Classification of satellite images with deep learning and machine learning

Remote sensing is of great importance in the detection of many natural differences such as disaster detection, climate changes and urban changes. As a result of the developments in the field of imaging in technology, remote sensing has become a very popular topic today. The biggest benefit provided by the developing technologies is that the data used for remote sensing has become easily accessible today. Remote sensing can be explained as the process of detecting the distinctive features of the environment, such as physical and spatial information. One of the most influential factors on success in detection processes is resolution. The resolution being lower than the required level causes the features of the objects to be differentiated to become incomprehensible and therefore constitutes a major obstacle to the detection process. Today, deep learning methods seem to be very popular and successful in classifying remote sensing data. Successful results are obtained in the field of learning image features and extracting the necessary features for classification using deep learning networks. Another difficulty in satellite image classification is the inability to obtain enough data for classification. In this study, deep learning and machine learning methods were used to classify satellite images. The analyzes were performed on EuroSAT, AID and a combined dataset created by combining two datasets under appropriate conditions. The main purpose of our data set merging process is to observe the changes in classification success according to the increasing number of classes in the satellite image classification process. In this thesis, two data sets were analyzed separately using deep learning networks. In addition, the combined data set was analyzed using both deep learning and machine learning methods. The purpose of this process is to compare machine learning and deep learning based classifiers on the created data set. As a result of the analysis, the highest classification success for the EuroSAT dataset was obtained using the DenseNet-201 deep learning network with an accuracy rate of 98.11%. The classifier with the highest accuracy for the AID dataset is the DenseNet-201 deep learning network with an accuracy rate of 96.10%. Among the classification methods applied in the analysis for the combined data set, the classifier with the highest accuracy is observed as the DenseNet-201 deep learning network, with an accuracy rate of 96.20%. It is seen that the success rate in the combined dataset is higher than in the AID dataset.