Classification of Harvesting Age of Mango Based on NIR Spectra Using Machine Learning Algorithms

The established assessment of post-harvest attributes, such as the age of harvesting day,
requires destructive sampling that the availability of fruit of trees can often limit and is
expensive. In contrast, non-destructive post-harvest attribute assessment utilizing the
NIR data spectrum is fast and reliable, especially for mango. However, NIR spectral
data frequently produce non-linearity with the reference dataset used. Therefore, this
study conducted research on using NIR spectral data to classify the harvesting age of
mango fruits using machine learning algorithms. A total of five supervised machine
learning algorithms were explored to generate the classification model, including
gradient boost (GB), k-nearest neighbor (k-NN), decision tree (DT), random forest
(RF), and linear discriminant analysis (LDA). In this study, 237 NIR spectral data from
mango fruits with Arumanis cultivars from orchard sites in the Garut district, West Java
Province (Indonesia) were measured to determine the appropriate harvest time using
NIR spectra 1000 to 2500 nm. The data sets were randomly divided into training and
testing datasets, 80% and 20%, respectively. Hyperparameter optimization was
performed using the GridSearchCV function from scikit-learn by observing the
evaluation of the confusion matrix. Generally, all machine learning algorithms can
show performance in classifying the harvest age of mango fruit based on NIR spectra
data. Based on the accuracy evaluation matrix, the best machine learning algorithm
arranged to classify the age of mango fruit harvest is DT>GB>LDA>RF>k-NN. Finally,
predictions generated using the DT algorithm from more established machine learning
algorithms as a training and testing set consistently yielded higher prediction accuracy
in classification models. This study provides a framework for understanding the
feasibility of machine learning algorithms on NIR data spectral to the accuracy of
classification prediction of the harvesting age of mango. In addition, this study presents
the importance of assessing the performance of the classification model using confusion
metrics.