From Presence-Only to Abundance Species Distribution Models Using Transfer Learning

Area of Science:

• Ecology
• Computational Biology
• Machine Learning

Background:

• Traditional Species Distribution Models (SDMs) face limitations in predicting species abundance due to small datasets.
• Deep learning models, like Convolutional Neural Networks (CNNs), show promise but require large datasets, hindering abundance studies.

Purpose of the Study:

• To enhance the performance of CNN-based SDMs for species abundance prediction.
• To address the challenge of limited sample sizes in species abundance datasets for deep learning models.

Main Methods:

• Utilized Convolutional Neural Network-Species Distribution Models (CNN-SDMs).
• Employed transfer learning combined with large presence-only species datasets.
• Applied the approach to Mediterranean coastal fish species abundance data.

Main Results:

• Significantly improved abundance prediction performance of CNN-SDMs by an average of 35% (D-squared score).
• Achieved average performance gains of 10% over classical SDMs in abundance prediction.
• Enhanced predictions for rare species and for widespread species in locally rare conditions.

Conclusions:

• CNN-SDMs, augmented with transfer learning and large datasets, effectively overcome sample size limitations for abundance modeling.
• This approach offers a significant improvement over traditional SDMs for predicting species abundance.
• The method provides valuable insights into the distribution of rare and locally rare species.