V0.0

Author: chathumal93

.DS_Store

@@ -0,0 +1,56 @@
+# Pytorch-UNet-Flood-Segmentation
+
+This repository contains a PyTorch implementation of a U-Net model for segmenting water areas (flood and permanent water) in Sentinel-1 satellite images. The U-Net architecture is a popular choice for image segmentation tasks, particularly for biomedical and satellite imagery.
+
+The model has been trained and evaluated on Cloud to Street - Microsoft flood dataset and it is composed of Sentinel-1 image chips with corresponding water labels.
+
+### Repository Structure
+The repository is structured as follows:
+
+* data/: contains datasets for model training
+* models/: contains the trained U-Net models
+* predictions/: contains the predictions of trained U-Net models
+* runs/ : contains tensorboard log files
+* make-dataset.py : the script extract and perform tiling (256*256) on sentinel-1 VV VH and water labels chips
+* gee-dem-data.py : the script to download the SRTM DEM tiles from Google Earth Engine
+* gee-pwater-data.py : the script to download the JRC permanent water tiles from Google Earth Engine
+* helpers.py : helper functions for gee-dem-data.py, gee-pwater-data.py and make-dataset.py
+* make-features.ipynb : the jupyter notebook to arrange data before model training
+* models.py : U-Net model
+* model-train.ipynb : the jupyter notebook to train the models
+* model-inference.ipynb : the jupyter notebook to inference the models on the test set
+* requirements.txt: lists the required Python packages to run the scripts
+
+### Preparation of data 
+#### Step 1
+* Download Sentinel-1 data for the Cloud to Street - Microsoft flood dataset (*c2smsfloods_v1_source_s1.tar.gz*) and Sentinel-1 water labels for the Cloud to Street - Microsoft flood dataset (*c2smsfloods_v1_labels_s1_water.tar.gz*) from https://mlhub.earth/data/c2smsfloods_v1.
+
+#### Step 2
+* Run the following scripts
+```bash
+python make-dataset.py --proj_dir path/to/project-directory --chips path/to/c2smsfloods_v1_source_s1.tar.gz  --labels path/to/c2smsfloods_v1_labels_s1_water.tar.gz
+```
+* make-dataset.py script will create data , data/chips/VV, data/chips/VH, and data/labels directories. (All the data is in the dimention of 256*256 with a resolution of 10 m)
+
+* Google Earth Engine cloud project id is needed to run the below scripts.
+```bash
+python gee-dem-data.py --cld_projid ee-xxxxxxx --in_dir data/chips/VV  --out_dir data
+```
+```bash
+python gee-pwater-data.py --cld_projid ee-xxxxxxx --in_dir data/chips/VV  --out_dir data
+```
+* gee-dem-data.py and gee-pwater-data.py scripts will create data/dem and data/pwater directories respectively.
+
+### Model information
+First the whole dataset was divided into two parts based on the water percentage (>=30% and <30%) of the Sentinel-1 water label chips. Then two U-Net models have been trained on those two datasets and final prediction was taken by combining the predictions of both models. 
+
+The combined prediction achieves an average Intersection over Union (IoU) of 0.877 on the test set.
+
+<p align="center">
+<img src="predictions/preds.gif" width="500" height="250" />
+</p>
+
+#### Acknowledgments
+The implementation is based on the original U-Net paper: https://arxiv.org/abs/1505.04597.
+
+This is a basic workflow of U-Net segmentation and the models require more improvements.

