"""Contains classes for image transformations specific for BreizhCrops dataset."""
import torch
import numpy as np
from ..base import BaseTransforms
BANDS = {
"L1C": [
"B1",
"B10",
"B11",
"B12",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B9",
"QA10",
"QA20",
"QA60",
"doa",
"label",
"id",
],
"L2A": [
"doa",
"id",
"code_cultu",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B11",
"B12",
"CLD",
"EDG",
"SAT",
],
}
SELECTED_BANDS = {
"L1C": [
"B1",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B9",
"B10",
"B11",
"B12",
"QA10",
"QA20",
"QA60",
"doa",
],
"L2A": [
"doa",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B11",
"B12",
"CLD",
"EDG",
"SAT",
],
}
[docs]class SelectBands(BaseTransforms):
"""
A class used to select and process spectral bands from satellite data.
:param level: satellite data level to be processed ("L1C" or "L2A")
:type level: str
.. note::
This class requires a level argument at initialization.
This should be one of the predefined satellite data levels ("L1C" or "L2A").
"""
configurables = ["level"]
def __init__(self, *args, **kwargs):
"""
Initialize the SelectBands class by setting the satellite data level.
"""
BaseTransforms.__init__(self, *args, **kwargs)
self.level = kwargs["level"]
# padded_value = PADDING_VALUE
self.sequencelength = 45
bands = BANDS[self.level]
if self.level == "L1C":
selected_bands = [
"B1",
"B10",
"B11",
"B12",
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B9",
]
elif self.level == "L2A":
selected_bands = [
"B2",
"B3",
"B4",
"B5",
"B6",
"B7",
"B8",
"B8A",
"B11",
"B12",
]
self.selected_band_idxs = np.array([bands.index(b) for b in selected_bands])
def __call__(self, input, target=None):
"""
Process the input and target data, apply transformation and return the result.
Transformation includes selecting bands, scaling, and replacing short seqences (if necessary).
:param input: input data to be processed
:type input: tuple
:param target: target data, defaults to None
:type target: tensor, optional
:return: processed input and target data
:rtype: tuple
"""
# x = x[x[:, 0] != padded_value, :] # remove padded values
# choose selected bands
x, y = input
x = x[:, self.selected_band_idxs] * 1e-4 # scale reflectances to 0-1
# choose with replacement if sequencelength smaller als choose_t
replace = False if x.shape[0] >= self.sequencelength else True
idxs = np.random.choice(x.shape[0], self.sequencelength, replace=replace)
idxs.sort()
x = x[idxs]
return torch.from_numpy(x).type(torch.FloatTensor), torch.tensor(
y, dtype=torch.long
)
