The development of science and its application in many practical fields are based on the artificial intelligence usage. Data engineering, computer analysis, fuzzy logic, machine perception, data mining and other are its main areas of interest. The methodological basis of the artificial intelligence methods are mathematical methods and intellectual models that are in constant development. Mathematical models and methods that are based on a large amount of data and incorporate them through deep learning methods are particularly active develop. Among the most important practical applications that use artificial intelligence based on deep learning using large amounts of data are studies of Earth's seismic activity, environmental monitoring, anomaly detection in cybersecurity, geospatial data analysis, and more.
One of the important aspects that has led to the mathematical deep learning methods development is the emergence of large volumes of available data and the computational power. Due to this, the field of big volume of multidimensional noise data analysis has been rapidly evolving in recent years. However, usage of the deep learning methods as they were designed for other well-known tasks is impossible for the task of big volume of multidimensional noise data analysis. Available publications and experiments on the deep learning usage in this field were conducted only for a small amount of data with uniform distribution in the input space, which is only a partial case that cannot be extended to the large volume of data with uneven distribution in the input space. Despite the great similarity between the mathematical formulation of the multidimensional noisy data analysis task and the formulation for traditional computer analysis and machine perception tasks, there are fundamental differences between them.
The most significant and successful results in the field of artificial intelligence with deep learning were obtained in the works of foreign authors Kunihiko Fukushima, Yann LeCun, Yoshua Bengio and Geoffrey Hinton, as well as in the works of Ukrainian authors O.G. Ivakhnenko, M.S. Zgurovsky, I.V. Sergienko, N.D. Pankratova, O.A. Pavlov, N.M. Kussul and others. However, the progress in computational resources and the emergence of the free accessed large volumes of multidimensional data make it actual to develop mathematical methods and models of artificial intelligence with deep learning to analyze large volumes of multidimensional noisy data in geospatial analysis and environmental monitoring.
The aim of the thesis is to develop and improve mathematical methods of deep learning, which are based on convolutional deep neural networks and differ in the initialization of the initial weights of networks with unlabeled data utilization based on sparse coding, which leads to improving accuracy in geospatial analysis tasks.
In the dissertation the following new scientific results were first received:
1. For the first time a mathematical method for unifying multidimensional noisy geospatial data has been developed based on the sparse input unlabeled data encoding that provides the possibility to build a single classification model for large volumes of input data, that allows to obtain higher classification accuracy.
2. Deep learning method based on convolutional neural networks has been improved, which, unlike the existing ones, not randomly initialize initial weights, but learn features extraction from the large volumes of unlabeled multidimensional noisy time series of data and provide significant enhancements of the overall classification accuracy.
3. For the first time filtration method for obtained classification maps of geospatial data has been developed for increasing their accuracy, based on object approach, in contrast to commonly accepted methods based on the sliding window principle, that allows to save the shape of objects on the map.
4. Proposed methods for noisy multidimensional geospatial data classifying have had further development through the implementation as a workflow using Amazon cloud based platform, which reduced data processing time through efficient data access and parallelization.