README.md

@@ -0,0 +1,39 @@
+vv,vh,dem,pwater,label,water_percent
+data\chips\VV\542_10.tif,data\chips\VH\542_10.tif,data\dem\542_10.tif,data\pwater\542_10.tif,data\labels\542_10.tif,1.271057129
+data\chips\VV\767_10.tif,data\chips\VH\767_10.tif,data\dem\767_10.tif,data\pwater\767_10.tif,data\labels\767_10.tif,2.137756348
+data\chips\VV\212_10.tif,data\chips\VH\212_10.tif,data\dem\212_10.tif,data\pwater\212_10.tif,data\labels\212_10.tif,6.518554688
+data\chips\VV\628_11.tif,data\chips\VH\628_11.tif,data\dem\628_11.tif,data\pwater\628_11.tif,data\labels\628_11.tif,6.608581543
+data\chips\VV\393_00.tif,data\chips\VH\393_00.tif,data\dem\393_00.tif,data\pwater\393_00.tif,data\labels\393_00.tif,10.46600342
+data\chips\VV\255_11.tif,data\chips\VH\255_11.tif,data\dem\255_11.tif,data\pwater\255_11.tif,data\labels\255_11.tif,12.84484863
+data\chips\VV\023_11.tif,data\chips\VH\023_11.tif,data\dem\023_11.tif,data\pwater\023_11.tif,data\labels\023_11.tif,15.48614502
+data\chips\VV\050_01.tif,data\chips\VH\050_01.tif,data\dem\050_01.tif,data\pwater\050_01.tif,data\labels\050_01.tif,18.87969971
+data\chips\VV\586_01.tif,data\chips\VH\586_01.tif,data\dem\586_01.tif,data\pwater\586_01.tif,data\labels\586_01.tif,19.18792725
+data\chips\VV\172_10.tif,data\chips\VH\172_10.tif,data\dem\172_10.tif,data\pwater\172_10.tif,data\labels\172_10.tif,29.30755615
+data\chips\VV\317_10.tif,data\chips\VH\317_10.tif,data\dem\317_10.tif,data\pwater\317_10.tif,data\labels\317_10.tif,30.6060791
+data\chips\VV\216_00.tif,data\chips\VH\216_00.tif,data\dem\216_00.tif,data\pwater\216_00.tif,data\labels\216_00.tif,30.63812256
+data\chips\VV\003_10.tif,data\chips\VH\003_10.tif,data\dem\003_10.tif,data\pwater\003_10.tif,data\labels\003_10.tif,38.21105957
+data\chips\VV\049_10.tif,data\chips\VH\049_10.tif,data\dem\049_10.tif,data\pwater\049_10.tif,data\labels\049_10.tif,45.8984375
+data\chips\VV\733_10.tif,data\chips\VH\733_10.tif,data\dem\733_10.tif,data\pwater\733_10.tif,data\labels\733_10.tif,46.51947021
+data\chips\VV\045_01.tif,data\chips\VH\045_01.tif,data\dem\045_01.tif,data\pwater\045_01.tif,data\labels\045_01.tif,48.24523926
+data\chips\VV\622_10.tif,data\chips\VH\622_10.tif,data\dem\622_10.tif,data\pwater\622_10.tif,data\labels\622_10.tif,49.96948242
+data\chips\VV\668_10.tif,data\chips\VH\668_10.tif,data\dem\668_10.tif,data\pwater\668_10.tif,data\labels\668_10.tif,51.18560791
+data\chips\VV\776_10.tif,data\chips\VH\776_10.tif,data\dem\776_10.tif,data\pwater\776_10.tif,data\labels\776_10.tif,53.50341797
+data\chips\VV\100_00.tif,data\chips\VH\100_00.tif,data\dem\100_00.tif,data\pwater\100_00.tif,data\labels\100_00.tif,54.81262207
+data\chips\VV\391_11.tif,data\chips\VH\391_11.tif,data\dem\391_11.tif,data\pwater\391_11.tif,data\labels\391_11.tif,62.42370605
+data\chips\VV\161_01.tif,data\chips\VH\161_01.tif,data\dem\161_01.tif,data\pwater\161_01.tif,data\labels\161_01.tif,63.17443848
+data\chips\VV\380_10.tif,data\chips\VH\380_10.tif,data\dem\380_10.tif,data\pwater\380_10.tif,data\labels\380_10.tif,65.66467285
+data\chips\VV\016_00.tif,data\chips\VH\016_00.tif,data\dem\016_00.tif,data\pwater\016_00.tif,data\labels\016_00.tif,67.06390381
+data\chips\VV\019_00.tif,data\chips\VH\019_00.tif,data\dem\019_00.tif,data\pwater\019_00.tif,data\labels\019_00.tif,67.14782715
+data\chips\VV\634_11.tif,data\chips\VH\634_11.tif,data\dem\634_11.tif,data\pwater\634_11.tif,data\labels\634_11.tif,67.67120361
+data\chips\VV\795_01.tif,data\chips\VH\795_01.tif,data\dem\795_01.tif,data\pwater\795_01.tif,data\labels\795_01.tif,70.97473145
+data\chips\VV\161_11.tif,data\chips\VH\161_11.tif,data\dem\161_11.tif,data\pwater\161_11.tif,data\labels\161_11.tif,73.01330566
+data\chips\VV\822_00.tif,data\chips\VH\822_00.tif,data\dem\822_00.tif,data\pwater\822_00.tif,data\labels\822_00.tif,81.56280518
+data\chips\VV\126_01.tif,data\chips\VH\126_01.tif,data\dem\126_01.tif,data\pwater\126_01.tif,data\labels\126_01.tif,81.84356689
+data\chips\VV\820_11.tif,data\chips\VH\820_11.tif,data\dem\820_11.tif,data\pwater\820_11.tif,data\labels\820_11.tif,82.4005127
+data\chips\VV\832_10.tif,data\chips\VH\832_10.tif,data\dem\832_10.tif,data\pwater\832_10.tif,data\labels\832_10.tif,85.61706543
+data\chips\VV\165_11.tif,data\chips\VH\165_11.tif,data\dem\165_11.tif,data\pwater\165_11.tif,data\labels\165_11.tif,88.02947998
+data\chips\VV\621_11.tif,data\chips\VH\621_11.tif,data\dem\621_11.tif,data\pwater\621_11.tif,data\labels\621_11.tif,91.5222168
+data\chips\VV\219_10.tif,data\chips\VH\219_10.tif,data\dem\219_10.tif,data\pwater\219_10.tif,data\labels\219_10.tif,92.97943115
+data\chips\VV\824_11.tif,data\chips\VH\824_11.tif,data\dem\824_11.tif,data\pwater\824_11.tif,data\labels\824_11.tif,96.04492188
+data\chips\VV\171_00.tif,data\chips\VH\171_00.tif,data\dem\171_00.tif,data\pwater\171_00.tif,data\labels\171_00.tif,96.85516357
+data\chips\VV\657_01.tif,data\chips\VH\657_01.tif,data\dem\657_01.tif,data\pwater\657_01.tif,data\labels\657_01.tif,99.83520508

