[Paper Review] Active Learning with Gaussian Processes for High Throughput Phenotyping

Sumit Kumar, Wenhao Luo+2 more|arXiv (Cornell University)|Jan 21, 2019

Smart Agriculture and AI21 references4 citations

TL;DR

This paper proposes an active learning framework using Gaussian Processes (GPs) that enables autonomous ground robots to efficiently map plant phenotypes in large agricultural fields by intelligently balancing high-accuracy static measurements and lower-quality mobile measurements collected during motion. The method achieves comparable predictive accuracy to exhaustive sampling with significantly reduced data collection time, demonstrating superior performance on real sorghum phenotype data.

ABSTRACT

A looming question that must be solved before robotic plant phenotyping capabilities can have significant impact to crop improvement programs is scalability. High Throughput Phenotyping (HTP) uses robotic technologies to analyze crops in order to determine species with favorable traits, however, the current practices rely on exhaustive coverage and data collection from the entire crop field being monitored under the breeding experiment. This works well in relatively small agricultural fields but can not be scaled to the larger ones, thus limiting the progress of genetics research. In this work, we propose an active learning algorithm to enable an autonomous system to collect the most informative samples in order to accurately learn the distribution of phenotypes in the field with the help of a Gaussian Process model. We demonstrate the superior performance of our proposed algorithm compared to the current practices on sorghum phenotype data collection.

Motivation & Objective

• To address the scalability bottleneck in high-throughput phenotyping (HTP) caused by exhaustive data collection in large fields.
• To develop an active learning algorithm that selects informative sampling locations to minimize data collection time while maintaining model accuracy.
• To incorporate both static (high-accuracy) and mobile (lower-accuracy, motion-based) measurements into a unified GP-based learning framework.
• To design an informative path planning algorithm that maximizes joint information gain from both static and mobile measurements.
• To evaluate the method on real-world sorghum phenotype data and compare it with conventional exhaustive sampling approaches.

Proposed method

• The system models the spatial distribution of a target phenotype (e.g., plant height) using a Gaussian Process (GP) with a kernel function that accounts for spatial correlation.
• Two types of measurements are modeled: static measurements (accurate, collected when robot stops) and mobile measurements (noisy, collected during motion), each with distinct noise levels.
• A utility function based on entropy reduction (MaxEnt) is used to select the next sampling locations that maximize information gain about the underlying phenotype distribution.
• An informative path planning algorithm computes the optimal trajectory that maximizes joint information gain from both static and mobile measurements along the path.
• The algorithm dynamically balances exploration (new informative plots) and exploitation (high-uncertainty areas) using a budget-constrained path search with slack length (ξ) to allow flexibility.
• The framework is evaluated using a real sorghum dataset collected by a robotic platform, with performance compared against exhaustive sampling and baseline active learning methods.

Experimental results

Research questions

• RQ1 Can an active learning approach with GPs reduce data collection time in HTP while maintaining predictive accuracy compared to exhaustive sampling?
• RQ2 How does the integration of mobile measurements (collected during motion) affect model accuracy compared to relying only on static measurements?
• RQ3 What is the optimal trade-off between measurement quality (static vs. mobile) and data collection efficiency in large-scale field phenotyping?
• RQ4 How does the inclusion of path planning slack (ξ) influence model performance and computational cost?
• RQ5 To what extent can a hybrid static-mobile measurement strategy achieve performance comparable to full-field scanning?

Key findings

• The proposed active learning algorithm achieved comparable predictive accuracy to exhaustive sampling, with a mean absolute error (MAE) of 4.22 on the test set, while significantly reducing data collection time.
• The model maintained high accuracy even when the mobile measurement noise ratio (σₘ/σₛ) reached up to 5, indicating that mobile data can be effectively leveraged without degrading performance.
• Increasing the path planning slack (ξ) improved model performance slightly, but the improvement was non-monotonic, suggesting that longer paths do not always yield better information due to spatial correlation in GP kernels.
• A small additional budget (ξ=5) provided meaningful performance gains with manageable computational cost, while larger budgets led to diminishing returns and increased computation.
• The results demonstrate that intelligent combination of static and mobile measurements enables accurate phenotype mapping with minimal resource use, making large-scale HTP feasible.
• The method outperformed current practices in HTP, showing that active learning with GPs can accelerate breeding pipelines by reducing the need for exhaustive field coverage.