Automatic classification of hyperspectral satellite images with deep learning methods in Python

Hyperspectral Imaging, which is one of the most advanced methods of digital imaging, differs from the well known RGB imaging by the much more electromagnetic bands it contains. Imbalanced datasets are frequently encountered in machine learning, and datasets of hyperspectral images are often imbalanced. The success of models created with imbalanced data sets is low. In order to prevent this performance loss, data sets should be balanced using special methods. The main goal in this thesis is to reveal the performance difference between the imbalanced original dataset and the new datasets balanced by using Smote, Adasyn, K-Means and Cluster methods. Xuzhou Hyspex dataset was used within the scope of the study. This data set, consisting of 9 different classes in total, was taken from IEEE-Dataport Machine Learning Repository. Convolutional Neural Networks was preferred for classification. In addition, the approximation approaches used for multi-class problems such as One vs All (OvA) and One vs One (OvO) were used. The dataset is splitted into training and testing parts using Hold-Out and Cross Validation methods. Confusion Matrix was created for the evaluation of the obtained results, in addition, Accuracy, Precision, Recall, F-Measure and Cohen's Kappa values were calculated. When the obtained results were evaluated, it was observed that the performance obtained with balanced data sets was higher than with the unbalanced data set. In this way, it has been demonstrated that data set balancing methods contribute positively to performance