Figure gallery for spatial data analysis

Standard, reusable figure designs from our team's published work. Search by analysis, figure type or method; open a figure to view it in full, or its caption to see why it is a standard design and which studies it suits.

Feature engineering

Generating GCF variables from original predictors

Fig. 5 in Song, Y., 2026. Generalized covariate field (GCF): spatial-pattern and neighbourhood-distribution feature expansion improves geospatial prediction. International Journal of Geographical Information Science, pp.1–29.pdfsource

How the generalized covariate field (GCF) expands 12 environmental predictors into spatial pattern features (996) and neighbourhood distribution features (2,268), with the nine derived features retained for predicting vascular plant species richness in the Southwest Australian Floristic Region.

Why it is a standard design

It is the standard way to show a feature expansion: the original predictors appear as a stack on the left, arrows split them into the derived feature families, and each branch states how many features it creates (for example 12 (x) × 9 (b) × 21 (τ) = 2268). The selected features are then shown as a grid of small maps with the same extent, a shared colour ramp per family (green for spatial pattern, orange for neighbourhood distribution), and the feature name in its formula notation above each map, so readers can see both how many variables were made and what the kept ones look like.

Reuse for: any study that derives new spatial covariates before modelling, such as focal statistics, texture measures or multi-scale features for machine-learning prediction. GCF is available in Heterra; see also the GCF blog series.

Model evaluation

Comparing prediction models on two metrics: SIE and R²

Combined from Figs. 8 and 14 in Song, Y., 2026. Spatial information explained by prediction models under block cross-validation. ISPRS Journal of Photogrammetry and Remote Sensing, 242, pp.984–1000.pdfsource

How ten machine-learning models predicting C4 natural grass area across Australia rank by accuracy (R²) and by the spatial information they explain (SIE), under spatial block and random cross-validation.

Why it is a standard design

It is the standard two-metric comparison plot: each model is one point with the new metric on the x-axis and the familiar one on the y-axis, error bars on both axes show uncertainty, and every point carries a direct label instead of relying only on the legend. Each model keeps the same colour and marker in both panels, and the panels place the two validation designs side by side with a shared legend, so readers can see at once where the two metrics agree, where they rank models differently, and how random cross-validation shifts every model up and to the right.

Reuse for: benchmarking several models or methods on two criteria, such as accuracy against interpretability, error against runtime, or one validation design against another. SIE is available in Heterra; see also the SIE blog series.

Prediction and uncertainty

Second dimension of spatial association (SDA): concept, predictions and uncertainty

Figs. 1, 7 and 9, combined here, in Song, Y., 2022. The second dimension of spatial association. International Journal of Applied Earth Observation and Geoinformation, 111, p.102834.pdfsource

The standard workflow and results of SDA, in two columns. The top diagrams contrast the first dimension of spatial association (FDA), which links a response sample to explanatory values at the same location or at other sample locations, with the SDA, which uses the explanatory information within the neighbourhood of each sample. Below them are prediction maps of soil Cr and Cu from MLR, random forest (RF) and RF kriging (RFK) with FDA and SDA variables, and the uncertainty of FDA-RFK and SDA-RFK as maps and density curves, with maps of their difference and pie charts of the share of locations where the uncertainty is reduced.

Why it is a standard design

It places each concept directly above its evidence: the two schematics use the same stacked response and explanatory layers and the same symbols for samples and for the data used in modelling, so the only visible difference is how explanatory information is taken; the prediction maps are then a grid of variable types (rows) by models (columns) on one colour scale per element, and the uncertainty is shown as paired maps on one scale, density curves with mean lines, and a difference map with a pie chart giving the share of locations that improve. A reader can follow the idea, the predictions and the gain in one figure.

Reuse for: introducing a new way of deriving spatial variables or features and showing its effect, where a schematic of the old and new idea is followed by predictions across several models and a map of where the uncertainty changes.

Prediction and uncertainty

Geographically optimal similarity: optimal similarity, predictions and uncertainty

Figs. 7, 9 and 11(a, b), combined here, in Song, Y., 2023. Geographically optimal similarity. Mathematical Geosciences, 55(3), pp.295-320.pdfsource

The standard result figure of geographically optimal similarity (GOS), in three parts. First, the optimal similarity: cross-validation RMSE against the share of samples used, showing that each prediction point uses only the most similar 4% to 5% of samples. Second, the spatial distribution of the predictions of soil Cu, Zn and Pb by GOS, next to ordinary kriging (OK), multiple linear regression (MLR), regression kriging (RK) and the similarity-based BCS model. Third, the spatial distribution of the corresponding uncertainty, as cumulative distributions and maps for BCS and GOS.

Why it is a standard design

It reports a new prediction model in the standard order of parameter, prediction and uncertainty: a tuning curve of cross-validation error against the key parameter with the optimum marked, prediction maps where the proposed model is shown large and the benchmark models small on the same colour scale for each variable, and uncertainty shown both as cumulative distributions across thresholds and as maps on the same study area. Readers can check that the parameter is optimal, compare the spatial detail of each model directly and see where the predictions can be trusted.

Reuse for: presenting any new spatial prediction or interpolation model against benchmark models, such as soil, air quality or climate surfaces, wherever a key parameter is tuned and prediction uncertainty can be mapped.

Workflow

A standard multi-component analysis workflow

Fig. 1 in Song, Y., Awange, J., Wang, J., Xu, D. and Wu, C., 2026. Spatially clustered global wildfire trends and their climate, fuel and human drivers (2001–2025): A spatio-temporal trend-cluster attribution approach. Geography and Sustainability, p.100530.pdfsource

How burned-area data and climate, fuel and human drivers move through three linked models: trend clustering, spatial attribution and temporal attribution.

