Mineral prospectivity mapping is a critical task in mineral exploration, effectively integrating which demands models capable of capturing both local and global geological features. While deep learning models excel in this domain, their "black-box" nature often limits the trust and insights geologists can derive from their predictions. This paper introduces a novel and interpretable hybrid model to explicitly address this challenge. Our architecture synergistically combines a deformable convolutional network for adapting spatially varying local mineralization features, such as geochemical anomalies, a Transformer module for modelling long-range global-scale spatial features governing mineral deposition. A pivotal innovation is the incorporation of an attribution branching network that generates significance scores for each input predictive factor to the final prospectivity probability. These scores not only provide a direct interpretation of factor relevance but are also fed back to dynamically modulate the key values in the Transformer's attention mechanism, effectively injecting prior geological knowledge into the local and global feature learning process. This design fosters a more geologically informed integration of local and global representations. The model's performance is evaluated against benchmark models including standalone deformable convolutional network, Transformer, and a hybrid model with deformable convolutional network and Transformer. Results demonstrate a superior predictive accuracy and more geologically plausible prospectivity maps. Furthermore, we provide a multi-faceted interpretation framework: the attribution branching network quantifies the contribution of each evidence layer, while gradient-weighted class activation mapping visualizes the discriminative local regions highlighted by the convolutional components and attention maps reveal the long-range feature relationships prioritized by the Transformer to elucidate how the model hierarchically integrates local and global features to arrive at its decisions. This dual-path interpretation strategy demystifies the model's decision-making process, offering geologists tangible insights into both "where" and "why" the model identifies high-potential zones, thereby bridging the gap between high-performing deep learning and actionable c intelligence.