The vulnerability of permafrost carbon pool : the investigation of abrupt thaw features in lowland settings

Abstract

Size of digital information have been increasing with developing technologies in last two decades. Today, the size of data that people obtain by using digital tools is quite high. This data can be categorized to different branches: Text, sound, image etc. Although it was quite difficult to process the obtained data manually at the past, it becomes almost impossible at the moment. For these reasons, it is necessary to use automated systems during the processing of data. Artificial intelligence systems are important automated tools for processing data and extracting meaning from it. Extraction of the information in images by performing different operations on images is called image processing. Processing can be done on a single image or on video images. Video images contain time information in addition to single image. Based on the time information, it is possible to have knowledge about the motion information of the objects in the images. Understanding of a movement, motion in images by using some image processing operators is called activity recognition in computer vision field. The operators used during activity recognition differ according to the methods used. In addition to traditional feature extraction methods, Convolutional Neural Networks are frequently used recently. Convolutional Neural Networks are very effective for solving problems such as classification of images, tracking objects and activity recognition in computer vision. No manual operation is performed during the feature extraction process. The whole process is automated by the CNN. However, the number of operations performed during the process of CNN is quite high. In addition, the training required for these systems to give proper results requires a high number of labeled data. Data processing can be performed on raw data as well as on compressed data. The most successful examples of data compression methods can be found in video compression techniques. As a result of video compression processes, it is ensured that the repetitive information in the videos is extracted and removed from the video. Thus, as a result of the compression process, only necessary and simple information is obtained. Raw videos and compressed videos can be used for the activity recognition problem. Activity recognition on data obtained from compressed videos is faster and more effective. In the light of the information and methods mentioned above, a study was carried out on the comparison of activation functions and neural network types on the compressed domain activity recognition system. According to this study, a low cost but effective neural network was sought instead of the backbone neural network used on the system. Trainings were carried out with different neural networks. When the results were examined, the advantages and weaknesses of the neural networks relative to each other were seen. ResNeXt neural network's number of parameter - training success data was sufficient according to the backbone neural network ResNet. A different study was carried out by expanding the analysis studies on activation functions. During the analysis, trainings were conducted with 5 different activation functions and the results were reported. Among the activation functions, those belonging to the ReLU family of functions showed a more successful result.