Feature selecti̇on for opti̇mal weather detecti̇on from meteorologi̇cal radar data and development of new methods

Especially in recent years, meteorological radar data processing and optimal weather forecasts in the direction of these data, have become an important research area. It is also focused on the detection of non-meteorological targets such as birds, insects and clutter using these radar data. Remote sensing systems as well as satellites have begun to be used in meteorological detection. Especially with polarimetric micro doppler radar, it is possible to detect the hydrometeors in the region by measurements. The most important reasons for choosing these radars instead of satellites can be gathered more detailed information about smaller areas and and quick measures can be taken according to the meteorological situation in the region with short term and instant detections. Within the scope of the thesis, a classifier on the detection of the hydrometeors in this region was constructed by using Hatay Meteorology Radar data and studies were carried out to improve the performance and the success of this classifier. In the first step, the data obtained from Hatay Radar were adapted to the formative studies and a support vector machine classifier was created using the extracted polarimetric radar features. The classifier works in two stages; firstly rainfall detection in radar images, then birds-bugs and clutter detection in regions where rainfall could not be detected. Since there is no ground truth information for classifier training in radar data, radar features are examined, regions visually detectable in hydrometeor classes are scanned, and a ground truth dataset for training is created. Classifier training was done using the ground truth dataset. Then a feature selection was made for the eight features from the radar and the features required for both classification stages were determined. Erroneous and noisy measurements in the radar data were taken using a masking method and the classification results were compared with the selected features. Finally, in order to minimize false precipitation detections, new texture features were extracted by using the spectral features measured from the radar, as well as the classification results, and they were reclassified using these features.As texture features, were used on the texture obtained from the classification results, energy and entropy values calculated using neighborhood relations between the classes. At the end of the thesis, in the classification process using the obtained texture features, the performance of the classifier performance in the precipitation detection was increased and the results of faulty rainfall classification, which was measured especially in the precipitation without precipitation, were largely eliminated.