Browsing by Author "Heipke, Christian"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    Detection of Forest Fire Consequences on Satellite Images Using a Neural Network
    (German Society for Photogrammetry, Remote Sensing and Geoinformation, 2023) Hnatushenko, Viktoriia V.; Hnatushenko, Volodymyr V.; Kashtan, Vita; Heipke, Christian
    ENG: The objective of this research is the detection of burnt forest areas from Sentinel-2 imagery. The proposed algorithm uses an approach based on convolutional neural networks (CNN). The functionality of the created system allows solving the task, starting from the moment of receiving the input data, image preprocessing and ending with the export of a hot-spot fire polygonal file describing the area that was burnt. These results are compared to methods based on the dNBR and a variant of BAIS2 called dBAIS2, which are generated from measurements in the near and middle IR channels of the Sentinel images. The proposed algorithm was tested on Sentinel satellite images acquired from June to September 2021for the Tizi Ouzou region, Algeria. We found it to have an overall accuracy of 97%, outperforming the results obtained from dNBR and dBAIS2 by large margins.
  • Thumbnail Image
    Item
    Enhancing the Quality of CNN-Based Burned Area Detection in Satellite Imagery through Data Augmentation
    (Copernicus GmbH (Copernicus Publications) on behalf of the International Society of Photogrammetry and Remote Sensing, 2023) Hnatushenko, Viktoriia V.; Hnatushenko, Volodymyr V.; Soldatenko, Dmytro V.; Heipke, Christian
    ENG: This study aims to enhance the quality of detecting burned areas in satellite imagery using deep learning by optimizing the training dataset volume through the application of various augmentation methods. The study analyzes the impact of image flipping, rotation, and noise addition on the overall accuracy for different classes of burned areas in a forest: fire, burned, smoke and background. Results demonstrate that while single augmentation techniques such as flipping and rotation alone did not result in significant improvements, a combined approach and the addition of noise resulted in an enhancement of the classification accuracy. Moreover, the study shows that augmenting the dataset through the use of multiple augmentation methods concurrently, resulting in a fivefold increase in input data, also enhanced the recognition accuracy. The study also highlights the need for further research in developing more efficient CNN models and in experimenting with additional augmentation methods to improve the accuracy of burned area detection, which would benefit environmental protection and emergency response services.
  • Thumbnail Image
    Item
    Machine Learning Approaches for Evaluating Forest Fire Impacts on Sentinel-2 Satellite Imagery Across Ukraine
    (WUST Publishing House, Wrocław, 2024) Hnatushenko, Viktoriia V.; Hnatushenko, Volodymyr V.; Soldatenko, Dmytro V.; Heipke, Christian
    ENG: Forest fires have long-term consequences and serious ecological, social, and economic implications. Utilizing multispectral imagery from the Sentinel-2 satellite, we propose an algorithm based on machine learning models for the detection of burnt forest areas. A new dataset on forest fires has been created, suitable for semantic segmentation models. The proposed algorithm uses an approach based on convolutional neural networks (CNN). The results are analyzed and compared in terms of the intersection over union (IoU) score. The proposed algorithm was tested on Sentinel satellite images acquired in October 2022 for the Kinburn Peninsula, Ukraine, to have an accuracy in terms of IoU of 95%.