Etwork. The formula on the channel the channel consideration is: NP
Etwork. The formula in the channel the channel interest is: NP Wx eWk x j N Zi = Xi +Z XReLUReLU((W(W e Wv2 W ( LN LN v1 (3) x j )) x j )) i i v2 v1 N N p (3) j=j11 Wx Wk xm e eP k j p k mm= m1Wk x j exactly where x where j j= ee W k xm represents the international pooling and Wv 2 ReLU( LN(( LN))) v1 ()) denotes represents the international pooling and Wv2 ReLU Wv1 ( (W denotes the botWm meWk x jek mthe bottleneck transform. The channel consideration moduleglobalglobal focus pooling to tleneck transform. The channel interest module utilizes makes use of interest pooling to model model the long-distance dependences and capture discriminative channel functions the rethe long-distance dependences and capture discriminative channel characteristics from from the redundant hyperspectral images. dundant hyperspectral images.Figure The architecture of 2D channel attention block. Figure 2.2. The architecture of 2D channel focus block.two.three.two. Spatial Consideration two.3.2. Spatial Attention A spatial attention block depending on the interspatial relationships of options is develA spatial focus block according to the interspatial relationships of features is developed, as inspired by CBAM [20]. Figure illustrates the structure on the spatial attention oped, as inspired by CBAM [20]. Figure 33 illustrates the structure of the spatial interest block. To produce an efficient feature descriptor, average-pooling and max-pooling operblock. To create an effective feature descriptor, average-pooling and max-pooling operaations are applied along the channel axis, and they concatenate them. Pooling operationsMicromachines 2021, 12,5 ofMicromachines 2021, 12, x FOR PEER REVIEW5 oftions are applied along the channel axis, and they concatenate them. Pooling operations along the channel axis are shown to become efficient at highlighting informative regions. Then, along the channel axis are shown to become efficient at highlighting informative regions. Then, a convolution layer is applied for the concatenated feature descriptor to make a spatial a convolution layer is applied to the concatenated feature descriptor to create a spatial interest map that specifies which features to emphasize or suppress. They are then interest making use of a normal convolution layer to make a Streptonigrin Purity two-dimensional spatial then conconvolvedmap that specifies which capabilities to emphasize or suppress. These are interest volved brief, spatial focus is calculated to make a map. In using a standard convolution layer as follows: two-dimensional spatial focus map. In quick, spatial interest is calculated as follows: 3 M( F ) = F )f [ AvgPool ( F);;MaxPool( F )]) F )]) M( ( ( f 33 [ AvgPool( F ) MaxPool ( (four) (four) = ( f three(([3F ([ F ; FF ])) ])) f 3 ;avgavg max maxwhere denoted the sigmoid function and f 3f 3 3 represents a convolution operation where denoted the sigmoid function and3 represents a convolution operation with using the filter size 3. the filter size of 3 f 3 three .(1, H, W)Spatial Attention(2, H, W)Con vMaxpool(1, H, W)CHc(1, H, W) pixel-wise Mu t iplicat ionW AvgpoolFigure 3. The architecture of 2D spatial attention block. Figure 3. The architecture of 2D spatial attention block.two.4. HSI Classification Determined by BSJ-01-175 Data Sheet MFFDAN two.4. HSI Classification Based on MFFDAN The architecture of MFFDAN depicted in Figure four. The University of of Pavia dataset The architecture of MFFDAN isis depicted in Figure 4. The University Pavia dataset is is useddemonstrate the the algorithm’s detailed process. raw data information are normali.