AI-Based Species Richness Prediction Models

Abstract

Machine learning and deep learning approaches offer transformative potential for predicting species richness from remote sensing, climate, and environmental data at spatial and temporal scales inaccessible to field surveys, yet systematic comparison of AI-based prediction model performance against traditional statistical approaches (generalised linear models, random forests) for species richness mapping is rare. This study benchmarked eight prediction models - linear regression, generalised additive models (GAM), random forest (RF), gradient boosting (XGBoost), convolutional neural network (CNN), long short-term memory recurrent network (LSTM), transfer learning (ResNet-50 fine-tuned on biodiversity imagery), and an ensemble model combining all eight -- for predicting species richness of four taxonomic groups (plants, birds, butterflies, beetles) from 2,840 site-year records at 42 European sites (2002-2022) using 48 environmental predictors from Sentinel-2, CHELSA v2.1, soil databases, and land cover data. The ensemble model (weighted average of all eight predictions) achieved the highest performance (r = +0.884; R2 = 0.782; RMSE = 8.4 species) significantly outperforming single models. XGBoost (R2 = 0.724) and CNN (R2 = 0.684) outperformed GAM (R2 = 0.584) and linear regression (R2 = 0.484) when all 48 predictors were included. However, XGBoost and CNN showed substantially higher overfitting risk than GAM in spatial cross-validation (mean CV R2 decrease: XGBoost -0.184; CNN-0.224; GAM -0.084), and GAM remained most interpretable for ecological hypothesis testing. An AI Model Selection Index (AMSI) quantifying the accuracy-interpretability-overfitting trade-off recommended the ensemble for operational prediction and GAM for mechanistic understanding. Transfer learning from satellite imagery showed particular promise for data-sparse regions where field survey data are insufficient for training local models.