Modeling Multivariate Spatio-Temporal Remote Sensing Data with Large Gaps

Qiang Lou, Zoran ObradovicIJCAI 2011 · OpenReview

Abstract

Prediction models for multivariate spatio-temporal functions in geosciences are typically developed using supervised learning from attributes collected by remote sensing instruments collocated with the outcome variable provided at sparsely located sites. In such collocated data there are often large temporal gaps due to missing attribute values at sites where outcome labels are available. Our objective is to develop more accurate spatio-temporal predictors by using enlarged collocated data obtained by imputing missing attributes at time and locations where outcome labels are available. The proposed method for large gaps estimation in space and time (called LarGEST) exploits temporal correlation of attributes, correlations among multiple attributes collected at the same time and space, and spatial correlations among attributes from multiple sites. LarGEST outperformed alternative methods in imputing up to 80% of randomly missing observations at a synthetic spatio-temporal signal and at a model of fluoride content in a water distribution system. LarGEST was also applied for imputing 80% of nonrandom missing values in data from one of the most challenging Earth science problems related to aerosol properties. Using such enlarged data a predictor of aerosol optical depth is developed that was much more accurate than predictors based on alternative imputation methods when tested rigorously over entire continental US in year 2005.

Introduction

In geosciences and remote sensing, missing data is a pervasive challenge. Satellite sensors are frequently obstructed by cloud cover or hardware malfunctions, resulting in significant "gaps" in multivariate spatio-temporal datasets. To construct effective predictive models—such as those estimating air quality from satellite imagery—researchers require "collocated" data, where satellite observations are paired with ground-truth measurements. However, when gaps occur at ground station locations during satellite overpasses, valuable training instances are lost. Traditional imputation methods often address only a subset of data dimensions (temporal, spatial, or cross-attribute), leading to performance degradation when missingness is high (e.g., 80%).

Methodology: The LarGEST Framework

The authors introduce LarGEST (Large Gaps Estimation in Space and Time), a framework designed to recover missing values by simultaneously exploiting temporal, spatial, and cross-attribute correlations. LarGEST integrates two probabilistic models via an Extended Expectation Maximization (EM) algorithm:

  1. Temporal & Attribute Model: Utilizes a Kalman Filter to capture the evolution of attributes at a single site over time and the correlations between different attributes (e.g., spectral bands).
  2. Spatial Model: Employs a linear-Gaussian latent variable model to capture correlations between attributes at neighboring sites.

The core innovation lies in the iterative "handshake" between these models. During the EM process, estimates from the temporal/attribute model refine the spatial model, and vice versa. This collaborative refinement allows LarGEST to maintain accuracy even under extreme data sparsity.

Experimental Evaluation

The framework was evaluated across three distinct scenarios:

  • Synthetic Data: A 2D temporal signal simulating noisy, shifted sensor data.
  • Fluoride Content: Simulated water quality data across a 36-node distribution network.
  • Aerosol Optical Depth (AOD): Real-world remote sensing data from the MODIS satellite (19 attributes) paired with ground-truth from 33 AERONET sites across the continental US.

Results and Discussion

LarGEST demonstrated remarkable robustness to data gaps. In synthetic tests, it maintained a low Mean Square Error (MSE) even at 80% missingness, whereas traditional methods like linear interpolation and Empirical Orthogonal Functions (EOF) exhibited sharp error spikes.

LarGEST maintains low MSE compared to alternative methods as the fraction of missing values increases.

In the AOD retrieval task, the primary benefit of LarGEST was the "enlargement" of the training set. By imputing missing satellite attributes at ground-truth sites, the usable training data increased from 805 to 4,112 instances. A neural network trained on this enlarged dataset achieved an R-square of approximately 0.74, a substantial improvement over the 0.58 achieved by the original (non-imputed) dataset.

R-square performance of AOD retrieval models, showing LarGEST variants significantly outperforming original and alternative imputation methods.

Furthermore, the model's optimal performance was reached when considering neighbors within a 36km radius, aligning with the physical spatial correlation range of aerosols. This suggests that the statistical framework effectively captures underlying physical phenomena.

Conclusion

LarGEST provides a robust solution for imputing massive gaps in multivariate spatio-temporal data. By leveraging three-way correlations, it enables the creation of significantly larger and more effective training sets for geoscience applications. Future research may explore the integration of deep generative models to handle highly non-linear dependencies and optimize computational scalability for global-scale datasets.

Summary updated

Abstract

Prediction models for multivariate spatio-temporal functions in geosciences are typically developed using supervised learning from attributes collected by remote sensing instruments collocated with the outcome variable provided at sparsely located sites. In such collocated data there are often large temporal gaps due to missing attribute values at sites where outcome labels are available. Our objective is to develop more accurate spatio-temporal predictors by using enlarged collocated data obtained by imputing missing attributes at time and locations where outcome labels are available. The proposed method for large gaps estimation in space and time (called LarGEST) exploits temporal correlation of attributes, correlations among multiple attributes collected at the same time and space, and spatial correlations among attributes from multiple sites. LarGEST outperformed alternative methods in imputing up to 80% of randomly missing observations at a synthetic spatio-temporal signal and at a model of fluoride content in a water distribution system. LarGEST was also applied for imputing 80% of nonrandom missing values in data from one of the most challenging Earth science problems related to aerosol properties. Using such enlarged data a predictor of aerosol optical depth is developed that was much more accurate than predictors based on alternative imputation methods when tested rigorously over entire continental US in year 2005.
BibTeX
@inproceedings{lou2011modeling,
  author = {Qiang Lou and Zoran Obradovic},
  title = {{Modeling Multivariate Spatio-Temporal Remote Sensing Data with Large Gaps}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2011},
  url = {https://openreview.net/forum?id=HJ4AIrzu-B},
}
Figure