Algorithm 1 Training, validation and testing steps

1: Read training, validation, testing, and mask set, process them into tensor

2: For epoch from 1 to 200

3: For iter_training from 1 to training_size/batch_size

4: pass the training data to the training system (FCN, Deeplabv3, U-Net), getting the train_result

5: accumulate the loss_train by the loss function between train_result and ground truth mask

6: update weights W_training

7: End iter training

8: For iter_val from 1 to val_size/batch_size

9: pass the validation data to the trained system, getting val_result

10: accumulate loss_val by the loss function between val_result and ground truth mask

11: End iter validation

12: print the average loss by loss_train/(epoch + 1) and loss_val/(epoch + 1)

13: End training and val

14: load the model with trained W_weights

15: pass testing data to the model and get test_result

16: compute the mIoU, and mPA

17: merge the images of raw data with test_result and ground truth