No abstract

We introduced a new classifier named Learning Word-vector Quantization (LWQ) to solve morphological ambiguities in Turkish, which is an agglutinative language. First, a new and morphologically annotated corpus, and then its datasets are prepared with a series of processes. According to datasets, LWQ finds optimal word-vectors positions by moving them in the Euclidean space. LWQ does morphological disambiguation in two steps: First, it defines all solution candidates of an ambiguous word using a morphological analyzer; second, it chooses the best candidate according to its total distances to neighbor words that are not ambiguous. To show LWQ's performance, we have conducted many tests on the corpus by considering the consistency of classification. In the experiments, we achieve 98.4% correct classification ratio to choose correct parse output, which is an excellent level for the literature.

In this article, a solution to the morphological ambiguity problem which occurs frequently in morphologically complex languages like Turkish is proposed. Generally, statistical methods are applicable for these tasks which maximize the information, obtained for a probable word order sequence in a sentence. The decision in selection of the method for calculation of the probabilities and the sequence selection method depends on the nature of the language. By using the co-occurrence statistics obtained from a semantic graph network which represents the lemmas of the sentences, the best word order sequence is selected from the alternatives. The non-ambiguous and free-word-order character of this network is helpful in determining the statistics independently. The probability values are obtained by using the Naive Bayes (NB) method and the selection of each word sequence is achieved by maximization, in the inspiration of the Viterbi algorithm.

No abstract

Generation of flood inundation maps is beneficial in flood risk assessment and evaluation. Flood inundation mapping can be achieved by many remote sensing techniques like change detection (CD) with thresholding and machine learning-based (ML) methods. Optical and synthetic aperture radar (SAR) imagery are widely used, provided by different satellite systems. This study used Sentinel-1 SAR and Sentinel-2 MSI satellite data in Google Earth Engine (GEE) with supervised ML algorithms. Gediz Plain, Turkey was selected as the study area, which is an agricultural area covered mostly by croplands. A flood event that occurred on February 2, 2021, was examined and flood inundation map for the study area was composed. Support Vector Machines (SVM), Random Forest (RF) and K-Nearest Neighbor (KNN) ML algorithms were selected and models were trained with manually created labelled data in GEE. Also, CD was applied on after and before event SAR images in a traditional approach. RF classifier performs best in Sentinel-2 MSI imagery with 94% overall classification accuracy where KNN classifier gives 93.3% accuracy value for Sentinel-1 SAR dataset, indicating the robustness of SAR imagery for all-weather conditions.