This repository documents a deep learning project developed as part of the Sabancı University PURE (Program for Undergraduate Research) and submitted to the PlantCLEF 2024 plant identification competition.
The goal of the project is to automatically identify plant species from images using deep learning–based object detection and classification techniques.
The system is built around a two-stage Faster R-CNN pipeline that first detects plant regions in images and then classifies them into plant species.
⚠️ Important Note
Due to the very large dataset and GPU requirements of the project, the experiments were conducted on Sabancı University laboratory machines with specialized infrastructure.
Therefore some training scripts and internal processing code are not included in this repository.
This project was conducted within the:
- PURE – Program for Undergraduate Research
- Sabancı University Computer Vision and Pattern Analysis Laboratory (VPA)
- PlantCLEF 2024 Competition
Team Members:
- Ayça Elif Aktaş
- Şimal Yücel
- Doruk Benli
The project explores automated plant identification using deep learning techniques, leveraging large-scale biodiversity datasets provided by the PlantCLEF challenge. :contentReference[oaicite:0]{index=0}
Plant identification is an important task in fields such as:
- agriculture
- environmental science
- biodiversity monitoring
- botany research
Traditional plant classification requires expert knowledge and manual inspection, which is time-consuming and difficult to scale.
Deep learning offers a promising solution by allowing models to learn visual patterns directly from plant images. :contentReference[oaicite:1]{index=1}
However, the PlantCLEF dataset introduces a unique challenge:
- The training dataset contains images with a single plant label
- But test images may contain multiple plant species
This creates a multi-instance detection problem, requiring both:
- Plant localization
- Species classification
To address this challenge, we designed a two-stage plant identification pipeline.
A pre-trained Faster R-CNN model with a ResNet50 backbone was used to:
- Detect plant regions
- Generate bounding boxes
- Filter irrelevant detections
- Select the most relevant plant region
Bounding boxes were determined using:
- minimum area filtering
- center weighting
- intersection-over-union (IoU) evaluation
These steps allowed automatic bounding box annotation across the dataset.
Once plant regions were extracted:
- The dataset was converted into COCO format
- Bounding boxes were used to train a fine-tuned Faster R-CNN model
- Species classification was learned using Stochastic Gradient Descent (SGD) optimization
Additional techniques used:
- Data augmentation
- Early stopping
- Training / validation split
This approach enabled the model to detect plant locations and classify species simultaneously.
The project uses the PlantCLEF dataset, one of the largest biodiversity datasets available.
Dataset characteristics:
- ~1.4 million plant images
- ~7,800 plant species
- Images collected from various ecosystems
Due to computational limitations, a subset of 1000 samples was selected for training experiments.
The dataset was split:
80% Training
20% Validation
The model achieved low training loss and stable validation performance, indicating effective species recognition. :contentReference[oaicite:2]{index=2}
The model training was performed using:
- PyTorch
- Faster R-CNN
- ResNet50 backbone
- CUDA GPU acceleration
- Anaconda environment management
Training large-scale datasets required high-performance lab machines, which is why some internal code and scripts are not included in this repository.
This project demonstrates:
- Automated plant detection using Faster R-CNN
- Bounding box generation for large biodiversity datasets
- Conversion of plant datasets to COCO format
- Deep learning–based plant species classification
- A scalable pipeline for automated plant recognition
This repository serves as a documentation and research overview of the project.
Due to:
- extremely large dataset size
- GPU infrastructure requirements
- university lab environment restrictions
some internal scripts used in the training pipeline are not included in this repository.
Potential improvements include:
- Training on the full PlantCLEF dataset
- Experimenting with Vision Transformers (ViT)
- Using EfficientNet or SE-ResNeXt architectures
- Multi-plant detection using improved region proposals
Key research and datasets used in this project include:
- PlantCLEF biodiversity challenge
- iNaturalist dataset
- Faster R-CNN object detection
- CNN-based plant classification research
More detailed references can be found in the original project report.
This repository is intended for academic and research documentation purposes.
Dataset usage follows PlantCLEF dataset licensing terms.