A Three-Dimensional Phenotyping Framework for Quantifying Soybean Resilience to Pest Stress in the Field

Problem Statement

Biotic stress, particularly from pests, is a major driver of global soybean yield losses, estimated at 20–40% annually. Identifying genotypes with high pest tolerance is critical for breeding but remains a bottleneck due to the limitations of conventional phenotyping. Traditional methods, such as manual pest counting and visual leaf-damage scoring, are labor-intensive, difficult to scale, and often lack the statistical replication required for large genotype panels. Furthermore, these methods are often unreliable and fail to capture the complex, quantitative nature of pest tolerance, which involves multiple genes and is influenced by spatial and temporal variability. While Unmanned Aerial Vehicle (UAV) imaging has emerged as a solution for large-scale monitoring, existing approaches relying on handcrafted vegetation indices (VIs) are limited by their reliance on pre-defined spectral combinations, which may not fully capture the structural and textural degradation caused by pest damage.

Methodology

The study proposes a three-dimensional (3D) UAV-based phenotyping framework designed to evaluate soybean genotype performance under natural pest pressure without direct pest quantification. The methodology follows an end-to-end pipeline:

1. Experimental Design:

   • Location: Anhumas Experimental Genetics Station, Piracicaba, Brazil (2024/25 season).
   • Material: 30 soybean genotypes (25 advanced breeding lines and 5 commercial checks) grown in two adjacent fields: a Controlled field (insecticide-protected) and a NoControl field (unprotected).
   • Design: Randomized Complete Block Design (RCBD) with three replicates per field.
   • Data Acquisition: Six UAV flights (DJI Mavic Air 2, RGB sensor) were conducted from late vegetative stages to physiological maturity (R8). Images were processed into georeferenced orthomosaics using Structure-from-Motion (SfM) and RTK-GNSS for precise plot segmentation.

2. Dual-Stream Feature Extraction:
   The framework extracts two complementary feature representations from the RGB imagery:

   • Vegetation Indices (VI): Seven RGB-based indices (e.g., ExG, NGRDI, VARI) were computed. Statistical descriptors (mean, standard deviation, median) were extracted to form 21-dimensional feature vectors.
   • Deep Visual Embeddings: Features were extracted using the self-supervised DINOv3 Vision Transformer (ViT-S/16) model. This model, trained on 1.7 billion images, generates 384-dimensional embeddings without domain-specific fine-tuning, capturing holistic structural and textural information.

3. Three-Dimensional Evaluation Space:
   For each genotype, a 3D space is constructed using three axes:

   • X-Axis (Feature Similarity): A cross-field similarity score ($\bar{s}_i$) calculated by averaging the normalized Euclidean distance between feature vectors of the Controlled and NoControl plots across all time points. High similarity indicates the genotype maintains feature integrity under pest stress.
   • Y-Axis (Phenological Timing): Number of Days to Maturity (NDM). Lower values indicate earlier maturity, potentially avoiding peak pest pressure.
   • Z-Axis (Performance): Two frameworks were used:
     • Framework A: Absolute grain yield under NoControl conditions.
     • Framework B: Yield Gain Rate ($g_i$), representing the relative yield advantage of the NoControl plot over the Controlled plot.

4. Zone Z8 Classification:
   The 3D space is partitioned into eight octants (Z1–Z8) using median thresholds. Zone Z8 is defined as the subspace containing genotypes with:

   • Above-median Feature Similarity (stability).
   • Below-median NDM (early maturity).
   • Above-threshold Performance (high yield or positive gain).

5. Ranking and Validation:
   Genotypes were ranked based on their Z8 occupancy frequency ($p_i$) across 24 distinct 3D scatter views (2 feature streams × 2 performance definitions × 6 time points). Consistency between the VI and DINOv3 streams served as an internal cross-validation to reduce false positives.

Key Results

• Temporal Dynamics: Genotype discrimination was minimal in early stages (T1–T2) but became distinct from T3 onward, coinciding with the pod-fill and grain-fill stages (R3–R7). This suggests a critical phenotyping window where pest-induced differences are most detectable.
• Feature Comparison: DINOv3 embeddings demonstrated earlier sensitivity to canopy stress signals compared to VI features. For instance, high-performing genotype RIL 145 entered the optimal Z8 zone earlier under DINOv3 features (T1) than under VI features (T4).
• High-Performance Candidates: Three genotypes were consistently identified as high-performance candidates across both feature streams and all time points:
  • RIL 145 (75.0% Z8 occupancy)
  • NS 6700 (66.7% Z8 occupancy)
  • LQ 008 (66.7% Z8 occupancy)
    These lines exhibited stable feature structures, early maturity, and either high absolute yield or positive yield gains under natural pest pressure.
• Susceptible Lines: Genotypes such as T 27-18, LQ 267, and LQ 174 consistently fell into low-similarity, late-maturity, and low-yield zones, confirming the framework's ability to discriminate susceptible lines.
• Yield Impact: The average field-level yield penalty due to pests was approximately 15% (1,190 kg/ha in Controlled vs. 1,010 kg/ha in NoControl).

Key Contributions

1. Unified 3D Framework: The paper introduces a novel metric that jointly quantifies productivity, feature-level similarity (as a proxy for tolerance), and phenological response. This replaces labor-intensive pest counting with a non-invasive, plant-centric response measurement.
2. Integration of Foundation Models: It demonstrates the utility of self-supervised vision foundation models (DINOv3) in agricultural phenotyping, showing they can capture structural degradation signals invisible to handcrafted spectral indices.
3. Dual-Stream Consistency Check: By requiring agreement between VI and DINOv3 feature streams, the framework provides a robust internal validation mechanism, reducing the risk of false-positive selections in breeding programs.
4. Operational Efficiency: The study identifies that a reduced campaign of 2–3 UAV flights during the critical pod-fill to grain-fill window is sufficient to capture performance-relevant variation, significantly improving phenotyping efficiency.

Significance and Claims

The authors claim this work establishes a scalable, field-ready paradigm for quantifying crop resilience to biotic stress. By shifting from pest-centric measurement (counting insects) to plant-centric response quantification (measuring physiological impact on feature structure), the framework offers a practical pathway to accelerate breeding for stable yields under real-world conditions.

The study emphasizes that the identified high-performance genotypes reflect robust phenotypic signals rather than representation-specific artifacts, supported by the strong concordance between spectral and deep learning features. The authors note that while the framework is robust, it was validated in a single environment and season; thus, multi-location validation is required before broad deployment. Additionally, the Z8 classification is relative to the evaluated population and requires recalibration for different germplasm panels. The work concludes by offering a non-invasive, operationally efficient solution that is readily transferable to other crop systems and compatible with genomic selection pipelines.