Fig 1.
The three major steps used of the study include (1) processing and segmenting the images, (2) applying the random forest classifier, and (3) evaluating and assessing the results.
Table 1.
The ground-truth classes of the AVIRIS and HYDICE datasets and the training and test sets used for the classes.
Fig 2.
AVIRIS Indian Pines (A) RGB composite image (channels 47, 24, and 14), (B) its reference map and (C) its training and test set. HYDICE Washington DC Mall (D) RGB composite image (channels 51, 41, and 22), (E) its reference map and (F) its training and test set.
Fig 3.
Schematic diagram of MDI applied on a sample spectral reflectance curve of a green vegetation (adapted from Salas and Henebry [12]).
Note that the number of points between LP and RP pivots can vary depending on the number of bands analyzed or the width of the pivot wavelength region.
Table 2.
Description of other spectral indices used as input predictor variables in this study.
Table 3.
Five sets of data were separately used as inputs in the object-based random forest classification.
Fig 4.
Contrasting the classification results of using different datasets for AVIRIS Indian Pines image: (A) dataset 1 without including any moment distance method, (B) dataset 2 with new MDIN, (C) dataset 3 with new MDRLR, (D) dataset 4 with new MDRRL, and (E) dataset 5 with original MDI. The maps were derived using object-based random forest classification.
Table 4.
Summary of classification accuracies (%) from five sets of data using AVIRIS Indian Pines image: set 1 (no MDI), set 2 (with MDIN), set 3 (with MDRLR), set 4 (with MDRRL), and set 5 (with original MDI).
Table 5.
Summary of classification accuracies (%) from five sets of data using HYDICE Washington DC Mall image: set 1 (no MDI), set 2 (with MDIN), set 3 (with MDRLR), set 4 (with MDRRL), and set 5 (with original MDI).
Fig 5.
Contrasting the classification results of using five datasets for HYDICE Washington DC Mall image: (A) dataset 1 without including any moment distance method, (B) dataset 2 with new MDIN, (C) dataset 3 with new MDRLR, (D) dataset 4 with new MDRRL, and (E) dataset 5 with original MDI. The maps were derived using object-based random forest classification.
Table 6.
Rankings of the 5 object features with maximum importance across classes in the RF model using AVIRIS Indian Pines image.
Table 7.
Rankings of the 5 object features with maximum importance across classes in the RF model using HYDICE Washington DC Mall image.
Fig 6.
Encircled and magnified sample portion of AVIRIS Indian Pines classification maps, showing the difference of the performances of using a dataset (A) without MDI, (B) with MDIN, (C) with MDRLR, and (D) with MDRRL for classes soybean-min and corn-notill.
Fig 7.
Encircled and magnified sample portion of AVIRIS Indian Pines classification maps, showing the difference of the performances of using a dataset (A) without MDI, (B) with MDIN, (C) with MDRLR, and (D) with MDRRL for classes grass/pasture and grass/trees.
Fig 8.
Encircled and magnified sample portion of HYDICE Washington DC Mall classification maps, showing the difference of the performances of using a dataset (A) without MDI, (B) with MDIN, (C) with MDRLR, and (D) with MDRRL for classes tree and grass.
Fig 9.
Sample ground reference photo for (A) corn-notill and (B) soybean-min taken at the AVIRIS image field site. Notill = no tillage; min = minimum tillage.
Fig 10.
MD algorithm applied to the spectral responses of corn-notill and soybean-min for (A) AVIRIS, and grass and tree for (B) HYDICE, and how MDI values varied moving the pivot from left to right, and vice versa for (C) AVIRIS and (D) HYDICE images. Maximum values are observed at maximum shape differences, usually occurring at the inclusion of a curve peak or dip. Note that differences in curve shape could mean discrimination between classes.