Variation in forest root image annotation by experts, novices, and AI

Area of Science:

• Ecology
• Plant Science
• Soil Science

Background:

• Manual root dynamics studies are time-intensive and subject to annotator bias.
• Artificial intelligence (AI) tools show promise for image analysis but require testing in heterogeneous, non-agricultural soils.
• Forest root dynamics are crucial for understanding the global carbon cycle.

Purpose of the Study:

• Quantify human annotator variance in root length measurement.
• Evaluate a convolutional neural network (CNN) for root length analysis in forest soil images.
• Assess AI model performance in complex, heterogeneous environments.

Main Methods:

• Collected minirhizotron images from a mature deciduous temperate forest.
• Assessed root length annotation variability among human annotators with different experience levels.
• Applied a CNN model trained on accessible software to the image dataset.

Main Results:

• Human annotator experience significantly impacted root length measurements.
• The CNN model overestimated root length compared to expert manual annotation (p=0.01).
• CNN root length change estimates were closer to manual but still showed variation.

Conclusions:

• Manual root annotation is subjective and depends on the individual.
• The tested CNN model lacks the accuracy and precision for ecological applications in complex forest soils.
• Further development and evaluation of accessible CNNs for natural ecosystems are necessary.