The detection of water-stressed plants have a significant impact in the agricultural field providing water conservation and improved crop yield. Deep Learning (DL) models can play a crucial role in this task based on the CNN models for image classification. A challenge is to build a dataset containing water-stressed plant images and non water-stressed plant images in order to train the classification model while the absence of such an open source dataset. In this paper, we compare between a two augmentation techniques: The first is based on Latent Diffusion Models (LDMs), and the second is based on a novel architecture entitled NST-SAM exploiting Segment Anything Model (SAM) and Neural Style Transfer (NST). In our comparative approach, we have created 4 datasets: The first composed of real plant images collected from the internet, the second and third are an augmentation of the first using respectively LDMs and NST-SAM, and the last with 1250 images containing the real and all augmentation images (LDMs and NST-SAM). The classification metrics indicates that when associating the LDMs and NST-SAM we augment the performances of the model presenting an accuracy of 84% using VGG16 and 85% with ResNet50.