Biodiversity loss is a major global challenge that affects the functioning of ecosystems and human well-being. Protected areas (PAs) are key strategies to conserve biodiversity, but climate change and land use change may reduce their effectiveness by altering the distribution of species. To achieve Sustainable Development Goal ۱۵ (SDG۱۵) and Aichi Target ۱۱, it is essential to identify and prioritize suitable areas for creating and expanding PAs, considering the shifts in species distributions due to these global changes. However, this task is complex due to the diversity and complexity of natural ecosystems and the limitations of existing biodiversity prediction models. We leveraged key datasets, including land use change data derived from the Landsat archive, bioclimatic data, and species occurrence data from the Global Biodiversity Information Facility (GBIF), to train a deep convolutional neural network (CNN) model. Unlike conventional methods, our approach explicitly incorporated species dispersal capacity, which plays a vital role in determining species' adaptive responses to environmental changes. We used the modeling outcomes in a gap analysis to characterize the effectiveness of PAs. To address the limitations of the binary nature of gap analysis, we applied a rescaling technique on pixel values associated with desirable habitat areas. This technique can help measure protection gaps more precisely in current conditions and various future scenarios. Furthermore, we projected future changes in biodiversity under a combination of different General Circulation Models (GCMs) under four representative concentration pathways (RCPs) of greenhouse gas emissions and four land use scenarios from the Shared Socioeconomic Pathways (SSPs). By doing so, we aimed to reduce the uncertainty associated with different climate projections and human activities. The results show that the area of desirable habitats inside the PAs will increase by ۴۱% in the future, but the rescale approach indicates a decreasing trend of ۲۹%. The CNN model (AUC = ۰.۹۶۴) outperformed the conventional models (AUC = ۰.۸) in predicting species distribution. Under the limited dispersal scenario, the species distribution was reduced by ۳۲% compared to the full dispersal scenario. Moreover, the results indicate that we reduced the uncertainty and variability in the predictions of biodiversity changes by combining different GCMs. Our research provides a comprehensive framework for a gap analysis that addresses some of the limitations and challenges in existing approaches. By incorporating species dispersal capacity, we account for ecological processes that influence species distribution and adaptation. By applying a technique that accounts for habitat suitability variability, we measure protection gaps more accurately. By combining different climate models, we reduce uncertainty and variability in biodiversity predictions.
