A spatio temporal spectral framework for plant stress phenotyping

Table 5 Mean cross validation accuracies for different machine learning algorithms on the dataset [2]

Method	Cross validated training accuracy			Test set accuracy
	Water	Nitrogen	Weeds	Water	Nitrogen	Weeds
Decision trees [49]	93.39	63.66	60.36	86.90	47.62	65.48
\(\hbox {LDA}^{\mathrm{a}}\) [50]	96.10	68.47	75.68	94.05	78.57	75.00
\({{SVM}}^{\mathrm{b}}\) [51]	93.09	75.68	83.18	95.24	80.95	77.38
\(\hbox {KNN}^{\mathrm{c}}\) [52]	92.79	62.16	69.97	92.86	55.95	65.48
BaggedTrees [53]	94.89	67.57	71.47	91.67	63.10	69.05
Subspace discriminant [54]	94.59	70.57	75.08	94.05	75.00	72.62
Subspace KNN [54]	93.39	60.66	64.26	97.62	66.67	72.62
\(\hbox {RUSBoostedTrees}^{\mathrm{d}}\) [55]	95.20	69.37	69.37	97.62	63.10	71.43

For detailed descriptions of the machine learning methods evaluated we refer the reader to the cited papers. The SVM classifier showed the best overall performance from the tested methods, on both the training and test data, indicating good generalization to novel inputs. The implementations for the classification methods provided by the MATLAB® Statistics and Machine Learning Toolbox were used. The specific parameters for each of the classifiers can be found within the MATLAB® functions provided in the accompanying software suite
\({}^{\mathrm{a}}\) Linear discriminant analysis
\({}^{\mathrm{b}}\) Support vector machine
\({}^{\mathrm{c}}\) \(k-\)nearest neighbor
\({}^{\mathrm{d}}\) Randomly undersampled boosted trees

ISSN: 1746-4811