MixChannel: Advanced Augmentation for Multispectral Satellite Images

Usage of multispectral satellite imaging data opens vast possibilities for monitoring and quantitatively assessing properties or objects interest on a global scale. Machine learning computer vision (CV) approaches show themselves as promising tools automatizing image analysis. However, there are limitations in using CV data. Mainly, the crucial one is amount available model training. This paper presents novel augmentation approach called MixChannel that helps to address this limitation improve accuracy solving segmentation classification tasks with images. The core idea utilize fact usually more than each location remote sensing tasks, extra can be mixed achieve robust performance trained models. proposed substitutes some channels original training from other images exact mix auxiliary technique preserves spatial features adds natural color variability probability. We also an efficient algorithm tune channel substitution probabilities. report method provides noticeable increase all considered models studied forest types problem.