Abstract

This study explores the potential of a novel hyperspectral snapshot mosaic camera forweed and maize classification. The image processing, feature engineering and machinelearning techniques were discussed when developing an optimal classification model forthe three kinds of weeds and maize. A total set of 185 spectral features includingreflectance and vegetation index features was constructed. Subsequently, the principalcomponent analysis was used to reduce the redundancy of the constructed features, andthe first 5 principal components, explaining over 95% variance ratio, were kept for furtheranalysis. Furthermore, random forests as one of machine learning techniques were builtfor developing the classifier with three different combinations of features. Accuracy-oriented feature reduction was performed when choosing the optimal number of fea-tures for building the classification model. Moreover, hyperparameter tuning wasexplored for the optimal selection of random forest model. The results showed that theoptimal random forest model with 30 important spectral features can achieve a meancorrect classification rate of 1.0, 0.789, 0.691 and 0.752 forZea mays,Convolvulus arvensis,RumexandCirsium arvense, respectively. The McNemar test showed an overall betterperformance of the optimal random forest model at the 0.05 significance level comparedto the k-nearest neighbours (KNN) model.