Source code for geobench_v2.datamodules.dynamic_earthnet

# Copyright (c) 2025 GeoBenchV2. All rights reserved.
# Licensed under the Apache License 2.0.

"""DynamicEarthNet DataModule."""

import os
from collections.abc import Callable, Sequence
from typing import Any, Literal

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 GeoBenchDynamicEarthNet

from .base import GeoBenchSegmentationDataModule
from .utils import TimeSeriesResize


# TODO add timeseries argument
[docs] class GeoBenchDynamicEarthNetDataModule(GeoBenchSegmentationDataModule): """GeoBench DynamicEarthNet Data Module.""" # TODO img_size will change to 512
[docs] def __init__( self, img_size: int = 512, band_order: Sequence[float | str] = GeoBenchDynamicEarthNet.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 DynamicEarthNet 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.dynamic_earthnet.GeoBenchDynamicEarthNet` """ super().__init__( dataset_class=GeoBenchDynamicEarthNet, 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 load_metadata(self) -> pd.DataFrame: """Load metadata file. Returns: pandas DataFrame with metadata. """ return tacoreader.load( [ os.path.join(self.kwargs["root"], f) for f in GeoBenchDynamicEarthNet.paths ] )
[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 """ if batch is None: 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 == 5: # BxCxTxHxW -> BxTxCxHxW batch[k] = v.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(4, batch_size) indices = torch.randperm(batch_size)[:n_samples] # Collect modality images and timesteps per modality modalities: dict[str, np.ndarray] = {} 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 # Layout: for each sample, stack timesteps vertically t_max = max(timesteps_per_mod.values()) if timesteps_per_mod else 1 num_columns = len(modalities) + 1 # +1 for mask fig, axes = plt.subplots( n_samples * t_max, num_columns, figsize=(num_columns * 4.0, 2.6 * n_samples * t_max), # slightly tighter gridspec_kw={ "width_ratios": num_columns * [1], "wspace": 0.02, "hspace": 0.02, }, ) if axes.ndim == 1: axes = axes.reshape(1, -1) # Add timestep row labels (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, ) # Prepare mask and legend masks = batch["mask"][indices] unique_classes = torch.unique(masks).cpu().numpy() unique_classes = sorted({int(c) for c in unique_classes if int(c) >= 0}) class_names = getattr(self, "class_names", None) if not class_names or max(unique_classes, default=0) >= len(class_names): # Fallback numeric names class_names = [ f"class {i}" for i in range(max(unique_classes, default=-1) + 1) ] cmap = plt.cm.tab20 colors = {i: cmap(i % 20) for i in unique_classes} class_cmap = plt.cm.colors.ListedColormap( [colors[i] for i in unique_classes] or [(0, 0, 0, 1)] ) legend_elements = [] for cls_id in unique_classes: legend_elements.append( plt.Rectangle( (0, 0), 1, 1, color=colors[cls_id], label=class_names[cls_id] if cls_id < len(class_names) else f"class {cls_id}", ) ) # Plot modalities for i in range(n_samples): for j, mod in enumerate(modalities.keys()): 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: img = mod_images[i, t] # h, w, c ax.imshow(percentile_normalization(img, lower=2, upper=98)) if i == 0 and t == 0: ax.set_title(f"{mod.upper()}", fontsize=13) else: ax.axis("off") ax.axis("off") # Mask column at the end (only first t shown) for t in range(t_max): row_idx = i * t_max + t ax = axes[row_idx, -1] if t == 0: mask_img = masks[i].squeeze(0).cpu().numpy() vmax = max(unique_classes) if unique_classes else 1 ax.imshow( mask_img, cmap=class_cmap, vmin=min(unique_classes) if unique_classes else 0, vmax=vmax, ) ax.set_title("Label", fontsize=13) 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_geospatial_distribution( self, split_column: str = "tortilla:data_split", buffer_degrees: float = 5.0, sample_fraction: float | None = None, scale: Literal["10m", "50m", "110m"] = "50m", alpha: float = 0.8, s: float = 10, ) -> plt.Figure: """Visualize the geospatial distribution of dataset samples on a map. Creates a plot showing the geographic locations of samples, colored by dataset split (train, validation, test, extra_test). This helps to understand the spatial distribution and potential geographic biases in the dataset. Args: split_column: Column name in the metadata DataFrame that indicates the dataset split. buffer_degrees: Buffer around the data extent in degrees. sample_fraction: Optional fraction of samples to plot (0.0-1.0) for performance with large datasets. scale: Scale of cartopy features (e.g., '10m', '50m', '110m'). alpha: Transparency of plotted points s: Size of plotted points. Returns: A matplotlib Figure object with the geospatial distribution plot. """ return super().visualize_geospatial_distribution( split_column=split_column, buffer_degrees=buffer_degrees, sample_fraction=sample_fraction, scale=scale, alpha=alpha, s=s, )
[docs] def setup_image_size_transforms(self) -> tuple[nn.Module, nn.Module, nn.Module]: """Setup image resizing transforms for train, val, test. Image resizing and normalization happens on dataset level on individual data samples. """ return ( TimeSeriesResize(self.img_size), TimeSeriesResize(self.img_size), TimeSeriesResize(self.img_size), )