A model-based stochastic simulation for image classification: feature-space indicator simulation

Authors: Qing Wang*, Southern Illinois University
Topics: Remote Sensing, Spatial Analysis & Modeling, Land Use and Land Cover Change
Keywords: New image classification method, Remote sensing, LULC
Session Type: Paper
Day: 4/14/2018
Start / End Time: 10:00 AM / 11:40 AM
Room: Galvez, , Marriott, 5th Floor
Presentation File: No File Uploaded


Traditional remote sensing image classification methods are design-based that make a strong assumption of a universal underlying distribution over classes and space. This limits their classification accuracy because data distributions are often complex. Unlikely, nonparametric methods do not consider statistical assumption about data distributions. Although this property makes them extremely flexible in dealing with data that have various distributions, the nonparametric methods are highly sensitive to the quantity of sample data. Since the samples in remote sensing are usually scarce, it is necessary to use the assumption to gain more information if such assumption hold. Model-based geostatistic algorithms, such as indicator-based classification, make a weaker assumption about the natural phenomena: spatial autocorrelation. However, the mathematical model of this assumption is highly sensitive to the sample distances. Other than that, for creating indicator variables, the traditional way of using thresholds to subset continuous spectral values is questionable. In this study, a novel model-based classification method, feature-space indicator simulation (FSIS), was proposed. FSIS considers the sample locations in a feature space, which circumvents the issues. As a stochastic simulation technique, FSIS not only provides the classification estimations but also reproduces the natural variability of the classification decisions, which can be further used for assessing the uncertainty of the FSIS classification. The proposed method was validated and assessed by comparison with maximum likelihood classification, support vector machine, and random forest in a case study. Additionally, an error propagation of classification was modeled.

Abstract Information

This abstract is already part of a session. View the session here.

To access contact information login