SpaceGAN: Geospatial Data Generation Model

• SpaceGAN is a generative model for spatial point-pattern data that extends conditional GANs by conditioning on neighboring feature vectors.
• It preserves local spatial autocorrelation using the Mean Moran’s I Error (MIE) to ensure synthetic samples accurately mimic true spatial dependencies.
• It supports ensemble learning for spatial regression, achieving up to 20% improvement in RMSE over conventional augmentation methods.

SpaceGAN is a generative model for geospatial domains introduced in “Augmenting correlation structures in spatial data using deep generative models” (Klemmer et al., 2019). It extends the conditional Generative Adversarial Network paradigm to spatial point-pattern data by conditioning each observation not merely on spatial coordinates but on the feature-neighborhood of its spatial neighbors. The model is designed to learn neighborhood structures through spatial conditioning, to preserve local spatial autocorrelation via a model-selection criterion based on local Moran’s $I$, and to support both synthetic spatial data generation and downstream ensemble learning for prediction. The reported use cases are artificially inflating sparse geospatial data and improving generalization of geospatial models.

1. Conceptual basis and data representation

SpaceGAN builds on the conditional GAN framework. A vanilla GAN consists of a generator $G(\cdot; \Theta_G)$, which maps noise $z \sim p_z(z)$ to a synthetic sample $\hat x$, and a discriminator $D(\cdot; \Theta_D)$, which maps a sample $x$ to a scalar $D(x) \in [0,1]$ estimating the probability that $x$ is real. In SpaceGAN, the relevant data domain is geospatial point-pattern data, where each datum $i$ is represented as

$d_i = (x_i, y_i, c_i),$

with $G(\cdot; \Theta_G)$0 an $G(\cdot; \Theta_G)$1-dimensional attribute vector, $G(\cdot; \Theta_G)$2 the target or response variable, and $G(\cdot; \Theta_G)$3 the two-dimensional spatial coordinates (Klemmer et al., 2019).

The motivation is explicitly spatial. An unconditional GAN would ignore the fact that nearby points are more highly correlated, invoking Tobler’s First Law. SpaceGAN therefore uses a conditional objective in which a context vector $G(\cdot; \Theta_G)$4 is injected into both $G(\cdot; \Theta_G)$5 and $G(\cdot; \Theta_G)$6. This design targets complex geospatial dependence structures that standard deep generative setups struggle to represent.

A common simplification in spatial machine learning is to treat coordinates as the sole conditioning signal. SpaceGAN rejects that simplification. Its premise is that spatial structure is more flexibly represented when each point is conditioned on neighboring observations rather than on raw coordinates alone. This suggests a shift from coordinate-based conditioning to neighborhood-structured conditioning as the central representational move.

2. Spatial conditioning through neighborhood features

The core mechanism of SpaceGAN is spatial conditioning via feature-neighborhoods. Let

$G(\cdot; \Theta_G)$7

denote the set of neighbors of observation $G(\cdot; \Theta_G)$8 under a spatial weight matrix $G(\cdot; \Theta_G)$9. The context vector for sample $z \sim p_z(z)$0 is then defined as

$z \sim p_z(z)$1

By feeding $z \sim p_z(z)$2 into both the generator and discriminator, the model can learn how attributes at location $z \sim p_z(z)$3 co-vary with attributes at neighboring locations (Klemmer et al., 2019). This is the defining feature of SpaceGAN: spatial dependence is represented through neighboring feature vectors rather than through coordinates used in isolation.

In practice, $z \sim p_z(z)$4 is arranged as a fixed-size tensor so that it can be processed by 1D-CNN layers in both $z \sim p_z(z)$5 and $z \sim p_z(z)$6. The summary specifies fixed-size realization through mechanisms such as zero-padding or fixed $z \sim p_z(z)$7-nearest-neighbor ordering. The architecture therefore presumes a neighborhood operator that can be rendered into a regular tensorized input.

This formulation is also closely tied to the choice of spatial weights. The summary names several possibilities for $z \sim p_z(z)$8, including $z \sim p_z(z)$9-nearest-neighbor structure and queen-adjacency matrices in the areal case. A plausible implication is that the model’s effective notion of locality is inseparable from the analyst’s choice of neighborhood graph, because the conditioning tensor is built directly from that graph.

