Harshini Suresha

TriggerNet

An Explainable AI Framework for Red Palm Mite Detection

Year2025
LocationAnna University
TagsExplainable AI, CNN, EfficientNet, YOLOv8

Grad-CAM, RISE, FullGrad and TCAV fused into a hierarchical interpretability pipeline across 11 plant species; 95.1% accuracy with EfficientNet + Random Forest and 94.4% detection with YOLOv8, served in a Streamlit app.

Nov 2024 – Feb 2025 at Anna University, Chennai, supervised by Sunil KC. Presented at 13th ICCET Conference.

Workflow figureOpen full size ↗
Figure for TriggerNet: An Explainable AI Framework for Red Palm Mite Detection
Methods
  1. 01
    Dataset & weak labels

    RGB images across 11 plant species, labelled into four classes (Healthy, Yellow Spots, Reddish Bronzing, Silk Webbing) using heuristic weak supervision to cut manual annotation.

    Snorkel, OpenCV
  2. 02
    Model comparison

    Compared deep and classical models side by side.

    CNN, EfficientNet, MobileNet, Vision Transformer, ResNet50, InceptionV3, Random Forest, SVM, KNN
  3. 03
    Detection

    Trained an object detector to localise infestation on leaves (94.4% detection accuracy).

    YOLOv8
  4. 04
    Explainability pipeline

    Fused four attribution methods in a hierarchical pipeline that switches explanation modules on when predictions are uncertain or classes are ambiguous.

    Grad-CAM, RISE, FullGrad, TCAV
  5. 05
    Results

    95.1% accuracy with EfficientNet + Random Forest and 94.2% with ResNet50 + SVM; disease F1 of 0.87 for Silk Webbing and 0.86 for Reddish Bronzing.

    Scikit-learn, Matplotlib
  6. 06
    App

    An interactive app for real-time infestation detection with explanations.

    Python, TensorFlow, Keras, Streamlit
← Back