The paper describes the cloud service of ISDCT SB RAS for machine learning research. The introduction discusses the relevance of creating a service. Further, the theoretical part is considered, which describes the component of the services and the principle of their interaction. Then the results of practical application and discussion are presented. In the conclusion, the results of the work are summarized.
Researchers often choose between existing methods and the development of new ones for solving
practical problems [
In our work, we use machine learning methods to classify the land cover with multispectral remote sensing images. On the remote sensing data the spectral characteristics can significantly differ from each other, since they depend on many factors so the same algorithms can give different results. Therefore, it is important to be able to apply the method and get the expected result on a custom dataset.
We tried to repeat the classification method used in work [
A cloud service for machine learning was created at the ISDCT SB RAS as part of an applied digital platform. The goal is to provide technical and software base for the development of new and testing of known methods. The service takes into account factors such as speed of deployment, customization flexibility, user preferences and scalability.
In the process of the service developing, the existing approaches were studied [
Description and role of the main components of the Geoportal ISDCT cloud service [
Figure 1 shows a general diagram of the cloud service components interaction.
The cloud service is focused on the development, implementation and testing of remote sensing methods. Distinctive features of the service from existing analogues are: • unlimited working time; • increased data storage; • interaction of an interactive map with the development environment; • the ability to run custom Docker images.
In the course of practical experiments, the algorithms for the land cover classification from [
On the technical side, the method is demanding on computational resources, so two Docker images were created: a regular one for writing algorithms and a high-performance one with support for CUDA technology, which is necessary for neural networks training. Switching between images occurs with KubeSpawner, a plug-in for Kubernetes integration in JupyterHub. KubeSpawner allows determine the number of processor cores, the amount of RAM, access to video cards and other parameters. Next, Kubernetes looks for a suitable node to run the image. The resulting images are used to solve machine learning tasks. Next, we consider the developed and implemented algorithms for solve the machine learning tasks.
Training sample balancing algorithm. The algorithm input is the path to the data divided by classes in various directories. Each directory contains N files - containers for training data. First, a list of pairs is formed (file name, sample serial number). Next, a list of characteristics for clustering is built (the arithmetic mean and standard deviation of the sample pixels for each band). Clustering is performed based on the list of characteristics. The result is ordered in ascending order of the number of elements in the cluster. After that, samples are taken from each cluster according to the rule – if number of cluster elements are less than a specified threshold value, then we take everything, otherwise with a certain step. This approach guarantees that rare samples will definitely be included in the training set, and frequent ones will be thinned out. Figure 2-3 shows results of the algorithm work.
Classification quality control algorithm. The algorithm input is two files, the result of the classification and the labeled dataset. Since it is not possible for a specialist to completely label the whole image, the check is carried out only in the areas corresponding to the markup. The PyCM library is used to calculate statistics. PyCM is a multi-class confusion matrix library written in Python that supports both input data vectors and direct matrix, and a proper tool for post-classification model evaluation that supports most classes and overall statistics parameters. ACC (Accuracy), PPV (Precision or positive predictive value) indicators were used as the main criteria. As a result of the check, it became possible to compare two versions of the method and choose the best one. The error matrix for each version of the method is shown in Figure 4.
The result of applying adapted method from work [
The purpose of this study is to create a cloud service for the ISDCT SB RAS Geoportal for machine learning tasks. We took into account the needs of users and the features of the tasks to be solved during the design stage. This service allows users to apply data processing methods to solve practical problems, develop and implement new methods. Unlike existing services, the proposed one has the following advantages: unlimited working time; increased data storage; connection of the interactive map with the development environment; the ability to run custom Docker images. During testing, we gained experience in adapting methods for specific tasks, taking into account the specifics of the processed data, repeating the method close to the original in the software part and partly in the technical one in the shortest possible time.
The results were obtained within the framework of the State Assignment of the Ministry of Education and Science of the Russian Federation for the project "Methods and technologies of cloudbased service-oriented platform for collecting, storing and processing large volumes of multi-format interdisciplinary data and knowledge based upon the use of artificial intelligence, model-guided approach and machine learning" (state registration number 121030500071-2). Results are achieved using the Centre of collective usage «Integrated information network of Irkutsk scientific educational complex».