Facebook Instagram Twitter RSS Feed PodBean Back to top on side

News

A helicopter with hyperspectral cameras and lidar sensors

Institute of Informatics SAS completed project HYSPED

10. 12. 2023 | 425 visits

In November 2023, the Institute of Informatics SAS (ÚI SAV) completed the HYSPED - Research on the application of artificial intelligence tools in the analysis and classification of hyperspectral sensing data. The Technical University in Zvolen (TUZVO) and VUJE, a. s., participated on a project supported by the European Regional Development Fund (ERDF) within the Operational Programme Integrated Infrastructure.

The goal of the researchers from the Institute of Informatics SAS in the HYSPED project was the preparation of a cloud infrastructure for data collection and the design of models based on artificial intelligence for the classification of tree species using hyperspectral images as part of aerial photography of overhead power lines in the locations of Visolaje, Sverepec, Závada and Počarová. During the implementation of the project, 28 flights were carried out by the project partner VUJE, a. s. Approximately 550 GB of data were obtained from each flight, which represented a total of up to 15 TB of data for further processing. The input data for the classification models were generated by the solvers from TUZVO.

The cloud infrastructure was represented by the open software cloud platform OpenStack, which provided 756 CPU cores, 11 GPU accelerators NVIDIA A100, 4 accelerators NVIDIA K20, RAM 3.9 TB and storage with a capacity of 400 TB.


 These resources were shared with users in the form of ready-made virtual machines or Docker containers. In this way, a serverless paradigm for artificial intelligence in cloud computing was secured.

Research on classification models was focused on pixel and object classification methods. Within the pixel classification, fully connected 1D forward neural networks (1D FFNN) and 1D convolutional networks (1D CNN) were applied without reducing the dimension of the feature space and, in the case of 1D CNN, also with dimension reduction based on principal component analysis (PCA). A 3D CNN with dimensionality reduction of the feature space using PCA was used for object classification. The classification accuracy ranged from 0.997 to 0.999. A 2D CNN model was also designed and implemented. The combination of a balanced dataset, extensive augmentation, a custom-designed model and an extensive learning process resulted in classification accuracies exceeding 99% for all and sparsely represented tree species. Another approach to the classification of hyperspectral data, developed within the HYSPED project, uses the proven GoogleLeNet model. The inception module of the original model was improved from 2D to 3D convolution, and parameterizable filters were introduced to reduce information loss and add wider receptive fields. The mentioned model actively participated in the Kaggle competition focused on the classification of hyperspectral images, specifically the detection of stripe rust on images of winter wheat. The presented model ranked second shared place among 27 teams.

The output of the HYSPED project of the Institute of Informatics SAS is a study of modelled dependencies in data and validated models with the support of a cloud infrastructure for the detection and classification of woody plants at the TRL 5 level.

 

Source: Institute of Informatics SAS

Photo: VUJE, a. s.

 

 

 

Related articles