Why it is a standard design

It follows the standard flowchart grammar: every element type has its own shape and colour, explained in a legend (yellow parallelograms for input data, pink boxes for processes, blue boxes for outcomes), arrows run one way from inputs through processes to outputs, and each model component is a lettered panel (a)–(d). Readers already know this grammar, so they can trace any path through the analysis without reading every box.

Reuse for: any study that chains several models, such as clustering → attribution → temporal analysis. The spatial attribution step uses OPGD, available in Heterra.

Review overview

A conceptual framework of GeoAI beyond mapping

Fig. 1 in Song, Y., Biljecki, F., Camps-Valls, G. and Atkinson, P.M., 2026. GeoAI: Beyond mapping earth and cities through explainability, adaptability, and sustainability. iScience, 29(1), p.114407.pdfsource

How eleven GeoAI studies move beyond traditional mapping toward shaping Earth and city systems, classified along three dimensions: explainability, adaptability and sustainability, each split into sub-dimensions.

Why it is a standard design

It is the standard framework matrix for a review: the studies form the columns, each with a short descriptive title and a thumbnail of its result, and the framework’s dimensions and sub-dimensions form the rows. Each dimension has its own colour, icon and marker shape (squares, hexagons, circles), and a filled marker means the study covers that sub-dimension while an empty one means it does not, so readers can scan across a row to see how often a property is addressed and down a column to see one study’s profile.

Reuse for: any review or perspective paper that classifies a set of studies, methods or tools against several criteria, such as data types used, validation designs, or which challenges each method addresses.

Review overview

A framework linking geospatial methods with global resource and sustainability challenges

Fig. 1 in Song, Y., Helmholz, P., Su, F., Zhou, C., Lotfata, A., Zohar, M., González Leonardo, M. and Sila-Nowicka, K., 2025. Advancing geospatial methods for addressing global resource and sustainability challenges. Resources, Conservation and Recycling, 223, p.108517.pdfsource

How the reviewed studies combine five groups of geospatial methods, from geospatial data and remote sensing to hybrid models for complex systems, with six challenges, from biodiversity and ecosystem services to supply chains and inequality.

Why it is a standard design

It is the standard two-axis framework for a review: one axis orders the methods by increasing complexity (from data acquisition to complex system modelling) and the other orders the application challenges (from ecosystems to human interventions), each with a labelled arrow. Every cell where the axes cross lists the studies that use that method for that challenge, empty cells show the gaps, and an illustrated landscape block along the challenge axis makes each challenge recognisable at a glance.

Reuse for: the framework figure of a review, special issue editorial or research agenda, whenever the literature can be organised by two ordered dimensions such as methods and applications, scales and sectors, or data and questions.

Review overview

Applications of GeoAI: where and what the reviewed studies map

Fig. 1 in Song, Y., Kalacska, M., Gašparović, M., Yao, J. and Najibi, N., 2023. Advances in geocomputation and geospatial artificial intelligence (GeoAI) for mapping. International Journal of Applied Earth Observation and Geoinformation, p.103300.pdfsource

Where the GeoAI and geocomputation studies in this review were carried out, and what each one mapped, grouped into buildings and infrastructure, natural environment and natural hazards, land use analysis, and social issues and human activities.

Why it is a standard design

It is the standard overview figure for a review or special issue: a world map in the centre marks every case study with a point coloured by application category, zoomed circles enlarge the regions where studies cluster (Europe and China), and leader lines connect each point to a thumbnail of that study’s result in a panel framed in the same colour. Readers can see at once where the research was done, which kinds of applications dominate, and what each study actually produced.

Reuse for: the opening figure of a review paper, a special issue editorial or a thesis introduction, whenever you need to summarise many studies by location and by topic in one view.

Review overview

Linking a research field to the Sustainable Development Goals

Fig. 3 in Song, Y. and Wu, P., 2021. Earth observation for sustainable infrastructure: A review. Remote Sensing, 13(8), p.1528.pdfsource

How Earth observation for sustainable infrastructure (EOSI) relates to every target of the 17 Sustainable Development Goals (SDGs), across five infrastructure sectors: energy, transport, water, solid waste and digital communication.

Why it is a standard design

It is the standard SDG linkage matrix: each goal keeps its official icon and colour, every target gets its own column, and the five sectors are fixed rows repeated under every goal. Each cell pairs two marks, a square for how infrastructure influences the target (direct or indirect) and a circle for what Earth observation can offer (an available, direct or indirect indicator, or irrelevant), with grey for no link, so readers can scan the whole SDG framework at once and see where a field contributes and where the gaps are.

Reuse for: any review or policy paper that maps a research field, data source or technology against the SDG targets, or against another standard framework of goals and indicators.

Spatio-temporal trends

Spatially varied temporal trends

Fig. 3 in Song, Y., Awange, J., Wang, J., Xu, D. and Wu, C., 2026. Spatially clustered global wildfire trends and their climate, fuel and human drivers (2001–2025): A spatio-temporal trend-cluster attribution approach. Geography and Sustainability, p.100530.pdfsource

Where burned area has significantly increased or decreased since 2001, and how each regional trend compares with the global trend.

Why it is a standard design

It uses the standard “map plus small multiples” design for spatio-temporal trends: the map locates each region, and every region gets a small time-series panel with the same layout (observations, a linear trend line with its confidence band, the slope ± error and p-value, and the global series in grey as a common baseline). One colour per trend class links each panel to its region on the map.

