Robust eggplant disease recognition using a learnable weighted deep ensemble with test-time augmentation

Abstract

One of the most extensively cultivated vegetables worldwide, eggplant is significantly affected by a wide range of diseases that reduce both yield and quality. Ensuring sustainable crop production and food security therefore requires early and reliable disease diagnosis. The rapid evolution of deep learning architectures and computer vision techniques has established convolutional neural networks as powerful tools for plant disease recognition, often outperforming traditional diagnostic approaches. In this study, three benchmark eggplant image datasets with distinct class structures and imbalance characteristics (6-class Eggplant1, 5-class Eggplant2, and 7-class Eggplant3) were utilized to develop a robust disease classification framework. The proposed methodology introduces a learnable weighted ensemble that adaptively integrates ConvNeXt, DenseNet, and EfficientNet architectures within an end-to-end trainable fusion scheme. The framework is further reinforced by systematic test-time augmentation and cross-validation to enhance inference stability. Experimental evaluation demonstrates that the ensemble model achieves accuracies of 0.9970, 0.9375, and 0.9557 on Eggplant1, Eggplant2, and Eggplant3, respectively, while maintaining balanced performance across heterogeneous class distributions. These results confirm the effectiveness of the adaptive ensemble fusion in capturing complementary feature representations and mitigating inter-class variability across datasets with differing levels of complexity. Overall, this work contributes a methodologically robust and interpretable ensemble-based framework that advances more dependable image-based plant disease diagnosis by addressing practical reliability concerns in precision agriculture applications.