Topographic Similarity (TopSim)

• Topographic Similarity (TopSim) is a framework that quantifies spatial resemblance by aligning structural and functional properties across adjacent regions in both geoscientific and computational contexts.
• In geospatial modeling, methods like the GEAR pipeline and MSG-Net leverage elevation profiles and morphological textures to achieve expert-consistent terrain analog retrieval.
• In neural networks, weight and activation similarity constraints promote local smoothness, enhancing robustness and mirroring biological topographic maps.

Topographic Similarity (TopSim) denotes a class of quantitative metrics, modeling techniques, and regularization objectives designed to ensure that spatially adjacent elements—whether in geospatial terrains or artificial neural maps—develop and maintain similar organizational features or representations. In earth sciences, TopSim quantifies terrain-structure and guides retrieval of geomorphologically homologous landscapes. In computational neuroscience and machine learning, TopSim refers to constraints or evaluation metrics that induce or assess local smoothness in learned features, providing computational parallels to the columnar and map-like structure observed in biological cortex.

1. Formulation of Topographic Similarity in Geospatial Modeling

TopSim operationalizes the intuitive notion of terrain resemblance by leveraging hierarchical, physics-informed pipelines. The GEAR framework exemplifies this, targeting the problem of cross-domain analog retrieval between the Mariana Trench and terrestrial valleys of the Qinghai-Tibet Plateau (Liu et al., 19 Mar 2026). The GEAR pipeline structures the TopSim assessment into three stages:

• Skeleton-Guided Screening and Clipping: Digital elevation models (DEMs) are coarsely filtered for valley-like structures based on valley mask extraction from local elevation deficit binarization, skeletonization via Zhang–Suen thinning, and strict morphological filters (length, mean squared error with best-fit lines).
• Topographic Waveform Comparator (TWC): Shape similarity is scored using Derivative Dynamic Time Warping (DDTW) across normalized, sectioned elevation profiles.
• Morphological Texture Module (MTM): Higher-order shape “texture” is captured using angle-based shape-function matrices, eigenshape decomposition (SVD), and principal component loadings, with final similarity assessed via cosine similarity in the reduced eigenshape basis.
• Graph-based Fine Recognition (MSG-Net): Candidate terrains are represented as graphs of sampled contour points, with each node characterized by five geomorphological metrics (Vector Ruggedness Measure, Arc-Chord Ratio, Slope, Contour Density, Direction Shannon Entropy). Morphology-integrated Siamese Graph Networks (MSG-Net) produce pairwise similarity scores via metric learning on graph embeddings.

This hierarchical approach enables tractable, expert-consistent quantification of topographic similarity between complex, high-dimensional terrain structures.

2. Topographic Similarity in Neural Network Models

In computational models, particularly topographic neural networks, TopSim is instantiated through spatial regularization constraints designed to impose local similarity among the parameters (or activations) of neighboring units. The prototypical TopSim implementation is the Weight Similarity (WS) constraint, which directly regularizes the incoming weight vectors of adjacent units on a pre-defined grid (Truong et al., 31 Jul 2025):

$$L_\mathrm{spatial}^{(\mathrm{WS})} = \frac{1}{|E|} \sum_{(i \to j) \in E} \lVert w_i - w_j \rVert_2$$

where $E$ are ordered pairs of neighboring units on the spatial grid and $w_i$ is the weight vector into unit $i$. The total loss function for model training is:

$$L_\mathrm{joint} = L_\mathrm{CE}(\hat{y}, y) + \lambda \cdot L_\mathrm{spatial}^{(\mathrm{WS})}$$

where $\lambda$ modulates the strength of the spatial constraint.

An alternative is Activation Similarity (AS), regularizing the similarity between activations across training batches using, for adjacent units, the average correlation distance. While AS encourages functional smoothness, WS directly shapes the filter geometry.

3. Metrics and Datasets for TopSim Assessment

In geospatial contexts, the expert-annotated Topographic Similarity Dataset introduced in (Liu et al., 19 Mar 2026) comprises 305 human-labeled valley–trench pairs (154 positive, 151 negative) using a protocol with substantial inter-rater agreement (Cohen’s κ = 0.698). Processing includes DEM normalization, morphological alignment, and eigenshape extraction.