Reuse for: gridded time series such as NDVI, land surface temperature, PM2.5 or night-time light, whenever trends differ between regions and you want to show both where and how.

Workflow

Explaining the steps of a spatial model: second-dimension outliers

Fig. 1 in Ren, K., Song, Y. and Yu, Q., 2026. Second-dimension outliers for spatial prediction. International Journal of Geographical Information Science, 40(6), pp.1915-1942.pdfsource

How the SDO model works, from data to model. The dependent variable and the stacked explanatory variables are drawn as point layers. Step 1 sets local ranges as buffers of increasing size (bα) around a sampled location u; Step 2 identifies local outliers within each range; Step 3 sums them into positive and negative outlier strength indices (OSI); and Step 4 relates the dependent variable to the explanatory variables and the OSIs, Y(u) = f(X, I(X, bα, v)).

Why it is a standard design

It follows the standard grammar for explaining a spatial method: data are shown as tilted map layers (one layer per variable, stacked for many variables), steps are numbered and named in bold and run top to bottom with the same arrow, each step is illustrated on the same small schematic map with the parameter changing between copies (α = 1, 2, …), the formulas appear in boxes at the step where they are defined, and every symbol is explained in a notation key beside the step that uses it. Readers can follow the theory without reading the equations first, and the same template fits most local or neighbourhood-based spatial methods.

Reuse for: explaining a new spatial model or algorithm in a methods paper, especially methods built on neighbourhoods, buffers, windows or local statistics around sample locations.

Workflow

Spatial multi-criteria decision making: from objective to scored alternatives

Fig. 4 in Song, Y., Thatcher, D., Li, Q., McHugh, T. and Wu, P., 2021. Developing sustainable road infrastructure performance indicators using a model-driven fuzzy spatial multi-criteria decision making method. Renewable and Sustainable Energy Reviews, 138, p.110538.pdfsource

The decision question (which indicator best describes the burden of road maintenance) is broken into three criteria (road, vehicles, climate), each with its sub-criteria drawn as stacked map layers. The MFSD method is used twice: (1) to compute a new MFSD-based indicator from four existing alternatives (deflection, curvature, roughness, rutting), and (2) to rank all five alternatives against the criteria and give each a score.

Why it is a standard design

It follows the standard hierarchy of a spatial decision problem, read from left to right: objective, criteria, sub-criteria and alternatives as labelled columns, criteria in colour-coded diamonds, every sub-criterion shown as a map layer in a stack of the same colour, and the method steps as numbered boxes with coloured arrows that show which inputs go where. Readers see the whole decision structure and the order of the steps at once, and anyone can copy it by swapping in their own objective, criteria, layers and method.

Reuse for: site selection, suitability or vulnerability assessment, indicator selection, and any GIS-based multi-criteria decision study that weights spatial layers and ranks alternatives.

Workflow

Redefining city boundaries from points of interest, step by step

Fig. 3 in Song, Y., Long, Y., Wu, P. and Wang, X., 2018. Are all cities with similar urban form or not? Redefining cities with ubiquitous points of interest and evaluating them with indicators at city and block levels in China. International Journal of Geographical Information Science, 32(12), pp.2447-2476.pdfsource

The process of delineating cities from POI data: POI points are turned into a density surface with a kernel density function at 500 m resolution; the cumulative distribution of POI density is used to choose the classification threshold (50 points/km²); the classified density map is converted to polygons; and polygons of at least 2 km² are kept as the redefined cities.

Why it is a standard design

It is a map-based workflow: every step is shown as the actual map of the same small area at that stage, arrows between the panels carry the operation and its parameter (resolution, threshold, minimum area), and the one decision that needs evidence, the density threshold, gets its own chart in the sequence. Readers can check each operation visually and repeat it with the same settings, which a box-and-arrow flowchart cannot show.

Reuse for: any GIS processing chain that turns raw spatial data into a product, such as delineating urban areas, hotspots, habitats or service areas from points, images or networks with density, threshold and area rules.

Spatial anomalies

Second-dimension outliers: spatial distribution of local anomalies

Figure in Ren, K., Song, Y. and Yu, Q., 2026. Second-dimension outliers for spatial prediction. International Journal of Geographical Information Science, 40(6), pp.1915-1942.pdfsource

The core result of second-dimension outliers (SDO): the spatial distribution of positive and negative local anomalies as the buffer distance grows from 100 to 700 km. Each column is one buffer distance; the national maps mark the study region, and the enlarged views above (positive, red) and below (negative, blue) show how the anomalies emerge and spread.

Why it is a standard design

It is the standard small-multiple design for a scale-dependent local result: one column per parameter value in increasing order, the same map extent in every column, positive and negative values split into two rows with their own sequential colour scales that meet at zero, and an overview map with a box linked to an enlarged view of the region of interest. Readers can see at a glance where anomalies occur, whether they are high or low, and at which scale they appear, and the layout can be rebuilt for any local statistic computed over several neighbourhood sizes.

Reuse for: mapping local outliers, hot and cold spots or any local statistic across several buffer distances, bandwidths or neighbourhood sizes, especially when positive and negative values carry different meanings.

Multi-scale mapping

Segmenting a road network at several scales: maps, profiles and evaluation

Figs. 1, 4, 5 and 8, combined here, in Song, Y., Wu, P., Gilmore, D. and Li, Q., 2021. A spatial heterogeneity-based segmentation model for analyzing road deterioration network data in multi-scale infrastructure systems. IEEE Transactions on Intelligent Transportation Systems, 22(11), pp.7073-7083.pdfsource

The standard results of the spatial heterogeneity-based segmentation (SHS) model, combined from four figures. Top left: the study area and road network in Western Australia, with a state locator. Bottom left: curvature and deflection mapped by project-level (100–500 m) and network-level (1–5 km and 10–50 km) segments, with zoomed insets of one road. Top right: nested step charts of the segment mean, median and 75% quantile along that road, zooming from 10–50 km segments down to the raw observations. Bottom right: homogeneity within segments against heterogeneity among segments for SHS and two other methods (CDA, MCV).

