The advancement in deep learning techniques has led to significant improvements in agriculture, particularly in the detection of diseases in crops. This research paper explores the performance of three extensively used deep learning architectures, MobileNetV2, VGG16, and ResNet50, in detecting cauliflower diseases. The study focuses on the impact of various hyperparameter variations like various values of learning rate, etc., on the accuracy of these models. A comprehensive experimental evaluation was conducted by using the dataset of cauliflower images, and the findings were equated between these three models. The results showed that both MobileNetV2, ResNet50, and VGG16 have the capability to detect cauliflower diseases with high accuracy. However, the optimal hyperparameters for each model were found to be different after taking different hyperparameter values, with some hyperparameter variations having a larger impact on one model compared to the other. This highlights the importance of thorough hyperparameter tuning when designing deep learning models for disease detection in agriculture. We have found that MobileNetV2 performs better than other two models. After training, pre-trained models were tested on real-time images which were taken by our smartphones and pre-processed according to model input parameter and compared results in this research. The outcomes of this study support to the progress of more efficient and effective solutions for monitoring and controlling diseases in crops, which is crucial for ensuring food security and sustainable agriculture.