RFAConv-KAN: A novel deep learning approach combining receptive-field attention convolution and Kolmogorov-Arnold Network for classifying durian ripeness from knocking sounds

The classification of durian maturity from acoustic signals has gained increasing attention due to its non-invasive and effective evaluation capabilities. Convolutional Neural Networks (CNNs) are commonly employed for this task because of their ability to extract complex audio features. However, conventional CNNs are limited by fixed receptive fields and a lack of attention mechanisms, which may hinder their performance when confronted with the inherent variability of knocking sounds. Similarly, multilayer perceptron (MLP)-based architectures often struggle to capture localized dependencies in audio signals effectively. To address these challenges, we introduce RFAConv-KAN, an innovative approach that integrates Receptive-Field Attention Convolution (RFAConv) with the Kolmogorov–Arnold Network (KAN). The RFAConv module adaptively adjusts its receptive fields to capture multi-scale, context-sensitive features, while KAN efficiently models nonlinear relationships within the data. Our methodology was evaluated on knocking-sound recordings from 52 Monthong durian specimens collected in an orchard in Chanthaburi Province using a handheld tape recorder. In a five-fold cross-validation experiment, RFAConv-KAN achieved 96.16% accuracy and 96.14% F 1 score, surpassing the strongest CNN baseline and outperforming the best traditional ML baseline (SVM) while using only 2.55 M parameters and approximately 5.98 M FLOPs. These results underscore the effectiveness and efficiency of RFAConv-KAN for non-invasive durian maturity classification based on knocking sounds. • Introduces RFAConv-KAN, a unified model combining receptive-field attention and Kolmogorov–Arnold networks. • Uses adaptive receptive fields and learnable nonlinear functions to capture subtle durian knocking acoustics. • Achieves state-of-the-art accuracy on our Monthong durian dataset for non-invasive maturity assessment.