Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 1 of 8 METHOD ARTICLE | SHORT Introducing INPOX: a method for informed point extraction from geological 2D surfaces exemplified on the Danish national hydrostratigraphic model Rasmus Bødker Madsen1* , Frederik Alexander Falk2 , Ingelise Møller1 , Anne-Sophie Høyer1 1Near-surface Land and Marine Geology, Geological Survey of Denmark and Greenland (GEUS), Aarhus, Denmark; 2Department of Geoscience, Aarhus University, Aarhus, Denmark Abstract This study presents a probabilistic method for extracting informed points from geological surfaces, named INPOX. The method generates a probability map from the existing surface by calculating the Laplacian at each location and combining it with a user-defined transfer function. A set of points from the surface is then extracted with a density proportional to the probability map. The method allows a de-coupling of the most informative points in the surface from points carrying less or even biased information. INPOX can be applied on any geological surface where the user needs to retrieve the structurally relevant parts and remove the information created by the initial interpolation. Here, we test INPOX on synthetic data, with and without supressing interpolation artifacts. In both cases, the informed points extracted with INPOX outperforms a uniform probability map in recreating the orig- inal features. We show that the method requires a minimum of points to be extracted for INPOX to be more informative than a uniform point retrieval. Finally, to showcase the strength of the method in both retrieving the relevant geological features and suppressing the existing interpolation artifacts, we apply INPOX to a real case surface from the Danish national hydrostratigraphic model. *Correspondence: rbm@geus.dk Received: 11 Dec 2023 Revised: 25 Feb 2024 Accepted: 12 Mar 2024 Published: 27 May 2024 Keywords: INPOX, Laplacian, decoupling information, interpolation artifacts, 3D layer models, artifact suppression Abbreviations FOHM: Danish national hydrostratigraphic model (Fælles Offentlig Hydrostratigrafisk Model) INPOX: INformed POint eXtraction LM: Laplacian map PM: Probability map RM: Random map MAD: Mean Absolute Deviation GEUS Bulletin (eISSN: 2597-2154) is an open access, peer-reviewed journal published by the Geological Survey of Denmark and Greenland (GEUS). This article is distributed under a CC-BY 4.0 licence, permitting free redistribution, and reproduction for any purpose, even commercial, provided proper citation of the original work. Author(s) retain copyright. Edited by: Julian Koch (GEUS, Denmark) Reviewers by: Jacob Skauvold (Norwegian Computing Center, Norway) Volume: 57 Funding: See page 7 Competing interests: See page 8 Additional files: None Introduction In 3D geological modelling, models are either constructed as voxel models (Van Der Meulen et al. 2013; Madsen et al. 2021a; Kawo et al. 2023) or layer models (Lelliott et al. 2006; Kaufmann & Martin 2008; Robins et al. 2008; Wycisk et al. 2009; Royse 2010; Madsen et al. 2022). In the case of layer mod- els, which are often used in subsequent hydrological modelling (Seifert et al. 2012; Enemark et al. 2023), geological units are defined by layer boundaries or surfaces, determined based on elevation on a 2D spatial grid. A common Tabular abstract Geographical coverage N/A Temporal coverage N/A Subject(s) covered Computational geoscience, informatics and remote sensing Method type The presented method is computational and is newly developed Method name INPOX Instruments and equipment used • Computer • Programming language Related publications N/A Potential application(s) for this method The method can be applied to retrieve the most informed points from a given geological surface while downplaying the significance of interpolation artifacts. This allows a re-interpolation or simulation of the surface without including the information on the initial interpolation. https://doi.org/10.34194/geusb.v57.8364 https://orcid.org/0000-0001-8538-7491 https://orcid.org/0009-0003-1023-5161 https://orcid.org/0000-0002-1154-3700 mailto:rbm@geus.dk https://creativecommons.org/licenses/by/4.0/deed.ast Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 2 of 8 GEUSBULLETIN.ORG method for generating such layer models is through manual interpretation where a series of interpretation points are placed in space based on the available geo- logical data or information and interpolation is used to estimate the surfaces between the interpreted points. Thus, the chosen interpolation method and the way the geologist interacts with the interpolation method influ- ence the characteristics of the final model and may also lead to interpolation artifacts on the surface (Groshong 2006; Yilmaz 2007; Setianto & Triandini 2015). Modellers often rely on commercial software