Search

In this study, the issue of intent classification in frequently asked questions for chatbots is discussed. There are many studies on this subject. However, most of the studies were conducted for the English language, and the issue of improving the intent classification results for small data sets has not been adequately examined. For this purpose, frequently asked questions data were used with subword semantic hashing method and its effects in Turkish and English languages were examined. According to the results, we observe an increase in classification performance of the subword semantic hashing method in Turkish language and the improvement is significantly higher compared to the English language. In addition, the classification performance results of pretrained the subword semantic hashing is better than the results of neural network models such as CLIP and FastText.

Remote sensing mainly focuses on information extraction from data acquired by sensors on satellite and aerial platforms. Here, one such area of interest is ground object detection and shape extraction. Recently launched satellites and conventional aerial platforms (such as commercial UAV and professional drones) have sensors leading to more detailed and rich data source for this purpose. From these, data most of the times come in the form of optical images and LiDAR measurements. Resolution of this acquired data has increased significantly such that most ground objects (as buildings, trees, ships, cars, airplanes) can be detected and analyzed in detail. Therefore, computer vision methods have become extremely useful in remote sensing applications such as building detection and shape extraction for urban planning; tree crown measurement for crop yield forecasting; ship detection for monitoring unlawful fishery; car detection for traffic flow monitoring and intelligent transportation; and airplane detection for military and commercial operations. Researchers proposed several methods to automate the mentioned applications since manually handling them is extremely hard and prohibitively time consuming. Unfortunately, the proposed methods focus on one object type most of the times. Therefore, there is no general method to handle all the mentioned applications using computer vision tools. To overcome this problem, we propose a general framework for object detection and shape extraction in remote sensing data. Our method is based on probabilistic representation inspired by our previous work and perceptual organization principles. Due to space limitations, we only focus on buildings, trees, ships, airplanes, and cars as objects of interest in this study. We test the proposed method on several optical images acquired by different satellites and LiDAR data obtained from an aerial platform. For all objects of interest, we provide test results on both object detection and shape extraction steps. We analyze the proposed method based on these tests and discuss its strengths and weaknesses. We also comment on possible future extensions of the proposed method.

LiDAR technology is advancing. As a result, researchers can benefit from high-resolution height data from Earth's surface. Digital terrain model (DTM) generation and point classification (filtering) are two important problems for LiDAR data. These are connected problems since solving one helps solving the other. Manual classification of LiDAR point data could be time consuming and prone to errors. Hence, it would not be feasible. Therefore, researchers proposed several methods to solve DTM generation and point classification problems. Although these methods work fairly well in most cases, they may not be effective for all scenarios. To contribute in this research topic, a novel method based on two-dimensional (2-D) empirical mode decomposition (EMD) is proposed in this study. Local, nonlinear, and nonstationary characteristics of EMD allow better DTM generation. The proposed method is tested on two publicly available LiDAR dataset, and promising results are obtained. Besides, the proposed method is compared with other methods in the literature. Comparison results indicate that the proposed method has certain advantages in terms of performance.

Crop yield forecasting in a region has become an important research area due to global warming and related climate changes. Although this can be performed by available statistical information, obtaining recent and up to date data to extract reliable statistical information is not easy. Very high resolution satellite images can be used for this purpose. However, manually processing these images acquired from large regions is neither feasible nor reliable. Therefore, automated methods are needed for this purpose. In this study, we propose a novel method to help forecasting the crop yield in an orchard. The number of trees in an orchard with the size and type of each tree crown gives an approximate crop that can be harvested. Therefore, we focus on both tree crown detection and delineation for this purpose. The proposed method for tree crown detection is based on probabilistic voting. For tree crown delineation, we propose a watershed segmentation based ellipse fitting method. We tested the proposed method on 17 satellite images containing 13,476 trees. We compared the method with the classical local maxima/minima filtering and a recent method in literature using three more test images. These tests indicate the strengths and weaknesses of the proposed method. © 2017 Informa UK Limited. 1,14e+201 This work is supported by TUBITAK through project no 114E199.

As satellite images cover wide areas and obtaining them has become easier, using these images in agriculture has become an important research area. Especially, satellite images can be used in seasonal crop estimation. Obtaining the number of trees in a region, with the size of each tree, gives the approximate amount of crop that can be harvested from that region. In this study, we propose a voting based method on grayscale satellite images. To test the proposed method, we picked eight satellite images containing 2668 trees. We summarized the obtained results in this study. © 2016 IEEE. 24th Signal Processing and Communication Application Conference, SIU 2016 -- 16 May 2016 through 19 May 2016 -- -- 122605

