Awesome

<h1 align='center'>Label-Pixels</h1>

Label-Pixels is the tool for semantic segmentation of remote sensing imagery using Fully Convolutional Networks (FCNs). Initially, this tool developed for road extraction from high-resolution remote sensing imagery and now, this tool can be used to extract various features. This is part of my MSc research work (Automatic Road Extraction from High-Resolution Remote Sensing Imagery using Fully Convolutional Networks and Transfer Learning), worked under Mr. Ashutosh Kumar Jha (IIRS) and Dr. Claudio Persello (ITC, University of Twente).

Tested on

python 3.7.15
Keras 2.9.0 
Tensorflow 2.9.2
gdal 2.2.3
scikit-learn 1.0.2

Checkout Demo in Google Colab

:point_right: Google Colab Notebook

Installation

Clone repository

git clone https://github.com/venkanna37/Label-Pixels.git
cd Label-Pixels

Setup environment

Create an environment and install above packages OR install with environment.yml file in Anaconda
Note: environment.yml file created with conda env export > environment.yml command in Anaconda (this file will clean and updated soon)

conda env create -f environment.yml
conda activate label-pixels
cd tools

Usage

Rasterize

Create labels with shapefiles
The projection of imagery and shapefiles should be same
Projection units should be in meters if you want to buffer line feature
--buffer not required for polygon
--labels_atr not required for line and single class
Output directory has to create to save label images Eg: ../data/spacenet/labels/

options	Description
--help	Print usage information
--raster_dir	Directory that contains raster image/images
--vector_dir	Directory that contains vector files with the same projection as raster data. And name of the vector and raster files should be same.
--raster_format	Raster format of the image/images
--vector_format	Vector format ex: shp, geojson, etc.
--output_dir	Output directory to save labels
--buffer	Buffer length for line feature. Not required for polygon
--buffer_atr	Attribute from the vector file, this attribute can be buffer width and it multiplies with `--buffer`.
--labels_atr	Attribute from the vector file, pixels inside the polygon will be assigned by its attribute value.

Examples:

python rasterize.py --raster_dir ../data/spacenet/raster/ --raster_format tif --vector_dir ../data/spacenet/vector/ --vector_format shp --buffer 2 --output_dir ../data/spacenet/labels/ --label_atr partialDec --buffer_atr lanes

python rasterize.py --raster_dir ../data/spacenet/raster/ --vector_dir ../data/spacenet/vector_multi/ --vector_format shp --output_dir ../data/spacenet/labels/ --label_atr value

Patch Generation

Generate patches (small image subset with fixed size) from images/tiles
To generate patches for train, test and validation sets, the command needs to be run three times or can create patches with single run and split dataset with split_dataset.py script
Name of input and label images should be same
Output directory has to create to save patches Ex: ../data/mass_patches/

options	Description
--image_folder	Folder of input images/tiles with directory
--image_format	Image format tiff/tif/jpg/png
--label_folder	Folder of label images with directory
--label_format	Label format tiff/tif/jpg/png
--patch_size	Patch size to feed network. Default size is 256
--overlap	Overlap between two patches on image/tile (units: pixels)
--output_folder	Output folder to save patches

Example:

python patch_gen.py --image_folder ../data/massachusetts/test/image/ --image_format tiff --label_folder ../data/massachusetts/test/roads_and_buildings/ --label_format tif --patch_size 256 --output_folder ../data/mass_patches/

CSV Paths

Save directories of patches in CSV file instead of reading patches from folders directly
Output directory has to create to save the csv files Eg: ../paths/

options	Description
--image_folder	Folder of image patches with directory
--image_format	Image format tif (patch_gen.py save patches in tif format)
--label_folder	Folder of label patches with directory
--label_format	Label format tif (patch_gen.py save patches in tif format)
--patch_size	Patch size to feed network. Default size is 256
--output_csv	csv filename with directory

