Skip to content
Snippets Groups Projects
viirs_surfrad.py 7.46 KiB
import numpy as np
import h5py
from util.util import get_grid_values, get_grid_values_all, is_night, is_day, compute_lwc_iwc, get_fill_attrs
import glob
import os
from aeolus.datasource import CLAVRx_VIIRS
from icing.moon_phase import *
from pathlib import Path


target_param = 'cloud_probability'
# target_param = 'cld_opd_dcomp'

group_name_i = 'super/'
group_name_m = 'orig/'

solzen_name = group_name_m + 'solar_zenith'

label_params = [group_name_i+target_param]
data_params = [group_name_m+'temp_11_0um', group_name_m+'refl_0_65um', group_name_m+target_param]


def keep_tile(param, param_s, tile):
    k = param_s.index(param)
    grd_k = tile[k, ].copy()

    if target_param == 'cloud_probability':
        grd_k = process_cld_prob_(grd_k)
    elif target_param == 'cld_opd_dcomp':
        grd_k = process_cld_opd_(grd_k)

    if grd_k is not None:
        tile[k, ] = grd_k
        return tile
    else:
        return None


def process_cld_prob_(grd_k):
    keep = np.invert(np.isnan(grd_k))
    num_keep = np.sum(keep)
    if num_keep / grd_k.size < 0.98:
        return None
    keep_clr = np.where(keep, grd_k < 0.20, False)
    frac_keep = np.sum(keep_clr)/num_keep
    if not (0.40 < frac_keep < 0.60):
        return None
    grd_k = np.where(np.invert(keep), 0, grd_k)  # Convert NaNs to 0
    return grd_k


def process_cld_opd_(grd_k):
    keep = np.invert(np.isnan(grd_k))
    num_keep = np.sum(keep)
    if num_keep / grd_k.size < 0.98:
        return None
    grd_k = np.where(np.invert(keep), 0, grd_k)
    keep = np.where(keep, np.logical_and(0.1 < grd_k, grd_k < 158.0), False)
    if np.sum(keep)/num_keep < 0.50:
        return None
    return grd_k


def run_all(directory, out_directory, day_night='ANY', pattern='clavrx_*.nc', start=10):
    cnt = start
    total_num_train_samples = 0
    total_num_valid_samples = 0
    num_keep_x_tiles = 8

    path = directory + '**' + '/' + pattern

    data_files = glob.glob(path, recursive=True)

    label_valid_tiles = []
    label_train_tiles = []
    data_valid_tiles = []
    data_train_tiles = []
    f_cnt = 0

    num_files = len(data_files)
    print('Start, number of files: ', num_files)

    for idx, data_f in enumerate(data_files):
        # if idx % 4 == 0:  # if we want to skip some files
        if True:
            try:
                h5f = h5py.File(data_f, 'r')
            except:
                print('cant open file: ', data_f)
                continue

            try:
                run(h5f, data_params, data_train_tiles, data_valid_tiles,
                    label_params, label_train_tiles, label_valid_tiles,
                    num_keep_x_tiles=num_keep_x_tiles, tile_width=64, kernel_size=7, day_night=day_night)
            except Exception as e:
                print(e)
                h5f.close()
                continue
            print(data_f)
            f_cnt += 1
            h5f.close()

            if len(data_train_tiles) == 0 and len(data_valid_tiles) == 0:
                continue

            if (f_cnt % 10) == 0:
                num_valid_samples = 0
                if len(data_valid_tiles) > 0:
                    label_valid = np.stack(label_valid_tiles)
                    data_valid = np.stack(data_valid_tiles)
                    np.save(out_directory + 'data_valid_' + str(cnt), data_valid)
                    np.save(out_directory + 'label_valid_' + str(cnt), label_valid)
                    num_valid_samples = data_valid.shape[0]

                num_train_samples = 0
                if len(data_train_tiles) > 0:
                    label_train = np.stack(label_train_tiles)
                    data_train = np.stack(data_train_tiles)
                    np.save(out_directory + 'label_train_' + str(cnt), label_train)
                    np.save(out_directory + 'data_train_' + str(cnt), data_train)
                    num_train_samples = data_train.shape[0]

