Assessment of ensemble classifiers using the bagging technique for improved land cover classification of multispectral satellite images

This study evaluates an approach for Land-Use Land-Cover classification (LULC) using multispectral satellite images. This proposed approach uses the Bagging Ensemble (BE) technique with Random Forest (RF) as a base classifier for improving classification performance by reducing errors and prediction variance. A pixel-based supervised classification technique with Principle Component Analysis (PCA) for feature selection from available attributes using a Landsat 8 image is developed. These attributes include coastal, visible, near-infrared, short-wave infrared and thermal bands in addition to Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI). The study is performed in a heterogeneous coastal area divided into five classes: water, vegetation, grass-lake-type, sand, and building. To evaluate the classification accuracy of BE with RF, it is compared to BE with Support Vector Machine (SVM) and Neural Network (NN) as base classifiers. The results are evaluated using the following output: commission, omission errors, and overall accuracy. The results showed that the proposed approach using BE with RF outperforms SVM and NN classifiers with 93.3% overall accuracy. The BE with SVM and NN classifiers yielded 92.6% and 92.1% overall accuracy, respectively. It is revealed that using BE with RF as a base classifier outperforms other base classifiers as SVM and NN. In addition, omission and commission errors were reduced by using BE with RF and NN classifiers.