- Research ArticleAssessment of the spatiotemporal prediction capabilities of machine learning algorithms on Sea Surface Temperature data: A comprehensive studySerkan KartalEngineering Applications of Artificial Intelligence, 2023
Spatiotemporal time series prediction plays a crucial role in a wide range of applications. However, in most of the studies, spatial information was ignored and predictions were carried out either on a few points or on average values. In this study, 37 different configurations of 4 traditional ML models and 3 Neural Network (NN) based models were utilized to provide a comprehensive comparison and evaluate the spatiotemporal data prediction capabilities of the ML models. Additionally, to reveal the importance of spatial data for the time series prediction process, the best configuration of each ML model was evaluated with and without using spatial information. The utilized models were: (i) Linear Regression (LR), (ii) K-Nearest Neighbors (KNN), (iii) Decision-Trees (DT), (iv) Support Vector Machine (SVM), (v) Multi-Layer Perceptron (MLP), (vi) Long Short-Term Memory (LSTM), and (vii) Gated Recurrent Unit (GRU). The study was performed on the Sea Surface Temperature (SST) data collected by satellite radiometers via infrared measurements. The models were evaluated according to their one-month ahead spatiotemporal SST prediction performance over the southern coasts of Turkey, and the effects of spatial information on model performance were presented. Results reveal that the spatial information increased the prediction performance by approximately 25%, in terms of RMSE. Additionally, acquired results show that the LSTM model outperforms all other ML models and gives the smallest prediction errors in all metrics.
- Research ArticleA decision tree-based measure-correlate-predict approach for peak wind gust estimation from a global reanalysis datasetS. Kartal, S. Basu, and S. J. WatsonWind Energy Science Discussions, 2023
Peak wind gust (Wp) is a crucial meteorological variable for wind farm planning and operations. However, for many wind farm sites, there is a dearth of on-site measurements of (Wp). In this paper, we propose a machine-learning approach (called INTRIGUE) that utilizes numerous inputs from a public-domain reanalysis dataset, and in turn, generates long-term, site-specific (Wp) series. Through a systematic feature importance study, we also identify the most relevant meteorological variables for (Wp) estimation. Even though the proposed INTRIGUE approach performs very well for nominal conditions compared to specific baselines, its performance for extreme conditions is less than satisfactory.
- Research ArticleRegion contrastive camera localizationMehmet Sarıgül, and Levent KaracanPattern Recognition Letters, 2023
Visual camera localization is a well-studied computer vision problem and has many applications. Recently, deep convolutional neural networks have begun to be utilized to solve six-degree-of-freedom (6-DoF) camera pose estimation via scene coordinate regression from a single RGB image and they outperform the traditional methods. However, recent works do not consider scene variations such as viewpoint, light, scale, etc due to the camera motion. In this work, we propose a region contrastive representation learning approach to alleviate these problems. The proposed approach maps image features from different camera views of the same 3D region to nearby points in the learned feature space. In contrast, it pushes visual features of other regions to distant points. Our method improves the existing camera localization methods and achieves state-of-the-art results on indoor 7-Scenes and outdoor Cambridge Landmarks datasets. Experimental results show that the proposed approach reduces the pose and angle errors and increases the average accuracy from 84.8% to 85.62% on the state-of-the-art baseline model. In addition, we perform an ablation study on a baseline network with different settings to demonstrate the efficiency of the proposed region contrastive camera localization method.
- Research ArticleA survey on digital video stabilizationMehmet SarıgülMultimedia Tools and Applications, Apr 2023
Shakes and jitters are an eventual result of involuntary camera movements during video recording. Digital video stabilization is the elimination of these errors with smart algorithms. This process is usually performed in three steps which are camera motion estimation, motion correction, and stable video synthesis. In the literature, methods differ by the way they perform these steps. The recent success of deep learning has pioneered learning-based video stabilization approaches. This paper provides a detailed explanation of video stabilization methods by analyzing and comparing the applied approaches from past to present.
