Automated classification of natural habitats using ground-level imagery

Area of Science:

• Ecology
• Computer Science
• Conservation Biology

Background:

• Terrestrial habitat classification is crucial for biodiversity conservation, ecological monitoring, and land use planning.
• Current methods often rely on satellite imagery and field validation, which can be resource-intensive and limited in scale.
• A need exists for more accessible and scalable habitat classification methodologies.

Purpose of the Study:

• To develop and validate a deep learning methodology for classifying terrestrial habitats using solely ground-level photographs.
• To create a classification system based on the 'Living England' framework, categorizing images into 16 distinct habitat classes.
• To assess the performance and scalability of this approach for ecological monitoring and conservation applications.

Main Methods:

• A deep learning classifier, based on the DeepLabV3-ResNet101 architecture, was developed and fine-tuned.
• Ground-level habitat photographs were pre-processed (resizing, normalization, augmentation) and resampled for class balance.
• Five-fold cross-validation was employed to evaluate the model's performance across 16 habitat classes.

Main Results:

• The deep learning model achieved a mean F1-score of 0.63 across all 16 habitat classes.
• High performance was observed for visually distinct habitats (e.g., Bare Sand, Coniferous Woodland with F1-scores > 0.87).
• Performance varied, with lower scores for visually mixed or ambiguous habitat classes.

Conclusions:

• Deep learning applied to ground-level imagery provides a robust and scalable method for terrestrial habitat classification.
• This approach enhances ecological monitoring capabilities by utilizing easily obtainable citizen science data.
• The developed methodology and accompanying web application support practitioners in habitat classification and conservation efforts.