As satellite images cover wide areas and obtaining them has become easier, using these images in agriculture has become an important research area. Especially, satellite images can be used in seasonal crop estimation. In this study, we focused on crop estimation from trees. The boundary of a tree is proportional to its age which gives information on the approximate crop that can be obtained from it. Obtaining the number of trees in a region, with the size of each tree, gives the approximate amount of crop that can be harvested from that region. In this study, we propose a method based on multiple filtering, watershed segmentation, and Otsu thresholding to detect trees and their boundaries. To test the proposed method, we picked three satellite images containing 6928 trees. These trees have diameters between 2 to 30 pixels. We compared the proposed method with two other methods in the literature. We summarized the obtained results in this study. © 2015 IEEE. Aselsan;e al.;HAVELSAN;Roketsan;TAI;TURKSAT 7th International Conference on Recent Advances in Space Technologies, RAST 2015 -- 16 June 2015 through 19 June 2015 -- -- 116912

Automatic extraction of bare-Earth LiDAR points to generate Digital Terrain Model (DTM) is still an ongoing problem. Even though there are several methods for ground filtering, automatic and adaptive methods are still a need due to the complexity of the environment. In this study, we address the ground filtering problem by applying Empirical Mode Decomposition (EMD) to the airborne LiDAR data. EMD is a data-driven method that adapts to the local characteristics of the signal. We benefit from EMD to extract the local trend of the LiDAR height data. This way, can extract a local adaptive threshold to filter ground and non-ground objects. We tested our method using the ISPRS LiDAR reference dataset and obtained promising results. We also compared the filtering results with the ones in the literature to show the improvements obtained. © 2015 IEEE. 2015 23rd Signal Processing and Communications Applications Conference, SIU 2015 -- 16 May 2015 through 19 May 2015 -- -- 113052

Crowd monitoring is an important task of security forces. If an emergency occurs during large events, authorities should take urgent measures to prevent causalities. Also understanding crowd dynamics such as tracking crowds or sparse people goups before an emergency occurs is a need. Therefore, crowd detection and analysis is a critical research area. There are several studies for crowd monitoring that use street or indoor cameras which may not be directly used for analyzing large crowds. In this study, we approach the problem using aerial images. We propose two novel methods. In the first method, we use first-order spatial point statistics. It uses the nearest neighbor relations for each person in the image to detect crowd regions. Our second method also uses the first order statistics with an additional sparse people group detection flexibility. We test the proposed methods on two aerial images and provide quantitative test results. © 2015 IEEE. Aselsan;e al.;HAVELSAN;Roketsan;TAI;TURKSAT 7th International Conference on Recent Advances in Space Technologies, RAST 2015 -- 16 June 2015 through 19 June 2015 -- -- 116912

LiDAR data provides valuable information for various remote sensing applications. For these, one important and challenging problem is ground filtering. This operation separates the bare earth and object data. Researchers proposed several methods to solve this problem. However, the complexity of the data limit the usability of these methods for all terrain types. Besides, the performance obtained in ground filtering should be improved further. In this study, we focus on this problem and propose a novel ground filtering method using Empirical Mode Decomposition (EMD). We tested the proposed method on the standard ISPRS data set and evaluate its strengths and weaknesses. We also compared the proposed method with the ones in the literature to show the improvements obtained.

Recent sensors give valuable data for remote sensing applications. Among these, building and change detection are important problems. Therefore, researchers worked on these problems using both 2D and 3D data. Some previous studies used only 2D data due to their availability. Yet others used either 3D data alone or 2D and 3D data in a joint manner. Besides, some studies only focused on building detection. Yet others used detected building information in change detection. In this study, we focus on 3D change detection based on building information. Therefore, we first detect buildings. At this step, we benefit from both 2D and 3D data. Then, we locate changes based on these detected buildings. We detect building pixels using panchromatic, multispectral, and Digital Surface Model (DSM) data using a decision tree classifier. Then, we refine the detected building pixels using morphological and shape based operations. Finally, we apply an object based hierarchical change detection method on the refined pixels. We tested our method on 780 buildings and obtained promising results. © 2014 IEEE. 2014 22nd Signal Processing and Communications Applications Conference, SIU 2014 -- 23 April 2014 through 25 April 2014 -- Trabzon -- 106053

In this paper, sea clutter radar plots are modeled by spatial point processes. A test procedure is proposed to analyze "Complete Spatial Randomness (CSR)" characteristics of radar plot locations. Plot intensity map is also constructed. This map is separated into two sub-regions; cutter region and moving target region. This map can be used as a reliability metric for target detection algorithms.

The oscillating and swinging parts of a target observed by radar cause additional frequency modulation and induce sidebands in the target's Doppler frequency shift (micro-Doppler). This effect provides unique features for classification in radar systems. In this paper, the micro-Doppler spectra and range-Doppler matrices of single bird and bird flocks are obtained by simulations for linear FMCW radar. Obtained range-Doppler matrices are compared for single bird and bird flock under several scenarios and new features are proposed for classification.

In this paper, statistical modeling of clutter data measured by a noncoherent S-band marine radar mounted on a fixed position is presented. Characterization is done by finding the best fitted density function to the clutter over eight candidate distribution. Real-time parameter estimation of the predetermined distribution and automatic threshold detection for Constant False Alarm Rate (CFAR) is provided.