Accuracy, Stability, and Computational Efficiency of Deep CNN Architectures for Apple Leaf Disease Classification

Abstract

Reliable apple leaf disease classification requires high predictive performance and transparent computational reporting under heterogeneous field conditions. This study presents an orchard-image benchmark of four ImageNet-initialized convolutional architectures—Xception, EfficientNetB0, DenseNet121, and ResNet50V2—using images acquired at two Moroccan orchard locations (Midelt and Azrou), four cultivars, and three mobile devices. From a curated four-class corpus of 20,000 RGB images, 18,000 images (4,500 per class) formed the five-fold cross-validation benchmark; 2,000 images were isolated before fold construction as an untouched challenge reserve for future independent confirmation and were not evaluated in the present manuscript. All models used a common 224×224 input resolution, a harmonized classification head, Adam optimization, L1/L2 regularization, learning-rate reduction, and validation-loss early stopping. Evaluation combined accuracy, cross-entropy loss, macro-precision, macro-recall, macro-F1, specificity, Cohen’s kappa, Matthews correlation coefficient, fold stability, parameter count, model size, GMACs/GFLOPs, and measured CPU training time. All reported values and figures derive from completed HPC-MARWAN runs under the common protocol. Xception achieved the highest observed mean validation accuracy (97.99±0.37%) and aggregate out-of-fold macro-F1 (97.99%), whereas EfficientNetB0 reached 97.33±0.68% mean validation accuracy with 0.78 GFLOPs per image, 5.29 million parameters, and the shortest CPU training time. The results define a descriptive accuracy–cost trade-off within the benchmark; no inferential superiority or device-level deployment latency is claimed.