- Research ArticlePrediction of MODIS land surface temperature using new hybrid models based on spatial interpolation techniques and deep learning modelsSerkan Kartal, and Aliihsan SekertekinEnvironmental Science and Pollution Research, Sep 2022
Land surface temperature (LST) prediction is of great importance for climate change, ecology, environmental and industrial studies. These studies require accurate LST map predictions considering both spatial and temporal dynamics. In this study, multilayer perceptron (MLP), long short-term memory (LSTM) and an integrated machine learning model, namely Convolutional LSTM (ConvLSTM), were utilized for one step ahead LST prediction. Data were gathered from 1-day (MYD11A1) and 8-day composite (MYD11A2) Moderate Resolution Imaging Spectroradiometer (MODIS) sensors, which have 1-km × 1-km spatial resolution. Considering the inability of MODIS sensors to provide LST data under cloudy conditions, Inverse DISTANCE WEIGHTING (IDW), natural neighbor (NN), and cubic spline (C) methods were used to overcome the missing pixel problem. The proposed methods were tested over the Northern part of Adana province, Turkey, and the performances of the models were quantitatively evaluated through performance measures, namely, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE). The selected datasets range from 01 January 2017 to 01 November 2020 and from 01 January 2015 to 01 November 2020 for daily LST and 8-day composite LST, respectively. While 60% of the datasets were used as training set, the remaining 40% were used as validation (20%) and test (20%) sets. RMSE maps were generated to evaluate the pixelwise performance of the proposed method. On the other hand, the best average RMSE and MAE for the daily test set were obtained from the combination of ConvLSTM and NN (NN-ConvLSTM) as 3.62 °C and 2.85 °C, respectively, while they were acquired 3.57 °C and 2.69 oC from the combination of MLP and NN (NN-MLP) for the 8-day composite LST test set. The results revealed that the proposed hybrid models could be used for one step ahead spatiotemporal prediction of LST data.
- Research ArticleDecoupled Adaptive Backstepping Sliding Mode Control of Underactuated Mechanical SystemsBaris Ata, and Ramazan CobanControl Engineering and Applied Informatics, Sep 2022
In this paper, a combination of sliding mode control and adaptive backstepping control with a decoupling algorithm is considered for controlling 2 degrees of freedom underactuated mechanical systems subject to parametric uncertainties and external disturbances. The stability of the system is assured by the design steps of the proposed decoupled adaptive backstepping sliding mode control which are based on the Lyapunov theorem. The effectiveness of the proposed decoupled adaptive backstepping sliding mode control method is compared against a decoupled sliding mode controller by testing on a real-life inverted pendulum on a cart system which is a classical testbed for underactuated mechanical systems. The experimental outcomes justify the proposed decoupled adaptive backstepping sliding mode controller provides a more satisfying performance compared to the conventional decoupled sliding mode controller. Besides the proposed method is able to handle parametric uncertainties contrarily to the decoupled sliding mode control. © 2022. All Rights Reserved.
- Research ArticleComparison of semantic segmentation algorithms for the estimation of botanical composition of clover-grass pastures from RGB imagesSerkan KartalEcological Informatics, Sep 2021
In dairy industry, estimation of the in-field clover-grass ratio is an important factor in composing feed ratios for cows. Accurate estimation of the grass and clover ratios enables smart decisions to optimize seeding density and fertilization, resulting in increased yield and reduced amount of chemicals used. In practice, this process is still primarily performed by human-eye, which is labor-intensive, subjective, and error-prone. Therefore, plant species ratio estimation using traditional methods is hardly possible and misleading. Modern semantic segmentation models on digital images offer a promising alternative to overcome these drawbacks. In this paper, an extensive comparison of Deep Learning (DL) models for estimating the ratio of clover, grass, and weeds in red, green, and blue (RGB) images is presented. Three DL architectures (Unet, Linknet, FPN) are combined with ten randomly initialized encoders (variations of VGG, DenseNet, ResNet, Inception and EfficientNet) to construct thirty different segmentation models. Evaluation of models was performed on a publicly available dataset provided by the Biomass Prediction Challenge. The best segmentation accuracy was reached by the FPN-Inceptionresnetv2 model by 76.7%. This result indicates the great potential in deep convolutional neural networks for the segmentation of plant species in RGB images. Furthermore, this study lays the foundation for our next set of experiments with DL to improve the benchmarks and will further the quest to identify phenotype characteristics from agricultural imagery collected from the field.
- Research ArticleMachine Learning-Based Plant Detection Algorithms to Automate Counting Tasks Using 3D Canopy ScansSerkan Kartal, Sunita Choudhary, Jan Masner, and 5 more authorsSensors, Sep 2021
This study tested whether machine learning (ML) methods can effectively separate individual plants from complex 3D canopy laser scans as a prerequisite to analyzing particular plant features. For this, we scanned mung bean and chickpea crops with PlantEye (R) laser scanners. Firstly, we segmented the crop canopies from the background in 3D space using the Region Growing Segmentation algorithm. Then, Convolutional Neural Network (CNN) based ML algorithms were fine-tuned for plant counting. Application of the CNN-based (Convolutional Neural Network) processing architecture was possible only after we reduced the dimensionality of the data to 2D. This allowed for the identification of individual plants and their counting with an accuracy of 93.18% and 92.87% for mung bean and chickpea plants, respectively. These steps were connected to the phenotyping pipeline, which can now replace manual counting operations that are inefficient, costly, and error-prone. The use of CNN in this study was innovatively solved with dimensionality reduction, addition of height information as color, and consequent application of a 2D CNN-based approach. We found there to be a wide gap in the use of ML on 3D information. This gap will have to be addressed, especially for more complex plant feature extractions, which we intend to implement through further research.
- Research ArticlePosition resolution study at high energies of a sampling electromagnetic calorimeter whose active material is a scintillator with Peroxide-cured polysiloxane baseGüral Aydın, Mehmet Sarıgül, and Hasan SarıgülNuclear Instruments and Methods in Physics Research Section A: Accelerators, Spectrometers, Detectors and Associated Equipment, Sep 2020
This study is based on the simulation for the position resolution performances of a sampling electromagnetic calorimeter with a Peroxide-cured polysiloxane based scintillator as an active material. Various algorithms and corrections were applied to reconstruct hit positions. Energy deposition in the center tower and neighboring towers were used to reconstruct the beam impact position in a detector tower module consisted of a 3 x 3 array matrix. Linear weights, corrected linear weights, logarithmic weights, and corrected logarithmic weights were the different algorithms to reconstruct beam hit positions through energy weighted tower positions. Moreover, the iterative weighting method based on the logarithmic weights were applied. The re-weighting algorithm based on the iterative weighting method was seen to improve the results necessarily at relatively low beam energies. Additionally, deep neural network structures were applied over the linear weights without using any logarithmic weights or correction to improve the position resolutions.
- Research ArticleSegmentation of Bean-Plants Using Clustering AlgorithmsSerkan Kartal, Sunita Choudhary, Michal Stočes, and 5 more authorsAgris on-line Papers in Economics and Informatics, Sep 2020
In recent years laser scanning platforms have been proven to be a helpful tool for plants traits analysing in agricultural applications. Three-dimensional high throughput plant scanning platforms provide an opportunity to measure phenotypic traits which can be highly useful to plant breeders. But the measurement of phenotypic traits is still carried out with labor-intensive manual observations. Thanks to the computer vision techniques, these observations can be supported with effective and efficient plant phenotyping solutions. However, since the leaves and branches of some plant types overlap with other plants nearby after a certain period of time, it becomes challenging to obtain the phenotypical properties of a single plant. In this study, it is aimed to separate bean plants from each other by using common clustering algorithms and make them suitable for trait extractions. K-means, Hierarchical and Gaussian mixtures clustering algorithms were applied to segment overlapping beans. The experimental results show that K-means clustering is more robust and faster than the others.
- Research ArticleDeep Convolutional Generalized Classifier Neural NetworkMehmet Sarigul, B. Melis Ozyildirim, and Mutlu AvciNeural Processing Letters, Mar 2020
Up to date technological implementations of deep convolutional neural networks are at the forefront of many issues, such as autonomous device control, effective image and pattern recognition solutions. Deep neural networks generally utilize a hybrid topology of a feature extractor containing convolutional layers followed by a fully connected classifier network. The characteristic and quality of the produced features differ according to the deep learning structure. In order to get high performance, it is necessary to choose an effective topology. In this study, a novel topology based hybrid structure named as Deep Convolutional Generalized Classifier Neural Network and its learning algoritm are introduced. This novel structure allows the deep learning network to extract features with the desired characteristics. This ensures high performance classification, even for relatively small deep learning networks. This has led to many novelties such as principal feature analysis, better learning ability, one-pass learning for classifier part, new error computation and backpropagation approach for filter weights. Two experiment sets were performed to measure the performance of DC-GCNN. In the first experiment set, DC-GCNN was compared with clasical approach on 10 different datasets. DC-GCNN performed better up to 44.45% for precision, 39.69% for recall and 42.57% for F1-score. In the second experiment set, DC-GCNN’s performance was compared with alternative methods on larger datasets. Proposed structure performed better than alternative deep learning based classifier structures on CIFAR-10 and MNIST datasets with 89.12% and 99.28% accuracy values.
