gasparfj/Opuntia-detector

Opuntia Detection from Google Street View

YOLOv8 object detection model trained to detect Opuntia invasive species from Google Street View imagery using high-resolution sliced inference workflows.

This model is part of the StreetViewSentinel project, focused on large-scale environmental monitoring and invasive species detection through computer vision.

---

Model Description

The model was trained using:

• YOLOv8
• SAHI sliced inference
• High-resolution Google Street View imagery
• Custom preprocessing and smart tiling pipeline

The objective is to improve detection performance on large panoramic images where invasive species may occupy small regions of the scene.

---

Intended Use

This model is intended for:

• Invasive species monitoring
• Environmental analysis
• Geospatial biodiversity studies
• Large-scale Street View image analysis
• Research applications

---

Training Data

Training data consists of manually annotated Google Street View imagery containing instances of Opuntia species.

Preprocessing included:

• Vertical image splitting
• Margin-based cropping
• Smart tiling
• Data augmentation

---

Inference

Example using Ultralytics:

from ultralytics import YOLO

model = YOLO("model.pt")

results = model.predict(
    source="image.jpg",
    conf=0.3
)

Example using SAHI sliced inference

from sahi import AutoDetectionModel
from sahi.predict import get_sliced_prediction

detection_model = AutoDetectionModel.from_pretrained(
    model_type="yolov8",
    model_path="model.pt",
    confidence_threshold=0.3,
    device="cpu"
)

result = get_sliced_prediction(
    "image.jpg",
    detection_model,
    slice_height=256,           # Best results obtained with
    slice_width=256,            # patches of 2000 by 2000
    overlap_height_ratio=0.2,
    overlap_width_ratio=0.2,
)

Evaluation Metrics

The model was first evaluated using standard object detection metrics on a held-out validation dataset.

Object Detection Metrics

Metric	Value
mAP50	XX
mAP50-95	XX
Precision	XX
Recall	XX

---

Classification-Oriented Evaluation

After object detection evaluation, the model was additionally evaluated on a separate test dataset using a binary image-level classification approach.

An image was considered positive if the detector produced at least one detection.

This evaluation strategy was designed to assess the model as a practical screening tool for detecting the presence of Opuntia invasive species in Google Street View imagery.

Classification Rule

Positive image  -> at least one detection
Negative image  -> no detections

Classification metrics

Metric	Value
Accuracy	XX
Precision	XX
Recall	XX

---

Inference Configuration

The reported metrics were obtained using:

• YOLOv8 object detector
• SAHI sliced inference
• Sliced size: 2000x2000
• Overlap ratio: 0.0
• Confidence treshold: 0.5

Limitations

Performance may vary under:

• Strong occlusions
• Extrem distance from target objects
• Panoramic deformations Model was trained specifically on Google Street View imagery and may not generalize to other image domains without fine-tuning.

Citation

If you use this model in research, please cite the corresponding project or repository.

Author

Developed as part of a Master's Thesis (TFM) in Bioinformatics and Biostatistics