# Copyright (c) 2025 GeoBenchV2. All rights reserved.
# Licensed under the Apache License 2.0.
"""PASTIS DataModule."""
import os
from collections.abc import Callable, Sequence
from typing import Any
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import tacoreader
import torch
import torch.nn as nn
from einops import rearrange
from torchgeo.datasets.utils import percentile_normalization
from geobench_v2.datasets import GeoBenchPASTIS
from .base import GeoBenchSegmentationDataModule
# TODO add timeseries argument
[docs]
class GeoBenchPASTISDataModule(GeoBenchSegmentationDataModule):
"""GeoBench PASIS Data Module."""
[docs]
def __init__(
self,
img_size: int = 128,
band_order: Sequence[float | str] = GeoBenchPASTIS.band_default_order,
batch_size: int = 32,
eval_batch_size: int = 64,
num_workers: int = 0,
collate_fn: Callable | None = None,
train_augmentations: nn.Module | None = None,
eval_augmentations: nn.Module | None = None,
pin_memory: bool = False,
**kwargs: Any,
) -> None:
"""Initialize GeoBench PASIS DataModule.
Args:
img_size: Image size
band_order: The order of bands to return in the sample
batch_size: Batch size during training
eval_batch_size: Evaluation batch size
num_workers: Number of workers
collate_fn: Collate function
train_augmentations: Transforms/Augmentations to apply during training, they will be applied
at the sample level and should include normalization. See :meth:`define_augmentations`
for the default transformation.
eval_augmentations: Transforms/Augmentations to apply during evaluation, they will be applied
at the sample level and should include normalization. See :meth:`define_augmentations`
for the default transformation.
pin_memory: Pin memory
**kwargs: Additional keyword arguments for:class:`~geobench_v2.datasets.pastis.GeoBenchPASTIS`.
"""
super().__init__(
dataset_class=GeoBenchPASTIS,
band_order=band_order,
img_size=img_size,
batch_size=batch_size,
eval_batch_size=eval_batch_size,
num_workers=num_workers,
collate_fn=collate_fn,
train_augmentations=train_augmentations,
eval_augmentations=eval_augmentations,
pin_memory=pin_memory,
**kwargs,
)
[docs]
def visualize_batch(
self, batch: dict[str, Any] | None = None, split: str = "train"
) -> tuple[Any, dict[str, Any]]:
"""Visualize a batch of data.
Args:
batch: Optional batch of data. If not provided, a batch will be fetched from the dataloader
split: One of 'train', 'validation', 'test'
Returns:
The matplotlib figure and the batch of data
Raises:
AssertionError: If bands needed for plotting are missing
"""
if split == "train":
batch = next(iter(self.train_dataloader()))
elif split == "validation":
batch = next(iter(self.val_dataloader()))
else:
batch = next(iter(self.test_dataloader()))
for k, v in batch.items():
orig_dim = v.dim()
if orig_dim == 4: # CxTxHxW -> TxCxHxW
batch[k] = batch[k].permute(1, 0, 2, 3)
elif orig_dim == 5: # BxCxTxHxW -> BxTxCxHxW
batch[k] = batch[k].permute(0, 2, 1, 3, 4)
if hasattr(self.data_normalizer, "unnormalize"):
batch = self.data_normalizer.unnormalize(batch)
batch_size = batch["mask"].shape[0]
n_samples = min(8, batch_size)
indices = torch.randperm(batch_size)[:n_samples]
# Collect modality images and determine timesteps per modality
modalities = {}
timesteps_per_mod: dict[str, int] = {}
for mod in self.band_order.keys():
mod_plot_bands = self.dataset_band_config.modalities[mod].plot_bands
missing_bands = [
band for band in mod_plot_bands if band not in self.band_order[mod]
]
if missing_bands:
raise AssertionError(
f"Plotting bands {missing_bands} for modality '{mod}' not found in band_order {self.band_order[mod]}"
)
# Get plot indices for bands that exist
mod_plot_indices = [
self.band_order[mod].index(band) for band in mod_plot_bands
]
tensor = batch[f"image_{mod}"]
if tensor.ndim == 5:
# time series data [B, T, C H, W] -> [b, t, h, w, c]
mod_images = tensor[indices][:, :, mod_plot_indices, :, :]
mod_images = (
rearrange(mod_images, "b t c h w -> b t h w c").cpu().numpy()
)
timesteps_per_mod[mod] = mod_images.shape[1]
else:
# single image data [B, C, H, W] -> [b, 1, h, w, c]
mod_images = tensor[indices][:, mod_plot_indices, :, :]
mod_images = rearrange(mod_images, "b c h w -> b 1 h w c").cpu().numpy()
timesteps_per_mod[mod] = 1
modalities[mod] = mod_images
# Global layout: for each sample, stack timesteps vertically
t_max = max(timesteps_per_mod.values()) if timesteps_per_mod else 1
num_modalities = len(modalities) + 1 # +1 for mask column
fig, axes = plt.subplots(
n_samples * t_max,
num_modalities,
figsize=(num_modalities * 4, 3 * n_samples * t_max),
gridspec_kw={"width_ratios": num_modalities * [1]},
)
if axes.ndim == 1:
axes = axes.reshape(1, -1)
