Semi-supervised Generative Adversarial Network for Identifying Multispectral crop images

Keywords: Generative Adversarial Network, Semi-supervised learning, Wasserstein distance, Farmland surveying
Abstract Type: Poster Abstract

Abstract

Current farmland surveying relies on labors and cause the waste of time. Many survey results become outdated before completion. This study focuses on how to catalyze the process of farmland surveying, ensures the accuracy of result as well. In recent years, deep learning has achieved several milestones in image recognition. Among several categories of algorithms, supervised learning-based Convolutional Neural Network (CNN) is a common approach to achieve image classification. However, supervised learning usually requires tons of labeled data marked by human, some data needs expertise to identify, which make data preparation quite inefficient.
In view of this situation, this study adapts the approach of semi-supervised learning-based Generative Adversarial Network (GAN), known as Semi-supervised GAN (SGAN), to classify crop images of mobile phones, and modify both the network structure and loss function based on the result proposed in Wasserstein GAN (WGAN), proposing a stable semi-supervised learning algorithm named Semi-Supervised Wasserstein GAN (SWGAN). By applying the proposed model to crop image classification, we can merge unlabeled data into training set, expanding the data distribution of training set, and taking the advantage of modification with WGAN, we are able to stabilize the training process of the model. On the other hand, this study will perform a benchmark test on the discriminator of SWGAN by setting up a CNN which has identical structure as discriminator. We can infer that if adversarial training really benefits the model performance, then make use of other fields.