pipelines/preprocessing_pipeline.py

from prefect import task
from prefect.logging import get_run_logger
import torch
import os
from typing import List, Any
import pickle
import yaml
import re
import rasterio
import xarray as xr
import pandas as pd
import numpy as np
from datetime import datetime
from tools.historical_weather import get_historical_weather_data


with open('config/params.yml') as file:
    config = yaml.safe_load(file)

datetime_col = config['datetime_col']


### Helper functions

@task(task_run_name="get_consortia")
def get_consortia() -> List[str]:
    """
    Retrieves the list of consortia from the configuration.

    Returns:
        List[str]: List of consortia names.
    """
    logger = get_run_logger()
    consortia = config['consortia']
    logger.info(f"We work with the following consortia: {consortia}")
    return consortia


@task(tags=['write_to_file_{df.name}'], retries=3)
def save_to_file(df: Any, output_file: str) -> None:
    """
    Saves DataFrames or GeoDataFrames to a file in the specified format.
    Args:
        df (Any): DataFrame or GeoDataFrame to save.
        output_file (str): Path to the output file.
        file_format (str, optional): Format to save ('parquet', 'geojson', 'zarr', 'pickle', 'netcdf'/'nc' and torch 'pt' are supported).
    """
    logger = get_run_logger()
    logger.info(f'Saving data to {output_file}...')

    file_format = output_file.split('.')[-1]

    logger.info(f'The output file has format {file_format}.')

    folder = '//'.join(output_file.split('//')[:-1])
    os.makedirs(folder, exist_ok=True)

    if file_format == 'parquet':
        df.to_parquet(output_file)
    elif file_format == 'geojson':
        df.to_file(output_file, driver="GeoJSON")
    elif file_format == 'zarr':
        df.to_zarr(output_file, mode='w')
    elif file_format == 'pickle':
        with open(output_file, 'wb') as handle:
            pickle.dump(df, handle)
    elif file_format == 'netcdf' or file_format == 'nc':
        df.to_netcdf(output_file, engine='netcdf4', format='NETCDF4')
    elif file_format == 'pt':
        torch.save(df, output_file)
    else:
        raise Exception('Format specified is not supported.')
    
    return


@task(task_run_name='read_satellite_data')
def read_satellite_data(consortium_name):
    logger = get_run_logger()
    logger.info('Collecting satellite data...')

    dir = f"data//01_raw//{config['consortia_data_folders'][consortium_name]}//satellite_data"
    file_paths = sorted([f for f in os.listdir(dir) if f.lower().endswith((".tif", ".tiff"))])

    data = []
    dates = []

    # Regex pattern for datetime in filenames
    pattern = r'\d{4}-\d{2}-\d{2}T\d{2}_\d{2}_\d{2}'

    # Read data from each file
    for file_path in file_paths:
        dates.append(re.search(pattern, file_path).group())
        with rasterio.open(os.path.join(dir, file_path)) as src:
            transform = src.transform
            bands = src.count
            band_data = src.read(range(1, bands + 1))  
            data.append(band_data)

    # Convert extracted dates
    datetimes = [datetime.strptime(date, '%Y-%m-%dT%H_%M_%S') for date in dates]

    ############################ CHANGE 1 ###################################
    # Use the last opened raster for spatial info (all files share same grid)
    height, width = band_data.shape[1], band_data.shape[2]
    
    ############################ CHANGE 2 ###################################
    # INSTEAD OF NESTED LOOP WE EXTRACT THE PIXEL CENTER COORDINATES based on the trasnformation
    x_coords = transform[2] + (np.arange(width) + 0.5) * transform[0]
    y_coords = transform[5] + (np.arange(height) + 0.5) * transform[4]
    
    ########################## CHANGE 3 #####################################
    # Build the main DataArray with CENTER COORDINATES ASSOCIATED TO EACH PIXEL 
    satellite_data = xr.DataArray(
        data,
        dims=['time', 'band', 'y', 'x'],
        coords={
            'time': datetimes,
            'x': x_coords,
            'y': y_coords,
        },
        name='satellite_data'
    )
    
    #### this is still creating what you wanted, although i dont think it is necessary (OPTIONAL)
    # it is not contributing to any potential calculation  
    x_edges = transform[2] + np.arange(width + 1) * transform[0]
    y_edges = transform[5] + np.arange(height + 1) * transform[4]

    left_edges = np.tile(x_edges[:-1], (height, 1))
    right_edges = np.tile(x_edges[1:],  (height, 1))
    top_edges = np.tile(y_edges[:-1][:, None], (1, width))
    bottom_edges = np.tile(y_edges[1:][:, None],  (1, width))

    # Build Dataset with all information
    dataset = xr.Dataset(
        {
            "satellite_data": satellite_data,
            "pixel_bounds_left": (("y", "x"), left_edges),
            "pixel_bounds_right": (("y", "x"), right_edges),
            "pixel_bounds_top": (("y", "x"), top_edges),
            "pixel_bounds_bottom": (("y", "x"), bottom_edges),
        }
    )

    return dataset


@task(task_run_name='update_historical_weather_data_{consortium_name}')
def update_historical_weather_data(consortium_name, location_ids):
    """
    Update historical weather data for a consortium.

    Parameters:
    -----------
    consortium_name : str
        Name of the consortium
    location_ids : pd.DataFrame
        DataFrame with columns: ['datastream_name', 'datastream_id', 'x', 'y']
        where x = latitude, y = longitude

    Returns:
    --------
    pd.DataFrame with historical weather data
    """

    logger = get_run_logger()

    start_date = config['start_date']
    end_date = config['end_date']

    childs = next(os.walk('data//01_raw//'))
    initial_file_exists = False
    for child in childs[-1]:
        if child.startswith(f'historical_weather_data_{consortium_name}_'):
            initial_file_exists = True
            current_dates = child.split(f'historical_weather_data_{consortium_name}_')[-1]
            current_dates = current_dates.split('.parquet')[0].split('_')
            current_start_date = datetime.strptime(current_dates[0], '%Y-%m-%d').date()
            current_end_date = datetime.strptime(current_dates[1], '%Y-%m-%d').date()

    if f'historical_weather_data_{consortium_name}_{start_date}_{end_date}.parquet' in childs[-1]:
        logger.info(f'Consortium {consortium_name}: historical weather data is already fully downloaded!')
        data = pd.read_parquet(
            f'data//01_raw//historical_weather_data_{consortium_name}_{current_start_date}_{current_end_date}.parquet')
    else:
        logger.info(f'Consortium {consortium_name}: some historical weather data need to be downloaded...')

        if not initial_file_exists:
            data = get_historical_weather_data(location_ids, start_date=start_date, end_date=end_date)
            data[datetime_col] = pd.to_datetime(data['datetime'].dt.tz_localize(None))
            data = data.drop(columns=['datetime'])
        else:
            if current_start_date > start_date:
                logger.info('Need to download some data in the past...')

                new_data = get_historical_weather_data(location_ids, start_date=start_date, end_date=current_start_date - datetime.timedelta(days=1))
                new_data[datetime_col] = pd.to_datetime(new_data['datetime'].dt.tz_localize(None))
                new_data = new_data.drop(columns=['datetime'])

                current_data = pd.read_parquet(
                    f'data//01_raw//historical_weather_data_{consortium_name}_{current_start_date}_{current_end_date}.parquet')

                data = pd.concat(
                    [new_data, current_data],
                    axis=0
                ).drop_duplicates(subset=[datetime_col]).sort_values('datetime')

            if current_end_date < end_date:
                logger.info('Need to download some more recent data...')

                new_data = get_historical_weather_data(location_ids, start_date=current_end_date, end_date=end_date)
                new_data[datetime_col] = pd.to_datetime(new_data['datetime'].dt.tz_localize(None))
                new_data = new_data.drop(columns=['datetime'])

                current_data = pd.read_parquet(
                    f'data//01_raw//historical_weather_data_{consortium_name}_{current_start_date}_{current_end_date}.parquet')

                data = pd.concat(
                    [new_data, current_data],
                    axis=0
                ).drop_duplicates(subset=[datetime_col, 'datastream_name']).sort_values(datetime_col)

            logger.info('Old historical weather data are being removed...')
            os.remove(
                f'data//01_raw//historical_weather_data_{consortium_name}_{current_start_date}_{current_end_date}.parquet')

        logger.info('New historical weather data are being saved...')
        save_to_file(df=data,
                     output_file=f'data//01_raw//historical_weather_data_{consortium_name}_{start_date}_{end_date}.parquet')
        logger.info('Done!')

    return data