previtus

Weighing the loss of the model during training by the MF (Matched Filter, any variant) methane enhancement product. (Somewhat akin to the way boundaries between cells were weighted up in the original U-Net paper.) This method has showned to work well for the STARCOP paper.


Code refs:
- in STARCOP code we cooked these weight files as extra product (now available as the "weight_mag1c.tif" file in https://zenodo.org/records/7863343): https://github.com/spaceml-org/STARCOP/blob/c4789268a3fa0395f92357429052f6f5fc748acb/starcop/data/feature_extration.py#L32
- in later work we compute these during training (as it's a very fast op), example here in mf_weighing_example.py

previtus

import matplotlib.pyplot as plt
import rasterio as rio
from georeader.geotensor import GeoTensor
from georeader.save import save_cog
import georeader
from georeader import read
from georeader.rasterio_reader import RasterioReader
from georeader.readers import emit
import numpy as np
import os

previtus

class DataNormalizerLogManual():
    def __init__(self):
        self.setup()
    def setup(self):
        # These were edited to work with the 10 bands we had in Wildfires project (FireCLR)
        # only use 10m resolution bands (10): Blue (B2), Green (B3), Red (B4), VNIR (B5),
        # VNIR (B6), VNIR (B7), NIR (B8), VNIR (B8a), SWIR (B11), SWIR (B12) combining
        self.BANDS_S2_BRIEF = ["B2", "B3", "B4", "B5", "B6", "B7", "B8", "B8A", "B11", "B12"]

previtus

for i in *.pdf ; do ps2pdf "$i" "${i%.*}c.pdf" ; done

# for examle this goes from 15.7 MB to 3.9 MB ... without really visibly loosing quality!

previtus

def calc_ime(plume_da):
    molar_volume = 22.4 # L/mol at STP
    molar_mass_ch4 = 0.01604 #kg/mol

    kg = plume_da * (1/1e6) * (60*60) * (1000) * (1/molar_volume) * molar_mass_ch4
    ime = np.nansum(kg)
    return ime

previtus

import os
import pytorch_lightning as pl
from typing import Optional, Tuple, List

import kornia.augmentation as K
from torch.utils.data import DataLoader, WeightedRandomSampler
from starcop.data import dataset
import pandas as pd
from . import feature_extration
import rasterio.windows

previtus

# more in https://github.com/spaceml-org/georeader/blob/main/notebooks/reading_overlapping_sentinel2_aviris.ipynb
from georeader.rasterio_reader import RasterioReader
from georeader import read

def same_location_as_source_from_target(source_tif, target_tif, show=False):
    
    src_reader = RasterioReader(source_tif)
    src_reader_in_memory = src_reader.load()

    target_reader = RasterioReader(target_tif)

previtus

# from an example file with a plume event
#    EMIT_L1B_RAD_001_20220823T074504_2223505_021.nc


EMIT_wavelenghts = ['381.00558', '388.4092', '395.81583', '403.2254', '410.638', '418.0536', '425.47214', '432.8927', '440.31726', '447.7428', '455.17035', '462.59888', '470.0304', '477.46292', '484.89743', '492.33292', '499.77142', '507.2099', '514.6504', '522.0909', '529.5333', '536.9768', '544.42126', '551.8667', '559.3142', '566.7616', '574.20905', '581.6585', '589.108', '596.55835', '604.0098', '611.4622', '618.9146', '626.36804', '633.8215', '641.2759', '648.7303', '656.1857', '663.6411', '671.09753', '678.5539', '686.0103', '693.4677', '700.9251', '708.38354', '715.84094', '723.2993', '730.7587', '738.2171', '745.6765', '753.1359', '760.5963', '768.0557', '775.5161', '782.97754', '790.4379', '797.89935', '805.36176', '812.8232', '820.2846', '827.746', '835.2074', '842.66986', '850.1313', '857.5937', '865.0551', '872.5176', '879.98004', '887.44147', '894.90393', '902.3664', '909.82886', '917.2913', '924.7538

previtus

previtus

The run.sh bash file should do everything for you, you just need to have an audio sample (in wav format) and the Nvidia docker installed (https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html#docker)

Download the run.sh command and give it executible rights:
chmod +x run.sh

Then simply run

./run.sh sample.wav 16

This will train a model for 16 epochs and then run an interactive real-time handler with it.