Model compression has been employed extensively recently to reduce the size of deep learning models, enabling efficient deployment in resource-constrained environments like edge devices. However, there is a trade-off between model performance and compression. In this study, we conducted several experiments to investigate the performances and weaknesses of three compression techniques: knowledge distillation, model pruning, and quantization on three different image classification tasks: pothole detection, paddy rice maturity detection, and soybean weed detection. Real-life datasets were collected and utilized solely for the purpose of this work. The experimental results reveal that model performance typically suffers after compression, which is the common notable drawback of model compression strategies. Therefore, we proposed a hybrid lossless-compressed model by combining pruning and quantization to reduce parameters and weights, resulting in a lightweight model. The proposed model is three times smaller than the vanilla CNN model. Moreover, the proposed compressed model achieved a state-of-the-art accuracy of 97% after compression, which shows the effectiveness of our approach.