Oversampling Versus Variational Autoencoders: Employing Synthetic Data for Detection of Heracleum Sosnowskyi in Satellite Images

Abstract

Detection of growth areas of hazardous invasive plants, such as Heracleum Sosnowskyi (HS), in satellite images, is an important application of machine learning and computer vision methods. There exists extensive literature on the problem of crop classification in images. However, sometimes, the hardest part is to gather qualitative labels and gather enough data to train models. Notably, this difficulty arises in the analysis of satellite images, where labeling the data is hard to perform manually. In this work, we’ve faced the same problem (lack of data) when trying to build a classification model for HS detection in images. The issue of lack of data can be solved by generating synthetic data using simple methods like oversampling (OS) to more complex techniques like Variational Autoencoders (VAE) and Generative Adversarial Networks. To the best of our knowledge, there exists no work that has compared the performance of using OS versus VAE to generate synthetic pixels to overcome the problem of data deficiency for the task of HS classification in images. Accordingly, in this work, we perform this comparison and present the evaluation results on our dataset.