3. Local spatial autocorrelation and the MIE criterion

SpaceGAN evaluates spatial fidelity through local Moran’s $\hat x$0. For observation $\hat x$1,

$\hat x$2

where $\hat x$3 is the number of observations, $\hat x$4, and $\hat x$5 is the spatial weight between $\hat x$6 and $\hat x$7.

To ensure that the generator faithfully reproduces the local spatial autocorrelation of $\hat x$8, the model introduces the Mean Moran’s $\hat x$9 Error (MIE) as a model-selection criterion: $D(\cdot; \Theta_D)$0 At training time, after each snapshot of the generator, the procedure draws $D(\cdot; \Theta_D)$1 synthetic samples, computes local $D(\cdot; \Theta_D)$2, averages over snapshots, and selects the generator instance $D(\cdot; \Theta_D)$3 with minimal MIE (Klemmer et al., 2019).

The role of MIE is specifically tied to convergence and stopping. Rather than relying only on adversarial loss trajectories, SpaceGAN uses loss in local spatial autocorrelation between real and generated data as a stopping criterion for parameterization. In the reported formulation, this is what ensures that the generator produces synthetic samples faithful to the spatial patterns observed in the input.

This criterion also clarifies a potential misconception. SpaceGAN is not described as optimizing Moran’s $D(\cdot; \Theta_D)$4 directly inside the adversarial objective. Instead, MIE is used for model selection and stopping among generator snapshots. The distinction matters: the adversarial game provides the basic training dynamics, while MIE governs which trained instance is retained.

4. Objective function, optimization, and implementation

SpaceGAN’s basic minimax objective is the standard cGAN loss,

$D(\cdot; \Theta_D)$5

Training proceeds through stochastic gradient updates in alternating fashion. For the discriminator, using a batch of $D(\cdot; \Theta_D)$6 real pairs $D(\cdot; \Theta_D)$7 with contexts $D(\cdot; \Theta_D)$8 and $D(\cdot; \Theta_D)$9 generated pairs,

$x$0

For the generator,

$x$1

Beyond these SGD steps, convergence is detected via the MIE stopping criterion. Every $x$2 steps, the current generator is stored as $x$3; then $x$4 full-field synthetic samples are drawn, $x$5 is computed, and the selected model is

$x$6

The implementation details reported for the experiments are uniform enough to define a characteristic configuration. SpaceGAN uses a 1D-CNN architecture for both $x$7 and $x$8, with one hidden convolutional layer each. Hidden nonlinearities are ReLU, the generator output is linear, and the discriminator output is sigmoid. The noise prior is $x$9, with dimension $D(x) \in [0,1]$0 for the toy datasets and $D(x) \in [0,1]$1 for California Housing. Batch size is $D(x) \in [0,1]$2, training lasts $D(x) \in [0,1]$3 steps, snapshots are stored every $D(x) \in [0,1]$4 steps, and the evaluation sample count is $D(x) \in [0,1]$5. Optimization uses SGD with learning rate $D(x) \in [0,1]$6, and both input $D(x) \in [0,1]$7 and target $D(x) \in [0,1]$8 are standardized by $D(x) \in [0,1]$9-score before training (Klemmer et al., 2019).

The summary also specifies dataset-dependent filter counts and kernel sizes. Filters per layer are $x$0 for Toy 1, $x$1 for Toy 2, $x$2 for Housing 15, and $x$3 for Housing 50. Kernel sizes are equal to neighborhood size: queen adjacency $x$4 for the toy settings and $x$5-NN or $x$6-NN for the housing settings.

5. Experimental design and empirical findings

SpaceGAN was evaluated on three datasets under 10-fold spatial cross-validation, with points near a test fold buffered out of the training fold (Klemmer et al., 2019). The datasets are summarized below.

Dataset	Size and geometry	Response construction or target
Toy 1	$x$7 points on a $x$8 regular grid in $x$9	$i$0, $i$1
Toy 2	$i$2 points on a finer grid	$i$3, $i$4, $i$5
California Housing	$i$6 house records with longitude and latitude	Response: median_house_value

For the toy datasets, queen-adjacency was used. For California Housing, two versions of spatial cross-validation were used, with $i$7-NN and $i$8-NN buffers. The housing attributes are listed as median_age, total_rooms, …, median_income.

