Generating Labeled Samples Based on Improved Cdcgan for Hyperspectral Data Augmentation: A Case Study of Drought Stress Identification of Strawberry Leaves

Deep learning has been increasingly adopted in analyzing the hyperspectral imaging (HSI) data, and a large amount of high-quality labeled dataset is indispensable for its superior modeling performance. However, the acquisition of large-scale annotated spectral data is very time-consuming and costly. To address this challenge, this study proposed an improved conditional Deep Convolutional Generative Adversarial Network (cDCGAN) model to generate labeled HSI samples for data augmentation. The identification of drought stressed strawberry leaves was taken as the study object. Seven small-sample training datasets were constructed with sample sizes of 6, 10, 20, 30, 40, 50, 70 and 100, respectively. Different cDCGAN architectures were tested by varying the network depth and label concatenation pattern. An Indicator of Generated Data Quality (IGDQ) was proposed to evaluate the quality of generated spectra for exploring the optimal architecture of cDCGAN. Then, on each of seven training datasets with limited samples, high-quality pseudo spectral data were generated using the proposed cDCGAN and merged to the original training datasets for data augmentation. Residual Network (ResNet) classifier was established respectively before and after data augmentation. Conventional machine learning classifiers, including Support Vector Machine (SVM) and Decision Tree (DT), were also constructed. Results showed that the accuracy of ResNet, SVM, and DT improved by an average of 6.9%, 3.4%, and 3.1%, respectively, after data augmentation. Moreover, the minimal sample size achieving effective data augmentation could be as low as 20, its augmented datasets achieved comparable or even superior accuracy than the original training dataset with 100 samples. The various aspects affecting the quality of generated spectral data were also discussed, including different model frameworks (cDCGAN and cWGAN) and architectures. The overall results demonstrated that the proposed cDCGAN model achieved satisfactory results on the small-sample datasets of drought stressed strawberry leaves. This method has great potential for the common scenario of imbalanced or small-sample datasets in the domain of plant science.