data/chip_info.csv

@@ -0,0 +1,67 @@
+import argparse
+import os
+import glob
+import concurrent.futures
+from google.auth.transport.requests import AuthorizedSession
+import ee
+import restee as ree
+from tqdm import tqdm
+from helpers import restgee_data
+
+def main(args):
+          
+    #DEM dir creation
+    os.makedirs(os.path.join(args.out_dir,'dem'), exist_ok=True)
+    
+    ee.Authenticate(auth_mode='notebook') 
+    session = AuthorizedSession(ee.data.get_persistent_credentials())
+    
+    # Creating a restee session
+    class EESessionContainer(ree.EESession):
+        def __init__(self, project, session):
+            self._PROJECT = project
+            self._SESSION = session
+
+    # Create an EESesssion object with the correct permissions
+    ee_session = EESessionContainer(args.cld_projid, session)
+
+    # Authenticate EE with the session credentials
+    ee.Initialize(ee_session.session.credentials, project=args.cld_projid)
+    
+    # SRTM dem data
+    elevation = ee.Image('NASA/NASADEM_HGT/001').select('elevation')
+    
+    # One of 2 polarization
+    chip_list = sorted(glob.glob(os.path.join(args.in_dir,'*.tif'), recursive = True))
+    with tqdm(total=len(chip_list),position=0, leave=True, desc="GEE data request progress: SRTM DEM") as pbar:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=15) as executor:
+            # Start the load operations and mark each future with its chip        
+            future_to_chip = {executor.submit(restgee_data,chip,elevation, 'elevation', os.path.join(args.out_dir,'dem'),ee_session): chip for chip in chip_list}          
+            for future in concurrent.futures.as_completed(future_to_chip):
+                chip = future_to_chip[future]
+                pbar.update(n=1)
+    
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Downloading corresponding SRTM DEM tile from GEE for a given raster chip')
+    parser.add_argument(
+        "--cld_projid",
+        required=True,
+        type=str,
+        help="Cloud project id",
+    )  
+    parser.add_argument(
+        "--in_dir",
+        type=str,
+        required=True,
+        help="Input tile directory",
+    )
+    parser.add_argument(
+        "--out_dir",
+        type=str,
+        required=True,
+        help="Output folder for corresponding dem tiles",
+    )
+    args = parser.parse_args()
+    
+    main(args)

data/chip_set1.csv

@@ -0,0 +1,69 @@
+import argparse
+import os
+import glob
+import concurrent.futures
+from google.auth.transport.requests import AuthorizedSession
+import ee
+import restee as ree
+from tqdm import tqdm
+from helpers import restgee_data
+
+def main(args):
+          
+    #DEM dir creation
+    os.makedirs(os.path.join(args.out_dir,'pwater'), exist_ok=True)
+    
+    ee.Authenticate(auth_mode='notebook') 
+    session = AuthorizedSession(ee.data.get_persistent_credentials())
+    
+    # Creating a restee session
+    class EESessionContainer(ree.EESession):
+        def __init__(self, project, session):
+            self._PROJECT = project
+            self._SESSION = session
+
+    # Create an EESesssion object with the correct permissions
+    ee_session = EESessionContainer(args.cld_projid, session)
+
+    # Authenticate EE with the session credentials
+    ee.Initialize(ee_session.session.credentials, project=args.cld_projid)
+    
+    # JRC water data
+    occurrence = ee.Image('JRC/GSW1_4/GlobalSurfaceWater').select('occurrence')
+    
+    
+    
+    # One of 2 polarization
+    chip_list = sorted(glob.glob(os.path.join(args.in_dir,'*.tif'), recursive = True))
+    with tqdm(total=len(chip_list),position=0, leave=True, desc="GEE data request progress: JRC Water") as pbar:
+        with concurrent.futures.ThreadPoolExecutor(max_workers=15) as executor:
+            # Start the load operations and mark each future with its chip        
+            future_to_chip = {executor.submit(restgee_data,chip,occurrence, 'occurrence', os.path.join(args.out_dir,'pwater'),ee_session): chip for chip in chip_list}          
+            for future in concurrent.futures.as_completed(future_to_chip):
+                chip = future_to_chip[future]
+                pbar.update(n=1)
+    
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description='Downloading corresponding JRC permanent water tile from GEE for a given raster chip')
+    parser.add_argument(
+        "--cld_projid",
+        required=True,
+        type=str,
+        help="Cloud project id",
+    )  
+    parser.add_argument(
+        "--in_dir",
+        type=str,
+        required=True,
+        help="Input tile directory",
+    )
+    parser.add_argument(
+        "--out_dir",
+        type=str,
+        required=True,
+        help="Output folder for corresponding permanent water tiles",
+    )
+    args = parser.parse_args()
+    
+    main(args)

