Session: ASME Undergraduate Student Design Expo

Paper Number: 173916

Comparison of Traditional and Gan-Based Data Augmentation Techniques for Deep Learning Classification of Lettuce Nutrient Stress 

Data scarcity is a significant barrier to the application of AI in non-data heavy contexts. One such context is the budding controlled environment agriculture (CEA) industry, where plants such as butter lettuce (L. sativa) are grown under specific conditions within tech-laden environments. AI driven automatic monitoring of factors such as nutrient stress has the potential to further streamline the specific growth processes, but very little public data is available for the training of AI models. In this research three different approaches are used to create more data for model training, those being conventional data augmentation, a feature conversion generative adversarial network (GAN), and a dual discriminator GAN. Various pretrained deep learning models are then used on the created datasets to classify nutrient stress.

 The dataset used in this research contains images of various nutrient deficiencies in L. sativa. It consists of four classes: nitrogen deficient (-N), phosphorus deficient (-P), potassium deficient (-K), and healthy (FN). Before any data augmentation the dataset has only 208 images distributed unevenly between the classes, with 58 N- images, 66 P- images, 72 K- images, and 12 FN images. To be an effective training set for deep learning networks, the content of the dataset must be expanded. 

For conventional data augmentation rotation, shear, scale, and gaussian noise were randomly applied to the images of each class, expanding the dataset to a total of 800 images, with 200 images per class. This 800-image dataset then served as the base dataset. For the second method, a feature conversion GAN was used to add an additional 416 images to each class. The feature conversion GAN takes two image classes at a time and, using the GAN discriminator-generator structure, generates progressively better fake images for each of the input classes based on images of the opposite class. The model used for this method is largely based on the model proposed by Lin et. al (2025). For the third method, a dual discriminator GAN was also used to add 416 images to each class. This method involves both a regular and high frequency discriminator to capture more information from the generated images, leading to more accurate fake images. The FWDH model proposed by Wang et. al (2025), a type of dual discriminator, was used as the basis of this model. 

After the various data augmentations are complete three deep learning models, ResNet50, VGG19, and Inception V3, all available through the TensorFlow python package pretrained on the ImageNet dataset, are to be used to classify the images in each of the three created datasets. These models were chosen both because they are available pretrained, but also because they vary in structure and performance. The VGG19 has a simple structure but a very high computational cost, with almost 6 times as many parameters compared to the other models. The ResNet50 and Inception V3 models both employ modern optimization techniques to achieve high accuracy at lower cost, but each also include a unique computation feature, with the ResNet50 model using gradient skip connections and the Inception V3 model using multi-sized kernels to extract various sized features in parallel. The classification accuracy of each model applied to each generated dataset will be compared. Using the 800-image dataset as the baseline for model accuracy and based on available literature, an increase in classification accuracy is expected for both novel augmentation approaches.

In this research, for one of the first times in the literature, novel data augmentation techniques and common deep learning models are used to expand the limited public L. sativa data for the classification nutrient stresses, with different approaches compared to identify the best methods. As L. sativa is one of the most grown crops in CEA facilities, the findings of this research could be used to automate nutrient stress detection, enable more efficient resource use, enhance yields, and overall streamline the production process.

Presenting Author: Luke Baxter Boston College

Presenting Author Biography: Luke Baxter is an undergraduate student at Boston College in the Departments of Engineering and Math. He has previously been involved with biochemical engineering research in the Brace Lab at Boston college, and was part of the 2025 NSF REU for controlled environment agriculture at University of Wyoming.

Authors: