| 07 Jul 2026
The impact of data preprocessing in machine-learning models for spatial continuity of fault detection using borehole data and Triangulated Irregular Networks
Abstract. Accurate fault detection is crucial for 3D subsurface structural modeling, yet the success of the method fundamentally depends on a critical property: the spatial continuity of fault-traces. In this study, we illustrated that spatial continuity strongly depends on spatial additivity of machine-learning-based classification. Spatial additivity dictates whether a machine-learning (ML) model yields consistent predictions when evaluating an entire structural domain versus its localized spatial subsets. While supervised ML can be used to map tectonic discontinuities, the impact of data preprocessing on the important properties (spatial additivity, spatial continuity) remains unaddressed. This study compares three data-scaling workflows influencing a Support Vector Machine (SVM) classifier: independent subset scaling, per-surface local scaling, and canonical ML scaling. Validated on structural data from the Kraków-Silesian Homocline, the results demonstrate that local and independent scaling frameworks violate spatial additivity, introducing potential artifacts. Conversely, the canonical global architecture satisfies spatial additivity, ensuring stable, continuous fault networks. Crucially, maintaining spatial additivity is vital for the incremental updating of geological models, enabling the seamless integration of newly acquired borehole data without triggering complete recalculations of the global feature space.
This manuscript investigates the influence of different data-scaling strategies on SVM-based fault classification using synthetic and borehole-derived TIN datasets. The topic is relevant to machine-learning-assisted geological modelling, particularly because preprocessing choices are often treated as a technical detail despite their potentially important influence on spatial predictions. The manuscript therefore addresses a useful methodological problem. However, in its present form, substantial revisions are required before the main conclusions can be fully supported.
I provide the following major comments.
1. Definition of spatial additivity and spatial continuity
The manuscript should distinguish more clearly between spatial additivity, prediction invariance, spatial continuity, and geological accuracy. The current subset experiment mainly evaluates whether predictions for the same observations change when those observations are processed within different spatial extents. This is a measure of prediction or spatial additivity, but it does not directly demonstrate that the resulting fault traces are geometrically or topologically continuous. I recommend introducing a formal mathematical definition of spatial additivity in the methods and using separate metrics for fault continuity.
2. Dataset description and reproducibility
Although the Study Area section provides enough geological information about the 810 boreholes, the actual dataset used in the study for the machine learning remains insufficiently described. A dedicated subsection should explain the spatial distribution and characteristics of the boreholes, the specific stratigraphic horizon extracted from each borehole, how these observations are converted to a TIN, the resulting number of vertices and triangles, and exactly what constitutes one observation supplied to the supplier.
The manuscript should also clearly separate the synthetic and real datasets. The number of synthetic surfaces, parameter ranges used for regional dip, dip direction, fault throw, noise, and other variables, the distributions from which these parameters are sampled, and the class proportions should be reported. A table summarizing these parameters would greatly improve reproducibility.
3. Relationship between synthetic and real data
The method assumes that the synthetic training surfaces represent the geometric characteristics of the real-borehole derived surface. This is a fundamental assumption and should be evaluated rather than only stated. I recommend comparing the distributions of the principal geometric attributes between synthetic and real datasets. Otherwise, differences in classification may partly result from synthetic-to-real domain shift rather than preprocessing alone.
4. Mathematical description of the three scaling methods
Since scaling architecture is the central methodological variable, the three approaches should be defined mathematically. This would make the distinction among independent subset scaling, per-surface local scaling, and training-derived global scaling much clearer than the current predominantly descriptive presentation.
The workflow figures should also be revised. Software-specific terms such as FIT_TRANSFORM and TRANSFORM are less informative for the general geological readership than explicitly showing where the mean and standard deviation are estimated and where those same parameters are subsequently reused. A single three-panel comparative figure may be more effective than separate figures.
6. Data-leakage discussion
The Data Leakage subsection is useful but relies heavily on a limited number of sources (Kumar and Narayan) and focuses mainly on conventional ML leakage. The discussion should be expanded to include spatial dependence and spatially correlated train/test observations, which are directly relevant to TIN-based geological data.
The authors should also distinguish classical data leakage from inconsistent preprocessing. Fitting a scaler using combined training and test data constitutes direct leakage, whereas independently fitting a transformation on the test or production dataset represents a different methodological problem. These situations should not be treated as conceptually identical.
7. Quantitative assessment of spatial additivity
The Results section currently relies too heavily on visual inspection. The number of changed classifications should be treated as a primary quantitative result rather than appearing mainly in the Discussion.
For each method, the authors should report at least the number and percentage of changed classifications, together with separate counts of fault-to-homocline and homocline-to-fault transitions. A formal disagreement metric could be defined as the proportion of observations within a subset whose predictions differ between complete-domain and subset processing.
The current analysis also appears to use only one selected rectangular subset. The result may depend strongly on subset size, position, point density, and local feature distribution. I therefore recommend repeating the experiment using multiple spatial subsets of different sizes and locations and reporting the mean, variability, and range of the disagreement metric for each preprocessing method.
Some of the tables appear in the Appendix instead of the Results, where they should be properly described.
8. Discussion
The Discussion requires substantial strengthening. Several parts currently repeat the Results, whereas greater emphasis should be placed on explaining the mechanism causing the observed behaviour, the conditions under which local scaling becomes problematic, the influence of subset size and feature distributions, and the geological implications of the classification patterns.
The suggestion that independently scaling real data could alleviate poor representation of reality by synthetic models should be treated cautiously. Independent standardization does not generally solve synthetic-to-real domain shift. Similarly, the suggestion that Method 2 may be preferable because it produces visually smoother patterns is not sufficiently supported, particularly because the manuscript also states that this method fails to identify some known faults. Visual smoothness should not be equated with geological accuracy.
The formal definition of spatial additivity currently introduced in the Discussion should instead appear in the Methods. Likewise, much of the literature discussion presented under “Bridging the research gap” would be more appropriate in the Introduction, where the novelty of the study is established.
Figure 1 appears two times one time after the abstract and one time in the methodology. I never understood why it appears two times. In addition, in the Introduction, 45-50 is not properly described for a non-expert reader. Also, in sentence 285 real data file word is not properly quoted; the commas used are wrong. Appendix A should go to the results and must be properly described and discussed. In addition, Appendix B.