Why it is a standard design

It is the standard way to present a segmentation of line or area data across scales: a locator and study area map first; then the same variables mapped at each scale side by side with the same insets and legends; then a nested zoom of profiles, linked by dashed lines, so each finer scale is seen inside the coarser one down to the raw data; and finally a two-criterion scatter (homogeneity within against heterogeneity among segments) that compares methods, with symbol colour for method, shape for variable and size for the number of segments. Readers can see what the segments look like, how they change with scale and why one method is better.

Reuse for: segmenting or zoning roads, rivers, pipelines, coastlines or regions at several scales, and any study that needs to show results at nested scales and compare zoning methods.

Multi-scale mapping

Global and local mapping of highly uneven distributions

Fig. 8 in Song, Y., Wu, P., Hampson, K., and Anumba, C., 2021. Assessing block-level sustainable transport infrastructure development using a spatial trade-off relation model. International Journal of Applied Earth Observation and Geoinformation, 105, 102585.pdfsource

Hotspots and cold spots of infrastructure development across Western Australia, where most blocks are packed into the Perth metro area and the rest are spread over a huge state.

Why it is a standard design

It uses the standard cartographic “overview plus detail” layout: one map shows the full extent, inset maps zoom into the dense areas, and a locator box or connecting line shows where each inset sits. Each map keeps its own scale bar and north arrow, and all maps share one legend, so the same classes can be read at both the regional and the local scale.

Reuse for: any variable concentrated in a few small areas, such as cities within a country or coasts within a continent, where a single map either hides the dense area or wastes the sparse one.

Spatio-temporal trends

Spatial and temporal variation shown as stacked monthly map layers

Fig. 2 in Song, Y., Tan, Y., Song, Y., Wu, P., Cheng, J.C., Kim, M.J. and Wang, X., 2018. Spatial and temporal variations of spatial population accessibility to public hospitals: a case study of rural–urban comparison. GIScience & Remote Sensing, 55(5), pp.718-744.pdfsource

Twelve monthly maps of Australia from July 2012 to June 2013, stacked as oblique layers along a time arrow: weekly precipitation on the left and the resulting variation in road travel speed on the right, which together drive the changes in population accessibility to public hospitals.

Why it is a standard design

It is a standard space–time layout for monthly or seasonal data: each time step is the same map drawn as a tilted layer, the layers are stacked in order along one labelled time arrow from the first to the last month, each variable keeps a single colour bar for all layers so any two months can be compared directly, and two variables are placed side by side so their timing can be matched layer by layer. Readers see where and when values change in one figure without a grid of separate panels.

Reuse for: showing how a spatial variable changes over months, seasons or years, such as rainfall, temperature, traffic or accessibility, and comparing it with a related variable over the same periods.

Geographical detector

Geographically optimal zones of soil moisture and the rules that define them

Figs. 3 and 6, combined here, in Luo, P., Song, Y., Huang, X., Ma, H., Liu, J., Yao, Y. and Meng, L., 2022. Identifying determinants of spatio-temporal disparities in soil moisture of the Northern Hemisphere using a geographically optimal zones-based heterogeneity model. ISPRS Journal of Photogrammetry and Remote Sensing, 185, pp.111-128.pdfsource

Top: monthly mean soil moisture at stations across the Northern Hemisphere in the unfrozen season, with a latitude profile fitted by LOESS and the seasonal time series. Bottom: the GOZH result. Nine geographically optimal zones (A–I) are mapped with regional insets; a tree gives the explanatory variables and thresholds that split the data into the zones, with circles sized by each zone’s mean soil moisture; and boxplots compare NDVI, precipitation, temperature and soil pH across the zones.

Why it is a standard design

It is the standard way to report a zone-based heterogeneity model: first the observed pattern of the response (map plus profile), then the zones on a map with one colour and symbol per zone and insets for dense or distant areas, then the splitting rules as a tree with the threshold of each split, and finally boxplots of the explanatory variables in the same zone colours. The shared zone colours link map, tree and boxplots, so readers can see where each zone is, why it was formed and how its conditions differ from the others.

Reuse for: geographical detector, regression-tree and other stratified heterogeneity studies that divide a study area into zones and need to show where the zones are, which variables and thresholds define them, and how the zones differ.

Geographical detector

Key technical settings of the geographical detector and spatial interaction

Fig. 1 in Song, Y. and Wu, P., 2021. An interactive detector for spatial associations. International Journal of Geographical Information Science, 35(8), pp.1676-1701.pdfsource

The four steps of the interactive detector for spatial associations (IDSA), with the recommended settings written into each box. (1) Optimal spatial discretization: either test 3–6 breaks with the quantile, equal, natural, geometric, standard deviation or manual method and keep the combination with the highest Q value, or use 3–22 quantile breaks, fit a LOESS curve to the 75% quantile of Q values and pick the break number where the increase rate falls below 0.05. (2) Power of determinant (Q values) of individual variables. (3) Power of interactive determinant: starting from the strongest variable, add one variable at a time by spatial fuzzy overlay (compared with plain spatial intersection) and keep it only while Q increases. (4) Model evaluation indicators.

