GEAR: Geography-knowledge Enhanced Analog Recognition Framework in Extreme Environments
Zelin Liu *, Bocheng Li *, Yuling Zhou *, Xuanting Li, Yixuan Yang, Jing Wang † , Weishu Zhao † , Xiaofeng Gao †, Shanghai Jiao Tong University · 2026
Read GEAR: Geography-knowledge Enhanced Analog Recognition Framework in Extreme Environments by Zelin Liu *, Bocheng Li *, Yuling Zhou *, Xuanting Li, Yixuan Yang, Jing Wang † , Weishu Zhao † , Xiaofeng Gao †, Shanghai Jiao Tong University (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-18626.txt.
Introduction
Liu and co-authors present GEAR, a system that finds land sites shaped like the Mariana Trench. Deep-sea sampling is costly, so matching sites on the Qinghai-Tibet Plateau could guide biology work. The search covers 2.5 million square kilometers in three stages. Skeleton screening picks candidate valleys first. Physics-aware filters then compare waveform and texture. A graph network makes the final call using landform metrics. Their expert-labeled dataset targets tectonic collision zones. Features from the network also correlate with biological data from the sites.
Catalog record
| CREATOR | Zelin Liu *, Bocheng Li *, Yuling Zhou *, Xuanting Li, Yixuan Yang, Jing Wang † , Weishu Zhao † , Xiaofeng Gao †, Shanghai Jiao Tong University |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 34965 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.18626
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GEAR: Geography-knowledge Enhanced Analog Recognition Framework in Extreme Environments Zelin Liu *, Bocheng Li *, Yuling Zhou *, Xuanting Li, Yixuan Yang, Jing Wang † , Weishu Zhao † , Xiaofeng Gao †, Shanghai Jiao Tong University Abstract The Mariana Trench and the Qinghai-Tibet Plateau exhibit significant similarities in geological origins and microbial metabolic functions. Given that deep-sea biological sampling faces prohibitive costs, recognizing structurally homologous terrestrial analogs of the Mariana Trench on the Qinghai-Tibet Plateau is of great significance. Yet, no existing model adequately addresses cross-domain topographic similarity retrieval, either neglecting geographical knowledge or sacrificing computational efficiency. To address these challenges, we present G eography-knowledge E nhanced A nalog R ecognition ( GEAR ) Framework, a three-stage pipeline designed to efficiently retrieve analogs from 2.5 million square kilometers of the Qinghai-Tibet Plateau: (1) Skeleton guided Screening and Clipping: Recognition of candidate valleys and initial screening based on size and linear morphological criteria. (2) Physics aware Filtering: The Topographic Waveform Comparator (TWC) and Morphological Texture Module (MTM) evaluate the waveform and texture and filter out inconsistent candidate valleys. (3) Graph based Fine Recognition: We design a M orphology-integrated S iamese G raph N etwork ( MSG-Net ) based on geomorphological metrics. Correspondingly, we release an expert-annotated topographic similarity dataset targeting tectonic collision zones. Experiments demonstrate the effectiveness of every stage. Besides, MSG-Net achieved an F1-Score 1.38 percentage points higher than the SOTA baseline. Using features extracted by MSG-Net, we discovered a significant correlation with biological data, providing evidence for future biological analysis. 1 1 footnotetext: Equal contribution. 4 4 footnotetext: Corresponding authors: [email protected], [email protected], [email protected]. 1 Introduction The Qinghai-Tibet Plateau and the Mariana Trench share structural origins driven by tectonic collision and exhibit over 90% similarity in functional metabolic modules of their microbial communities [ 30 , 9 , 18 ] . This similarity may be caused by their shared V-shaped topography (Fig. 1 (a)), in which local slopes influence nutrient accumulation through gravity-driven processes. As deep-sea biological sampling faces prohibitive costs, identifying structurally homologous terrestrial analogs on the Qinghai-Tibet Plateau provides a promising way to roughly estimate the biological conditions of certain regions in the Mariana Trench and guide the selection of future sampling sites. Thus, a framework to recognize terrestrial analogs across the Qinghai-Tibet Plateau is required. Figure 1 : Topology and microbial comparison between the Tibetan Plateau (TP) and the Mariana Trench (MT).(a) V-shaped topographic profile comparison.(b) Location and DEM grids of the TP.