Slum Mapping

4. Get Started

4.1 Data

4.1.1 Vector Ground Truth (ArcGIS)

Find here: OneDrive

Created by ArcQIS, saved with extension xx.kmz. Take Shenzhen.kmz as an example in following contents.

4.1.2 Raster Ground Truth (QGIS)

Add Vector Layer

Layer -> Add Layer-> Add Vector Layer
Source: path/to/Shenzhen.kmz
Select Vector Layers to Add: Geometry type: Polygon

Save shapefile to "./src/ori/cities/Shenzhen/shapefile/shenzhen.shp", CRS in EPSG:3857 - WGS84 (meters instead of degrees).

Add Google Satellite Layer

Find Google Maps URL

Plugins -> Tile+ -> Tile+

Add Google Satellite by URL, zoom level information here.

Browser -> XYZ Tiles -> New Connection -> xxx

Extract Google Satellite Imagery

Extract target Google Satellite Imagery based on related shapefile

XYZ Tiles：Google Satellite -> Export layer -> To file ；or XYZ Tiles: New conection-> URL：http://www.google.cn/maps/vt?lyrs=s@189&gl=cn&x={x}&y={y}&z={z} 

Save to "./src/ori/cities/Shenzhen/dataset/2.0/image_full/Shenzhen.tif"

no Create VRT
Calculate from Layer -> Shenzhen (as example)
Resolution (current: userdefined) -> 2 (as example)

4.1.3 Clip Into Small Pieces

Load shapefile with related Google Satellite imagery in GIS (just for visualization).

Google Satellite imager dir: "./src/ori/cities/Shenzhen/dataset/2.0/image_full/Shenzhen.tif"

Shapefile dir: "./src/ori/cities/Shenzhen/shapefile/shenzhen.shp"

Clip Google Satellite imagery and related ground truth (shapefile) into small pieces

Run:

$ python ./utils/preprocess.py

Arguments:

# city name
CITY = 'Shenzhen' 
# resolution
RESO = '2.0'
VRT_CLIP = True
# tile shape size
if VRT_CLIP:
    tile_size_x = 5000
    tile_size_y = 5000
    
shapefile_root = 'to/your/path/slum-mapping/src/ori/cities/%s/shapefile/%s.shp'%(CITY, CITY)
dataset_dir = 'to/your/path/slum-mapping/src/ori/cities/%s/dataset/%s/'%(CITY, RESO)

Dataset saved in "./src/ori/cities/Shenzhen/dataset/2.0/", directory tree:

.
├── anno
│   ├── Shenzhen_0.tfw
│   ├── Shenzhen_0.tif
│   ├── ...
│   ├── Shenzhen_41.tfw
│   └── Shenzhen_41.tif
├── anno_full
│   ├── Shenzhen.tfw
│   └── Shenzhen.tif
├── image
│   ├── Shenzhen_0.tfw
│   ├── Shenzhen_0.tif
│   ├── ...
│   ├── Shenzhen_41.tfw
│   └── Shenzhen_41.tif
└── image_full
    └── Shenzhen.tif

4.1.4 Upload in GPU Server

SSH connect to reedbush server, run

$ ssh -l p75001 reedbush.cc.u-tokyo.ac.jp

Make training data dirs in server, run

$ cd /lustre/gp75/p75001/Work/slum-mapping/src
$ mkdir Shenzhen_2.0_train && cd Shenzhen_2.0_train
$ mkdir image label
$ touch train.txt test.txt

Upload dataset from local via sftp, run on local terminal

$ sftp [email protected]
sftp> lcd xx # Change the local directory
sftp> cd xx # Change the remote directory
sftp> put # # Transfer file from local root to remote server

an example for uploading annotation, run on local terminal

$ sftp [email protected]
sftp> lcd ./src/ori/cities/Shenzhen/dataset/2.0/anno/
sftp> cd /lustre/gp75/p75001/Work/slum-mapping/src/Shenzhen_2.0_train/label/
sftp> put

an example for uploading image, run on local terminal

$ sftp [email protected]
sftp> lcd ./src/ori/cities/Shenzhen/dataset/2.0/image/
sftp> cd /lustre/gp75/p75001/Work/slum-mapping/src/Shenzhen_2.0_train/image/
sftp> put

4.2 Model Training

In server

4.2.1 Data Extraction

Choose training tiles names, and edit train.txt and test.txt in "./src/shenzhen_2.0_train/";

Edit "./utils/run_extractor.sh" for your own requirement

#!/bin/sh
#PBS -q l-debug
#PBS -W group_list=gp75
#PBS -l select=4:mpiprocs=8:ompthreads=4
#PBS -l walltime=00:10:00
cd $PBS_O_WORKDIR
. /etc/profile.d/modules.sh
module purge
module load anaconda3/2019.10 cuda10/10.0.130 intel openmpi/3.1.4/intel
export PYTHONUSERBASE=/lustre/gp75/p75001/packages
export LD_LIBRARY_PATH=/lustre/app/acc/anaconda3/2019.10/lib
python ./extractor.py -data Shenzhen_2.0_train -data_usage train -mode slide-rand -nb_crop 400

run run_extractor.sh in ./utils/

$ qsub run_extractor.sh

extracted data saved in "./dataset/shenzhen_2.0_train_rand/";

4.2.2 Training

Edit ./run_FPN.sh, choose training dataset and related modes:

#!/bin/sh
#PBS -q l-regular
#PBS -W group_list=gp75
#PBS -l select=4:mpiprocs=8:ompthreads=4
#PBS -l walltime=100:00:00
cd $PBS_O_WORKDIR
. /etc/profile.d/modules.sh
module purge
module load anaconda3/2019.10 cuda10/10.0.130 intel openmpi/3.1.4/intel
export PYTHONUSERBASE=/lustre/gp75/p75001/packages
export LD_LIBRARY_PATH=/lustre/app/acc/anaconda3/2019.10/lib
python ./FPN.py -train_data Shenzhen_2.0_train-rand -terminal 1200

Run

$ qsub ./run_FPN.sh

Check running status:

$ rbstat

Log results saved in "./logs/"; trained model saved in ./checkpoint/

4.2 Inference

Take testing Guangzhou_2.0 as an example;

Prepare Guangzhou_2.0_test as introduced in 4.2.1;

Edit "./run_inference.sh", debug args parameters:

#!/bin/sh
#PBS -q l-regular
#PBS -W group_list=gp75
#PBS -l select=4:mpiprocs=8:ompthreads=4
#PBS -l walltime=00:40:00
cd $PBS_O_WORKDIR
. /etc/profile.d/modules.sh
module purge
module load anaconda3/2019.10 cuda10/10.0.130 intel openmpi/3.1.4/intel
export PYTHONUSERBASE=/lustre/gp75/p75001/packages
export LD_LIBRARY_PATH=/lustre/app/acc/anaconda3/2019.10/lib
python ./vissin_Area.py -data Guangzhou_2.0_test -checkpoints FPN_epoch_300_May11_00_16.pth

Run:

$ qsub run_inference.sh

Results will bed saved in "./result/area-binary".

Convert into Transparent Mode

$ python ./utils/transparent.py

Final results saved in "./result/Guangzhou/2.0_trans/"

4.3 Visualization

Confusion Matrix	Predicted (True)	Predicted (False)
Actual (True)	Green	Red
Actual (False)	Blue	No Color