Spatial Classification With Limited Observations Based On Physics-Aware Structural Constraint

Arpan Man Sainju, Wenchong He, Zhe Jiang, Da Yan, Haiquan Chen

Abstract

Spatial classification with limited feature observations has been a challenging problem in machine learning. The problem exists in applications where only a subset of sensors are deployed at certain spots or partial responses are collected in field surveys. Existing research mostly focuses on addressing incomplete or missing data, e.g., data cleaning and imputation, classification models that allow for missing feature values or model missing features as hidden variables in the EM algorithm. These methods, however, assume that incomplete feature observations only happen on a small subset of samples, and thus cannot solve problems where the vast majority of samples have missing feature observations. To address this issue, we recently proposed a new approach that incorporates physics-aware structural constraint into the model representation. Our approach assumes that a spatial contextual feature is observed for all sample locations and establishes spatial structural constraint from the underlying spatial contextual feature map. We design efficient algorithms for model parameter learning and class inference. This paper extends our recent approach by allowing feature values of samples in each class to follow a multi-modal distribution. We propose learning algorithms for the extended model with multi-modal distribution. Evaluations on real-world hydrological applications show that our approach significantly outperforms baseline methods in classification accuracy, and the multi-modal extension is more robust than our early single-modal version especially when feature distribution in training samples is multi-modal. Computational experiments show that the proposed solution is computationally efficient on large datasets.

Introduction

In many real-world spatial applications, researchers face the challenge of "limited observations." For example, in flood mapping using drone imagery, high-quality spectral data may only be available for a small fraction of the region due to operational constraints. Standard machine learning models typically assume missing data is the exception; when missingness is the majority, traditional imputation or simple classification leads to significant errors. Furthermore, spectral data alone can be misleading—a phenomenon known as "class confusion"—where tree canopies look identical regardless of whether the ground beneath is dry or flooded.

Illustration of the problem: (a) elevation field, (b) sparse feature observations, and (c) ground truth classes.

Instead of treating physical features like elevation as simple inputs, this research uses them to define the model's structure. This physics-aware approach acts as a powerful regularizer, guiding the model when spectral data is missing and helping overcome class confusion by considering the physical context, such as downstream relationships.

Methods

The proposed framework integrates physics-aware structural constraints into a probabilistic graphical model to infer labels across the entire spatial field.

  1. The Split Tree: Using a digital elevation model (DEM) available for the entire region, the authors construct a "split tree." This tree encodes the physical law of gravity: water flows from high to low elevations. If a location is flooded, its downstream neighbors are likely flooded as well.
  2. Expectation-Maximization (EM): Because class labels for most pixels are unknown, they are treated as hidden variables. The EM algorithm iteratively estimates these labels (E-step) and updates model parameters (M-step).
  3. Message Propagation: To ensure scalability to millions of pixels, the authors employ belief propagation (sum-product and max-sum algorithms) on the split tree, allowing for efficient inference of hidden classes.
  4. Multi-modal Extension: To handle complex real-world environments where feature distributions are not simple bell curves, the model supports Gaussian Mixture Models (GMMs). This allows the feature distribution of a class to have multiple "modes," increasing robustness.
Probabilistic graphical model structure showing hidden class nodes and partially observed feature nodes.

Datasets and Results

The model was evaluated using real-world hydrological data from Hurricane Matthew (2016) in North Carolina, specifically the Kinston ($1743 \times 1349$ pixels) and Grimesland ($2757 \times 3853$ pixels) regions. Data included 2-meter resolution RGB aerial imagery and LiDAR-based DEMs.

  • Accuracy: On the Kinston dataset, the proposed model achieved an average F-score of 0.96-0.97, vastly outperforming Gradient Boosted Models (0.74) and standard EM without structural constraints (0.66).
  • Robustness: In multi-modal training scenarios, the "EM-Structure-Multi" version maintained high performance, while the single-modal version's F-score dropped to 0.34.
  • Efficiency: The algorithm classified approximately 10 million pixels in about 2 minutes on a standard workstation, demonstrating near-linear scalability with the size of the test region.

Conclusion

This physics-aware structural constraint framework effectively addresses spatial classification with limited observations. While the model's success depends on the accuracy of the elevation-based potential field, it provides a robust scaffold for mapping in data-poor environments. Future work includes integrating these constraints into deep learning frameworks like Graph Convolutional Networks and extending the model to handle noisy volunteered geographic information.

Summary updated

Abstract

Spatial classification with limited feature observations has been a challenging problem in machine learning. The problem exists in applications where only a subset of sensors are deployed at certain spots or partial responses are collected in field surveys. Existing research mostly focuses on addressing incomplete or missing data, e.g., data cleaning and imputation, classification models that allow for missing feature values or model missing features as hidden variables in the EM algorithm. These methods, however, assume that incomplete feature observations only happen on a small subset of samples, and thus cannot solve problems where the vast majority of samples have missing feature observations. To address this issue, we recently proposed a new approach that incorporates physics-aware structural constraint into the model representation. Our approach assumes that a spatial contextual feature is observed for all sample locations and establishes spatial structural constraint from the underlying spatial contextual feature map. We design efficient algorithms for model parameter learning and class inference. This paper extends our recent approach by allowing feature values of samples in each class to follow a multi-modal distribution. We propose learning algorithms for the extended model with multi-modal distribution. Evaluations on real-world hydrological applications show that our approach significantly outperforms baseline methods in classification accuracy, and the multi-modal extension is more robust than our early single-modal version especially when feature distribution in training samples is multi-modal. Computational experiments show that the proposed solution is computationally efficient on large datasets.
BibTeX
@inproceedings{sainju2020spatial,
  author = {Arpan Man Sainju and Wenchong He and Zhe Jiang and Da Yan and Haiquan Chen},
  title = {{Spatial Classification With Limited Observations Based On Physics-Aware Structural Constraint}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i01.5436},
}
Figure