The first experimental focus was reproducing spatial patterns. The reported result is that SpaceGAN samples accurately match the ground-truth $i$9 fields and their local Moran’s $d_i = (x_i, y_i, c_i),$0 maps. Against a vanilla Gaussian-Process RBF smoother used for augmentation, SpaceGAN attained significantly lower MIE on all four tasks: Toy 1, Toy 2, Housing 15, and Housing 50.

The second focus was augmentation for prediction. Across all datasets, the SpaceGAN-MIE ensembles achieved the lowest out-of-sample RMSE, with up to approximately $d_i = (x_i, y_i, c_i),$1 improvement over GP-Bagging and approximately $d_i = (x_i, y_i, c_i),$2 over spatial bootstrap. The summary also states more generally that SpaceGAN is successfully applied for data augmentation, outperforms compared to other methods of synthetic spatial data generation, and empirically shows superiority over conventional ensemble learning approaches while rivaling spatial data augmentation methods (Klemmer et al., 2019).

6. Ensemble learning via “Ganning”

Once $d_i = (x_i, y_i, c_i),$3 has been selected, SpaceGAN is used to generate $d_i = (x_i, y_i, c_i),$4 full datasets of $d_i = (x_i, y_i, c_i),$5 points each. For each $d_i = (x_i, y_i, c_i),$6, noise $d_i = (x_i, y_i, c_i),$7 is drawn for every spatial index $d_i = (x_i, y_i, c_i),$8, synthetic targets are generated as

$d_i = (x_i, y_i, c_i),$9

the original $G(\cdot; \Theta_G)$00 values are kept fixed, and a base learner $G(\cdot; \Theta_G)$01 is trained on $G(\cdot; \Theta_G)$02. Predictions for a new point are then averaged across the ensemble members $G(\cdot; \Theta_G)$03 (Klemmer et al., 2019).

The paper terms this framework “Ganning.” In the reported experiments, the ensembles are formed by drawing $G(\cdot; \Theta_G)$04 SpaceGAN synthetic realizations and training $G(\cdot; \Theta_G)$05 regression-tree base learners on them. The comparisons include SpaceGAN ensembles with MIE-based convergence, SpaceGAN ensembles with RMSE-based convergence, GP-Bagging based on sampling from the GP posterior, and spatial bootstrap ensembles.

The rationale given is the variance-reduction principle: as long as base learners are lowly correlated and not excessively biased, ensemble MSE drops. Empirically, SpaceGAN ensembles outperform both GP-bagging and spatial bootstrap ensembles by margins of $G(\cdot; \Theta_G)$06–$G(\cdot; \Theta_G)$07 in RMSE, while the broader summary reports the strongest performance for SpaceGAN-MIE in out-of-sample RMSE across all datasets. Taken together, these results position the generative model not only as a simulator of spatial fields but also as an augmentation mechanism for supervised geospatial learning.

7. Interpretation, scope, and reported extensions

The reported contribution of SpaceGAN has three linked components: conditioning a cGAN on neighbor feature vectors, enforcing local-autocorrelation fidelity through the MIE stopping rule, and leveraging generated samples to build ensemble learners for spatial regression tasks (Klemmer et al., 2019). In this formulation, the model addresses two problems at once: synthetic generation that respects spatial correlation structures and downstream prediction under sparse or spatially structured data regimes.

The summary emphasizes that the method augments spatial representation beyond mere spatial coordinates. This directly counters the assumption that coordinate-aware models are necessarily sufficient for spatial dependence. SpaceGAN instead treats neighborhoods as the operative conditioning object. A plausible implication is that its representational advantage depends on whether relevant spatial signal is captured by local feature neighborhoods rather than by smoother global trends alone.

The paper’s reported findings suggest that SpaceGAN can be used as a tool for artificially inflating sparse geospatial data and improving generalization of geospatial models. At the same time, the implementation summary explicitly notes possible experimental variations: one may also experiment with alternative GAN variants such as Wasserstein loss, different neighborhood definitions, or deeper $G(\cdot; \Theta_G)$08 architectures. These are presented as implementation directions rather than evaluated claims.

Within the boundaries of the reported evidence, SpaceGAN is therefore best understood as a conditional adversarial framework for geospatial point data in which spatial dependence is encoded through neighborhood-conditioned generation and validated through local Moran’s $G(\cdot; \Theta_G)$09-based model selection. Its empirical role is dual: reproducing spatial correlation structures and furnishing synthetic training distributions for ensemble-based spatial prediction.