DL-SDM

DL-SDM models species distributions using a deep neural network to predict species presence/absence from environmental predictors and species interactions, addressing geographic and taxonomic sampling biases.

Key Features:

• Deep Learning Architecture: Implements a deep neural network (DNN) with four hidden layers and drop-out regularization between each layer to reduce overfitting.
• Binary Classification Model: Functions as a binary classifier distinguishing species presence from absence based on environmental predictors.
• Performance Comparison with MaxEnt: Benchmarked against MaxEnt, showing comparable performance for species with larger sample sizes and reduced accuracy for species with smaller datasets.
• Enhanced Sampling and Predictors: Utilizes modified sampling of negative instances (absences) and increased numbers of environmental predictors, including species interactions, to improve performance across sample sizes.
• Handling Correlated Features: Directly incorporates species interactions and manages correlated input features to represent complex ecological relationships.

Scientific Applications:

• Ungulate niche modeling: Models ecological niches and distributions of wild and domesticated ungulate species using abiotic environmental predictors and species interactions.
• Comparative species-distribution analyses: Facilitates comparative studies among species and addresses geographic and taxonomic sampling biases that limit biodiversity inference.

Methodology:

Uses a pre-existing dataset of world ungulates with abiotic environmental predictors previously used in MaxEnt models; trains a DNN with four hidden layers and drop-out regularization as a binary presence/absence classifier; benchmarks performance against MaxEnt; applies modified negative-instance sampling and increases environmental predictors, including species interactions, as part of regularization and enhanced sampling methods.