Deep transfer learning reveals severe habitat collapse and limited upslope migration of Persian Oak in the Zagros Mountains under climate change

Predicting the impacts of climate change on biodiversity is hindered by data scarcity, particularly for semiarid endemic species. While Deep Learning (DL) offers powerful modeling capabilities, overfitting risks limit its application to rare species. We propose a Deep Transfer Learning (TL) framework to project the distribution of the data-scarce Persian Oak (Quercus brantii) in the Zagros Mountains of western Iran. We pre-trained a Deep Neural Network (DNN) on European Oak (Quercus robur) data and fine-tuned it using limited Q. brantii records. To mitigate spatial autocorrelation, we employed rigorous Spatial Cross-Validation (SCV). The TL model significantly outperformed a tuned Random Forest benchmark (Mean Spatial AUC: 0.88 vs. 0.75), demonstrating superior spatial generalization. We projected distributions under two emission scenarios (SSP2-4.5 and SSP5-8.5) using an ensemble of three CMIP6 models. Projections for 2070, based on the identified critical annual mean temperature threshold of 19.8°C, reveal severe sensitivity to global mitigation: the suitable habitat is projected to contract by 50.1% under the moderate scenario, intensifying to 67.4% under the high-emission scenario. Under both scenarios, suitable distribution ranges showed a modest upslope redistribution (approx. +57.5 m). This limited tracking response, combined with the massive habitat collapse, suggests that the pace of regional warming may exceed the species' natural migration capacity, leading to a significant "migration debt" and increased extinction risk.