with built-in interpo- lation routines, which may lead to a lack of awareness regarding specific interpolation routines and parame- terisation (Wellmann & Caumon 2018). This is true for both explicit and implicit geological modelling as some form of interpolation is a prerequisite in both modelling techniques. Ultimately, the final product of a 3D geological mod- elling exercise will be the model itself. Specifically for a layer model, the output is the layers, while the initial interpretation points and information about the inter- polation are considered by-products and are in many cases lost. This makes it difficult, if not impossible, to redo the interpolation to test alternative parameterisa- tions or to perform geostatistical simulation based on point information. It is therefore necessary to establish practical methods that can reverse engineer the process of interpolation to obtain a set of points from which the interpolation could be redone. The simplest way to sample from any surface is to draw a subset of the grid points using a uniform dis- tribution. To solve issues of underrepresentation of some areas by the uniform point extraction, sampling methods have been developed that ensure a more even spatial coverage (Brus et al. 2006; Walvoort et al. 2010). Within the scientific field of computer vision, the problem of surface recreation has been studied exten- sively with the purpose of rendering primary structural features and suppressing noise (see e.g. Agrawal et al. 2006; Boissonnat & Oudot 2006; Harker & O’Leary 2015). However, to our knowledge no method or algo- rithm is currently available that considers the spatial properties of the surface when sampling points from a (geological) surface, although Hansen (2021) and Madsen et al. (2021b) showed that information content differs between points. Thus, the selection of points matters in terms of recreating the essential features of the surface. In this short article, we address this problem and present a method, INformed POint eXtraction (INPOX), for extracting points of high geological importance from a geological surface while suppressing interpolation artifacts. The method is based on quantifying the curva- ture by calculating the Laplacian for the entire surface to assess its spatial properties and subsequently drawing points based on these values. The method is tested in a synthetic case to test the validity of INPOX in a con- trolled setting as well as demonstrated in an applied case with a cut-out from a pre-existing 2D surface from the Danish national hydrostratigraphic model (FOHM; Miljøstyrelsen 2023). Resources required The following resources are required for this method: 1. A computer 2. A programming language 3. Data in the form of (a) geological surface(s) Methodological protocols INPOX requires three steps: (1) construction of a Lapla- cian map (LM), (2) a probability map (PM) based on a transfer function between probability and Laplacians and (3) a random point retrieval based on the PM. These steps are explained in detail next. Laplacian To inform the point extraction in INPOX, the user must quantify the spatial characteristics of the surface. Here, the Laplacian is utilised to quantify the rate of change of the gradient in the surface, that is its cur- vature. The Laplacian (∆) defines the divergence (∇) of the gradient (∇z) of the elevation (z) of the given sur- face as: L z x y z x y z x y x z x y y , . , , ,2 2 2 2( ) ( ) ( ) ( ) = ∆ = ∇ ∇ = ∂ ∂ + ∂ ∂ (1) where x and y are the two spatial coordinates in the grid. To compute the Laplacian map, the operator is simply applied to all z in the geological surface producing a Laplacian map, with higher values relating to areas with more variation in the surface. Transfer functions To select areas with geologically relevant features, and to suppress the effect of unwanted simulation artifacts in the surface, a transfer function f(L) must be constructed that provides a probability draw- ing for  the current location (x,y) as a function of the Laplacian. A basis probability level P0 is first set in INPOX. P0 ensures that all areas of the surface have some probability of being drawn and that a desired frac- tion of the surface is drawn. For instance, if 5% of the surface should be drawn, P0= 0.05 is selected. In the case that, f(L) = P0, the result will be a random https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 3 of 8 GEUSBULLETIN.ORG selection from a uniform distribution. To manipulate the basis probability, a series of probability functions (f1 (L), f2 (L), f3 (L),…) can be introduced that either increases or decreases the basis probability depend- ing on the Laplacian. Madsen et al. (2021b) and Enemark et al. (2023) argue that major changes in the elevation of a geo- logical surface in manual interpretation models are related to geologically relevant features. This is argued since variations in elevation mainly arise due to active choices made by the interpreter, for exam- ple, when mapping rivers, buried valleys and faults. We introduce f1(L), which increases the probability of drawing a point from the surface as a function of the Laplacian L: f L CP L AL g 2 tanh ( ) 11 0 0( )( ) = − + (2) where CP0 is the maximum probability added to P0 and thereby sets importance of the informed point retrieval in contrast to the base level. g is the maximum derivative of f1 (L) with respect to L. L0 is a reference Laplacian and AL0 is the value of L where the steepest rise of f1 (L) occurs. A and g thereby define how quickly and how abruptly f(L) becomes (1+C)P0 rather than P0 when L increases. A graphical representation of f1 (L) can be seen in Fig. 1 (cyan shading, left hand side). Mathematical interpolation of geological point infor- mation results in spatial continuous surfaces. Hence, larger discontinuities can stem from either an active choice made by the modeller (e.g. introducing faults) or from the interpolation routines (e.g. using a limited neighbourhood of points locally instead of all points to make the algorithms more computationally feasible). In a system where there are no faults mapped or the Laplacian is significantly different than the faults, a sec- ond function f2(L) can be designed to reduce the base probability and minimise the effect of interpolation artifacts: f L C D P L BL EL DP ( ) ( ( ) ) 1 2 0 0 2 0 0( ) = + − + − (3) where DP0 is the maximum probability subtracted from P0 , and ELO functions as a dampener going from (1+C) P0 to (1−D)P0. BL0 is the Laplacian defining the switch between domain f1 (L) and f2 (L). A graphical represen- tation of f2 (L) can be seen in the purple part of Fig. 1. In INPOX, the following transfer function is chosen as a combination of f1 (L) (eq. 2) and f2 (L) (eq. 3) with BLO as the separation criteria between two: f L f L P L BL f L P L BL 1 0 0 2 0 0 ( ) ( ) ( )= + < + ≥     (4) (L >= BL0): f2(L) P0 BL0 (L < BL0): f1(L) Pr ob ab ilit y Laplacian (L) AL0 EL0 DP0 CP0 (1+C)P0 (1–D)P0 AL0 BL0 (B+E)L0 g (B–A)L0 f(L) = f1(L) + f2(L) + P0 L0 Fig. 1 Proposed transfer function f (L) in INPOX. The red line shows the probability (on the y-axis) as a function of ∆z (on the x-axis). The parameter B divides the values where f1 (L) and f2 (L) are used, shown by cyan and purple shading, respectively. Other parameters are explained in the main text. https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 4 of 8 GEUSBULLETIN.ORG Random point retrieval Once a PM is constructed, the task of drawing the points is trivial. Firstly, the PM is scaled such that the average probability of the entire grid becomes P0. A random map (RM) of similar size to the PM is constructed with val- ues from an uncorrelated uniform distribution between 0 and 1. At each location where the RM value is lower than the PM value, a point is extracted from the surface grid. Validation using synthetic data A synthetic reference (surface) model is created (Fig. 2A) with different features that need to be emphasised in (A) Reference model 0 25 50 75 100 Y –40 –30 –20 –10 0 Elevation (m a.s.l.) (B) Laplacian map 0 25 50 75 100 0 5 10 15 20 Laplacian 0 5 10 15 20 25 30 Laplacian 0 0.05 0.1 0.15 0.2 0.25 Pr ob ab ilit y (C) Transfer functions T1 T2 T3 Transfer functions (D) T1 Probability map 0 25 50 75 100 Y 0 0.05 0.1 0.15 0.2 0.25 Probability (H) T2 Probability map 0 25 50 75 100 Y (L) T3 Probability map X 0 25 50 75 100 Y 0 25 50 75 100 (E) T1 draw: 489 Points 0 25 50 75 100 (I) T2 draw: 502 Points X 0 25 50 75 100 (M) T3 draw: 498 Points (F) T1 interp. (MAD = 1.1214) 0 25 50 75 100 (J) T2 interp. (MAD = 1.2368) 0 25 50 75 100 (N) T3 interp. (MAD = 1.3885) X 0 25 50 75 100 (G) T1 Repeat (MAD = 1.0592) 0 25 50 75 100 (K) T2 Repeat (MAD = 1.1046) 0 25 50 75 100 (O) T3 Repeat (MAD = 1.2278) X 0 25 50 75 100 T1 T2 T3 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 0 10025 50 75 Fig. 2 Results from the synthetic case study. A: the reference model. B: the corresponding Laplacians. C: the transfer functions (T1, T2 and T3) used in the experiment. Results are arranged by transfer functions. T1: emphasises all large Laplacians. T2: emphasises all large Laplacians while downplaying the square. T3: uses a uniform distribution. Each row contains: a probability map (D, H, L); a single draw of point extraction (E, I, M); a linear interpo- lation of the drawn points (F, J, N), 100 repeats of drawing 500 points and interpolating (G, K, O). The Mean Absolute Deviation (MAD) is shown for each interpolation. On average 500 points are drawn from the three probability maps, but the number of points drawn in a single realisation varies as shown in panels E, I and M. https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 5 of 8 GEUSBULLETIN.ORG the point extraction. At X = 50 and Y = 75, a square is placed that is highly discontinuous with the surround- ing surface, while the other features are less discontinu- ous. The LM is calculated from the reference model and is presented in Fig. 2B. Figure 2C shows two different transfer functions (T1 and T2) to translate the LM into PMs as well as a transfer function where f(L) = P0 (red dotted line) that enables a comparison to random point retrieval (T3, orange line). For T1 (blue line), we treat all changes in the surface as relevant information and emphasise the full spectrum of Laplacians found in Fig. 2B. For T2 (red line), we treat the highly discontinuous square as an interpolation artifact, and only values < 10 are emphasised while values > 10 are suppressed. The parameterisation of each transfer function is presented in Table 1. Three PMs are obtained by applying the three transfer functions. The results are shown in Fig. 2D, H, L, where it is clearly visible that changes in elevation are reproduced using the INPOX strategy and that dif- ferent points are subsequently drawn based on these probabilities (Fig. 2E, I, M). In the third column (Fig. 2F, J, N), the points are linearly interpolated to create a new surface. The biggest difference between T1 and T2 and the uniform draw (T3) is the lack of dark red and black areas in T3 and thus a lack of structural informa- tion and a smoother output surface. Between T1 and T2, the square is more clearly defined in T1 as signifi- cantly more points were drawn to help define its shape. The Mean Absolute Deviation (MAD) is calculated in all cases. The MAD confirms the visual validation as T1 and T2 interpolations are closer to the reference model than a uniform draw. Thus, the extracted points carry more of the structurally important information for recreat- ing the reference model. The MAD for T2 is higher than T1 because the suppression of the square extends the difference to the reference model at that location. To ensure that these results did not arise because of a fluke draw, we repeat the experiment of drawing a set of points from the PM and interpolating between these 100 times to form a series of new surfaces. The average surface for each transfer function (displayed in Fig. 2G, K, O) confirms that T1 most accurately recreates the refer- ence model on average, while T2 recreates the reference model while downplaying the significance of the square. To assess the effect of the number of points being drawn in INPOX, we repeat the experiment, generating 100 surfaces from 100 set of points at various basis draw probabilities (here denoted as the draw percent- age) of the total grid size. Figure 3 shows the results for draw percentages in the interval between 1% and 13% of the surface being extracted, including the experiment shown in Fig. 2G, K, O that corresponds to a draw per- centage approximating 5%. In general, as the draw per- centage increases, the informed point retrievals from T1 and T2 decrease the MAD significantly more than T3. At approximately 3%, there are similar levels of MAD, and for very sparse data extraction (< 3%), the pattern is reversed, such that the uniform distribution out- performs T1 and T2. Because the informed extraction tends to cluster the drawn points, some areas may end up with a very low density of points when only a few points are extracted, which makes it difficult to recre- ate the surface in these areas. In contrast, the uniform extraction always secures an evenly spatial distribution of drawn points and is more robust in cases where data extraction is sparse. Naturally, these numbers vary depending on the choice of reference model and transfer function and should not be used directly as decisive indicators of where INPOX is a relevant point-extraction strategy. Instead, in a broader sense, these results indicate that INPOX can be a relevant strategy when a certain number of points need to be extracted, but some minimum of point draw percentage is needed for the method to be effective. Table 1 Transfer-function parameters and basis probability (draw percentage) for each of the transfer functions used in the study. T1–T3 are used in the synthetic case study, while T4 is used in the real-world case. Transfer function P0 [−] L0 [m -1] A[−] B[−] C[−] D[−] E[−] g[m] T1 0.05 1 1.5 25 4 1 1 2 T2 0.05 1 1.5 6.5 4 1 2 2 T3 0.05 - - - - - - T4 0.02/0.20 1 1.5 3 5 1 0.8 2 0 2 4 6 8 10 12 Ave. draw percent age (%) 0.5 1 1.5 2 2.5 3 3.5 4 M AD (1 00 d ra w s) MAD vs. draw percentage T1 T2 T3 Fig. 2 experiment Fig. 3 MAD for 100 point extractions and subsequent interpolations cal- culated as a function of draw percentage for the three transfer functions T1, T2 and T3. The experiment run in Fig. 2 is shown as a vertical black dotted line. https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 6 of 8 GEUSBULLETIN.ORG Demonstration using a real-world application We apply INPOX on a surface from the Danish national hydrostratigraphic model (FOHM; Fælles Offentlig Hydrostratigrafisk Model). This model was created by correlating and interpolating across multiple local models. This has led to many interpolation artifacts that reside amongst valuable geological features. This is exemplified in Fig. 4A through an excerpt of a Qua- ternary surface from the FOHM model covering 40.4 km2. A river-like structure passes through an area of the surface that is clearly affected by interpolation artifacts. These artifacts can be seen as round shapes (marked with arrows) due to interpolation distances set in the interpolation algorithm. We apply INPOX to extract points from the surface with the input param- eters shown in Table 1 as T4. From the LM (Fig. 4B), we assess that most of the geologically relevant fea- tures have Laplacians in the interval of 2 and 4, while the artifacts typically have Laplacians larger than ten but can go down to values around 5. Thus, we set B = 3 and C = 5P0 to emphasise the features and E = 0.8 to ensure that all Laplacians above 5 are downplayed in significance. In Fig. 4C, D, we show the result from a single draw with a draw percentage of 2% and 20%, respectively. They likely represent two extremes in terms of point extraction considering that the total number of inter- pretation points divided by the surface extent is a 9.9% coverage. In the case of the 2% extraction, the greatest concentration of points occurs along the river-like struc- ture and near the highest elevation peaks in the terrain. Conversely, point density is minimal in the low-lying areas, particularly to the south and west. For the 20% draw, the river-like structure is almost completely drawn out, while the artifacts are still difficult to detect visually. This demonstrates that INPOX does exactly as intended in a real-world setting. (A) FOHM-model layer 6.14 6.145 6.15 6.155 YU TM 106 –60 –40 –20 0 20 40 60 Elevation (m a.s.l.) (B) Laplacian map 5.35 5.4 5.45 5.5 XUTM 105 6.14 6.145 6.15 6.155 YU TM 106 0 10 15 20 Laplacian (C) Probability map (2%) 0 0.2 0.4 0.6 0.8 1 Probability (D) Probability map (20%) 5.35 5.4 5.45 5.5 XUTM 105 (E) 2% points drawn 5.35 5.4 5.45 5.5 XUTM 105 (F) 20% points drawn Int. artifacts Geol. feature 5 Fig. 4 Results from the real-world case study. A: an excerpt from a surface in the Danish national hydrostratigraphic model. B the corresponding Laplacian map. C: 2% draw from the surface using INPOX. D: 20% draw from the surface using INPOX. E: Random point extraction using P0 = 0.02. F: Random point extraction using P0 = 0.20. Geol.: geological. Int.: interpolation. https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 7 of 8 GEUSBULLETIN.ORG Discussion and outlook In this study we have suggested a method, INPOX, based on calculating the Laplacians at all locations and defining a suitable transfer function to emphasise and suppress certain structures in the surface. Thus, INPOX, to some extent, makes it possible to reverse engineer the process of geological interpolation by extracting the relevant geological features and making it possible to re-interpo- late the information. We have successfully validated and demonstrated the applicability of the method in both a theoretical and practical case where geological features are identified and are able to guide the point extraction while interpolation artifacts are removed. Obviously, INPOX cannot recreate the exact location of points placed by the geologist during interpretation and thus some of the interpolation information cannot be removed using INPOX. By setting the draw percent- age, the user can find a suitable trade-off between (1) removing all interpolation information while removing some geological information or (2) keeping all geo- logical information while keeping some interpolation information. If some geologically relevant points are available from for example borehole information, it is trivial to include these in the draw by setting the draw probability in the PM to 1. Madsen et al. (2021b) suggest using the gradient as a measure to calculate the points that are most relevant for describing the surface. As was the case for the Laplacians, the user would look for areas with a steep gradient. Although the gradient will likely provide a reasonable indication of geologically relevant areas of the model, it is diffi- cult to keep information on sharp transitions in the surface. Our tests, although not shown here, confirm this, as the MAD is always worse when using the gra- dient approach rather than the Laplacian approach. Since the Laplacian is the divergence of the gradient, it is low where the gradient is steep, and high adjacent to these areas. This is more logical because it is more informative to have points where a surface structure begins and ends rather than having the point informa- tion on the decline and then recreating the start and end-points with interpolation. The design of the transfer function matters for the resulting point extraction. In the current setup, INPOX uses a baseline probability that is manipulated with two sepa- rate transfer functions, one with the ability to enhance probability and one being able to decrease probability. As demonstrated, one can easily manipulate the param- eters to fit the surface features of interest and problems at hand. Furthermore, since all steps in the algorithm have negligible execution times, even on large surfaces, one can fine-tune the parameters of the transfer function in an iterative process until a satisfactory point-extraction strat- egy is found. For larger surfaces, it might be necessary to construct region-specific transfer functions, however one could also apply one global conservative transfer function that only suppresses extreme Laplacian values. It is important to stress that INPOX is not limited either by the specific parameterisation or the applied transfer function. One can freely choose another more specific or intricate transfer function if it provides a mapping between the Laplacian and draw probability. For instance, and not shown here, we adapted a piecewise linear func- tion mimicking the T2 transfer-function to a degree that very similar MAD results could be obtained. To conclude, the practitioner should complete a problem-specific assessment of the (spatial) distribution of Laplacians and then make an informed decision on the transfer-function that makes most sense geologically. Finally, our results on the effect of the draw percent- age indicate that choosing INPOX over any other possible point extraction method, as exemplified with the uni- form distribution, follows the well-known ‘no-free-lunch’ theorem (Mosegaard 2012; Shalev-Shwartz & Ben-David 2014). This states that the effectiveness of INPOX will be problem dependent. Thus, in some scenarios it could be better to implement another point extraction strategy. If the surface is rapidly fluctuating between neighbouring points, the entropy is high and the information content between points is little to non-existent. In this case, no point extraction method would be preferable. In general, the basic requirement for INPOX is that spatial correla- tions exist in the surface and furthermore that the surface curvature relates to geological information or artifacts. We see two main scenarios where INPOX can be a most valuable tool: (1) when dealing with national mod- els, where quality control is often limited to assessing the behaviour of the surface at local scale through cross-sections or possibly 2D maps of the surface and (2) in areas where models are created using different strategies or operate at different scales depending on the resources available for model creation. Under these circumstances, INPOX can be a valuable tool to homogenise what the points represent between differ- ent models and even within a single model. Acknowledgements The authors wish to thank Geocenter Denmark and the project DECODE-3D for providing funding for writing the manuscript. The method development was carried out in a consultancy project (Nret 2022–2024) conducted for the Danish Environmental Protection Agency (Miljøstyrelsen). Additional information Funding statement The method was developed during consultancy work for the Danish EPA, while the hours for turning the results into a scientific contribution and writing the manuscript was provided by Geocenter Denmark through the project DECODE3D. https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ Madsen et al. 2024: GEUS Bulletin 57. 8364. https://doi.org/10.34194/geusb.v57.8364 8 of 8 GEUSBULLETIN.ORG Authors’ contribution RBM: Conceptualisation, software, investigation, writing – original draft preparation FAF: Conceptualisation, methodology, software, investigation, writing – reviewing and editing IM: Conceptualisation, supervision, investigation, writing – reviewing and editing ASH: Project administration, supervision, writing – reviewing and editing Competing interests The authors declare no competing interests. Additional files None provided References Agrawal, A., Raskar, R. & Chellappa, R. 2006: What is the range of sur- face reconstructions from a gradient field? In: Leonardis, A., Bischof, H. & Pinz, A. (eds): Computer Vision – ECCV 2006. ECCV 2006. Lec- ture Notes in Computer Science 3951, 578–591. Springer. https://doi. org/10.1007/11744023_45 Boissonnat, J.D. & Oudot, S. 2006: Provably good sampling and meshing of Lipschitz surfaces. In: SCG ’06: Proceedings of the Twenty-second Annual Symposium on Computational Geometry, 337–346. https:// doi.org/10.1145/1137856.1137906 Brus, D.J., de Gruijter, J.J. & van Groenigen, J.W. 2006: Chapter 14 design- ing spatial coverage samples using the k-means clustering algorithm. Developments in Soil Science 31, 183–192. https://doi.org/10.1016/ S0166-2481(06)31014-8 Enemark, T. et al. 2023: Incorporating interpretation uncertainties from deterministic 3D hydrostratigraphic models in groundwater models. Hydrology and Earth System Sciences [preprint/in press]. https://doi. org/10.5194/hess-2023-74 Groshong, R.H. 2006: 3-D structural geology: a practical guide to quantitative surface and subsurface map interpretation. 2nd ed. 400 pp. Springer. Hansen, T.M. 2021: Entropy and information content of geostatis- tical models. Mathematical Geosciences 53, 163–184. https://doi. org/10.1007/s11004-020-09876-z Harker, M. & O’Leary, P. 2015: Regularized reconstruction of a sur- face from its measured gradient field: algorithms for spectral, Tikhonov, constrained, and weighted regularization. Journal of Mathematical Imaging and Vision 51, 46–70. https://doi.org/10.1007/ s10851-014-0505-4 Kaufmann, O. & Martin, T. 2008: 3D geological modelling from bore- holes, cross-sections and geological maps , application over former natural gas storages in coal mines. Computers & Geosciences 34, 278–290. https://doi.org/10.1016/j.cageo.2007.09.005 Kawo, N.S., Korus, J. & Gulbrandsen, M.L. 2023: Multiple-point statis- tical modeling of three-dimensional glacial aquifer heterogeneity for improved groundwater management. Hydrogeology Journal 31, 1525–1546. https://doi.org/10.1007/s10040-023-02658-x Lelliott, M.R., Bridge, D.M., Kessler, H., Price, S.J. & Seymour, K.J. 2006: The application of 3D geological modelling to aquifer recharge assessments in an urban environment. Quarterly Journal of Engi- neering Geology and Hydrogeology 39, 293–302. https://doi. org/10.1144/1470-9236/05-027 Madsen, R.B., Høyer, A.-S., Andersen, L.T., Møller, I. & Hansen, T.M. 2022: Geology-driven modeling: a new probabilistic approach for incorporating uncertain geological interpretations in 3D geological modeling. Engineering Geology 309, 106833. https://doi.org/10.1016/j. enggeo.2022.106833 Madsen, R.B., Kim, H., Kallesøe, A.J., Sandersen, P.B.E., Vilhelmsen, T.N., Hansen, T.M., Christiansen, A.V., Møller, I. & Hansen, B. 2021a: 3D multiple-point geostatistical simulation of joint subsurface redox and geological architectures. Hydrology and Earth System Sciences 25, 2759–2787. https://doi.org/10.5194/hess-25-2759-2021 Madsen, R.B., Møller, I. & Hansen, T.M. 2021b: Choosing between Gauss- ian and MPS simulation: the role of data information content – a case study using uncertain interpretation data points. Stochastic Environ- mental Research and Risk Assessment 35, 1563–1583. https://doi. org/10.1007/s00477-020-01954-2 Mosegaard, K. 2012: Limits to nonlinear inversion. In: Jónasson, K. (ed.): Applied parallel and scientific computing, 11–21. Springer. https://doi. org/10.1007/978-3-642-28151-8_2 Miljøstyrelsen 2023: FOHM – Fælles offentlig hydrologisk model (Dan- ish national hydrostratigraphic model). https://data.geus.dk/geusma- p/?mapname=fohm&lang=en#baslay=baseMapDa&optlay=&ex- tent=168222.2222222222,6019065.972222222,873777.7777777778, 6430934.027777778 Robins, N.S., Davies, J. & Dumpleton, S. 2008: Groundwater flow in the South Wales coalfield: historical data informing 3D modelling. Quar- terly Journal of Engineering Geology and Hydrogeology 41, 477–486. https://doi.org/10.1144/1470-9236/07-055 Royse, K.R. 2010: Combining numerical and cognitive 3D modelling approaches in order to determine the structure of the Chalk in the London Basin. Computers & Geosciences 36, 500–511. https://doi. org/10.1016/j.cageo.2009.10.001 Seifert, D., Sonnenborg, T.O., Refsgaard, J.C., Højberg, A.L. & Troldborg, L. 2012: Assessment of hydrological model predictive ability given mul- tiple conceptual geological models. Water Resources Research 48, 1–16. https://doi.org/10.1029/2011WR011149 Setianto, A. & Triandini, T. 2015: Comparison of Kriging and inverse dis- tance weighted (Idw) interpolation methods in lineament extraction and analysis. Journal of Applied Geology 5, 21–29. https://doi. org/10.22146/jag.7204 Shalev-Shwartz, S. & Ben-David, S. 2014: Understanding machine learn- ing: from theory to algorithms. Cambridge University Press. 