evaluate the accuracy of object detection and do not assess practical factors. This paper introduces three metrics that highlight different aspects relevant for real-world deployment of precision weeding and demonstrates their utility through experimental results.

Motivation & Objective

• To address the gap in evaluating machine learning models for precision agriculture weed spraying beyond standard accuracy metrics.
• To introduce and validate three new practical metrics—weed coverage rate (WCR), area sprayed, and inference speed—for assessing real-world feasibility of object detection models.
• To demonstrate that models with lower mAP can offer superior practical performance in herbicide reduction and processing speed, enabling efficient on-the-fly spraying.
• To provide a decision framework for selecting optimal models based on trade-offs between detection accuracy, herbicide efficiency, and computational cost.

Proposed method

• The authors developed a system using a tractor-mounted sprayer with 5 cameras and nozzles, capturing real-time images during field operation.
• They applied state-of-the-art object detection models—including YOLOv5m, YOLOv3, Faster R-CNN, and SSD—on two datasets: BB (broadleaf weeds) and LB (low-visibility weeds).
• The weed coverage rate (WCR) was calculated as the percentage of weeds that could be targeted by the sprayer based on predicted bounding boxes and nozzle resolution.
• The area sprayed metric was computed as the total area covered by spray zones, assuming rectangular spray patterns overlapping detected weeds.
• Inference speed was measured in frames per second (FPS) on a single GPU to assess real-time processing feasibility.
• The models were evaluated across varying batch sizes (1 to 44) to simulate multi-camera processing and assess scalability on limited onboard hardware.

Experimental results

Research questions

• RQ1 How do traditional mAP-based metrics compare to practical metrics like WCR and area sprayed in evaluating object detection models for precision weed spraying?
• RQ2 Which object detection model achieves the best balance between weed coverage (WCR) and herbicide efficiency (area sprayed) in real-world field conditions?
• RQ3 What is the trade-off between model accuracy (mAP) and inference speed, and how does this affect real-time deployment on embedded systems in agricultural vehicles?
• RQ4 How does increasing batch size affect frame rate and feasibility of processing multiple camera streams in real time?
• RQ5 To what extent do non-contiguous image frames and simplified spray geometry assumptions limit the accuracy of WCR and area sprayed estimates?

Key findings

• Faster R-CNN (ResNet-50-FPN) achieved a WCR of 85.78% on the BB dataset while spraying only 17.08% of the area, significantly outperforming YOLOv5m in herbicide efficiency despite lower mAP.
• On the LB dataset, SSD achieved 96.94% WCR with only 35.36% area sprayed, demonstrating a strong trade-off between coverage and waste, outperforming YOLOv5m (68.6% area sprayed) despite slightly lower WCR.
• YOLOv3 achieved the fastest inference speed at 75.18 FPS (batch size 1), making it suitable for real-time processing, though with lower detection accuracy than YOLOv5m.
• With batch size 44, YOLOv5m could achieve 277–333 FPS on a single GPU, suggesting that 3–4 GPUs would make real-time spraying feasible at 15 mph, though still below the required 968–1027 FPS.
• The highest WCR on the BB dataset was 88.2% (YOLOv5m, 3 nozzles), but this came at the cost of spraying 33.03% of the area, indicating high waste.
• The study found that models with the highest mAP (e.g., YOLOv5m) do not necessarily offer the best practical performance, as WCR and area sprayed reveal more meaningful trade-offs for real-world deployment.