Why it is a standard design

It is a method flowchart that doubles as a settings guide: each numbered step is a framed block, inputs, methods, conditions and results have their own shapes explained in a legend, and the empirical parameter values (break numbers, discretization methods, thresholds) are written directly in the boxes, with a small example chart or map beside each decision to show what the rule does. Readers can repeat the analysis step by step without searching the text for the settings, which is what makes a detector study reproducible.

Reuse for: geographical detector and other stratified heterogeneity studies that need to report how variables were discretized and how interactions were tested, and any method paper that should show its recommended parameter values alongside the workflow.

Geographical detector

Optimal parameters-based geographical detector (OPGD): the full set of results

Fig. 10 in Song, Y., Wang, J., Ge, Y. and Xu, C., 2020. An optimal parameters-based geographical detector model enhances geographic characteristics of explanatory variables for spatial heterogeneity analysis: cases with different types of spatial data. GIScience & Remote Sensing, 57(5), pp.593-610.pdfsource

The standard result figure of an OPGD model, in seven panels: (a) the optimal size of spatial unit, chosen from Q values of all variables at several unit sizes; (b) the optimal discretization of a continuous variable, comparing methods and numbers of intervals, with the chosen breaks on its histogram; (c) the factor detector, ranking variables by Q value; (d) the risk detector, showing mean incidence in each interval of the most important variable and the resulting risk map; and (e–g) matrices of the significance of risk differences, the interaction detector and the ecological detector.

Why it is a standard design

It reports a geographical detector analysis in the standard order of parameter optimisation followed by the four detectors: the spatial scale and discretization are justified first with Q-value curves, then the factor detector is a ranked bar chart with the leading variable highlighted, the risk detector pairs interval means with a map of high, medium and low risk, and the significance of risk differences, interactions and ecological differences are given as compact triangular matrices with Y/N cells or sized dots. Each result has a fixed panel type, so readers can compare OPGD studies directly and others can rebuild the same layout from the model output.

Reuse for: any geographical detector or spatial stratified heterogeneity analysis, for example of disease, land use or environmental change, where parameter choices need to be justified and all four detectors reported.

Geographical detector

Interaction detector of the geographical detector: Q values of all variable pairs

Fig. 8 in Song, Y., Wright, G., Wu, P., Thatcher, D., McHugh, T., Li, Q., Li, S.J. and Wang, X., 2018. Segment-based spatial analysis for assessing road infrastructure performance using monitoring observations and remote sensing data. Remote Sensing, 10(11), p.1696.pdfsource

The interaction detector result for 24 explanatory variables of road infrastructure performance. Every pair of variables is a bubble in a triangular matrix: the bubble size is the Q value of their interaction, the colour is the pair of variable categories (C1 to C4), and the numbers mark the ranks of the strongest interactions.

Why it is a standard design

It is the standard way to show all pairwise interactions at once: a lower-triangular matrix holds each pair exactly once, variables are ordered by category so the colour blocks show which groups of factors interact, bubble size encodes the Q value on a shared legend, and rank numbers written inside the largest bubbles point straight to the most important interactions. Readers can scan hundreds of pairs in one view, and the layout fits any number of variables.

Reuse for: reporting interaction detector results or any pairwise measure between many variables, such as interaction strengths, correlations or joint effects, especially when the variables fall into several groups.

Local explanatory power

Local indicator of stratified power: local power of variables and their interactions

Figs. 5 and 6, combined here, in Hu, J., Song, Y. and Zhang, T., 2025. A local indicator of stratified power. International Journal of Geographical Information Science, 39(4), pp.925-943.pdfsource

The standard result of a local indicator of stratified power (LISP) model. For each of eight explanatory variables (top) and four interactions (bottom), a pair of maps shows the local power (ρ, or ω for interactions) at every location and its significance. Violin plots summarise the distribution of the local power of each variable and interaction, with the global value marked as a red dot.

Why it is a standard design

It follows the standard layout for local explanatory results: for each variable, a small-multiple map of the local value in one shared classed colour scale sits directly above a map of its significance (p < 0.01, p < 0.05, not significant), so the strength and reliability of each effect are read together at the same locations; a violin plot with jittered points then compares the distribution of local values across all variables against their global values, and the same design is repeated for interactions so single factors and their combinations can be compared directly.

Reuse for: reporting any local measure of association or explanatory power, such as local geographical detector results, geographically weighted regression coefficients or local spatial autocorrelation, whenever values and their significance need to be mapped for many variables and compared across them.

Local explanatory power

Local pathways of association (LPA): path coefficients that vary from place to place

Fig. 5 in Hu, J., Qu, R., Song, Y. and Wu, P., 2025. Local pathways of association. International Journal of Applied Earth Observation and Geoinformation, 139, p.104531.pdfsource

The standard result of the LPA model. A path diagram links the observed variables (precipitation, land surface temperature, water content and NDVI) to the latent climate, soil and plant factors. Beside each path, a pair of maps shows its local coefficient (λ1 to λ7) and the significance at every location, and boxplots compare the distribution of the seven coefficients, showing that the same pathway is strong, weak, positive or negative in different places.

Why it is a standard design

It turns a path diagram into a map of local effects: the usual circles and arrows of a path model keep the causal structure readable, and each arrow is labelled with its own pair of small maps, the local coefficient in one shared diverging classed scale above its significance (p < 0.01, p < 0.05, not significant), while a boxplot panel with a zero line summarises every coefficient on one axis. Readers can follow the structure, see where each path holds and in which direction, and compare paths, and the layout works for any local path or structural equation model.

