Home

Computational Ecology and Software, 2026, 16(2): 198-209
[XML] [EndNote] [RefManager] [BibTex] [ Full PDF (408K)] [Comment/Review Article]

Article

Hybrid physics-informed deep learning with explainable graph neural networks for climate-driven biodiversity forecasting: A multi-scale approach

Mohamed Mazloum Salem
Department of Computer Science, Mansoura University, Mansoura 33561, Egypt

Received 10 November 2025;Accepted 8 December 2025;Published online 20 December 2025;Published 1 June 2026
IAEES

Abstract
Climate change poses unprecedented threats to global biodiversity, necessitating advanced computational frameworks for predicting species distributions and ecosystem responses under future climate scenarios. Traditional species distribution models (SDMs) and mechanistic approaches lack the capacity to capture complex nonlinear ecological dynamics while remaining interpretable to conservation practitioners (Christin et al., 2019). This paper presents a novel hybrid framework integrating Physics-Informed Neural Networks (PINNs) with Graph Neural Networks (GNNs), enhanced by multi-scale attention mechanisms and Bayesian uncertainty quantification (Wesselkamp et al., 2024). Our approach embeds mechanistic ecological constraints directly into neural architectures while explicitly modeling species interaction networks as graph-structured data (Anakok et al., 2025). We evaluate the hybrid PINN-GNN model against traditional SDMs, standard deep neural networks, and standalone PINN/GNN approaches using a multi-regional dataset spanning 225 species across diverse ecosystems. Results demonstrate superior predictive performance: 92% accuracy (vs. 81% for standard DNNs and 72% for traditional SDMs), RMSE of 0.08 (71% improvement over traditional methods), and AUC-ROC of 0.95. Explainable AI analysis via SHAP values (He et al., 2022) identifies temperature (0.42), habitat fragmentation (0.35), and precipitation (0.28) as the most influential environmental drivers. Climate change projections under SSP2-4.5 and SSP5-8.5 scenarios predict range shifts of 82-195 km by 2100, with 73% of species experiencing net range contractions. Bayesian uncertainty quantification reveals growing epistemic uncertainty (0.03-0.08) as ecosystems enter novel climates (Olivier et al., 2021). This research advances computational ecology by providing an interpretable, mechanistically-grounded, uncertainty-aware framework suitable for biodiversity conservation planning in the Anthropocene.

Keywords physics-informed neural networks;graph neural networks;species distribution modeling;climate change;explainable AI;deep learning;biodiversity forecasting;uncertainty quantification.


International Academy of Ecology and Environmental Sciences. E-mail: office@iaees.org
Copyright © 2009-2025 International Academy of Ecology and Environmental Sciences. All rights reserved.
Web administrator: office@iaees.org, website@iaees.org; Last modified: 2025/12/27


Translate page to: