<p>Geological target recognition is crucial in mineral exploration, and the fusion of multi-source geophysical data based on machine learning has emerged as a predominant method for target recognition. Due to substantial variations in tectonic contexts and exploration stages among different regions, coupled with exploration cost constraints, the quantity of labeled samples in training datasets varies significantly. Employing only a single learning mode (e.g., supervised learning) to address datasets under diverse scenarios will significantly diminish a model's accuracy and generalization performance. To address these limitations, this paper proposes a multi-mode adaptive prediction system (MAPS). The core innovation involves integrating supervised, semi-supervised, and unsupervised learning modes, alongside designing an adaptive paradigm-switching mechanism based on label coverage, thereby effectively decreasing the dependency on labeled data under various conditions and enhancing model generalizability. During this process, a genetic algorithm is utilized to adaptively optimize key hyperparameters across learning modes, significantly minimizing human intervention and achieving better model configurations. Multiple simulation scenarios indicate that MAPS consistently exhibits prediction accuracy and stability superior or comparable to those of traditional single-mode algorithms across varying label coverage levels. Moreover, in two practical mineral exploration cases (achieving borehole verification accuracies of 92.11% and 81.19%), the system likewise exhibited excellent applicability and reliability. In summary, MAPS not only markedly reduces dependency on large-scale labeled datasets but also streamlines algorithm deployment, thereby offering a flexible and efficient pathway for advanced resource exploration.</p>