Master Thesis about Change Detection in Satellite Imagery using Deep Learning
Héloïse BAUDHUIN - Antoine LAMBOT
The aim of this thesis is to create a CNN capable of detecting new constructions in satellite imagery. More precisly, taking two satellite images taken at different times and output a segmentation mask with the contour of the new buildings. The model takes two RGB images of 650 x 650 as input and outputs a semgmentation mask of the same size.
First input: before image
Second input: After image
Segmentation mask with the changes
Mask output by the CNN
The CNN architecture we are using is the UNet++ with layers removed and the filter size reduced. This way the model fits on a GPU with 6G of memory. As you can see on the above pictures, our model manages to find the changes and produce a segmentation mask with their approximated shape.
You need all the Earth_* (1-32) instances from the dataset. They need to be in the DATA folder.
Link to the dataset : https://drive.google.com/drive/folders/1rd1vseWiFSqQc5-93XSRQW9Bzzcgqc6H?usp=sharing
We provide two instances in the DEMO folder. You may add your own instances. You will need two images of 650x650 pixels, perfectly aligned, placed in a subfolder of the DEMO folder.
If you want to train our already trained Light UNet++ with the -reload argument, you need the following files :
- Weights/last.pth
- Loss/last.pth
- Loss/last_metrics.pth
If you want to predict masks with our already trained Light UNet++, you will need the following files :
- Weights/last.pth
Links to the files :
https://drive.google.com/drive/folders/1qbZm-b4gdhzzMCP09XwWx2wJKxsSXBJL?usp=sharing https://drive.google.com/drive/folders/1-DdCZxCv7OInvpUnbbT-4p2Uhc_v6ztI?usp=sharing
- PyTorch (1.3.1+)
- numpy
- scikit-learn
- matplotlib
- imageio
- Pillow
- imgaug
- tqdm
It is strongly recommended to use this code on a computer equipped with a GPU. The execution time on a CPU is manageable for a prediction but not for the training. Predicting takes ~2 sec per instance on a GPU, ~20 sec on a CPU.
python3 predict.py
[-h]
[--input INPUT]
[--output OUTPUT]
[--threshold THRESHOLD]
[--color COLOR]
optional arguments:
-h, --help show this help message and exit
--input INPUT, -i INPUT
path to the input directory, containing instance
directories, each instance directory should contain
before.png and after.png 650x650 images
--output OUTPUT, -o OUTPUT
path to the output directory, where the change masks
will be saved, can be the same as the input directory
--threshold THRESHOLD, -t THRESHOLD
a value between 0 and 1, to classify each pixel, if
not given the mask pixels will have continuous values
between the two classes
--color COLOR, -c COLOR
background color of the generated masks, can be 'red',
'blue' or 'black'
DATA_TO_BE_PREDICTED
│
└───instance_0
| ├── before.png
| └── after.png
└───instance_1
| ├── before.png
| └── after.png
└───my_instance
| ├── before.png
| └── after.png
...
Given an input instance of two images, predict will generate a change mask and store it in a png file. A single instance is a directory containing two files before.png and after.png, aligned satellite images of the same place at different times, each 650 by 650 pixels, at a resolution of around 50cm/pixel. To run predict on one or multiple instances, use the folder containing the instance as input parameter. See the demo for more details.
How to use our model to predict masks :
- Download the project
- Follow the requirements for the prediction
- run predict.py with the correct arguments. The input directory is the DEMO folder if you haven't added images of your own. Play around with the thresholds and see the influence on the output masks. You can try the following :
python3 predict -i DEMO -o DEMO -c red -t 0.1
python3 predict -i DEMO -o DEMO2
python3 predict -i DEMO/instance_1 -o DEMO/instance_1 !! This will NOT work !!
python3 train.py
[-h]
[--epochs EPOCHS]
[--learning_rate LEARNING_RATE]
[--n_data_augm N_DATA_AUGM]
[-reload]
[-save]
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS number of epochs the model will run (1 by default)
--learning_rate LEARNING_RATE
starting learning rate (0.001 by default)
--n_data_augm N_DATA_AUGM
number of data augmentation instances to generate per
original instance (2 by default)
-reload reload the model weights and metrics from the last run
(disabled by default)
-save save the weights and metrics of the model when it has
finished running (disabled by default)
The MIT License (MIT)
Copyright (c) 2020 <Héloïse Baudhuin and Antoine Lambot>
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE