search

qichaoliu / CNN_Enhanced_GCN

Q. Liu, L. Xiao, J. Yang and Z. Wei, "CNN-Enhanced Graph Convolutional Network With Pixel- and Superpixel-Level Feature Fusion for Hyperspectral Image Classification," in IEEE Transactions on Geoscience and Remote Sensing, doi: 10.1109/TGRS.2020.3037361.
145 stars 29 forks source link

SGsN module on IP data set #3

Open immortal13 opened 3 years ago

immortal13 commented 3 years ago

Hi, hope you are doing great! When I remove the PAsN module and test on PU data set, it is still effective and get a high OA. But when I conducted this ablation experiment on IP data set, I found that training process cannot converge and get poor performance.

qichaoliu commented 3 years ago

Thanks! Yes, you are right. In fact, different datasets have different patterns, which leads to the different performance of CNN and GCN. On the Salinas, GCN can get very high OA, but fail on the Indian Pines. In contrast, CNN can perform better on the Indian Pines, while getting only not good OA on the Salinas. This phenomenon above urges us to joint GCN and CNN together. In my understanding, pixel-level features (CNN) are required in every dataset, while superpixel-level features (GCN) only play an auxiliary role. If the graph structures of HSIs could be learned in the network automatically instead of designing it manually, the GCN could fully replace CNN to complete HSI classification. Efficient, controllable, and differentiable graph clustering (pooling) may be a promising direction in the future.

immortal13 commented 3 years ago

Thank you very much!!! cannot agree with you more!