Session

Technical Poster Session II

Location

Utah State University, Logan, UT

Abstract

§The advent of small satellites and hybrid constellations have made multiple types of sensors and image products available.

§In classification problems, these diverse data sources can be used as inputs (i.e., variables) to perform categorization tasks.

§Using the optimal number of variables is key because:

§Too little (under-fitting) may result in poor accuracy.

§Too many (over-fitting) may increase computation time and yield a classifier that is too specific to the dataset.

§Common techniques for selecting variables include:

§Metrics such as Mean Decrease in Accuracy (MDA) and Mean Decrease in Gini (MDG).

§Methods such as Principal Component Analysis (PCA), Support Vector Machines (SVM), and genetic algorithms.

§However, the aforementioned variable selection techniques have some limitations, including: §Inability to define arbitrary groups of variables and determine the importance of each group as a unit.

§No indication provided as to why a particular variable or group of variables is ranked as more or less important than another variable or group of variables.

§No consideration of interactions between variables (for example, correlated or dependent variables).

§For MDA and MDG, difficulty interpreting the importance values (both provide scaled numbers).

Share

COinS
 
Aug 1st, 12:00 AM

Using Shapley Values and Game Theory to Measure the Effectiveness of Different Satellite Image Products in Hybrid Constellations

Utah State University, Logan, UT

§The advent of small satellites and hybrid constellations have made multiple types of sensors and image products available.

§In classification problems, these diverse data sources can be used as inputs (i.e., variables) to perform categorization tasks.

§Using the optimal number of variables is key because:

§Too little (under-fitting) may result in poor accuracy.

§Too many (over-fitting) may increase computation time and yield a classifier that is too specific to the dataset.

§Common techniques for selecting variables include:

§Metrics such as Mean Decrease in Accuracy (MDA) and Mean Decrease in Gini (MDG).

§Methods such as Principal Component Analysis (PCA), Support Vector Machines (SVM), and genetic algorithms.

§However, the aforementioned variable selection techniques have some limitations, including: §Inability to define arbitrary groups of variables and determine the importance of each group as a unit.

§No indication provided as to why a particular variable or group of variables is ranked as more or less important than another variable or group of variables.

§No consideration of interactions between variables (for example, correlated or dependent variables).

§For MDA and MDG, difficulty interpreting the importance values (both provide scaled numbers).