- Research ArticleDifferential convolutional neural networkM. Sarıgül, B.M. Ozyildirim, and M. AvciNeural Networks, Mar 2019
Convolutional neural networks with strong representation ability of deep structures have ever increasing popularity in many research areas. The main difference of Convolutional Neural Networks with respect to existing similar artificial neural networks is the inclusion of the convolutional part. This inclusion directly increases the performance of artificial neural networks. This fact has led to the development of many different convolutional models and techniques. In this work, a novel convolution technique named as Differential Convolution and updated error back-propagation algorithm is proposed. The proposed technique aims to transfer feature maps containing directional activation differences to the next layer. This implementation takes the idea of how convolved features change on the feature map into consideration. In a sense, this process adapts the mathematical differentiation operation into the convolutional process. Proposed improved back propagation algorithm also considers neighborhood activation errors. This property increases the classification performance without changing the number of filters. Four different experiment sets were performed to observe the performance and the adaptability of the differential convolution technique. In the first experiment set utilization of the differential convolution on a traditional convolutional neural network structure made a performance boost up to 55.29% for the test accuracy. In the second experiment set differential convolution adaptation raised the top1 and top5 test accuracies of AlexNet by 5.3% and 4.75% on ImageNet dataset. In the third experiment set differential convolution utilized model outperformed all compared convolutional structures. In the fourth experiment set, the Differential VGGNet model obtained by adapting proposed differential convolution technique performed 93.58% and 75.06% accuracy values for CIFAR10 and CIFAR100 datasets, respectively. The accuracy values of the Differential NIN model containing differential convolution operation were 92.44% and 72.65% for the same datasets. In these experiment sets, it was observed that the differential convolution technique outperformed both traditional convolution and other compared convolution techniques. In addition, easy adaptation of the proposed technique to different convolutional structures and its efficiency demonstrate that popular deep learning models may be improved with differential convolution.
- Research ArticleDecoupled Backstepping Sliding Mode Control of Underactuated Systems with Uncertainty: Experimental ResultsBaris Ata, and Ramazan CobanArabian Journal for Science and Engineering, Mar 2019
In this paper, a decoupled backstepping sliding mode control method is proposed to control underactuated systems under uncertainties and disturbances. The sliding mode control technique and the backstepping control technique are combined owing to their merits. Since the design methodology is based on the Lyapunov theorem, the stability of the system is guaranteed. The effectiveness of the proposed method is verified by the experimental results of the controller which is applied to a nonlinear, underactuated inverted pendulum system. The experimental results show that the decoupled backstepping sliding mode control achieves a satisfactory control performance rather than the decoupled sliding mode controller and the proposed method provides a robust performance to overcome parametric uncertainties where the decoupled sliding mode control fails. © 2019, King Fahd University of Petroleum & Minerals.
- Research ArticleEstimation of daily global solar radiation using deep learning modelKazım Kaba, Mehmet Sarıgül, Mutlu Avcı, and 1 more authorEnergy, Mar 2018
Solar radiation (SR) is an important data for various applications such as climate, energy and engineering. Because of this, determination and estimation of temporal and spatial variability of SR has critical importance in order to make plans and organizations for the present and the future. In this study, a deep learning method is employed for estimating the SR over 30 stations located in Turkey. The astronomical factor, extraterrestrial radiation and climatic variables, sunshine duration, cloud cover, minimum temperature and maximum temperature were used as input attributes and the output was obtained as SR. The datasets of 34 stations, spanning the dates from 2001 to 2007, were used for training and testing the model, respectively, and simulated values were compared with ground-truth values. The overall coefficient of determination, root mean square error and mean absolute error were calculated as 0.980, 0.78 MJm−2day−1 and 0.61 MJm−2day−1, respectively. Consequently, DL model has yielded very precise and comparable results for estimating daily global SR. These results are generally better than or they are comparable to many previous studies reported in literature, so one can conclude that the method can be a good alternative and be successfully applied to similar regions.