# Add row labels for timesteps (t=0, t=1, ...)
for i in range(n_samples):
for t in range(t_max):
row_idx = i * t_max + t
ax_label = axes[row_idx, 0]
ax_label.text(
-0.06,
0.5,
f"t={t}",
transform=ax_label.transAxes,
va="center",
ha="right",
fontsize=10,
)
masks = batch["mask"][indices]
unique_classes = torch.unique(masks).cpu().numpy()
unique_classes = [
int(cls) for cls in unique_classes if cls < len(self.class_names)
]
# use tab20 colormap to color the unique classes found
cmap = plt.cm.tab20
colors = {
i: cmap(i) for i in range(len(self.class_names)) if i in unique_classes
}
class_cmap = plt.cm.colors.ListedColormap(colors.values())
# Build legend handles once
legend_elements = []
for cls in unique_classes:
if cls < len(self.class_names) and cls in colors:
legend_elements.append(
plt.Rectangle(
(0, 0),
1,
1,
color=colors[cls],
label=f"{self.class_names[cls]}",
)
)
# Plot
modality_list = list(modalities.keys())
for i in range(n_samples):
for j, mod in enumerate(modality_list):
mod_images = modalities[mod] # [b, t, h, w, c]
t_len = timesteps_per_mod[mod]
for t in range(t_max):
row_idx = i * t_max + t
ax = axes[row_idx, j]
if t < t_len:
plot_img = mod_images[i, t]
# Special handling for SAR style if applicable
if mod in ["s1_asc", "s1_desc"] and plot_img.shape[-1] >= 2:
vv = plot_img[..., 0]
vh = plot_img[..., 1]
vv = percentile_normalization(vv, lower=2, upper=98)
vh = percentile_normalization(vh, lower=2, upper=98)
ratio = np.divide(
vv, vh, out=np.zeros_like(vv), where=vh != 0
)
vv = np.clip(vv / 0.3, a_min=0, a_max=1)
vh = np.clip(vh / 0.05, a_min=0, a_max=1)
ratio = np.clip(ratio / 25, a_min=0, a_max=1)
img = np.stack((vv, vh, ratio), axis=2)
else:
img = percentile_normalization(plot_img, lower=2, upper=98)
ax.imshow(img)
if i == 0 and t == 0:
ax.set_title(f"{mod} image", fontsize=16)
else:
ax.axis("off")
ax.axis("off")
# Mask column (last)
for t in range(t_max):
row_idx = i * t_max + t
ax = axes[row_idx, -1]
if t == 0:
mask_img = masks[i].cpu().numpy()
ax.imshow(
mask_img,
cmap=class_cmap,
vmin=0,
vmax=max(unique_classes) if unique_classes else 1,
)
ax.set_title("Mask", fontsize=16)
else:
ax.axis("off")
ax.axis("off")
# Compute legend layout
n_classes = len(legend_elements)
ncols = min(6, max(1, n_classes))
legend = fig.legend(
handles=legend_elements,
loc="upper center",
bbox_to_anchor=(0.5, 1.0),
ncol=ncols,
fontsize=12.0,
title="Classes",
title_fontsize=14,
frameon=False,
# tighter legend paddings reduce vertical space
borderaxespad=0.0,
handlelength=0.9,
handletextpad=0.3,
columnspacing=0.8,
labelspacing=0.2,
)
# Render once to get accurate legend bbox
fig.canvas.draw()
renderer = fig.canvas.get_renderer()
bbox = legend.get_window_extent(renderer=renderer)
fig_w, fig_h = fig.get_size_inches()
legend_h_frac = (bbox.height / (fig_h * fig.dpi)) + 0.006
fig.subplots_adjust(left=0.02, right=0.98, bottom=0.02, top=1.0 - legend_h_frac)
return fig, batch
[docs]
def visualize_geolocation_distribution(self) -> None:
"""Visualize the geolocation distribution of the dataset."""
pass