For neural networks, TopSim effectiveness is quantified via:

• Clean and perturbed test accuracy: On benchmark datasets (MNIST, CIFAR-10), measuring robustness to both weight and input noise.
• Second-order isomorphism (prototype RSM similarity): Cosine similarity of class-prototype response similarity matrices (RSM) before and after synthetic perturbations.
• Spatial autocorrelation: Moran’s I quantifies spatial smoothness of activation patterns across the grid.
• Functional localization: Euclidean distances among highly co-activated grid units.
• Single-unit statistics: Activation entropy, fraction of zeros (PoZ), and diversity in orientation/eccentricity tuning.

These metrics jointly probe not only predictive fidelity but also the structural organization and robustness imparted by TopSim constraints.

4. Comparative Performance and Robustness

Empirical results for GEAR (Liu et al., 19 Mar 2026) demonstrate that MSG-Net, utilizing a graph-based TopSim approach, achieves an F1-score of 86.0%—1.38 percentage points higher than the best non-graph baseline—on the annotated TopSim dataset. For baseline models including SANI-SSL, ResNet50, and GCN-DP, F1-scores ranged from 74.0% to 84.7%. Coarse filtering stages (TWC/MTM) show that valleys rejected by similarity thresholds almost always register low (<0.22) MSG-Net similarity, underpinning the validity of multistage filtering.

In neural networks with topographic constraints (Truong et al., 31 Jul 2025), WS consistently yields superior robustness to both weight and input noise. On MNIST, WS models maintain higher RSM similarity (0.75–0.90) and smaller accuracy drop (5–10%) compared to AS (0.65–0.77; 10–20%) or controls. For CIFAR-10, the advantage is preserved, with WS drop ≈½ of AS. Functionally, WS maps also display increased spatial autocorrelation (Moran’s I up to 0.6), tighter functional clustering, elevated activation entropy, and distinct representational biases such as increased symmetry or peripheral orientation selectivity.

5. Biological and Ecological Relevance

The application of TopSim to geomorphology reveals ecologically meaningful correlations: the Mantel correlation between MSG-Net–derived topographic distances and microbial Bray–Curtis dissimilarities across terrain pairs is significantly positive (r_M > 0, p ≤ 0.01) (Liu et al., 19 Mar 2026). This evidences that high topographic similarity predicts similar microbial communities, validating the use of TopSim metrics for biologically informed terrain analog discovery.

In computational neuroscience, the induction of smooth, clustered, and robust feature maps by the WS constraint plausibly mirrors topographic map formation in sensory cortices (Truong et al., 31 Jul 2025). WS-driven topographic layers exhibit representational eccentricity and symmetry tuning profiles reminiscent of biological visual maps, indicating relevance to biophysically inspired modeling.

6. Interpretation and Implications

TopSim, via constraints such as WS in neural layers or graph metric learning in geospatial analysis, operationalizes spatial smoothness and functional locality. In machine learning, simple local topographic regularizers offer a low-complexity mechanism to greatly enhance model robustness and introduce structured, interpretable representations. These effects are robust to moderate trade-offs in unconstrained accuracy and require minimal architectural modification.

In geoscientific applications, TopSim metrics enable efficient, expert-consistent analog retrieval at continental scales, providing a scalable, reproducible substitute for labor-intensive field matching. The observed correlation with ecological ground truth (microbial communities) underscores the functional validity of such approaches.

7. Limitations and Theoretical Considerations

TopSim imposition by WS does not guarantee optimal accuracy under all data regimes; excessive constraint (large λ) can decrease clean set accuracy by 1–3% (Truong et al., 31 Jul 2025). Local activation similarity (AS) may fail to induce stable topographic clusters, instead producing patterned but spatially incoherent representations unless global smoothing is introduced. For geospatial analog retrieval, performance depends on the quality of initial morphometric filtering and the granularity of annotated datasets (Liu et al., 19 Mar 2026).

A plausible implication is that refinement of TopSim objectives (global vs local, weighted vs uniform adjacency, selective metric fusion) will yield further improvements in both representational fidelity and biological realism. The persistent increase in robustness and organization, despite minor accuracy trade-offs, positions TopSim as a valuable tool across domains where spatial coherence and error tolerance are essential.