Scholarlink

Purpose <p>The discrimination of grapevine cultivars is an important yet underexploredtopic in precision viticulture. This study evaluated three aspects of grapevinecultivar classification and model training strategies: (1) the influence ofseasonal variability and data collection strategy; (2) model generalisabilityacross sample types; and (3) the combined impact of temporal and sample-typevariation on classification performance.</p> Methods <p>Spectral data from leaf and canopy samples were used to classify six grapevinecultivars: Currant, Merbein, Muscat, Selma Pete, Sugra-39, and Sultana. Elevendataset configurations were evaluated using ten machine learning and deeplearning classifiers. Performance was assessed using F1 score, balancedaccuracy (BACC), Matthews Correlation Coefficient (MCC), and Area Under theReceiver Operating Characteristic curve (AUC). The Sum of Ranking Differences(SRD) method identified the most robust classifiers across training strategies.</p> Results <p>Leaf spectra collected in December—coinciding with the fruit set phenologicalstage—provided the highest classification accuracy. Models showed limitedgeneralisability across data scales (i.e., sample types), with substantialdeclines in accuracy when trained on one type and tested on another. Combiningleaf and canopy spectra in the training data improved performance but remainedlower than when models were trained and tested on the same sample type. SRDanalysis identified Support Vector Machine (SVM) and 1D Convolutional NeuralNetwork (CNN) as the most robust classifiers. SVM models achieved F1 scores of0.67–1.00, BACC of 0.74–0.98, MCC of 0.65–0.90, and AUC values of 0.87–0.94.The 1D CNN models also showed high performance (F1: 0.62–0.98; BACC: 0.79–0.99;MCC: 0.57–0.97; AUC: 0.93–1.00).</p> Conclusion <p>These findings highlight the importance of temporal and sample-typeconsiderations when developing spectral classification models for grapevinecultivars. Overall, the results provide a foundation for a scalable andcommercially viable approach to cultivar mapping, supporting more precisevineyard management.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Robust Spectral Classification Under Sample Type and Seasonal Variability: A Proximal Remote Sensing Approach for Grapevine Cultivar Discrimination

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月18日

查看全文
  • Kyle Loggenberg,
  • Albert Strever,
  • Zahn Münch

摘要

Purpose

The discrimination of grapevine cultivars is an important yet underexploredtopic in precision viticulture. This study evaluated three aspects of grapevinecultivar classification and model training strategies: (1) the influence ofseasonal variability and data collection strategy; (2) model generalisabilityacross sample types; and (3) the combined impact of temporal and sample-typevariation on classification performance.

Methods

Spectral data from leaf and canopy samples were used to classify six grapevinecultivars: Currant, Merbein, Muscat, Selma Pete, Sugra-39, and Sultana. Elevendataset configurations were evaluated using ten machine learning and deeplearning classifiers. Performance was assessed using F1 score, balancedaccuracy (BACC), Matthews Correlation Coefficient (MCC), and Area Under theReceiver Operating Characteristic curve (AUC). The Sum of Ranking Differences(SRD) method identified the most robust classifiers across training strategies.

Results

Leaf spectra collected in December—coinciding with the fruit set phenologicalstage—provided the highest classification accuracy. Models showed limitedgeneralisability across data scales (i.e., sample types), with substantialdeclines in accuracy when trained on one type and tested on another. Combiningleaf and canopy spectra in the training data improved performance but remainedlower than when models were trained and tested on the same sample type. SRDanalysis identified Support Vector Machine (SVM) and 1D Convolutional NeuralNetwork (CNN) as the most robust classifiers. SVM models achieved F1 scores of0.67–1.00, BACC of 0.74–0.98, MCC of 0.65–0.90, and AUC values of 0.87–0.94.The 1D CNN models also showed high performance (F1: 0.62–0.98; BACC: 0.79–0.99;MCC: 0.57–0.97; AUC: 0.93–1.00).

Conclusion

These findings highlight the importance of temporal and sample-typeconsiderations when developing spectral classification models for grapevinecultivars. Overall, the results provide a foundation for a scalable andcommercially viable approach to cultivar mapping, supporting more precisevineyard management.