data/chip_set2.csv

@@ -0,0 +1,97 @@
+import numpy as np
+import os
+import rasterio
+from rasterio.warp import calculate_default_transform, reproject, Resampling,transform_bounds,transform_geom
+from rasterio.transform import Affine
+import restee as ree
+
+def tile256(raster_path,output_dir):    
+    """
+    Create raster tiles of 256 by 256
+    Arguments:
+        raster_path: input raster path
+        output_dir: output folder path   
+    Returns:
+        256 by 256 raster tiles    
+    """
+    #Make outdir if is not created
+    os.makedirs(output_dir, exist_ok=True)
+    
+    
+    with rasterio.open(raster_path) as src: 
+        width = src.width 
+        height = src.height 
+        res =  src.transform[0]
+        UL_c,UL_r = src.transform * (0, 0)   
+        band = src.read(1)       
+        kwargs = src.meta.copy()
+                   
+        for col in range(int(np.round_(width/256))):           
+            for row in range(int(np.round_(height/256))):
+                tf_tile = Affine(res, 0.0,UL_c+(256*res*col),0.0,-1*(res),UL_r+(256*(-1*(res))*row))   
+                id = int(raster_path.split(os.sep)[-2].split('_')[-1])
+                tile_name = f'{id:03d}_{row:01d}{col:01d}.tif'            
+                np_arr = np.zeros((256, 256), dtype='float32')                
+                tile_band = band[256*row:256*row+256,256*col:256*col+256]              
+                np_arr[:tile_band.shape[0], :tile_band.shape[1]] = tile_band
+                
+                kwargs.update({
+                    'transform': tf_tile,
+                    'width': 256,
+                    'height': 256,
+                    'compress': 'lzw',
+                    'dtype':'float32',})
+                                
+                with rasterio.open(os.path.join(output_dir,tile_name), "w", **kwargs) as dst:
+                    dst.write(np_arr,1)
+
+        np_arr = None 
+        tile_band = None
+        band = None
+            
+def restgee_data(input_tile,ee_img,ee_img_band,output_dir,restee_session):
+    """
+    Download GEE objectecs correspond to raster tile extent using EE REST API(restee package)
+    Arguments:  
+        input_tile: Input calibrated raster tile 
+        ee_img: GEE object
+        ee_img_band: GEE object band name
+        output_dir: output directory path
+        restee_session : rest ee session
+    Returns:
+        GEE objectecs as tiles correspoding to calibrated raster tiles
+    """
+    with rasterio.open(input_tile) as src:
+        kwargs = src.meta.copy()
+        tile_tranform = src.transform
+        tile_bounds = src.bounds
+        tile_crs = src.crs     
+        gee_domain = ree.Domain((tile_bounds[0],
+                                 tile_bounds[1],
+                                 tile_bounds[2],
+                                 tile_bounds[3]),
+                                resolution= src.transform[0],
+                                crs =str(tile_crs))
+        
+        band_utm = np.int32(ree.img_to_ndarray(restee_session, gee_domain, image=ee_img, bands=ee_img_band))
+        gdal_mask = src.dataset_mask()
+        gdal_mask = np.int32(np.where(gdal_mask == 0, gdal_mask, 1))
+        band_utm = band_utm*gdal_mask
+        
+        dst_tranform = rasterio.transform.from_bounds(tile_bounds.left, 
+                                                      tile_bounds.bottom,
+                                                      tile_bounds.right, 
+                                                      tile_bounds.top, 
+                                                      band_utm.shape[1],
+                                                      band_utm.shape[0])      
+        out_file_path = os.path.join(output_dir,os.path.basename(input_tile))
+        with rasterio.open(out_file_path, "w", **kwargs) as dst:
+            reproject(
+                source=band_utm,
+                destination=rasterio.band(dst,1),
+                src_transform=dst_tranform,
+                src_crs=tile_crs,
+                dst_transform=tile_tranform,
+                dst_crs=tile_crs,
+                resampling=Resampling.nearest)            
+