Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Active Informative Planning for UAV-based Weed Mapping using Discrete Gaussian Process Representations

Jacob Swindell, Marija Popović, Riccardo Polvara

TL;DR

The paper addresses efficient UAV weed mapping by modeling weed density as a continuous 2D GP and applying informative path planning to collect data within a flight-time budget. It systematically evaluates six discrete GP representations (Quadtree, Wedgelets, BSP LSE, BSP Region, Hexagon, Voronoi) and embeds them in an online receding-horizon IPP, comparing offline map fidelity against high-resolution ground truth and online mission-level performance using a four-step planning horizon and a 40-minute budget. The results show that discretisation is a major design choice, with Quadtree excelling in early coverage and some fidelity metrics, while BSP Region and Voronoi perform best under dominance or highly heterogeneous distributions; Hexagon and Wedgelet offer context-dependent benefits. The work provides practical guidelines for selecting GP discretisation strategies based on weed distribution characteristics, and emphasizes the need to co-design discretisation with planners and GP inference to maximize information gain within resource limits, with extensions to more datasets and altitude-aware planning planned for future work.

Abstract

Accurate agricultural weed mapping using unmanned aerial vehicles (UAVs) is crucial for precision farming. While traditional methods rely on rigid, pre-defined flight paths and intensive offline processing, informative path planning (IPP) offers a way to collect data adaptively where it is most needed. Gaussian process (GP) mapping provides a continuous model of weed distribution with built-in uncertainty. However, GPs must be discretised for practical use in autonomous planning. Many discretisation techniques exist, but the impact of discrete representation choice remains poorly understood. This paper investigates how different discrete GP representations influence both mapping quality and mission-level performance in UAV-based weed mapping. Considering a UAV equipped with a downward-facing camera, we implement a receding-horizon IPP strategy that selects sampling locations based on the map uncertainty, travel cost, and coverage penalties. We investigate multiple discretisation strategies for representing the GP posterior and use their induced map partitions to generate candidate viewpoints for planning. Experiments on real-world weed distributions show that representation choice significantly affects exploration behaviour and efficiency. Overall, our results demonstrate that discretisation is not only a representational detail but a key design choice that shapes planning dynamics, coverage efficiency, and computational load in online UAV weed mapping.

Active Informative Planning for UAV-based Weed Mapping using Discrete Gaussian Process Representations

TL;DR

The paper addresses efficient UAV weed mapping by modeling weed density as a continuous 2D GP and applying informative path planning to collect data within a flight-time budget. It systematically evaluates six discrete GP representations (Quadtree, Wedgelets, BSP LSE, BSP Region, Hexagon, Voronoi) and embeds them in an online receding-horizon IPP, comparing offline map fidelity against high-resolution ground truth and online mission-level performance using a four-step planning horizon and a 40-minute budget. The results show that discretisation is a major design choice, with Quadtree excelling in early coverage and some fidelity metrics, while BSP Region and Voronoi perform best under dominance or highly heterogeneous distributions; Hexagon and Wedgelet offer context-dependent benefits. The work provides practical guidelines for selecting GP discretisation strategies based on weed distribution characteristics, and emphasizes the need to co-design discretisation with planners and GP inference to maximize information gain within resource limits, with extensions to more datasets and altitude-aware planning planned for future work.

Abstract

Accurate agricultural weed mapping using unmanned aerial vehicles (UAVs) is crucial for precision farming. While traditional methods rely on rigid, pre-defined flight paths and intensive offline processing, informative path planning (IPP) offers a way to collect data adaptively where it is most needed. Gaussian process (GP) mapping provides a continuous model of weed distribution with built-in uncertainty. However, GPs must be discretised for practical use in autonomous planning. Many discretisation techniques exist, but the impact of discrete representation choice remains poorly understood. This paper investigates how different discrete GP representations influence both mapping quality and mission-level performance in UAV-based weed mapping. Considering a UAV equipped with a downward-facing camera, we implement a receding-horizon IPP strategy that selects sampling locations based on the map uncertainty, travel cost, and coverage penalties. We investigate multiple discretisation strategies for representing the GP posterior and use their induced map partitions to generate candidate viewpoints for planning. Experiments on real-world weed distributions show that representation choice significantly affects exploration behaviour and efficiency. Overall, our results demonstrate that discretisation is not only a representational detail but a key design choice that shapes planning dynamics, coverage efficiency, and computational load in online UAV weed mapping.
Paper Structure (27 sections, 8 equations, 5 figures, 5 tables)

This paper contains 27 sections, 8 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Methodology
  4. Representations
  5. Dataset / Data Processing
  6. Continuous Weed Distribution
  7. Discrete Representations
  8. Evaluation Metrics
  9. Exploration Pipeline
  10. Pipeline Overview
  11. Gaussian Process
  12. Planning Algorithm
  13. Evaluation Metrics
  14. Experimental Evaluation
  15. Representations
  16. ...and 12 more sections

Figures (5)

  • Figure 1: Methodological overview of the adaptive weed mapping pipeline. A UAV incrementally acquires observations that update a Gaussian process model of weed distribution. The continuous field is discretised using different spatial representations, which generate candidate viewpoints that inform a planning module for autonomous exploration. Evaluation is performed on a fixed grid, allowing representation-independent comparison in terms of coverage, uncertainty reduction, and reconstruction accuracy.
  • Figure 2: Representations for 000_gt orthomosaic in the WeedMap dataset. Bright spots show a high weed presence, while dark spots show low weed presence.
  • Figure 3: Comparison of mission performance metrics over time for the IPP system using Quadtree and Voronoi spatial representations for candidate view generation.
  • Figure 4: Comparison of UAV trajectories generated by the IPP framework using Quadtree and Voronoi representations for candidate view generation. The mission start and end locations are denoted by green and red markers, respectively.
  • Figure 5: Breakdown of the time taken on different tasks for each representation. Graphs show time taken for: Generating / Training the GP, Generating the Discrete Representation, Planner choosing the next position and Travel Time assuming a constant speed of 2m/s.