Advanced Kiwifruit Soft Rot Detection with Deep Learning

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.May 17 2024

In an article published in the journal Nature, researchers explored the accurate detection of kiwifruit soft rot using hyperspectral imaging (HSI) and deep learning. They introduced a dual-branch selective attention capsule network (DBSACaps) to improve classification accuracy.

By separately extracting spectral and spatial features and then fusing them with an attention mechanism, the proposed method achieved superior performance compared to existing methods, with an overall accuracy (OA) of 97.08% and 97.83% accuracy for soft rot, aiding in smart agriculture.

Background

Kiwifruit, renowned for its taste and nutritional value, faces significant postharvest losses due to diseases like soft rot. Early detection is vital but challenging, as initial symptoms are subtle and rapidly progress to fruit decay. HSI offers promise, capturing both spectral and image data to reveal hidden defects.

Previous studies utilized classical machine learning for disease detection, lacking efficient feature extraction and uniform classification criteria. Recent advancements in deep learning show promise, yet many methods overlook the unique properties of HSI. Current approaches often struggle to fully leverage spatial and spectral information, limiting classification accuracy.

This paper addressed these gaps by proposing DBSACaps tailored for kiwifruit soft rot detection. The network segregated feature extraction into spectral and spatial branches to mitigate interference and incorporated a selective attention module to enhance meaningful feature contribution. By employing capsule networks, the model gained robustness and achieved superior classification accuracy compared to conventional methods.

This work aimed to bridge the gap between existing methodologies and the unique challenges of kiwifruit disease detection using HSI. By harnessing the power of deep learning and optimizing network architecture for HSI, the proposed DBSACaps offered a promising solution for early and accurate soft rot detection, potentially revolutionizing postharvest treatment and storage practices in the kiwifruit industry.

Methods and Materials 

The researchers intended to detect soft rot in kiwifruits using HSI combined with a novel deep-learning approach. The dataset comprised 'Yunhai No.1' kiwifruits inoculated with Diaporthe eres to simulate soft rot under different conditions. Images were acquired daily using a portable hyperspectral imager, and a block segmentation method was employed for analysis. The dataset was augmented to enhance diversity, resulting in over 11,000 samples.

For feature extraction, a dual-branch approach was adopted, separating spectral and spatial features to prevent mutual interference. Spectral features were extracted using a three-dimensional (3D) convolutional neural network (CNN), while spatial features were obtained via a two-dimensional (2D) CNN. A selective attention module was introduced to fuse these features, assigning different weights based on their importance for classification.

Subsequently, a capsule network was employed for classification, leveraging the unique properties of capsules to capture spatial relationships between features. The proposed network architecture, termed DBSACaps, demonstrated robustness and efficiency in soft rot detection.

The experimental evaluation compared DBSACaps with several existing methods, showcasing its superior performance in terms of classification accuracy. The network was implemented using PyTorch and trained on a workstation equipped with an Intel Xeon central processing unit (CPU) and Nvidia GeForce graphics processing unit (GPU).

The study adhered to ethical guidelines regarding research involving plants and complied with relevant policies. Kiwifruits used in the study were sourced from the Chuyang orchard in Wuhan, Hubei Province, China, in accordance with regulations governing species conservation and trade.

Analyzing Kiwifruit Soft Rot Classification Results and Methodology Efficacy

The classification results of various networks on the test set were thoroughly analyzed. The proposed DBSACaps network exhibited superior performance, achieving an OA and average accuracy (AA) of 97.08% and 97.64%, respectively, surpassing other methods by notable margins. CapsuleNet and HS-CNN also demonstrated commendable performance, attributable to their robust feature extraction capabilities tailored to the dataset.

Comparative analysis revealed that the performance of dual-branch multi-attention mechanism network (DBMA), hybrid spectral convolutional neural network (HybridSN), LMFN, and spectral partitioning residual network (SPRN) varied, with LMFN showing relatively better results, possibly influenced by dataset characteristics. Interestingly, the inclusion of the spatial attention module (HPDM) in SPRN did not yield improved accuracy, contrary to expectations, possibly due to dataset differences.

Further investigations validated the effectiveness of the proposed dual-branch feature extraction. Networks exclusively utilizing spectral or spatial branches demonstrated lower accuracy compared to the DBSACaps network, indicating the significance of feature fusion in enhancing classification outcomes.

Experimentation with reconstruction networks provided insights into their role. While DBSACapsWithRec1 suffered from parameter overload, DBSACapsWithRec2, and DBSACapsWithRec3 demonstrated marginal improvements, albeit with reduced parameters. However, none surpassed the performance of DBSACaps, suggesting that reconstruction networks may not be necessary for this specific classification task.

Conclusion

In conclusion, the researchers pioneered the integration of HSI and deep learning for kiwifruit soft rot detection. The proposed DBSACaps model, featuring dual-branch classification with attention and capsule networks, outperformed existing methods, achieving 97.08% accuracy. Results underscored the efficacy of feature fusion and capsule networks for accurate classification.

Future work will focus on diversifying sample sources and optimizing spectral band selection for broader applicability. Despite challenges like data storage, this approach holds promise for revolutionizing fruit disease detection and advancing smart agriculture practices.