Fig 1.
Preprocessing of HSI.
Fig 2.
Proposed multibranch CNN architecture.
Table 1.
Segmentation details of the datasets: Indian Pines (IP), Salinas Scene (SC), and The Pavia University (PU). Seg-L denotes the dataset is split into L highly correlated subgroups; in Seg-3, the dataset is split into 3 sub-groups.
Table 2.
Configuration of deep neural network used in feature learning procedure.
Table 3.
Number of pixels corresponding to their land cover categories for SC dataset.
Table 4.
Number of pixels corresponding to their land cover categories for the PU dataset.
Table 5.
Number of pixels corresponding to their land cover categories for the IP dataset.
Fig 3.
Band-to-band correlation matrix for the datasets, (a) SC, (b) PU, (c) IP, and (d) Band-to-band correlation matrix show the segmentation of SC when L=3.
Table 6.
Outcomes of the proposed model for different segmentations of datasets.
Table 7.
Comparison of the proposed model with the state of the arts for SC dataset.
Table 8.
Comparison of the proposed model with the state of the arts for the PU dataset.
Table 9.
Comparison of the proposed model with the state of the arts for the IP dataset.
Fig 4.
Overall accuracy comparison of (a) SC, (b) PU, and (c) IP datasets for different training samples.
Table 10.
Computational parameter comparison and analysis.
Table 11.
Class-wise accuracies (%) for the SC dataset with 10% training samples.
Table 12.
Class-wise accuracies (%) for the PU dataset with 10% training samples.
Table 13.
Class-wise accuracies (%) for the IP dataset with 10% training samples.
Fig 5.
Loss and accuracy curves of all datasets for training samples, (a) SC loss, (b) PU loss, (c) IP loss, (d) SC accuracy, (e) PU accuracy, and (f) IP accuracy.
Table 14.
Outcomes of the ablation studies.