Conventional MFCC and phase-difference MFCC interaction network for durian days-after-anthesis estimation

Information from the abstract

Precise assessment of durian maturity is important for harvest scheduling, whereas conventional methods are often destructive or dependent on human judgment. Although farmers commonly use knocking sounds in practice, systematic estimation of Days After Anthesis (DAA) from acoustic recordings remains limited. Most existing acoustic approaches for durian maturity assessment have focused on discrete ripeness classification or single-feature representations, which provide limited information for fine-grained harvest scheduling. To address this problem, this paper proposes a dual-stream acoustic regression framework, named the Conventional MFCC and Phase-Difference MFCC Interaction Network (CPDMIN), for estimating continuous durian DAA from knocking sounds. The framework is based on conventional Mel-Frequency Cepstral Coefficients (MFCC) and Phase-Difference Mel-Frequency Cepstral Coefficients (PDMFCC). The MFCC stream is extracted using the standard MFCC procedure from the original spectral representation. PDMFCC is introduced not as a replacement for conventional cepstral analysis, but as an additional phase-aware representation of the same acoustic signal. Specifically, PDMFCC is derived from a phase-difference-modulated spectral representation formed by combining spectral magnitude with adjacent-bin phase-difference information prior to mel filtering and discrete cosine transformation. By jointly modeling magnitude-related cepstral information and phase-difference-informed cepstral cues, the proposed method aims to provide a more informative acoustic representation than conventional single-stream MFCC-based approaches. The proposed CPDMIN framework employs two parallel encoders, a bidirectional interaction module, and an adaptive gated fusion module to learn and integrate the two feature streams before final DAA prediction. The proposed framework was evaluated against eight representative deep-learning baselines, including ResNet, EfficientNet, MobileNet, ConvNeXt, and vision-transformer-based architectures, under the same experimental protocol. Under the present experimental setting, CPDMIN yielded the most favorable overall results among the compared baselines, with a mean absolute error (MAE) of 2.33 ± 0.24 , a root mean square error (RMSE) of 3.82 ± 0.45 , a mean absolute percentage error (MAPE) of 2.18 ± 0.23 , and an R 2 of 0.965 ± 0.010 . These results suggest that jointly modeling MFCC and PDMFCC may be useful for non-destructive durian DAA estimation within the current dataset and acquisition conditions.

Why this record is monitored

This record has an Impact Signal of 72/100 based on recency, source, collaboration, and bibliographic signals. It prioritizes monitoring and is not a judgment of research quality.

Related topics: Smart Agriculture and AI · Plant Surface Properties and Treatments · Spectroscopy and Chemometric Analyses