Phenotyping of genotypes and diagnosis of water status in cowpea using thermographic images and machine learning

Technical Summary: Phenotyping of Genotypes and Diagnosis of Water Status in Cowpea Using Thermographic Images and Machine Learning

Problem Statement
Contemporary climate change has intensified drought frequency, posing significant challenges to cowpea (Vigna unguiculata (L.) Walp.) production. While cowpea is generally hardy, water stress during critical phenological stages—specifically the vegetative (V3) and reproductive (R2) stages—can severely limit leaf expansion, root establishment, and pod filling. Traditional methods for assessing water status, such as leaf water potential and stomatal conductance measurements, are destructive and impractical for large-scale phenotyping. Consequently, there is a need for rapid, non-invasive techniques capable of simultaneously monitoring water status and distinguishing genetic tolerance among multiple genotypes under varying irrigation regimes.

Methodology
The study employed a controlled experiment at the Federal University of Campina Grande (UFCG), Brazil, involving ten cowpea genotypes: five landrace varieties (Corujinha, Paulistinha, Sempre Verde, Pintado, Rabo de Tatu) and five improved cultivars (BRS Novaera, BRS Pajeú, IPA 206, BRS Tapaihum, BRS Miranda). These were subjected to four water regimes (25%, 50%, 75%, and 100% of crop evapotranspiration, ETc) in a randomized block design.

Thermographic images were captured using a Flir® C5 camera (160 x 120 pixels) at two distinct phenological stages: V3 (fourth node) and R2 (first flower anthesis). For each genotype, 100 images were acquired from multiple angles (nadir and 45°) to encompass canopy architecture.

The data processing pipeline utilized Orange Data Mining v. 3.37.0. The workflow involved:

1. Feature Extraction (Embedding): Images were vectorized using four pre-trained Deep Learning architectures: InceptionV3, SqueezeNet, VGG16, and VGG19.
2. Classification: Six machine learning algorithms were trained on the extracted vectors to classify both genotypes and water stress levels: k-Nearest Neighbors (k-NN), Decision Tree, Random Forest, Support Vector Machine (SVM), Neural Network (Multi-Layer Perceptron), and AdaBoost.
3. Validation: Models were evaluated using cross-validation and a test-on-test data method, where 25 images per genotype were withheld from training for independent validation. Performance metrics included Accuracy (CA), F1-score, Matthews Correlation Coefficient (MCC), Area Under the ROC Curve (AUC), and LogLoss.

Key Results

• Phenological Stage Influence: The vegetative stage (V3) proved more effective for diagnosis than the reproductive stage (R2). At V3, thermal signatures were more heterogeneous and discriminative, whereas R2 images showed greater homogenization due to canopy closure and physiological complexity.
• Model Performance (V3 Stage):
  • The Support Vector Machine (SVM) combined with the VGG16 embedder achieved the highest performance, with an accuracy (CA) of 0.910, an F1-score of 0.910, and an MCC of 0.907.
  • The Artificial Neural Network (ANN) with VGG16 also performed well (CA: 0.897).
  • Among solvers for the ANN, the Adam optimizer with a logistic activation function yielded the best results (CA: 0.886).
  • For SVM, the polynomial kernel outperformed linear, RBF, and sigmoid kernels.
  • Tree-based models (Random Forest, Decision Tree) and the Sigmoid kernel showed significantly lower performance.
• Model Performance (R2 Stage):
  • Performance metrics decreased compared to V3, reflecting the increased difficulty in distinguishing genotypes at this stage.
  • The SVM with a polynomial kernel and VGG16 remained the top performer (CA: 0.804), followed by the ANN (CA: 0.760).
  • In the validation phase, the SVM maintained superior stability (CA: 0.782 at V3; CA: 0.636 at R2) compared to the ANN.
• Confusion Patterns: Errors in classification were predominantly concentrated between the same genotype under different irrigation levels rather than between distinct genotypes. This indicates that the models successfully preserved varietal identity, with water status acting as a modulating factor that influenced thermal proximity but did not obscure genetic identity.

Key Contributions and Significance
The paper claims that the integration of infrared thermography with machine learning, specifically using deep learning architectures for feature extraction and SVMs for classification, is an effective tool for high-throughput digital phenotyping. The study demonstrates that:

1. Non-Destructive Diagnosis: Thermal imaging can accurately identify cowpea genotypes and diagnose water stress without destructive sampling.
2. Stage Sensitivity: The vegetative stage (V3) offers superior thermal contrast for discrimination compared to the reproductive stage (R2), though both stages yield viable results with appropriate modeling.
3. Algorithmic Superiority: The combination of VGG16 for feature extraction and SVM (with polynomial kernels) provides robust classification capabilities, outperforming traditional tree-based models and other deep learning configurations in this specific context.
4. Practical Application: The approach supports the selection of drought-tolerant materials and aids in precision irrigation management by enabling rapid, simultaneous monitoring of water status across diverse genotypes.

The authors conclude that while phenological advancement introduces complexity and reduces thermal contrast, the proposed methodology remains a promising, non-invasive solution for optimizing genotype selection and water management in precision agriculture.