Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Challenges in automatic and selective plant-clearing

Fabrice Mayran de Chamisso, Loïc Cotten, Valentine Dhers, Thomas Lompech, Florian Seywert, Arnaud Susset

TL;DR

The paper tackles automatic selective plant-clearing in forestry, aiming to detect saplings while avoiding weeds under variable weather, terrain, and lighting. It compares multispectral and RGB sensing and implements a semi-autonomous workflow combining a YOLOv8 detector with a fuzzy-annotation UNet++ segmentation, augmented by temporal stabilization and lifelong learning. RGB imagery with high-resolution sensors achieved high accuracy (near 98%) and high AUROC on large training sets, while fuzzy segmentation provided advantages with limited data; multispectral data underperformed due to background variability. The study demonstrates a practical path toward rugged, low-maintenance autonomous forestry tools, and proposes a hybrid YOLOV8–UNet approach as a future enhancement.

Abstract

With the advent of multispectral imagery and AI, there have been numerous works on automatic plant segmentation for purposes such as counting, picking, health monitoring, localized pesticide delivery, etc. In this paper, we tackle the related problem of automatic and selective plant-clearing in a sustainable forestry context, where an autonomous machine has to detect and avoid specific plants while clearing any weeds which may compete with the species being cultivated. Such an autonomous system requires a high level of robustness to weather conditions, plant variability, terrain and weeds while remaining cheap and easy to maintain. We notably discuss the lack of robustness of spectral imagery, investigate the impact of the reference database's size and discuss issues specific to AI systems operating in uncontrolled environments.

Challenges in automatic and selective plant-clearing

TL;DR

The paper tackles automatic selective plant-clearing in forestry, aiming to detect saplings while avoiding weeds under variable weather, terrain, and lighting. It compares multispectral and RGB sensing and implements a semi-autonomous workflow combining a YOLOv8 detector with a fuzzy-annotation UNet++ segmentation, augmented by temporal stabilization and lifelong learning. RGB imagery with high-resolution sensors achieved high accuracy (near 98%) and high AUROC on large training sets, while fuzzy segmentation provided advantages with limited data; multispectral data underperformed due to background variability. The study demonstrates a practical path toward rugged, low-maintenance autonomous forestry tools, and proposes a hybrid YOLOV8–UNet approach as a future enhancement.

Abstract

With the advent of multispectral imagery and AI, there have been numerous works on automatic plant segmentation for purposes such as counting, picking, health monitoring, localized pesticide delivery, etc. In this paper, we tackle the related problem of automatic and selective plant-clearing in a sustainable forestry context, where an autonomous machine has to detect and avoid specific plants while clearing any weeds which may compete with the species being cultivated. Such an autonomous system requires a high level of robustness to weather conditions, plant variability, terrain and weeds while remaining cheap and easy to maintain. We notably discuss the lack of robustness of spectral imagery, investigate the impact of the reference database's size and discuss issues specific to AI systems operating in uncontrolled environments.
Paper Structure (11 sections, 8 figures, 1 table)

This paper contains 11 sections, 8 figures, 1 table.

Table of Contents

  1. Introduction - challenges
  2. Vision system design
  3. Sensor choice: multispectral versus classical imagery
  4. Plant segmentation with deep learning
  5. Data handling and learning
  6. Results and analysis
  7. Temporal aspects
  8. Temporal stabilization
  9. Tool control
  10. Lifelong learning
  11. Conclusion

Figures (8)

  • Figure 1: Top: Typical existing tractor with retractable cutting tool. Bottom: Ruggedized acquisition system attached to a tractor for data collection.
  • Figure 2: Saplings before planting measure between a few centimeters and a few dozens of centimeters.
  • Figure 3: Evolution of acquisition systems. a) initial acquisition system, consisting of two manually triggered IMEC multispectral cameras and one manually triggered industrial global shutter FLIR camera with a SONY IMX sensor. Multispectral cameras have a resolution of 512x256x16 in the visible spectrum and 409x216x25 in the near infrared spectrum. This setup was complemented with a customer grade compact camera and a smartphone sensor. b) database construction system idea, with a sensor mounted on a belt-worn device, a feedback screen and a control panel. In the end, this system was simplified to a wrist-strapped $10$inch monitor, a hand-held camera and a single button to take a picture or a video. c) the ruggedized acquisition system used for all-terrain all-season operation, with a builtin illumination system for low-light operation.
  • Figure 4: In the laboratory vs reality. Top: in a controlled environment, spectral segmentation shows promising results. Bottom: in reality, intraclass variability is similar to interclass variability, with uncontrolled lighting.
  • Figure 5: Without fixing the white balance, sensors may produce weird looking images. There is no purple plant or soil here in reality.
  • ...and 3 more figures