                label_valid_tiles = []
                label_train_tiles = []
                data_valid_tiles = []
                data_train_tiles = []

                print('  num_train_samples, num_valid_samples, progress % : ', num_train_samples, num_valid_samples, int((f_cnt/num_files)*100))
                total_num_train_samples += num_train_samples
                total_num_valid_samples += num_valid_samples
                print('total_num_train_samples, total_num_valid_samples: ', total_num_train_samples, total_num_valid_samples)
                print('--------------------------------------------------')

                cnt += 1

    print('** total_num_train_samples, total_num_valid_samples: ', total_num_train_samples, total_num_valid_samples)


#  tile_width: Must be even!
#  kernel_size: Must be odd!
def run(h5f, param_s, train_tiles, valid_tiles, lbl_param_s, lbl_train_tiles, lbl_valid_tiles,
        num_keep_x_tiles=8, tile_width=64, kernel_size=3, day_night='DAY'):

    border = int((kernel_size - 1)/2) + 1  # Need to add for interpolation with no edge effects

    param_name = param_s[0]

    num_lines = h5f[param_name].shape[0]
    num_pixels = h5f[param_name].shape[1]  # Must be even

    if day_night != 'BOTH':
        solzen = get_grid_values(h5f, solzen_name, 0, 0, None, num_lines, num_pixels)

    grd_s = []
    for param in param_s:
        try:
            grd = get_grid_values(h5f, param, 0, 0, None, num_lines, num_pixels)
            grd_s.append(grd)
        except Exception as e:
            print(e)
            return
    data = np.stack(grd_s)

    grd_s = []
    for param in lbl_param_s:
        try:
            grd = get_grid_values(h5f, param, 0, 0, None, num_lines*2, num_pixels*2)
            grd_s.append(grd)
        except Exception as e:
            print(e)
            return
    label = np.stack(grd_s)

    tile_width += 2 * border

    i_skip = tile_width
    j_skip = tile_width
    i_start = int(num_pixels / 2) - int((num_keep_x_tiles * tile_width) / 2)
    j_start = 0

    num_keep_y_tiles = int(num_lines / tile_width) - 3

    num_y_valid = int(num_keep_y_tiles * 0.15) + 1
    num_y_train = num_keep_y_tiles - num_y_valid - 1

    for j in range(num_y_train):
        j_a = j_start + j * j_skip
        j_b = j_a + tile_width

        for i in range(num_keep_x_tiles):
            i_a = i_start + i * i_skip
            i_b = i_a + tile_width

            if day_night == 'DAY' and not is_day(solzen[j_a:j_b, i_a:i_b]):
                continue
            elif day_night == 'NIGHT' and is_day(solzen[j_a:j_b, i_a:i_b]):
                continue

            nda = data[:, j_a:j_b, i_a:i_b]
            nda_lbl = label[:, j_a*2:j_b*2, i_a*2:i_b*2]
            nda_lbl = keep_tile(group_name_i+target_param, lbl_param_s, nda_lbl)

            if nda_lbl is not None:
                train_tiles.append(nda)
                lbl_train_tiles.append(nda_lbl)

    j_start = num_y_train * tile_width + 2*tile_width
    for j in range(num_y_valid):
        j_a = j_start + j * j_skip
        j_b = j_a + tile_width

        for i in range(num_keep_x_tiles):
            i_a = i_start + i * i_skip
            i_b = i_a + tile_width

            if day_night == 'DAY' and not is_day(solzen[j_a:j_b, i_a:i_b]):
                continue
            elif day_night == 'NIGHT' and is_day(solzen[j_a:j_b, i_a:i_b]):
                continue

            nda = data[:, j_a:j_b, i_a:i_b]
            nda_lbl = label[:, j_a * 2:j_b * 2, i_a * 2:i_b * 2]
            nda_lbl = keep_tile(group_name_i+target_param, lbl_param_s, nda_lbl)

            if nda_lbl is not None:
                valid_tiles.append(nda)
                lbl_valid_tiles.append(nda_lbl)