Comparison of data mining algorithms in remote sensing using Lidar data fusion and feature selection

Application of data mining techniques defines the basis of land use classification. Even though multispectral images can be very accurate in classifying land cover categories, using spectral reflectivity alone sometimes fails to distinguish between landcover types that share similar spectral signatures such as forest and wetlands. The problem aggravates owing to interpolation of neighbourhood pixel values. In this paper, we present a comparison of four classification and clustering algorithms and analyze their performance. These algorithms are applied both on spectral reflectivity values alone and along with Lidar data fusion. Experiments were performed in the Carlton County of Minnesota. Accuracy estimation was conducted for all models. Experiments indicate that accuracy increases when Lidar data is used to complement the spectral reflectivity values. Random Forest Classification and Support Vector Machines yield good results consistently due to their ensemble learning methods and the ability to represent non-linear relationship in the dataset, respectively. Maximum likelihood shows significant improvement with Lidar data fusion and ISODATA clustering approach has the lowest accuracy rate.