Predicting and Interpolating Spatiotemporal Environmental Data: A Case Study of Groundwater Storage in Bangladesh

TL;DR

The paper tackles joint spatiotemporal estimation of environmental fields from point observations, using groundwater storage changes in Bangladesh as a case study. It systematically compares grid-to-grid, grid-to-point, and spatiotemporal deep learning approaches (including Unet variants, CNN-CNN-LSTM stacks, Kriging, and DeepKriging), across 2D and 3D configurations with temporal cross-validation. A key finding is that spatial interpolation is substantially harder than temporal prediction, with interpolation performance highly sensitive to local hydrogeology and time-series dynamics; Kriging-based interpolation and grid-based inpainting can degrade accuracy at observed points. Despite interpolation challenges, the data-driven models outperform the physics-based GLDAS 2.2 CLSM baseline, and the authors recommend exploring clustering-informed interpolation and neural-process frameworks for scalable, accurate spatiotemporal predictions in complex environments; code is available at the provided GitHub repository.

Abstract

Geospatial observational datasets are often limited to point measurements, making temporal prediction and spatial interpolation essential for constructing continuous fields. This study evaluates two deep learning strategies for addressing this challenge: (1) a grid-to-grid approach, where gridded predictors are used to model rasterised targets (aggregation before modelling), and (2) a grid-to-point approach, where gridded predictors model point targets, followed by kriging interpolation to fill the domain (aggregation after modelling). Using groundwater storage data from Bangladesh as a case study, we compare the effcacy of these approaches. Our findings indicate that spatial interpolation is substantially more difficult than temporal prediction. In particular, nearest neighbours are not always the most similar, and uncertainties in geology strongly influence point temporal behaviour. These insights motivate future work on advanced interpolation methods informed by clustering locations based on time series dynamics. Demonstrated on groundwater storage, the conclusions are applicable to other environmental variables governed by indirectly observable factors. Code is available at https://github.com/pazolka/interpolation-prediction-gwsa.

Figures (4)

• Figure 1: Distribution of train, validation & prediction test points, and interpolation test (holdout) points. Groundwater dynamics in Bangladesh are characterised by a high spatial heterogeneity.
• Figure 2: Employed spatial (2D) and spatiotemporal (3D) models, as described in Section \ref{['methods']}. 2D models are fed with gridded predictors ($C=5$) from the same timestep; 3D models additionally use lagged predictors $t...t-4$. The optimal temporal lag was found empirically during preliminary analysis. Results of all approaches are reported in Fig. \ref{['fig:boxplots']}. Spatiotemporal DeepKriging was implemented according to nag_spatio-temporal_2023. Note that Stage 2 of DeepKriging only uses past predictors, as opposed to all other models.
• Figure 3: Performance on $\Delta$GWS dataset from Bangladesh on data-driven models, R$^2$ over 10-fold temporal CV. Refer to Tab. \ref{['tab:performance']} for mean performance scores.
• Figure 4: Performance of the employed models at three example locations: (a) RJ092 (b) RA-22 (c) DI030 from the interpolation holdout set; last CV fold (95%-5%-5%).