A two-stage Deep Convolutional Generative Adversarial Network-based data augmentation scheme for agriculture image regression tasks

Recently, Deep Learning models, such as Deep Convolutional Neural Networks (CNNs), have shown remarkable performance on various Computer Vision tasks. Unfortunately, many application domains, such as agriculture image analysis, do not have access to large datasets. In this study, we are interested in the prediction of soil moisture dissipation rates from aerial images using a CNN model. This entails a regression task where the CNN, given an aerial input image, has to predict a vector of regression values associated with the dissipation rates at different locations of the agricultural field. CNNs, however, require large datasets to successfully train the deep network. To this end, we present a Deep Convolutional Generative Adversarial Network (DCGAN) to generate fake agriculture images, that are realistic enough with the objective to augment the original dataset, for better generalization capabilities and more accurate predictions by the trained CNN. Unlike existing GAN-based augmentation approaches used for segmentation or classification, where only images have to be generated to augment the training set, the regression task requires both an image and its associated regression vector. In our approach, we use a two-stage scheme where the baseline CNN trained on the original dataset is harnessed to predict, for each image generated by the DCGAN, the regression vector that will serve as its ground truth in the augmented dataset. The results show that the CNN trained on the augmented dataset in this way allows a relative reduction of the Mean Absolute Error by a percentage of 20.8 w.r.t to the CNN trained on the original dataset. We believe that our approach can be generalized to any image augmentation task involving a regression task.