Remote sensing data prove to be an effective resource for constructing a data-driven predictive model of mineral prospectivity. Nonetheless, existing deep learning models predominantly rely on neural networks that necessitate a substantial number of samples, posing a challenge during the early stages of exploration. In order to predict mineral prospectivity using remotely sensed data, this study introduced deep forest (DF), a non-neural network deep learning model. Mainly based on ASTER multispectral imagery supplemented by Sentinel-2 and geological data, gold ore in Hamissana area, NE Sudan was used to test the DF predictive model capability. In addition to four geological-based evidential layers, 20 remote sensing-based evidential layers were generated using remote sensing enhancing techniques, forming the predictor variables of the proposed model. The applicability of the DF was thoroughly examined including its accuracy for delineating prospective areas, sensitivity to amount of training samples, and adjustment of hyperparameters. The results demonstrate that DF model outperformed conventional machine learning models (i.e., support vector machine, artificial neural network, and random forest) with AUC of 0.964 and classification accuracy of 93.3%. Moreover, the sensitivity analysis demonstrated that the DF model can be trained with a limited number (i.e., < 15) of mineral occurrences. Therefore, the DF algorithm has great potential and proves to be a viable solution for data-driven prospectivity mapping, particularly in scenarios with data availability constraints.

遥感数据被证明是构建数据驱动的矿产前景预测模型的有效资源。尽管如此，现有的深度学习模型主要依赖于神经网络，需要大量样本，这在探索的早期阶段构成了挑战。为了利用遥感数据预测矿产前景，本研究引入了深度森林（DF）这种非神经网络深度学习模型。以ASTER多光谱影像为主，辅以Sentinel-2和地质数据，以苏丹东北部Hamissana地区的金矿石为基础，测试了DF预测模型的能力。除了 4 个基于地质的证据层之外，还使用遥感增强技术生成了 20 个基于遥感的证据层，形成了所提出模型的预测变量。对 DF 的适用性进行了彻底检查，包括其描绘预期区域的准确性、对训练样本数量的敏感性以及超参数的调整。结果表明，DF模型优于传统机器学习模型（即支持向量机、人工神经网络和随机森林），AUC为0.964，分类准确率为93.3%。此外，敏感性分析表明，DF 模型可以用有限数量（即 < 15）的矿物出现进行训练。因此，DF算法具有巨大的潜力，并被证明是数据驱动的前景图的可行解决方案，特别是在数据可用性受限的场景中。