Spatially-Explicit Bayesian Information Entropy Metrics for Calibrating Landscape Transformation Models

Kostas Alexandridis,Bryan Pijanowski

doi:10.3390/e15072480

Kostas Alexandridis, Bryan Pijanowski

Open Access

https://doi.org/10.3390/e15072480

Copy DOI

Abstract

Assessing spatial model performance often presents challenges related to the choice and suitability of traditional statistical methods in capturing the true validity and dynamics of the predicted outcomes. The stochastic nature of many of our contemporary spatial models of land use change necessitate the testing and development of new and innovative methodologies in statistical spatial assessment. In many cases, spatial model performance depends critically on the spatially-explicit prior distributions, characteristics, availability and prevalence of the variables and factors under study. This study explores the statistical spatial characteristics of statistical model assessment of modeling land use change dynamics in a seven-county study area in South-Eastern Wisconsin during the historical period of 1963–1990. The artificial neural network-based Land Transformation Model (LTM) predictions are used to compare simulated with historical land use transformations in urban/suburban landscapes. We introduce a range of Bayesian information entropy statistical spatial metrics for assessing the model performance across multiple simulation testing runs. Bayesian entropic estimates of model performance are compared against information-theoretic stochastic entropy estimates and theoretically-derived accuracy assessments. We argue for the critical role of informational uncertainty across different scales of spatial resolution in informing spatial landscape model assessment. Our analysis reveals how incorporation of spatial and landscape information asymmetry estimates can improve our stochastic assessments of spatial model predictions. Finally our study shows how spatially-explicit entropic classification accuracy estimates can work closely with dynamic modeling methodologies in improving our scientific understanding of landscape change as a complex adaptive system and process.

Highlights

The transformation of landscapes by humans is well known to be a complex phenomenon directed by a host of socioeconomic drivers interacting at multiple spatial and temporal scales [1]
The analysis provided in this paper aims to provide a suite of spatially explicit metrics that go beyond traditional statistical calibration and parameter estimation of model predictions, but attempts to provide a statistically sound and comprehensive insight into information-theoretic aspects of the modeled phenomena under study, and explore the parameter space and confines of our predictions
The analysis described above reveals the magnitude and multi-dimensionality of the spatial complexity involved in modeling landscape transformations in mixed and asymmetric landscapes in terms of amount and distribution of change

Summary

Introduction

The transformation of landscapes by humans is well known to be a complex phenomenon directed by a host of socioeconomic drivers interacting at multiple spatial and temporal scales [1]. The spatial patterns of land use land cover that result from human activity on the landscape are in turn complex and very heterogeneous [2]. Traditional statistical accuracy assessment techniques of land change models, essential for validating observed and historical land use changes, fail to capture the stochastic character of the dynamics that are simulated [11], especially across multiple scales. We propose new metrics and indicators that provide the modeler the ability to extract higher-order information dynamics from landscape transformation simulations. Much progress has been made in this area in the last decade, it is well recognized that adequate measures of the simulation itself and model output still need development [15]

Objectives

Methods

Results

Conclusion