Reuse for: reporting local or geographically weighted path analysis and structural equation models, or any model with direct and indirect effects whose strength varies across space.

Nonlinear relationships

Nonlinear relationships and thresholds between two variables

Fig. 6 in Song, Y., Wu, P., Hampson, K., and Anumba, C., 2021. Assessing block-level sustainable transport infrastructure development using a spatial trade-off relation model. International Journal of Applied Earth Observation and Geoinformation, 105, 102585.pdfsource

How the quality of transport infrastructure responds to its quantity, including where extra investment gives increasing, marginal or negative returns.

Why it is a standard design

It follows the standard sequence for a nonlinear relationship: show the raw joint distribution first (density plot with marginal distributions), then the fitted curve with its confidence band, then its first derivative (marginal effect) on a second axis, and finally mark thresholds where the derivative peaks or reaches zero. Labelling the regimes between thresholds (increasing, marginal and negative returns) turns the curve into a result that can be stated in words.

Reuse for: saturation, threshold or diminishing-return effects, such as greenspace and health, density and accessibility, or investment and performance.

Nonlinear relationships

Joinpoint regression of road roughness against pavement age

Fig. 6 in Song, Y., Wu, P., Li, Q., Liu, Y. and Karunaratne, L., 2022. Hybrid nonlinear and machine learning methods for analyzing factors influencing the performance of large-scale transport infrastructure. IEEE Transactions on Intelligent Transportation Systems, 23(8), pp.12287-12300.pdfsource

How road roughness changes with pavement age, fitted by joinpoint regression: the mean and standard error of roughness at each age are split by two joinpoints (33 and 63 years) into three groups, each with its own fitted segment, equation and age range, and a histogram above shows how many road segments there are at each age.

Why it is a standard design

It is the standard result figure of a joinpoint (segmented) regression on a time-related variable: the data are shown as means with error bars, the estimated joinpoints are marked, points in each segment share one colour with the fitted line of that segment, the legend gives every segment's equation, range and significance, and a histogram on the same x-axis shows where the data are dense or sparse. Readers can see when the trend changes, how steep each phase is and how reliable each part of the curve is.

Reuse for: any trend over time or age that changes slope, such as infrastructure deterioration, disease rates over years or land-cover change, where the breakpoints and the rate in each phase need to be reported.

Nonlinear relationships

Standard results of a generalized additive model: smooth effects

Figs. 9 and 10, combined here, in Song, Y.Z., Yang, H.L., Peng, J.H., Song, Y.R., Sun, Q. and Li, Y., 2015. Estimating PM2.5 concentrations in Xi’an City using a generalized additive model with multi-source monitoring data. PLoS ONE, 10(11), p.e0142149.pdfsource

The fitted smooth functions of a GAM for PM2.5 in Xi’an. Top: the one-dimensional smooths s1–s4 of NO2, log(O3), log(AOT) and temperature, each with its confidence band and a rug of the observed values. Bottom: two two-dimensional smooths, the joint effect of the wind components t1(Wc, Ws) as a 3D surface and the spatial trend t2(Lo, La) as a contour map with the monitoring stations.

Why it is a standard design

It is the standard way to report a GAM: one panel per smooth term, the y axis labelled with the term itself (s1(NO2), …) on a shared scale so effect sizes can be compared, a shaded confidence band around the fitted curve, and a rug on the x axis showing where the data are, so readers know which parts of the curve to trust. Two-variable smooths are shown as a perspective surface or a contour map, and a spatial smooth is mapped over longitude and latitude with the observation sites, so non-linear, interaction and spatial effects are all read in the same way.

Reuse for: any GAM or other additive model with smooth terms, such as environmental exposure, species distribution or price models, where you need to show the shape and uncertainty of each non-linear effect and a smooth spatial trend.

Nonlinear relationships

The nonlinear relationship between spatial data and its geocomplexity

Fig. 5 in Zhang, Z., Song, Y., Luo, P. and Wu, P., 2023. Geocomplexity explains spatial errors. International Journal of Geographical Information Science.pdfsource

How the geocomplexity of income changes with income level across spatial units in Australia: it rises to a peak at middle incomes and falls at high incomes, with each unit coloured by its Gini coefficient.

Why it is a standard design

It is the standard way to report a geocomplexity result: one point per spatial unit, the variable’s value on the x-axis and its geocomplexity (bounded from 0 to 1) on the y-axis, and a smooth fitted curve with a confidence band that makes the nonlinear shape, here an inverted U, readable at a glance. A third attribute is mapped to a Spectral colour scale on the points, so readers can also see how a related property (inequality) moves along the curve without adding another panel, and a plain background without gridlines keeps the eye on the points and the curve.

Reuse for: any spatial variable whose geocomplexity, spatial heterogeneity or local variability you want to relate to its own level or to a driver, such as land surface temperature, PM2.5, house prices or vegetation indices. Geocomplexity is available in Heterra and the R package geocomplexity.

Relationship change paths

How the relationship between spatial indicators changes along a third variable

Fig. 5 in Lin, G., Song, Y., Xu, D., Swapan, M.S.H., Wu, P., Hou, W. and Xiao, Z., 2024. Interpreting differences in access and accessibility to urban greenspace through geospatial analysis. International Journal of Applied Earth Observation and Geoinformation, 129, p.103823.pdfsource

How the relationship between two urban greenspace indicators, access (α) and accessibility (β), shifts for each region of Perth as the travel-time threshold grows from a 3-minute walk to 30 minutes.

