Abstract—Class imbalance remains a critical challenge in agricultural image classification, where standard Cross-Entropy (CE) training tends to favor frequent classes at the expense of rare but agronomically important disease categories. This study presents a controlled empirical comparison of four loss functions: Cross-Entropy (CE), Weighted Cross-Entropy (W-CE), Class-Balanced Loss (CB), and Focal Loss (FL) for five-class cassava leaf disease classification on the iCassava 2019 dataset (imbalance ratio = 8.41). Five Convolutional Neural Network (CNN) backbones (LeNet, ResNet-50, EfficientNet-B0, MobileNetV2, and ConvNeXt-Base) share a common Multilayer Perceptron (MLP) classifier head and follow matched data splits, augmentation policies, and optimization schedules to reduce confounding factors in the loss-function comparison. On the held-out test set (N = 1885), ConvNeXt-Base with focal loss and horizontal-flip Test-Time Augmentation (TTA) attains the highest observed Macro-F1, 89.39%, together with 91.67% accuracy, improving minority-class F1 for Cassava Bacterial Blight (CBB) (+2.57%) and Healthy (+2.38%) relative to the CE baseline while maintaining stable Cassava Mosaic Disease (CMD) performance. Class-balanced loss achieves a nearly identical result on ConvNeXt-Base (89.37% Macro-F1), whereas both W-CE and CB degrade performance on compact backbones (LeNet: −6.5 to −7.3% Macro-F1). These results suggest that explicit class weighting can make optimization less stable in low-capacity feature spaces. Within the scope of a single dataset and a single random seed, the findings indicate that the effectiveness of loss-function reweighting depends on backbone capacity: sample-level modulation through focal loss generally provides more consistent improvements than fixed class-level weighting, with the clearest gains on higher-capacity architectures.
 
Keywords—cassava leaf disease, imbalanced learning, focal loss, class-balanced loss, plant disease detection, agricultural image classification


Copyright © 2026 by the authors. This is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited (CC BY 4.0).