(c) Location of the MT and sampling sites. However, existing methods present limitations in measuring cross-domain topographic similarities. Firstly, traditional geographic methods lack strong representational capacity [ 26 ] . Secondly, Computer Vision (CV)-based approaches often disregard 3D structural trends when processing Digital Elevation Models (DEMs) [ 36 ] . Thirdly, recent Geospatial AI (GeoAI) models, such as GCN-DP [ 15 ] , remain primarily designed for specific regional analysis and lack the ability to support terrain matching across domains. To address these limitations, we propose the G eography-knowledge E nhanced A nalog R ecognition ( GEAR ) Framework, a three-stage pipeline designed to recognize valleys with similar sizes and structures to a reference trench in the Mariana Trench from an area of 2.5 million square kilometers of the Qinghai-Tibet Plateau (QTP): (1) Coarse Screening : We extract approximately 30,000 candidate valleys in the QTP via a proposed Skeleton-guided Screening and Clipping (SSC) approach. Specifically, we proposed a Local Elevation Deficit Binarization (LEDB) strategy to identify valley areas. Using the Zhang-Suen thinning algorithm [ 34 ] , we extract the skeletons of these areas. Since the Mariana Trench exhibits a straight morphology under regional tectonic stress fields [ 37 , 23 ] , we apply linear regression fitting to extract landforms with similar straightness and crop the corresponding rectangular regions. (2) Physics-aware Filtering: To match overall topographies efficiently, we introduce the Topographic Waveform Comparator (TWC) and the Morphological Texture Module (MTM). We first decompose the candidate valleys into slice sequences. TWC measure structural similarity via Derivative Dynamic Time Warping (DDTW [ 14 ] ) based on morphological trends. Subsequently, MTM extracts terrain texture by Eigenshape Analysis [ 20 ] , ranking candidates by cosine similarity. This stage efficiently compresses the candidate pool from roughly 10^{4} to 10^{3} . (3) Graph-based Fine Recognition : A Morphology-integrated Siamese Graph Network (MSG-Net) is employed to calculate the topological similarity between candidate valleys and the reference trench through shared-weight graph convolutions incorporating traditional topographic knowledge. This process resolves the topological distance between samples, retrieving most similar candidate valley to the reference Trench. This architecture resolves the limitations of existing methods. First, to overcome the limited capacity of traditional indices, GEAR captures the overall morphological structure. Second, unlike CV approaches that process DEMs as planar grids, the pipeline preserves physical topography by integrating geographical priors. Third, diverging from GeoAI classification, our matching process measures structural similarities across extremely different heights. Our key contributions are summarized as follows: (1) We achieve fast, topography-aware cross-regional terrain analog retrieval in extreme environments, overcoming the 2D flattening of CV models and the extreme height barriers that fail current GeoAI methods. (2) We introduce an annotated open-source topographic similarity dataset targeting tectonic collision zones 1 1 1 The dataset and source code are provided for double-blind review at: https://anonymous.4open.science/r/GEAR-Blind . Experimental evaluations indicate that GEAR achieves an accuracy of 83.93%, a recall of 98.05%, and an F1-score of 86.04%, outperforming the highest performing baseline, SANI-SSL (which achieved 79.82% accuracy and 86.49% recall). (3) We validate the topography-microbiome correlation in the Mariana Trench and Mount Everest, establishing an interdisciplinary methodology that maps macro-geomorphology to biological constraints. Utilizing AI-matched QTP analogs, we provide predictive, actionable guidance for deep-sea ecological sampling and functional transfer research. [ 9 , 18 ] . 2 Problem Formulation 2.1 Geographical Analysis and Data Processing Geographical analysis focuses on two extreme environmental regions: (1) the Qinghai-Tibet Plateau (QTP), and (2) the Mariana Trench. Qinghai-Tibet Plateau (QTP). The QTP region is between latitude 67 ∘ 40 E \sim 104 ∘ 40 E and longitude 25 ∘ 59 N \sim 40 ∘ 01 N with an irregular boundary. We use the Digital Elevation Model (DEM) data from the ASTER GDEM dataset 2 2 2 https://data.tpdc.ac.cn/zh-hans/data/bdacc6b7-aff0-49b1-b666-cb97d846d072 , which comprises 922 coordinate-based grids in .tif format, denoted as \mathcal{\overline{\overline{X}}}=\{\overline{\overline{X}}_{1},\overline{\overline{X}}_{2},\dots,\overline{\overline{X}}_{m^{\prime\prime}}\} , and m^{\prime\prime}=922 . Each grid \overline{\overline{X}}_{i} , depicts a unit sub-region of latitude 1^{\circ} \times longitude 1^{\circ} , has a size of 24 MB, with a spatial resolution of 30 m . \overline{\overline{X}}_{i} consists of 3601 \times 3601 pixels, whose value represents the elevation of the corresponding sub-region of 30m\times 30m , as shown in Fig. 1 (B). To avoid mismatches along grid boundaries, we concatenate \mathcal{\overline{\overline{X}}} into a new dataset \mathcal{\overline{X}}=\{\overline{X}_{1},\overline{X}_{2},\dots,\overline{X}_{m^{\prime}}\} . Each \overline{X}_{j} , denoted as a tile , is composed of 2\times 2 neighboring \overline{\overline{X}}_{i} ’s with zero-padding borders. Adjacent tiles \overline{X}_{j} and \overline{X}_{j+1} share two duplicated grids; thus, m^{\prime}=m^{\prime\prime}=922 . We further extract valleys from the DEM tiles \mathcal{\overline{X}} , denoted as \overline{\overline{\mathcal{V}}}=\{\overline{\overline{V}}_{1},\cdots,\overline{\overline{V}}_{m}\} . Each valley \overline{\overline{V}}_{i} has a region of 5km\sim 40km , representing a topographic valley in real world. Mariana Trench (MT). The DEM data for the Mariana Trench is a strict square region between 140.5963 ∘ E \sim 143.1206