Lychee (Litchi chinensis Sonn.) is a critical economic activity in tropical areas with serious challenges through foliar diseases that may lead to loss of yields of over 30-40%. Conventional visual diagnosis is usually subjective, time-consuming, and fallible. To overcome this, we introduce LycheeCare-Net, a lightweight deep learning ensemble model, which is suitable for the accurate and efficient classification of the diseases of the lychee leaves. We compared the benchmarks on various lightweight architectures with the new and popular lightweight Lychee Leaf Disease Dataset (2025): the architectures of EfficientNet-B0, MobileNetV2, and SqueezeNet. The results of the experiments show that LycheeCare-Net reaches the highest validation accuracy of 96.10 and macro F1-score of 0.959, which is considerably higher than that of single-model baselines. In particular, the ensemble decreases biased understanding by about 31 percent relative to the supreme single model (EfficientNet-B0), efficiently clearing up confusion amid the biologically comparable classes, e.g., Fungal Leaf Spot and Bituminous Leaf. Moreover, we used Post-Training Static Quantization (INT8), which minimized the model size to less than 20 MB and allowed us to make inferences with low latency that can be deployed on low-resource mobile devices. This research sets a new standard in the existing performance of lychee disease classification, providing a useful precision farming instrument of high accuracy in the sustainable use of the technology and preventing disease in time.