Data: Datasets, Samplers, Transforms, Losses

keras_climate.datasets / keras_climate.samplers / keras_climate.transforms / keras_climate.losses are a multi-backend Keras port of torchgeo's data-loading stack — the same geospatial dataset framework and catalog of 190+ concrete datasets (EuroSAT, So2Sat, Sentinel, SpaceNet, Copernicus-Bench, ...), reworked so sample values are Keras tensors (keras.ops.convert_to_tensor) instead of torch.Tensors, and so datasets work unchanged under the TensorFlow, JAX, and PyTorch Keras 3 backends.

Channels-last, not channels-first

Unlike torchgeo (C x H x W), image/mask samples here use the channels-last layout (H x W x C) to match the rest of keras_climate and Keras's own default Conv2D data format.

Two kinds of dataset

• NonGeoDataset (and NonGeoClassificationDataset) — a plain index-based dataset with no geospatial metadata, e.g. a folder of pre-chipped tiles. Index directly with dataset[i], same as any torch.utils.data.Dataset-style class.
• GeoDataset (and RasterDataset / VectorDataset / XarrayDataset) — a dataset backed by georeferenced files (rasters, vector layers, xarray-backed cubes), indexed by a BoundingBox query rather than a plain integer, so a single dataset can span an arbitrary tiled/mosaicked area of interest. IntersectionDataset/UnionDataset compose two GeoDatasets (e.g. imagery + a label raster) by their spatiotemporal extent.

Every sample is a dict (typically with "image" and, for labeled datasets, "mask" or "label" keys) of Keras tensors.

Classification example (NonGeoDataset)

from keras_climate.datasets import EuroSAT

train_ds = EuroSAT(root="data/eurosat", split="train", download=True)
sample = train_ds[0]
print(sample["image"].shape, sample["label"])  # (64, 64, 13), a scalar int64 tensor
print(train_ds.classes)

Tiled/geospatial example (GeoDataset + samplers)

keras_climate.samplers yields BoundingBox queries over a GeoDataset's spatiotemporal index — purely index arithmetic (pandas/geopandas/shapely), with no dependency on the active Keras backend at all:

from keras_climate.samplers import RandomGeoSampler
from keras_climate.datasets import stack_samples

sampler = RandomGeoSampler(some_raster_dataset, size=256, length=1000)

samples = [some_raster_dataset[bbox] for bbox in list(sampler)[:8]]
batch = stack_samples(samples)  # {"image": (8, 256, 256, C), ...}

GridGeoSampler tiles a region of interest exhaustively (e.g. for inference over a full scene); RandomBatchGeoSampler yields whole batches of bounding boxes at once. See Samplers API for the full list, including temporal samplers (RandomTimestampSampler, ...) for SpatioTemporalSampler-backed datasets.

Feeding model.fit()

Neither keras_climate.datasets nor keras_climate.samplers ships a keras.utils.PyDataset/tf.data wrapper — you write a thin one (or a generator) around the sample-fetching loop above, same as you would for any other custom data source:

import numpy as np
import keras

class MySequence(keras.utils.PyDataset):
    def __init__(self, dataset, batch_size=16, **kwargs):
        super().__init__(**kwargs)
        self.dataset, self.batch_size = dataset, batch_size

    def __len__(self):
        return len(self.dataset) // self.batch_size

    def __getitem__(self, idx):
        samples = [self.dataset[i] for i in range(idx * self.batch_size, (idx + 1) * self.batch_size)]
        batch = stack_samples(samples)
        return keras.ops.convert_to_numpy(batch["image"]), keras.ops.convert_to_numpy(batch["label"])

model.fit(MySequence(train_ds), epochs=10)

See Finetune a pretrained Sentinel-2 backbone on EuroSAT for a complete, runnable version of this pattern.

Transforms

keras_climate.transforms operate on the same dict samples, expecting channels-last input, and are meant to be composed inside a dataset's own transforms= callback:

from keras_climate.transforms import AppendNDVI

# appends an NDVI band computed from the given red/NIR band indices
ndvi = AppendNDVI(index_red=3, index_nir=7)
train_ds = EuroSAT(root="data/eurosat", split="train", transforms=ndvi, download=True)

Other families: band-index transforms for other indices (AppendNDWI, AppendNBR, ...), LeeFilter/lee_filter for SAR speckle reduction, SatSlideMix for a segmentation-aware mixing augmentation, and Rearrange for axis reordering. See Transforms API.

Losses

keras_climate.losses ships two loss functions with no direct Keras built-in equivalent, both from crop-type/time-series classification literature: EarlyRewardLoss (the ELECTS loss for early classification of time series — rewards a correct and early stopping decision) and QRLoss/RQLoss (a forward/backward pair for calibrating class probabilities against predictions). See Losses API.

Full dataset catalog

See Datasets API for every dataset class (190+, grouped by SpaceNet / Copernicus-Bench / the general catalog), each with parameters and citation info pulled directly from its docstring.