Investigating the Performance of the Ensemble Learning Methods using the Feature Selection Method Approach, for the Integration of Reflective and Thermal Classifiers to Identify the Cloud, Cirrus Clouds and Cnow/Ice in MODIS Satellite Images

Almost all MODIS images have cloudy regions. Cloud detection and discriminating it from similar objects like snow/ice is a necessary pre-processing step before extracting accurate information. Features for cloud detection can be divided into two categories; textural and spectral features. Using textural features in the visible bands make cloud and snow/ice pixels separable, while cloud and snow can have similar temperature. Thermal features have been used for cloud detection in different heights in MODIS cloud mask (MOD35). Recently variety of studies using ensemble learning methods for land cover classification have been done. In the studies, ensemble learning methods used for classification but in this study, a new application has been introduced; for fusion of classifiers having reflectance and thermal features. Also, in previous studies, the effect of changing feature selection method on the performance of ensemble learning methods has not been examined. So, the purpose of this study is the comparison of the performance of fusion of the reflectance and thermal classifiers using two kinds of ensemble learning methods including boosting and Random Forest (RF) for detection of cloud, cirrus and snow/ice pixels, based on the feature selection method applied. First, some of the Visible-Infrared bands (VIR), 1, 2, 8 and 26 in addition to the thermal bands including 20, 22, 31, 32 and 35 of Terra MODIS were calibrated and the reflectance and Brightness Temperature (BT) values were extracted. Also, three indexes, Normalized Difference Vegetation Index (NDVI), Normalized Difference Snow Index (NDSI) and the ratio of bands b1/b2 were computed. Gray Level Co-occurrence Matrix (GLCM) textural features of the above mentioned bands and some of the BT differences including, BT31-BT32, BT33-BT31, BT31-BT20, BT20-BT32, BT22-BT32 were added to the reflectance and thermal input features sets. For selecting the suitable reflectance and thermal features in different kinds of boosting methods including Adaboost.M1, AdaboostSVM, Logitboost and Totalboost, S criteria and Genetic Algorithm (GA) were used, and in RF algorithm in addition to these methods, Recursive Feature Elimination (RFE) and correlation matrix were applied too. After feature selection step, training data were selected manually for cloud, cirrus and snow/ice and fed into the reflectance and thermal classifiers. Classifiers were fused in decision level using the majority vote method. The performance of different cases was compared using producer accuracy, user accuracy and kappa coefficient indexes. For almost all ensemble methods, ignoring which one of the feature selection methods applied, a high cloud and cirrus producer accuracy achieved. RFE and correlation matrix methods in RF algorithm, resulted in 99% and 100% cloud producer accuracy. These values are higher than S criteria and GA. Boosting methods, ignoring the kind of feature selection method, got higher snow/ice producer accuracy than RF algorithm by assigning higher weights to the class that has less training data among classes. Also, boosting methods resulted in higher cirrus user accuracy than RF. Among feature selection methods applied in RF, correlation matrix achieved 91% value for cirrus user accuracy. Finally, agreement to the reference map produced from MOD35 calculated. RF algorithms showed higher agreement with the reference map in comparison to the boosting methods. Highest agreement resulted from RF-RFE with the value of 76% and lowest agreement from Logitboost-GA with the value of 42%.