Use of Logistic Regression in Land-Cover Classification with Moderate-Resolution Multispectral Data

Abstract

The current study highlights the use of binary logistic regression for land-use land-cover (LULC) classification. The moderate-resolution Sentinel-2 multispectral data was used for LULC map generation for the post-monsoon season. The main focus of this study is to present a simple and precise approach for image classification using binary logistic regression (BLR) technique. The study was carried out in cropland, fallow land, forest and water body dominated subtropical region of India located in the eastern coastal region. The cropland and fallow lands are mostly dependent on the monsoon and reciprocal land covers. A large number of training and testing data points were collected viewing the image in a standard false-color composite. ArcGIS, Microsoft Office Excel and R software were used for classification. In addition to BLR, the training and testing data points were also used to perform the classification with ‘random forest’ classifier in R. We observed higher classification accuracy for spectrally pure classes and pixels and lower for closely associated mix-pixels. Lower user’s and producer’s accuracies (< 90%) were observed for fallow land, water body and grassland class during training and model building and for fallow land and forest during accuracy assessment, whereas the accuracies were more than 90% for the rest of classes during both training and testing. Misclassifications were mostly observed between forest, fallow land, grassland and water body during training, which were forest and fallow land in testing, due to their lower spectral difference with reference to classified classes. However, the overall accuracy and kappa value during training and testing were more than 94% and 0.98, respectively. Similar accuracies and misclassification were also obtained with the results of random forest model, validating the adopted methodology. Regardless of the seasonal variations in cropland and fallow land, the field observations (52 locations) also corroborated the estimated classification accuracy. The easy implementation and comparatively higher classification accuracy with the binary logistic technique are believed to increase its intense use in land-cover classification.