Why it is a standard design

It is the standard trajectory (connected scatter) plot: the two indicators define the axes, each region becomes a path of points ordered by the changing variable (here travel time, labelled 3 min to 30 min along one path), and arrows show the direction of change. Further indicators are encoded in the marks themselves (size for |α|, shape for the sign of α), dashed zero lines split the plane into interpretable quadrants, and each region keeps one colour that matches an inset map, so readers see where each unit is, which way it moves and how far.

Reuse for: any relationship between two or more spatial variables that changes with time, distance, scale or a threshold, such as land-use change over years, accessibility across travel modes, or model results across spatial scales.

Model evaluation

Spatial validation of prediction models with the degree of spatial interpretability (DSI)

Fig. 6 in Liu, H., Song, Y. and Yi, W., 2026. Degree of spatial interpretability. International Journal of Geographical Information Science, 40(9), pp.3068–3088.pdfsource

How much of the spatial structure in the data each of ten prediction models explains, shown as a DSI range from θ1 to θ3 with θ2 marked inside it.

Why it is a standard design

It is the standard way to report a metric that comes as a range rather than a single value, much like a forest plot: one row per model, models ranked from top to bottom, all on one shared x-axis. Each triangle’s base spans θ1 to θ3 and its dashed apex marks θ2, every value is printed next to its point, and a light gridline runs under each row, so readers can compare both the position and the width of each model’s range without looking anything up.

Reuse for: any comparison of models, sites or scenarios where each result is an interval (lower, central and upper values), such as uncertainty ranges, bootstrap intervals or scenario bounds. DSI is available in Heterra.

Model evaluation

Locally optimal models identified using the degree of geocomplexity (DG)

Fig. 8 in Liu, H., Song, Y., Yi, W. and Zhang, P., 2026. Degree of geocomplexity for diagnosing spatial pattern loss beyond prediction accuracy. GIScience & Remote Sensing, 63(1), p.2657087.pdfsource

Which of ten prediction models best preserves the local spatial pattern at each sample location, judged by the degree of geocomplexity (DG): (a) a map of the model with the highest DG at every sample, (b, c) enlarged views of two dense regions, and (d) the number of samples where each model is locally optimal.

Why it is a standard design

It is the standard design for local model selection: every location gets one categorical symbol for its best model, with a distinct colour and shape per model listed in a single legend, red boxes on the overview map lead to enlarged views of dense areas so overlapping points stay readable, and a bar chart sorted by count in the same model colours gives the overall ranking. Readers see that no single model is best everywhere and where each model wins, and the layout works for any local accuracy or diagnostic metric.

Reuse for: comparing several prediction or interpolation models location by location, or mapping which method, parameter or data source performs best in each place.

Prediction and uncertainty

Comparing data-driven and knowledge-driven risk predictions

Fig. 13 in Wu, P., Wang, P., Chi, H.L., Zhong, Y. and Song, Y., 2022. Exploring factors affecting transport infrastructure performance: data-driven versus knowledge-driven approaches. IEEE Transactions on Intelligent Transportation Systems, 23(12), pp.24714-24726.pdfsource

Road defect risks across the state road network of Western Australia predicted by (a) a data-driven method and (b) a knowledge-driven method, each with an inset of the Perth area, and the association between the two predictions (c) across the whole network and (d) in each of eight regional road networks, with the correlation coefficient in every panel.

Why it is a standard design

It is the standard layout for comparing two predictions of the same quantity: the two maps sit side by side with identical classes, colours, scale bars and insets, so differences are spatial and not cartographic; below them a density scatter of one prediction against the other, with a fitted line and R, shows the overall agreement, and small multiples repeat the scatter for each region on the same axes. Readers can see where the two methods agree, how strongly, and whether the agreement holds in every region.

Reuse for: comparing two models, two data sources or two methods (for example data-driven and expert-based, satellite and ground, past and present) that predict the same variable across space, overall and by region.

Sensitivity analysis

Sensitivity analysis of the key parameters of a spatial model

Figure in Ren, K., Song, Y. and Yu, Q., 2026. Second-dimension outliers for spatial prediction. International Journal of Geographical Information Science, 40(6), pp.1915-1942.pdfsource

How the cross-validation RMSE of the SDO model changes when each key parameter is moved away from its chosen value while the others are fixed: (a) the buffer threshold, (b) the buffer interval and (c) the z-score threshold for outliers, with (d) the percentage change in RMSE for each comparison (Δ1 to Δ6).

Why it is a standard design

It is the standard one-at-a-time sensitivity layout that any spatial analysis can report: each important parameter gets its own panel, the chosen value is highlighted in a contrasting colour between neighbouring values, the fixed settings are written in the axis title, every comparison is labelled (Δ1, Δ2, …), and a final panel puts all comparisons on one percentage-change axis with symmetric limits. Readers can see which parameters matter, in which direction, and that the chosen setting is robust.

Reuse for: justifying parameter choices in any spatial model, such as bandwidths, buffer sizes, numbers of strata, thresholds or neighbourhood definitions, by showing how the accuracy changes around the chosen values.

Sensitivity analysis

Sensitivity of a spatial decision to the variables and the methods used

Fig. s 9 and 10, combined here in Song, Y., Thatcher, D., Li, Q., McHugh, T. and Wu, P., 2021. Developing sustainable road infrastructure performance indicators using a model-driven fuzzy spatial multi-criteria decision making method. Renewable and Sustainable Energy Reviews, 138, p.110538.pdfsource

How the weights of the road, vehicle and climate criteria change when the decision model is rerun without one input at a time. Top: each sub-criterion (variable) is left out in turn. Bottom: each model used to build the weights (statistical, machine learning and spatial models) is left out in turn. Left panels show the spread of the weights, with the full-model weight as a red diamond and outlier runs labelled; right panels show the percentage change in each criterion weight for every run.

