satellite_segmentaion

Introduction:

My final thsis is about satellite segmentaion. I need to extract road、builidings、water or plants from the images.

About the dataset:

The dataset must conation images and masks. The dataset is shared below about five kinds of objects.

baidu cloud：https://pan.baidu.com/s/1TcQuMAY2aEiVrFHJd5YIAA code：vfps

google dirve：https://drive.google.com/open?id=1gE6WeoSiXTPEr-mYH8uggXiJEISgBHYZ

About the mothod:

I use three models to predict the result. Unet、PSPnet and Segnet. And then I vote for a final result by combining the models' result. Then I do some post_processing for voted images. And then there can be some little improvement.

About the Result:

I just extract road from the models for now. So now I just show the Iou result about the road. I train them for 50 epochs and learning rate is1e-4. Adam optimizer. Here are the result.

And here is the performance in the different validation area which contaion a large area. As we can see the performance after voting can imporve a lot.

If you want use this code. you can just clone it . I will add the requirments.txt soon.

First step: prepare you own dataset:

You need to give the right path for images and masks in dataset_processing.py

image_path = Path('./BDCI2017-seg/CCF-training-Semi')/f'{i}.png'
img_class_path = Path('./BDCI2017-seg/CCF-training-Semi')/ f'{i}_class_vis.png'

Note： This step is for the big picture no for less than 1000*1000 shape. If you have hundreds of images, you can just jump to step two.

Second step: train

NOTE: If you want to train on your own dataset, the image name and the mask name must be the same. and the name suffix must be png format.

give the path of dataset and alter the params you want to use.

device = 'cuda'
path = '/home/shiyi/beshe/kinds_dataset/'
learning_rate = 5e-3
num_epochs = 50
num_classes = 5
optimizer = torch.optim.Adam(net.parameters(), lr=learning_rate)

train_dataset = SlippyMapTilesConcatenation(
os.path.join(path, "training", "images"), os.path.join(path, "training", "labels"), transform,debug = False
)

val_dataset = SlippyMapTilesConcatenation(
os.path.join(path, "validation", "images"), os.path.join(path, "validation", "labels"), transform,debug = False
)

Third step: Prediction

The code in predict.py maybe hard to understand. You can just use it. You need use the model you trained in step two.

model = torch.load(f'model/0514pspnet_50_epoch.pth')
# give the picture you want to predict
file_name = f'/home/shiyi/beshe/gaoxin_map/second_dataset/part1_500.png'
# give the name you want to store
save_dir = '0514predict1.png' 

Fourth Step: vote and post_processing

The post_processing progress can be interesting. You can different kinds of combanation.

# you can just run 
python post_deal.py

​