|
This study examined the utility of polytomous logistic regression in pixel classification of remotely sensed images by the growth stage of forests. For a population of grouped continuous categories, the assumption of normal distribution of independent variables, which is often required in multivariate classification methods, may not be appropriate. Two types of polytomous logistic regression procedures, multinomial and cumulative logistic regression, were used to classify Landsat TM data by growth stage (regeneration-immature, intermediate, and mature) of loblolly pine (Pinus taeda L.) forest in the east central Mississippi. Multinomial logistic regression is typically used for analysis of unordered categorical data. Cumulative logistic regression is one of the most commonly used methods of ordinal logistic regression which is generally preferred to analyze ordered categorical data, although, it imposes restrictions on the data. Three hundred sample points were located randomly throughout the study site and vectors of pixel values of four bands of Landsat TM data were used to predict growth stage at each sample location. The results were compared to that of parametric and nonparametric discriminant analysis, k-nearest neighbor method. Non-normal distribution of independent variables indicated a violation of the assumptions for parametric discriminant analysis. Classification with cumulative logistic regression using four bands was performed first. However, the assumption of the model was not met. So, the classification was also performed using only band 4 which appeared to meet the assumption. The error rate of cumulative logistic regression was 39.12% with all the bands and 37.70% with band 4 alone. Although error rate with cumulative logistic regression with band 4 alone resulted in the lowest error rate, the improvement over other methods was marginal. The error rate of k-nearest neighbor method varied from 38.68 to 48.06% depending on choice of the value of k.
|
