RUDN University engineers compared four machine learning methods that are used to process radar data. The researchers named the most effective and fastest methods. The study is published in the European Journal of Remote Sensing.
Newswise — Images of the surface of the Earth and other planets are obtained using a synthetic aperture radar (SAR). The radar is placed on a spacecraft or a carrier aircraft. It scans the surface and simultaneously tracks its position. As a result, detailed maps of the surface are obtained, and their quality does not depend on either the weather or the time of day. The most common type of such radars is PolSAR. Machine learning methods are used to process radar data. Due to the differences in the algorithms, they work with different accuracy and speed. Therefore, with an incorrectly selected algorithm, calculations turn out to be less accurate or require more time for calculations. RUDN engineers compared the four most popular methods and found out which of them is the most effective.
"The PolSAR data classification is always one of the favourite topics for remote sensing researchers. A large range of algorithms is used for this. The most well-known of them is the support vector machine (SVM), which is widely used to classify PolSAR data. However, there has been no research on the use of some extended versions of SVM so far. We compared these methods for classifying PolSAR data," said professor Yury Razoumny, the Head of Department for Mechanics and Mechatronics, Head of Academy of Engineering at RUDN University.
RUDN engineers together with their foreign partners compared four methods: the support vector Machine (SVM) and its three modifications - the least squares method of support vectors (LSSVM), the relevance vector machine (RVM) and the import vector machine (IVM). Their work was tested on three data sets obtained from PolSAR: images of Flevoland (Netherlands), Foulum (Denmark) and Winnipeg (Canada). The first and third data sets included extensive agricultural areas. Foulum's images were mostly forest, agricultural fields and populated areas. The machine learning algorithms had to determine how each piece of land is used (where wheat is grown, where the forest grows, where the river flows, and so on). Algorithms were trained on 5%, 10%, 50% and 90% of the data, and the remaining ones were used to test their perfomance. The effectiveness of the algorithms was evaluated by an indicator varying from 0 to 1, with the ideal classification corresponding to one, as well as the time required for learning according to the algorithm.
LSSVM turned out to be the fastest - for any amount of training data and for all three districts. For example, for a Foulum with 50% of the data given for training, LSSVM took less than 0.5 seconds, and the rest of the algorithms took 12-15 times longer. However, SVM turned out to be the most effective. It showed the highest learning rate for almost all data volumes for Winnipeg and Foulum: 0.78 for Foulum and 0.69 for Winnipeg. The second place in both cases was taken by IVM - 0.76 and 0.68, respectively.
"SVM has proven to be more efficient, more accurate, and more stable when classifying two of the three datasets. Another conclusion that we have made is the amazing speed of LSSVM compared to other methods. LSSVM produces comparable accuracy at a rate 12 times faster than SVM and about 15 times faster than RVM and IVM. Therefore, LSSVM can be considered as a worthy modification of SVM with acceptable accuracy and higher speed," said Javad Hatami Afkoueieh, PhD student at the RUDN Engineering Academy.