RACEGNet for Adaptive Reliability Aware Expert Fusion in Durian Disease Classification

Abstract Plant disease recognition under real-field scenarios is still a challenging task because of high variations in terms of symptom presentation, affected plant organs, lighting, background, and severity of diseases. Existing deep learning methods typically utilize single feature representation or fixed ensemble strategies that limits their ability for adaptation to heterogeneous disease manifestations. In addition, existing durian disease classification studies have mainly focused on organ-specific recognition, such as leaf or trunk disease classification, whereas real-field disease recognition requires handling diverse manifestations across different plant organs and visual conditions. These limitations highlight the need for an adaptive framework that can exploit complementary representations from heterogeneous architectures while dynamically determining their relevance for each input sample. Motivated by these challenges, This study proposes Reliability-Aware Cross-Expert Gating Network (RACEGNet) for durian disease classification. RACEGNet is a heterogeneous expert fusion framework that integrates four architecturally distinct pretrained experts including ResNet-50, ConvNeXt-Tiny, MaxViT-T, and DeiT-Small/16 and consists of three key modules including Cross-Expert Representation Learning (CERL), Reliability Estimation Module (REM), and Adaptive Top-k routing. It offers an adaptive recognition approach through the combination of heterogeneous CNN and Transformer-based representations and the dynamic adjustment of expert contributions based on the reliability of each sample rather than a single model or a fixed fusion method. This framework was evaluated on a real field durian disease dataset consisting of ten classes where individual disease classes include symptoms manifested on multiple plant organs, such as leaves, trunks, and fruits etc. The experimental results show that RACEGNet obtained 96.30% test accuracy and 96.34% Macro F1-scores, which outperformed several individual expert models and traditional ensembling techniques. An ablation study proves the effect of cross-expert contextualization, reliability estimation, and adaptive routing on improving classification results. The statistical analysis confirmed that RACEGNet significantly outperformed its constituent expert models. Since the framework is designed to address the challenges of visual heterogeneity in plant disease recognition, including variations in symptom appearance, affected organs, and field conditions, RACEGNet has potential for extension to other crops. The findings demonstrate that collaboration of heterogeneous DL experts based on reliability is a suitable solution for addressing visual diversity challenges of agriculture disease recognition.