Project Case Study

AI-Powered Medicinal Plant Identification

A deep learning system identifying 41 medicinal species native to Kashmir. By digitising traditional botanical wisdom, this project preserves critical herbal knowledge through state-of-the-art computer vision.

99.58%
Validation Accuracy
Kashmiri medicinal flora
Problem Context

Addressing Expert Dependency and Biodiversity Loss

The heavy reliance on expert botanists and high visual similarity between species creates a bottleneck for conservation. Without automated solutions, traditional knowledge risks being lost.

Classes Identified

41

Balanced species categorisation

Proposed Solution

Transfer Learning + CBAM

Deep learning architecture with focus-enhancing attention mechanisms.

System Pipeline

The architecture optimises feature extraction through a balanced dataset pipeline and a custom attention-integrated backbone.

database
Balanced Dataset
5,535 Images · 135/Class
psychology
Xception Core
ImageNet Transfer Learning

01. Dataset

All images were manually captured from botanical gardens, herbal farms, and field stations maintained by SKUAST-Kashmir, Wadoora Sopore, and natural habitats across Kupwara.

Each photograph was taken with a Samsung Galaxy S23 under natural daylight and varied backgrounds, ensuring the dataset reflects real-world conditions.

Every specimen was identified and verified by local botanical experts prior to inclusion, making this a scientifically validated resource.

02. Architecture

A pretrained Xception model with Depthwise Separable Convolutions, integrated with a Convolutional Block Attention Module (CBAM) to focus on critical leaf markers like vein patterns and margin serration.

Training Parameters

Adam (LR=1e-4)Batch Size 32Sparse Categorical Crossentropy25 Epochstf.data AUTOTUNE
menu_book

Data Publication & Citation

"Bhat, Naveed (2026), "Kashmiri Medicinal Plant and Leaf Dataset", Mendeley Data, V2, doi: 10.17632/ck4rfmrdym.2"

Performance Benchmarks

Comparative analysis against standard architectures demonstrates the superiority of the Xception + CBAM approach.

Validation Accuracy

99.58%

Training: 100%

Precision

0.995

Recall

0.996

F1-Score

0.995

Competitive Model Comparison

ArchitectureValidation AccuracyPerformance Delta
Proposed (Xception + CBAM)99.58%Baseline
MobileNetV299.50%-0.08%
VGG1697.28%-2.30%
ResNet5092.12%-7.46%

Challenges & Optimizations

Visual Similarity

Overcome through the CBAM attention mechanism, specifically capturing spatial nuances in leaf margins and venation patterns.

Environmental Noise

Background noise and varied lighting mitigated via intensive data augmentation and normalisation layers applied at every stage.

Impact & Heritage

  • check_circle
    Conservation Support:

    Providing automated tools for tracking endangered medicinal flora across Kashmir's diverse ecological zones.

  • check_circle
    Research Enablement:

    Empowering non-experts to identify plants for sustainable harvesting and ecological surveys.

  • check_circle
    Healthcare Heritage:

    Digital preservation of ethnobotanical knowledge using advanced AI, safeguarding traditional medicinal wisdom.

Future Roadmap

smartphone

TFLite Deployment

Quantizing the model for real-time inference on edge devices in remote Kashmiri regions without connectivity.

center_focus_weak

YOLO Integration

Transitioning from classification to real-time object detection for active-scanning field surveys.

query_stats

Grad-CAM Insights

Implementing interpretability heatmaps to visualise the diagnostic features the model uses to classify leaves.

“Encoding botanical heritage into the neural landscape.”
Next Projectarrow_forwardBack to Portfolio

Up next — Tyrads Production SDK