Deep learning models have been highly successful in the image-based plant phenotyping applications such as disease detection and classification. However, one of the main challenges in achieving this success is the requirement of large amount of labeled data. Data annotation could be costly, laborious, time consuming and needs domain expertise for many plant phenotyping tasks. To overcome this challenge, recently many active learning algorithms have been proposed to reduce the amount of labeling needed by deep learning models for achieving high performance.Active learning methods adaptively selects samples for annotation using an acquisition function to achieve maximum classification performance under a fixed labeling budget. In this work, we train the MobileNetV2 model developed by Sandler et al., 2018 for the classification of 8 different soybean stresses (biotic and abiotic) and healthy leaves. The dataset consists of 16,573 RGB images of healthy and stressed soybean leaves captured under controlled conditions. We empirically evaluate the performance of four different active learning methods with random sampling-based annotation for soybean stress classification.