Data augmentation for improving few-shot learning on ResNet50

Abstract

With the onset of rapid climate change and declining biodiversity, forest recovery management is becoming increasingly important. Tree inventory keeping and species identification are two necessary aspects to this, which can be very labour intensive. To alleviate this, a way to automate these tasks using machine learning models can be very helpful. However, due to the nature of how tree data is captured, there is very little usable data to train an image learning model, leading to very low classification accuracy. Because of this, finding a few-shot learning method that can yield high accuracies is of utmost importance. Data augmentation is a machine learning technique that can help in cases where data is scarce. In the case of image-based learning, small modifications would be made to the image, presenting a wider array of variations of the data during training and can help the model to generalise better.