Effective Methods Based on Distinct Learning Principles for the Analysis of Hyperspectral Images to Detect Black Sigatoka Disease.

Abstract

Current chemical methods used to control plant diseases cause a negative impact on the environment and increase production costs. Accurate and early detection is vital for designing effective protection strategies for crops. We evaluate advanced distributed edge intelligence techniques with distinct learning principles for early black sigatoka disease detection using hyperspectral imaging. We discuss the learning features of the techniques used, which will help researchers improve their understanding of the required data conditions and identify a method suitable for their research needs. A set of hyperspectral images of banana leaves inoculated with a conidial suspension of black sigatoka fungus ( Pseudocercospora fijiensis ) was used to train and validate machine learning models. Support vector machine (SVM), multilayer perceptron (MLP), neural networks, N-way partial least square-discriminant analysis (NPLS-DA), and partial least square-penalized logistic regression (PLS-PLR) were selected due to their high predictive power. The metrics of AUC, precision, sensitivity, prediction, and F1 were used for the models' evaluation. The experimental results show that the PLS-PLR, SVM, and MLP models allow for the successful detection of black sigatoka disease with high accuracy, which positions them as robust and highly reliable HSI classification methods for the early detection of plant disease and can be used to assess chemical and biological control of phytopathogens.

Plant phenotyping relevance

バナナ葉の病徴をハイパースペクトル画像と複数の機械学習モデルで早期検出し、モデル性能を評価する手法研究であり、植物病害状態の取得・推定が中心です。

Code and data availability

The paper's hyperspectral banana-leaf training/validation datasets and the authors' analysis source code (PLS-PLR, NPLS-DA, SVM, MLP) are explicitly stated as publicly available on the authors' GitHub repository.