The research focuses on the noninvasive identification of deficiencies known as NPK macronutrients using images of plant leaves. It involves data, divided as a balanced dataset for equitable representation across classes and an imbalanced dataset for real-world scenarios with diverse class proportions. The CNN, VGG19, and ResNet50 models are trained on both datasets. The trained models’ outputs are harmonized through a primary ensemble phase, consolidating forecasts using a weighted aggregation technique. Quantization techniques are applied to each model to optimize computational resources while preserving accuracy. The quantized model outputs converge in a secondary ensemble phase, yielding a refined output. The last phase focuses on evaluation and analysis of the models, which includes measuring and testing the performance of the ensembled and quantized models on other datasets and their generalization ability, exposing metrics of interest including accuracy, precision, recall, and F1 score. Weighted ensembling is shown to be effective in the study, \(91.54{\%}\) , showcasing its potential as a robust approach for model integration in the targeted application.