397 pp. https://doi.org/10.1017/CBO9781107298019 Van Der Meulen, M.J. et al. 2013: 3D geology in a 2D country: Perspec- tives for geological surveying in the Netherlands. Geologie en Mijn- bouw/Netherlands Journal of Geosciences 92, 217–241. https://doi. org/10.1017/S0016774600000184 Walvoort, D.J.J., Brus, D.J. & De Gruijter, J.J. 2010: An R package for spatial coverage sampling and random sampling from compact geographical strata by k-means. Computers & Geosciences 36, 1261–1267. https:// doi.org/10.1016/j.cageo.2010.04.005 Wellmann, J.F. & Caumon, G. 2018: 3-D Structural geological models: concepts, methods, and uncertainties. Advances in Geophysics 59, 121 pp. Elsevier Inc. https://doi.org/10.1016/bs.agph.2018.09.001 Wycisk, P., Hubert, T., Gossel, W. & Neumann, C. 2009: High-resolution 3D spatial modelling of complex geological structures for an environ- mental risk assessment of abundant mining and industrial megasites. Computers & Geosciences 35, 165–182. https://doi.org/10.1016/j. cageo.2007.09.001 Yilmaz, H.M. 2007: The effect of interpolation methods in surface defini- tion: an experimental study. Earth Surface Processes and Landforms 32, 1346–1361. https://doi.org/10.1002/esp.1473 https://doi.org/10.34194/geusb.v57.8364 http://www.geusbulletin.org/ https://doi.org/10.1007/11744023_45 https://doi.org/10.1007/11744023_45 https://doi.org/10.1145/1137856.1137906 https://doi.org/10.1145/1137856.1137906 https://doi.org/10.1016/S0166-2481(06)31014-8 https://doi.org/10.1016/S0166-2481(06)31014-8 https://doi.org/10.5194/hess-2023-74 https://doi.org/10.5194/hess-2023-74 https://doi.org/10.1007/s11004-020-09876-z https://doi.org/10.1007/s11004-020-09876-z https://doi.org/10.1007/s10851-014-0505-4 https://doi.org/10.1007/s10851-014-0505-4 https://doi.org/10.1016/j.cageo.2007.09.005 https://doi.org/10.1007/s10040-023-02658-x https://doi.org/10.1144/1470-9236/05-027 https://doi.org/10.1144/1470-9236/05-027 https://doi.org/10.1016/j.enggeo.2022.106833 https://doi.org/10.1016/j.enggeo.2022.106833 https://doi.org/10.5194/hess-25-2759-2021 https://doi.org/10.1007/s00477-020-01954-2 https://doi.org/10.1007/s00477-020-01954-2 https://doi.org/10.1007/978-3-642-28151-8_2 https://doi.org/10.1007/978-3-642-28151-8_2 https://data.geus.dk/geusmap/?mapname=fohm&lang=en#baslay=baseMapDa&optlay=&extent=168222.2222222222,6019065.972222222,873777.7777777778, 6430934.027777778 https://data.geus.dk/geusmap/?mapname=fohm&lang=en#baslay=baseMapDa&optlay=&extent=168222.2222222222,6019065.972222222,873777.7777777778, 6430934.027777778 https://data.geus.dk/geusmap/?mapname=fohm&lang=en#baslay=baseMapDa&optlay=&extent=168222.2222222222,6019065.972222222,873777.7777777778, 6430934.027777778 https://data.geus.dk/geusmap/?mapname=fohm&lang=en#baslay=baseMapDa&optlay=&extent=168222.2222222222,6019065.972222222,873777.7777777778, 6430934.027777778 https://doi.org/10.1144/1470-9236/07-055 https://doi.org/10.1016/j.cageo.2009.10.001 https://doi.org/10.1016/j.cageo.2009.10.001 https://doi.org/10.1029/2011WR011149 https://doi.org/10.22146/jag.7204 https://doi.org/10.22146/jag.7204 https://doi.org/10.1017/CBO9781107298019 https://doi.org/10.1017/S0016774600000184 https://doi.org/10.1017/S0016774600000184 https://doi.org/10.1016/j.cageo.2010.04.005 https://doi.org/10.1016/j.cageo.2010.04.005 https://doi.org/10.1016/bs.agph.2018.09.001 https://doi.org/10.1016/j.cageo.2007.09.001 https://doi.org/10.1016/j.cageo.2007.09.001 https://doi.org/10.1002/esp.1473 Introducing INPOX: a method for informed point extraction from geological 2D surfaces exemplified on the Danish national hydrostratigraphic model Introduction Resources required Methodological protocols Laplacian Transfer functions Random point retrieval Validation using synthetic data Demonstration using a real-world application Discussion and outlook Acknowledgements Additional information References Figures Fig. 1 Proposed transfer function f (L) in INPOX. The red line shows the probability (on the y-axis) as a function of Δz (on the x-axis). The parameter B divides the values where f1 (L) and f2 (L) are used, shown by cyan and purple shading, respectively. Other parameters are explained in the main text. Fig. 2 Results from the synthetic case study. A: the reference model. B: the corresponding Laplacians. C: the transfer functions (T1, T2 and T3) used in the experiment. Results are arranged by transfer functions. T1: emphasises all large Laplacians. T2: emphasises all large Laplacians while downplaying the square. T3: uses a uniform distribution. Each row contains: a probability map (D, H, L); a single draw of point extraction.... Fig. 3 MAD for 100 point extractions and subsequent interpolations calculated as a function of draw percentage for the three transfer functions T1, T2 and T3. The experiment run in Fig. 2 is shown as a vertical black dotted line. Fig. 4 Results from the real-world case study. A: an excerpt from a surface in the Danish national hydrostratigraphic model. B the corresponding Laplacian map. C: 2% draw from the surface using INPOX. D: 20% draw from the surface using INPOX. E: Random point extraction using P0 = 0.02. F: Random point extraction using P0 = 0.20. Geol.: geological. Int.: interpolation. Tables Table 1 Transfer-function parameters and basis probability (draw percentage) for each of the transfer functions used in the study. T1–T3 are used in the synthetic case study, while T4 is used in the real-world case.