Spectral-Spatial exploration for hyperspectral image semantic segmentation via the fusion of Fully Convolutional Networks

Training-test information leakage problem: As shown in Figure (b), both patches represent ‘3’. If one of them is selected as training data and the other one is selected as the test data, the evaluation results cannot demonstrate the real discrimination ability of the obtained model. Although the spatial features is vital for HSI analysis, the previous usage of spatial information in patch-wise classification might be inappropriate.

Data partitioning method： Take Salinas Valley for example, we split the multi-class blocks into 9 folds and the order of samples in the k-th fold is [k, 9+k, . . . , 9N+k] where N represents the number of samples in this fold. We select a single fold as the training set, the other one as the validation set, and the remaining 7 folds as the test. See article for details

Two branches network proposed in paper: The SS3FCN architecture used for HSI semantic segmentation. (a) the basic unit of SS3FCN; (b) and (c) the structure of the proposed SS3FCN. Prior to the fusion of two branches, we employ a convolution layer with large stride (1110 for Salinas Valley and Indian Pines, 116 for Pavia University and Houston University). The features from 3D branch and 1D branch are concatenated into 512 feature maps, and forwarded to the last convolution layer.

Methods of training and testing SS3FCN:

1. Modify data_name,fold_num_train in dataset. Py

2. Change data_name in model3d.py

3. Change data_name in main.py

Important training Settings:

1.python 3.6

2. Tensorflow - gpu 1.9.0

3. Keras 2.2.4

4.spectral 0.19

5. Other basic python package

Note: due to the limitation of upload file size, you need to put the dataset into the "data" folder