Nested Transformers for Hyperspectral Image Classification

Figure 4

Illustration of the GradCAT for HSI. (a) The schematic graph of the GradCAT. Given an input image, get the Dim dimension feature maps with the size of , in which each node is the activation value of the patch from the original data in the corresponding dimension. (b) A concrete example. We use a single-layer maximum strategy to select and visualize the blocks which have the greatest impact on classification.

(a)

(b)