Fig 1.
Batch normalization.
Fig 2.
Group normalization (g = 2).
Fig 3.
Visualization of whitening effectiveness.
(a) Original data. (b) PCA whiten data. (c) ZCA whiten data.
Fig 4.
The architecture of proposed SCNet.
Fig 5.
The structure of SCConv with residual connection.
Fig 6.
The architecture of spatial reconstruction unit.
Fig 7.
The architecture of channel reconstruction unit.
Fig 8.
Different depths contain varying quantities of feature information.
(a) Feature maps of Conv. (b) First SCConv module’s feature maps. (c) Second. (d) Third.
Table 1.
The class names of UP dataset along with the number of training, validation, and test samples for each class.
Table 2.
The class names of KSC dataset along with the number of training, validation, and test samples for each class.
Table 3.
The class names of WHLK dataset along with the number of training, validation, and test samples for each class.
Table 4.
The class names of IP dataset along with the number of training, validation, and test samples for each class.
Fig 9.
The FLOPs and OA on four datasets with different split rRESEARCHARTICLEatios in SCConv.
Fig 10.
Classification results for varying patch sizes on four datasets.
Table 5.
Network configuration of the SCNet model on Indian Pines dataset.
Fig 11.
Classification maps for the UP dataset using 1% training samples.
(a) RGB image. (b) Ground-truth (GT). (c–h) The classification maps with disparate algorithms.
Table 6.
The classification results for the UP dataset with 1% training samples.
Fig 12.
Classification maps for the KSC dataset using 5% training samples.
(a) False-color image. (b) Ground-truth (GT). (c–h) The classification maps with disparate algorithms.
Table 7.
The classification results for the KSC dataset with 5% training samples.
Fig 13.
Classification maps for the WHLK dataset using 0.2% training samples.
(a) False-color image. (b) Ground-truth (GT). (c–h) The classification maps with disparate algorithms.
Table 8.
The classification results for the WHLK dataset with 0.2% training samples.
Fig 14.
Classification maps for the IP dataset using 5% training samples.
(a) False-color image. (b) Ground-truth (GT). (c–h) The classification maps with disparate algorithms.
Table 9.
The classification results for the IP dataset with 5% training samples.
Table 10.
Parameter, FLOPs and runing time(s) of different methods for the four data sets.
Fig 15.
Accuracy of different methods with different numbers of training samples on UP.
(a) OA; (b) AA; (c) Kappa.
Table 11.
Ablation experimental results of different sequential spatial channel combinations on the UP dataset.
Table 12.
Ablation experimental results of different numbers of channels on the UP dataset.
Fig 16.
The training accuracy and loss curves of SRU using different normalization methods on the UP dataset.
(a,b) BN. (c,d) LN. (e,f) IN. (g,h) SN. (i,j) GN.
Table 13.
The training time of SRU using different normalization methods on the UP dataset.