Multimodal Heterogeneous Data Fusion Framework for Landforms and Genesis Detection Based on Knowledge-Driven

Effectively distinguishing landforms and their genesis from images is a critical task in Earth sciences. While the natural environmental conditions surrounding landforms provide essential context for their interpretation, efficiently integrating the corresponding Earth observation data into the detection tasks and subsequent digital mapping processes remains challenging. Here, we proposed a novel knowledge-driven multimodal hybrid fusion framework tailored for merging heterogeneous geospatial data to landforms and genesis detection. This framework involved various new features: (1) a feature fusion method was introduced to transform single-channel images of the same data type into fusion images exploiting the complementarity of diverse diverse features; (2) a novel heterogeneous graph deep learning approach was developed to extract richer graph-based terrain structural information guided by the knowledge from terrain analysis; (3) a novel hybrid fusion module was developed to integrate different feature representations from various geospatial data based on the knowledge assumption that landform and genesis are intrinsically interdepent. This module incorporates both feature-level and decision-level fusion to generate a more enriched and comprehensive joint representation. Numerous meticulously designed experiments validate the effectiveness of our method from multiple perspectives, demonstrating its superior performance relative to conventional methods. On this basis, this study elucidates the synergistic mechanisms among different geospatial data modalities, as well as the relationship between the dependence of model performance on data quantities used. The proposed method was also applied in several case studies on Martian landform detection, demonstrating its potential as a promising approach for advanced planetary landform detection.