Why it is a standard design

It is a standard leave-one-out sensitivity layout for decision and weighting models: one row per source of uncertainty (variables, then methods), the same two-panel design in each row, a boxplot of the resulting weights with the full-model value marked and the influential runs named, and a diverging bar chart centred on zero with one colour per criterion and dashed lines separating the groups. Readers can see whether the decision is stable, and exactly which variable or method moves it and in which direction.

Reuse for: testing the robustness of multi-criteria decisions, composite indices, ensemble models or variable weights to the choice of input variables and to the choice of methods.

Result interpretation

Road distress and performance with and without maintenance: scenarios to 2030

Figs. 15 and 16, combined here, in Song, Y., Wu, P., Li, Q., Liu, Y. and Karunaratne, L., 2022. Hybrid nonlinear and machine learning methods for analyzing factors influencing the performance of large-scale transport infrastructure. IEEE Transactions on Intelligent Transportation Systems, 23(8), pp.12287-12300.pdfsource

How road distress (top) and road performance (bottom) would change with and without maintenance. Each chart shows observations, a scenario that follows the deterioration model with maintenance from 2007 to 2030, and the worst scenario with no maintenance after 2012. The shaded band marks the maintenance period; small arrows label the change from each maintenance round and from each period of deterioration, and the long vertical arrows give the total difference between scenarios.

Why it is a standard design

It is the standard design for comparing scenarios over time: observations and each scenario are separate coloured lines on one time axis, the intervention period is shaded, every change is labelled with an arrow and its value (gains in one colour, losses in another), and dashed reference lines with vertical arrows give the gap between scenarios at key years. Two related indicators are stacked on the same time axis, so the same interventions can be read on both. Readers can see what each intervention achieved and what would happen without it, which turns a model into evidence for planning.

Reuse for: showing the effect of interventions or policies over time, such as maintenance, restoration, emission controls or land-use plans, against observations and a no-action scenario.

Result interpretation

From predicted traffic to road maintenance burden of local governments

Fig. 11 in Song, Y., Wang, X., Wright, G., Thatcher, D., Wu, P. and Felix, P., 2019. Traffic volume prediction with segment-based regression kriging and its implementation in assessing the impact of heavy vehicles. IEEE Transactions on Intelligent Transportation Systems, 20(1), pp.232-243.pdfsource

A further statistical analysis of predicted traffic volumes to answer a practical question: which local government areas (LGAs) carry the heaviest road maintenance burden from heavy vehicles. (A) LGAs are grouped by the main highway they lie along; (B) each LGA is a bubble placed by its percentage of heavy vehicle volume and mass, coloured by highway group and sized by road length.

Why it is a standard design

It is a standard way to carry spatial results to the units that make decisions: the predicted values are summarised for each administrative unit, the units are grouped by a meaningful category and shown on a map in group colours, and the same colours and labels are reused in a bubble chart that places each unit by two summary indicators with a third as bubble size. Readers can match every point to a place, see which groups and units stand out, and use the ranking directly for planning.

Reuse for: turning a predicted surface or model output into indicators for councils, districts or catchments, and comparing them on two or three measures to find the units that need attention.

Result interpretation

Determinants in four groups of counties and their statistical summaries

Fig. 9 in Ge, Y., Song, Y., Wang, J., Liu, W., Ren, Z., Peng, J. and Lu, B., 2017. Geographically weighted regression-based determinants of malaria incidences in northern China. Transactions in GIS, 21(5), pp.934-953.pdfsource

A further statistical analysis of local regression results to explain malaria incidence: (a) counties are grouped into four regions by their pattern of local coefficients, and (b–e) violin plots summarise the pseudo-t values of six determinants in each group, with red lines marking the significance thresholds, so the dominant determinants of each region can be read off directly.

Why it is a standard design

It is a standard way to summarise thousands of local results into a few interpretable regions: local model outputs are grouped into spatially coherent clusters shown on one map with group colours and counts, and each group then gets its own panel of distributions in the same colour, with the same variables in the same order and reference lines for significance. Readers can compare groups side by side and see which factors matter where, which turns a local model into region-specific conclusions.

Reuse for: interpreting the results of any local model, such as geographically weighted regression, local geographical detectors or clustering, by grouping locations and comparing the distributions of their coefficients or effects.

Method evolution

How a new spatial method evolves from earlier methods

Fig. 1 in Luo, P., Song, Y., Zhu, D., Cheng, J. and Meng, L., 2023. A generalized heterogeneity model for spatial interpolation. International Journal of Geographical Information Science, 37(3), pp.634–659.pdfsource

How the generalized heterogeneity model (GHM) for spatial interpolation builds step by step on ordinary kriging and kriging with external drift (OK and KED), stratified kriging (StK), P-MSN and ATAK, and which theory and type of geographical information each step adds.

Why it is a standard design

It is the standard staircase diagram of method evolution: each step up is one method, ordered from the simplest (spatial observations) to the new one (GHM) at the top, and a small schematic on each step shows what that method assumes about the data. Symbols explained in a legend separate the inputs to each step (circles for theoretical bases, triangles for types of geographical information, boxes for parallel models), and arrows show exactly which idea flows into which method, so readers see both the lineage and what is new at every step.

Reuse for: the introduction or review section of any methods paper, to place a new method among the methods it extends, such as regression models, spatial clustering methods, or prediction models from statistics to machine learning.

Figures are reproduced from the cited articles for reference and teaching. Copyright remains with the original publishers; please cite the original article when you adapt a design.