Example

python csv_paths.py --image_folder ../data/mass_patches/image/ --label_folder ../data/mass_patches/label/ --output_csv ../paths/data_rnb.csv

Training

Training FCNs (UNet, SegNet, ResUNet and UNet-Mini) for semantic segmentation
For Binary classification, --num_classes = 1
For Binary classification with one-hot encoding, --num_classes = 2
For multi class classification, --num_classes = number of target classes (>1)

options	Description
--model	Name of the FCN model. Existing models are unet, unet_mini, segnet and resunet
--train_csv	CSV file name with directory, consists of directories of image and label patches of training set.
--valid_csv	CSV file name with directory, consists of directories of image and label patches of validation set.
--input_shape	Input shape of model to feed patches (patch_size patch_size channels)
--batch_size	Batch size, depends on GPU/CPU memory
--num_classes	Number of classes in labels data
--epochs	Number of epochs
--rs	Radiometric resolution of the input images to rescale (Eg: 8, 12 and etc.)
--rs_label	The value for rescaling label images (Eg: `--rs_label 255` for converting 0 & 255 values to 0 & 1
--weights	Pretrained weights file for fine tuning the model

Example

python train.py --model unet_mini --train_csv ../paths/data_rnb.csv --valid_csv ../paths/data_rnb.csv --input_shape 256 256 3 --batch_size 4 --num_classes 3 --epochs 100

Accuracy

Calculates the accuracy using different accuracy metrics
IoU, F1-Score, Precision and Recall

options	Description
--input_shape	Input shape of model (patch_size, patch_size, channels)
--weights	Trained model with directory
--csv_paths	CSV file name with directory, consists of directories of image and label patches of test set.
--num_classes	Number of classes in labels data

Example

python accuracy.py --model unet_mini --input_shape 256 256 3 --weights ../trained_models/unet_mini_256_100_23_10_22.hdf5 --csv_paths ../paths/data_rnb.csv --num_classes 3

Prediction

Predicts the entire image/tile with trained model
Output directory has to create to save the predicted images Eg: ../data/predictions/

options	Description
--model	Name of the FCN model. Existing models are unet, unet_mini, segnet and resunet
--input_shape	Input shape of model (patch_size, patch_size, channels)
--weights	Trained model with directory
--image_folder	Folder of input images/tiles with directory
--image_format	Image format tiff/tif/jpg/png
--output_folder	Output folder to save predicted images/tiles
--num_classes	Number of classes in labels data
--rs	Radiometric resolution of the input images (Eg: 8, 12 and etc.)

Example:

python tile_predict.py --model unet_mini --input_shape 256 256 3 --weights ../trained_models/unet_mini_256_100_07_11_22.hdf5 --image_folder ../data/massachusetts/test/image/ --image_format tiff --output_folder ../data/predictions/ --num_classes 3

Summary of the Model

Summary of FCNs
Useful to check the configuration of FCNs, number of parameters in each layer and total
Replace unet, segnet and resunet with unet_mini to check configuration of all networks

options	Description
--model	Name of FCN model. Existing models are unet, unet_mini, segnet and resunet
--input_shape	Input shape of model to feed (patch_size patch_size channels)
--num_classes	Number of classes to train

Example

python summary.py --model unet_mini --input_shape 256 256 3 --num_classes 3

For examples of other scripts, check Google Colab Demo Notebook

split_dataset.py for splitting entire dataset into train, test and validation sets
patch_predict.py for predicting all patches and save in the directory
osm_data.py for downloading OpenStreetMap data to generate labels

Example Outputs

Benchmark datasets

Massachusetts Benchmark datasets for Roads and Buildings extraction https://academictorrents.com/browse.php?search=Volodymyr +Mnih
List of Benchmark datasets for semantic segmentation, object detection from remote sensing imagery https://github.com/chrieke/awesome-satellite-imagery -datasets

Any problem with code?

Please open the issue