Перегляд за Автор "Okhrimenko, Anton"

Зараз показуємо 1 - 4 з 4
  • Ескіз
    ДокументВідкритий доступ
    Generative adversarial network augmentation for solving the training data imbalance problem in crop classification
    (2023) Shumilo, Leonid; Okhrimenko, Anton; Kussul, Nataliia; Drozd, Sofiia; Shkalikov, Oleh
    Deep learning models offer great potential for advancing land monitoring using satellite data. However, they face challenges due to imbalanced real-world data distributions of land cover and crop types, hindering scalability and transferability. This letter presents a novel data augmentation method employing Generative Adversarial Neural Networks (GANs) with pixel-to-pixel transformation (pix2pix). This approach generates realistic synthetic satellite images with artificial ground truth masks, even for rare crop class distributions. It enables the creation of additional minority class samples, enhancing control over training data balance and outperforming traditional augmentation methods. Implementing this method improved the overall map accuracy (OA) and intersection over union (IoU) by 1.5% and 2.1%, while average crop type classes’ user accuracy (UA) and producer accuracies (PA), as well as IoU, were improved by 11.2%, 6.4% and 10.2%.
  • Ескіз
    ДокументВідкритий доступ
    KNN-Based Algorithm of Hard Case Detection in Datasets for Classification
    (2023) Okhrimenko, Anton; Kussul, Nataliia
    The machine learning models for classification are designed to find the best way to separate two or more classes. In case of class overlapping, there is no possible way to clearly separate such data. Any ML algorithm will fail to correctly classify a certain set of datapoints, which are surrounded by a significant number of another class data points at the feature space. However, being able to find such hardcases in a dataset allows using another set of rules than for normal data samples. In this work, we introduce a KNN-based detection algorithm of data points and subspaces for which the classification decision is ambiguous. The algorithm described in details along with demonstration on artificially generated dataset. Also, the possible usecases are discussed, including dataset quality assessment, custom ensemble strategy and data sampling modifications. The proposed algorithm can be used during full cycle of machine learning model developing, from forming train dataset to real case model inference.
  • Ескіз
    ДокументВідкритий доступ
    Unbalanced Datasets Management for the Problem of Segmentation of Satellite Images
    (IEEE, 2024-10) Okhrimenko, Anton; Kussul, Nataliia
    Unbalanced datasets pose significant challenges in the segmentation of satellite images for crop classification tasks. This paper proposes a method to mitigate the bias of models trained on such datasets without the need for additional data collection. The approach involves using spatial weight masks to modify the loss function during model training, assigning higher or lower importance to pixels based on their reliability. The reliability of pixels is determined using algorithms like K-Nearest Neighbors (KNN), and the weight masks undergo various transformations, including morphological operations and Gaussian blur, to refine their texture and ensure smoother transitions between weight factors. Experiments conducted on a dataset of satellite imagery from the Kyiv region demonstrate notable improvements in overall accuracy and Intersection over Union (IoU) when using weight masks that favor normal pixels and apply morphological dilation and Gaussian blur. The proposed method proves particularly beneficial for underrepresented crop types. Additionally, the paper explores the incorporation of synthetic data generated by a Generative Adversarial Network (GAN) alongside real data, revealing slight improvements in recognizing less common crops. Comparative analysis shows that using weight masks offers similar accuracy gains to GAN-based augmentation while being more costeffective by eliminating the need for training an intermediary model and generating additional data. The results suggest that the proposed method of using spatial weight masks is a viable and efficient approach to managing unbalanced datasets in satellite image segmentation for crop classification tasks.
  • Ескіз
    ДокументВідкритий доступ
    Сrop classification synthetic training data generation with use of generative adversarial network
    (Leaving Planet Symposium, 2022) Kussul, Nataliia; Okhrimenko, Anton; Shkalikov, Oleh; Shumilo, Leonid