<p>The use of deep learning (DL) algorithms for data-driven mineral prospectivity mapping (MPM) has emerged as a prominent and rapidly growing area of research in mineral exploration. However, DL models based on a neural network approach have high complexity that necessitates a large number of labeled samples. These models often struggle with generalization under data-scarce conditions, which is common in real-world exploration scenarios. This highlights that the field of mineral prospectivity modeling still lacks exploration of fundamentally different model architectures that may offer more robust performance. In this context, deep forest (DF) is a deep ensemble learning model that avoids the use of neurons and backpropagation, offering an alternative deep architecture to address complexity issues. This study utilized 12 feature variables derived from geochemical, geological, and remote sensing data to benchmark the performance of DF against convolutional neural network (CNN), long short-term memory (LSTM), gated recurrent unit (GRU), and self-attention neural network (SANN) models for gold prospectivity mapping in the Abidiya region of Sudan. The results show that the DF model performed on par with GRU in terms of classification accuracy (0.979), while maintaining high prediction performance, with an area under the receiver operating characteristic curve (AUC) of 0.984, only 0.004 below CNN and GRU, and 0.002 below LSTM. Moreover, the DF model also demonstrated significant advantages in training efficiency, robustness with limited samples, and reduced exploration risk, as evidenced by success-rate curve prediction-rate curve. The findings underscore the application of DF as a viable and innovative approach to mineral prospectivity modeling, offering practical advantages in terms of predictive accuracy, geological interpretability, and robustness under limited data conditions.</p>