Table 1 Accuracy metrics relating to three DL models: VGG9 trained from scratch, DenseNet-121 trained from scratch, and pretrained DenseNet-121

DL model scratch/pretrained (bands)	T max R²	T min RMSE	T mean R²	T mean RMSE	T sd R²	T sd RMSE	V max R²	V min RMSE	V mean R²	V mean RMSE	V sd R²	V sd RMSE
VGG9 scratch (RGB)	0.581	1.114	0.503	1.216	0.046	0.056	0.654	1.056	0.487	1.228	0.101	0.115
VGG9 scratch (RGNir)	0.615	1.067	0.532	1.179	0.066	0.083	0.716	0.958	0.520	1.185	0.101	0.110
Dense Net-121 scratch (RGB)	0.717	0.916	0.659	1.007	0.045	0.066	0.655	0.977	0.577	1.117	0.077	0.112
Dense Net-121 scratch (RGNir)	0.745	0.869	0.695	0.952	0.037	0.059	0.696	0.958	0.627	1.049	0.058	0.063
Dense Net-121 pretrained (RGB)	0.925	0.476	0.903	0.537	0.016	0.044	0.841	0.693	0.801	0.767	0.021	0.039
Dense Net-121 pretrained (RGNir)	0.950	0.383	0.886	0.576	0.038	0.101	0.822	0.734	0.789	0.790	0.021	0.027

For each model, two experiments were conducted with different band combinations (RGB and RGNir). The metrics include the maximum, mean and standard deviation of R² (coefficient of determination) and RMSE (root mean square error), obtained on the training folds (represented with T) and on the validation fold (V).

Quick links

Search