Projection of Land Use Change Patterns using Kernel Logistic Regression

Abstract

Change analysis is probably a natural step following the detection of changes using remote sensing data. One significant topic in change analysis is to model the changes in relation to their driving factors and to project future land-use patterns. While logistic regression (LR) has been widely used in change modeling, this paper presents an improved method, kernel logistic regression (KLR), to model the nonlinear relationship between land-use change and various causal factors such as population, distance to road and facilities, surrounding land-use, and others. Traditional KLR models contain one coefficient for each training sample, rendering it inappropriate for applications of land-use change analysis with more than a few thousand samples. A feature vectors selection method for the KLR model has therefore been proposed to impose sparsity and control complexity. To test the effectiveness of KLR, a case study was implemented to model rural-urban land-use conversion in the city of Calgary, Canada during the periods 1985 to 1990 and 1990 to 1999. The KLR model was compared with a commonly used LR model in terms of the Percentage of Correct Prediction (PCP), Area under Curve (AUC), and McNamara’s test, and the results consistently demonstrated the better performance of KLR.