Spatial Reconstructability Analysis: Information-Theoretic Modeling of Categorical Raster Data

Author

• David Percy, Portland State UniversityFollow

Sponsor

Portland State University. Systems Science Ph. D. Program

First Advisor

Martin Zwick

Term of Graduation

Summer 2026

Date of Publication

8-24-2026

Document Type

Dissertation

Degree Name

Doctor of Philosophy (Ph.D.) in Systems Science

Department

Systems Science

Language

English

Subjects

Categorical data, Geographic information systems, Information theory, Interpretable machine learning, Land cover change, Reconstructability analysis

Physical Description

1 online resource (xviii, 139 pages)

Abstract

This research focuses on the application of Reconstructability Analysis (RA), an information-theoretic machine learning (ML) methodology, to categorical spatial data in Geographic Information Systems (GIS). RA was developed in the systems community with applications including classification, prediction, pattern recognition, and decision analysis. RA is implemented in OCCAM (Organizational Complexity Computation and Modeling), a software suite developed at Portland State University that analyzes discrete data in tabular form, with observations in rows and variables in columns. While RA has been successfully applied to a variety of domains, no systematic application to georeferenced raster data arranged in regular grids has been previously accomplished.

This dissertation has two aims. The primary aim is to demonstrate RA’s applicability to spatial problems in three environmental science domains. Three distinct spatial data extraction techniques were developed and applied: (1) the cardinal stack for predicting forest management cycles using temporal land cover data in the Pacific Northwest, (2) an isotropic stack using nested rings for modeling wildfire size prediction, tested across 13 pyromes spanning the Pacific Northwest and Northern California, and (3) a vertical stack for landslide susceptibility mapping using categorical soil taxonomy, lithologic classifications, and geomorphic descriptions across 3,500 km² of the Oregon Coast Range, USA. Each technique represents a different strategy for transforming georeferenced raster data into the tabular format that OCCAM expects, potentially addressing different types of spatial relationships.

Across the three studies, RA achieved balanced accuracies from 66% to 83%. In Papers 2 and 3, its accuracy was close to Random Forest and the other ML comparison methods. RA models also produce conditional probability tables for every unique model state that can be read directly.

RA Search identifies the variables and interactions in a selected model, RA Fit then reports the probability of each outcome for the observed combinations of variable states. In the clearcuts study, RA captured the temporal signature of harvest/regrowth/re-harvest cycles operating on approximately 40-year intervals, quantifying how shrubs predict forest, forest predicts grass, and grass predicts shrubs. In the wildfire study, historical grass and shrub presence at 5-year & 10-year lags showed elevated probability of large fires (62–75%) compared to forest and other cover types. This temporal signal appeared after stratification by ecoregion (pyrome), in 6 out of 13 pyromes, but was masked in unstratified analysis. In the landslides study, lithology-based stratification into volcanic and sedimentary domains revealed fundamentally different model structures that emphasize different failure controls: where the rock sits in volcanic terrain versus what the rock has weathered into in sedimentary terrain. Pareto-optimal model selection produced composite probability tables spanning a 91-percentage-point range (4% to 95%) of landslide susceptibility from mostly descriptive environmental variables using their native classification encoding. The selected landslide models used detailed soil, lithologic, and geomorphic descriptive classes directly.

In this work, interpretability is the goal: the conditional probability tables produced by RA provide specific, named combinations of environmental conditions tied to quantified outcome probabilities with sample counts. These outputs can be directly inspected and used in decision support & planning. RA models without loops identify which variables predict the outcome, much as a feature-selection or dimensionality-reduction step would. Models with loops go much further, potentially fusing several predictive relations into a single distribution, which is where RA exhibits the ability to detect non-linear and multi-component interactions.

The secondary aim is to make Reconstructability Analysis generally available by successfully porting the OCCAM software to current Python (3.x), creating PyOccam, a Python package that can be imported into any modern Python environment. PyOccam wraps the OCCAM C++ core code in Python with an API that enables much easier usage. RA can therefore use the same train/test splits and evaluation code as other ML algorithms such as Random Forest, XGBoost, and Logistic Regression. The analyses can be scripted, modified, and rerun to test different configurations.

PyOccam introduces these information-theoretic methods into standard data science workflows, so that analyses can be reproduced and compared fairly against other ML methods. The ability to script PyOccam enables the sweeping of parameters to test out different hypotheses by running it in a loop with different settings. PyOccam is available on the major public software delivery platforms and includes sample data sets, scripts, and notebooks that demonstrate its usage.

Rights

In Copyright. URI: http://rightsstatements.org/vocab/InC/1.0/ This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s).

Persistent Identifier

https://archives.pdx.edu/ds/psu/45130

Recommended Citation

Percy, David, "Spatial Reconstructability Analysis: Information-Theoretic Modeling of Categorical Raster Data" (2026). Dissertations and Theses. Paper 7200.