Jackfruit (Artocarpus heterophyllus) is an economically and nutritionally significant fruit crop in Bangladesh, even though the growth and yield are influenced by foliar stress conditions, which limit productivity and quality. Jackfruit leaf health and growth stages could be accurately recognized and thus cultivated in a sustainable manner by early and reliable identification. In this study, we organized a balanced dataset of 2,500 images in five categories, namely, Dried, Healthy, Leaf Miner, Senescence, and Young, and benchmarked four ImageNet-pretrained CNN backbones, ConvNeXtBase, NASNetMobile, MobileNetV1, and InceptionV3. From single models, MobileNetV1 was the most accurate 98.97%, and NASNetMobile and InceptionV3 were both 98.45% with slightly lower recall on both Dried and Healthy classes. To address this, we proposed the NMI Ensemble, which combines NASNetMobile, MobileNetV1, and InceptionV3 using soft voting at the probability level. The ensemble significantly achieved 99.74% accuracy with almost perfect precision, recall, and F1-scores with all classes. The training-validation curves showed rapid convergence and a small generalization gap. A resolution ablation study on the input resolutions further revealed that 224×224 is the best resolution in terms of accuracy and efficiency. The findings make the NMI Ensemble a very robust and computationally feasible model to conduct jackfruit leaf health monitoring and growth stage classification and have great potential to be implemented in a real-world precision agriculture system.