Geological lithological interpretation is a key focus in Earth observation research, with applications in resource surveys, geological mapping, and environmental monitoring. Although deep learning (DL) methods has significantly improved the performance of lithological remote sensing interpretation, its accuracy remains far below the level achieved by visual interpretation performed by domain experts. This disparity is primarily due to the heavy reliance of current intelligent lithological interpretation methods on remote sensing imagery (RSI), coupled with insufficient exploration of sensitive features (SF) and prior knowledge (PK), resulting in low interpretation precision. Furthermore, multi-modal SF and PK exhibit significant spatiotemporal heterogeneity, which hinders their direct integration into DL networks. In this work, we propose the sensitive feature self-aggregation deep fusion network (SFA-DFNet). Inspired by the visual interpretation practices of domain experts, we selected the five most commonly used SF and one type of PK as multi-modal supplementary information. To address the spatiotemporal heterogeneity of SF and PK, we designed a self-aggregation mechanism (SA-Mechanism) that dynamically selects and optimizes beneficial information from multi-modal features for lithological interpretation. This mechanism has broad applicability and can be extended to support any number of modal data. Additionally, we introduced the cross-modal feature interaction fusion module (CM-FIFM), which enhances the effective exchange and fusion of RSI, SF, and PK by leveraging long-range contextual information. Experimental results on two datasets demonstrate that differences in lithological genesis and types are critical factors affecting interpretation accuracy. Compared with seven SOTA DL models, our method achieves more than a 3% improvement in mIoU, showcasing its effectiveness and robustness.

中文翻译：

---

通过敏感特征自聚集深度融合网络对地质岩性进行遥感影像解释


地质岩性解释是地球观测研究的重点，可应用于资源调查、地质测绘和环境监测。尽管深度学习 （DL） 方法显著提高了岩性遥感解释的性能，但其准确性仍远低于领域专家进行的目视解释所达到的水平。这种差异主要是由于目前智能岩性解释方法严重依赖遥感影像 （RSI），加上敏感特征 （SF） 和先验知识 （PK） 探索不足，导致解释精度低。此外，多模态 SF 和 PK 表现出显著的时空异质性，这阻碍了它们直接集成到 DL 网络中。在这项工作中，我们提出了敏感特征自聚合深度融合网络 （SFA-DFNet）。受领域专家视觉解释实践的启发，我们选择了 5 种最常用的 SF 和一种类型的 PK 作为多模态补充信息。为了解决 SF 和 PK 的时空异质性，我们设计了一种自聚集机制 （SA-Mechanism），该机制从多模态特征中动态选择和优化有益信息以进行岩性解释。此机制具有广泛的适用性，并且可以扩展以支持任意数量的模态数据。此外，我们还引入了跨模态特征交互融合模块 （CM-FIFM），该模块通过利用远程上下文信息增强了 RSI、SF 和 PK 的有效交换和融合。两个数据集的实验结果表明，岩性成因和类型的差异是影响解释准确性的关键因素。 与 7 个 SOTA DL 模型相比，我们的方法将 mIoU 提高了 3% 以上，展示了其有效性和稳健性。