Classification of plant leaf diseases through automated computers is critical in enhancing the yield of agricultural products, and most algorithms based on deep learning use computationally intensive architectures at the expense of laboratory-collected datasets, restricting their realworld use. This paper presents a proposal of HibiLite-CNN; a task-specific lightweight convolutional neural network to perform multi-class classification of hibiscus leaf diseases based on a single dataset that was captured in a natural environment. The dataset has 1,101 images in eight disease types and the processing is done with a class-balanced augmentation with avoiding data leakage. HibiLite-CNN combines depthwise separable convolutions, residual connections, and channel-wise attention to attain a successful trade-off between accuracy and efficiency. Experimental findings indicate that the proposed model can achieve a macroaveraged accuracy, precision, recall and F1-score of 0.976 using 0.23 M parameters, making it more computationallyefficient than a number of transfer learning baselines. Stable convergence and high generalization are ensured by three-fold cross-validation. Explainability with SHAP and LIME shows that the model pays attention to disease-relevant regions that increase interpretability and reliability towards practical agricultural use.