Hyperspectral selection based algorithm for plant classification

Abstract

Abstract-The popularity of using hyperspectral imaging systems in studying and monitoring plant properties and conditions has increased lately. This increase has been driven by both financial and environmental advantages of such systems. Using a nondestructive hyperspectral imaging system improves the breeding process, increases profit, and reduces the usage of herbicide, thus reducing side effects to plants and environment. This paper is concerned with the use of hyperspectral image analysis for differentiating different plant species as well as their conditions. The main contribution of the work lies in the use of feature selection for choosing relevant, discriminant spectral information as the input to the classifier (e.g. SVM), as compared to the use of empirical spectral indices. Two independent hyperspectral datasets, captured by different instrumentations, were used in evaluation. Experimental results show significant improvements in classification accuracy with several feature selection algorithms compared to with the spectral vegetation and disease indices. The study shows that systematically selection of wavelength features can shed light on attributes that differentiate plants and their conditions.

Spectral-texture approach to hyperspectral image analysis for plant classification with SVMs

“…It should be stated that no comparison between the selection algorithm and existing empirical indices (i.e. vegetation and disease) in the literature was performed since this comparison was already demonstrated in our previous work [13]. In addition, 50% of the samples were used for training with 10-fold cross validation and the remaining samples for testing in the conventional SVM classifier, while 60% of the normal samples were used for training and the remaining 40% and all the abnormal samples were used for testing and validation.…”

“…A good review of feature selection processes, models, and algorithms can be found in [11]- [13]. In short, the k-th feature (x k ) from a feature space x = {x 1 , .…”

“…, x n } for a specific dataset X can be considered as relevant, if it can be used to distinguish between two specified different classes, Y = {y1, y2}, in X. Several algorithms have been proposed in the literatures (a good list demonstrated in [13]) with the main goal of finding an optimal subset of features S o which maximises the evaluation criteria J(. ):…”

“…This study is limited to the CFS feature selection algorithm since it mainly focuses on comparing the performance of conventional and one-class classifiers, not the performance of different selection algorithms. A performance comparison of several selection algorithms can be found in our previous study [13]. The CFS is used in this study due to its ability to discard irrelevant and redundant features, i.e.…”

Feature-Ensemble-Based Novelty Detection for Analyzing Plant Hyperspectral Datasets

“…Recently, researchers have shown an increased interest in developing efficient and effective analysis tools for hyperspectral images or data, since a large amount of HSI data is being generated and it is difficult, if not impossible, to analyze the information directly from the pixel values [16], [17]. Increasingly machine learning techniques are applied due to the scale and complexity of the problems.…”