dl4eo 0.4.1

Deep Learning for Earth Observation — automated training-dataset builder for EO segmentation tasks

dl4eo

dl4eo is a Python package for building multi-source Earth Observation training datasets and training segmentation models end-to-end. It automates the full pipeline from raw satellite data to model checkpoint:

• Sentinel-2 (L2A, cloud-filtered, spectral indices)
• Sentinel-1 RTC (VV + VH, batched by date)
• Copernicus DEM (elevation + slope, per-scene mosaic)
• Segmentation masks from any vector label file
• Train-ready PyTorch dataset with global normalization
• Model training with UNet, DeepLabV3+, SegFormer, ViT, and more

---

Installation

# Pipeline only (no PyTorch required)
pip install dl4eo

# Pipeline + training stack
pip install dl4eo[train]

Requires Python ≥ 3.8.

---

Quick Start

1 — Build a dataset

import dl4eo

dl4eo.generate_dataset(
    base_dir="/data/glacial_lakes",
    aoi_shapefile_dir="/data/aoi/",           # folder with AOI.shp (study area polygon)
    feature_shapefile="/data/lake_boundaries.shp",  # label polygons
    date_range="2021-06-01/2021-08-31",
    cloud_cover=20,
    patch_size=256,           # pixels
    overlap=0.0,
    spectral_index="NDWI",    # NDWI | NDSI | NDVI | NDRE | EVI | None
    skip_sentinel1=False,
    skip_dem=False,
    normalize=False,          # recommended: normalize at load time via PatchDataset
    n_jobs=8,
)

2 — Quality control, splits, statistics

# Filter bad patches (nodata, no foreground, constant bands)
valid = dl4eo.qc.validate("/data/glacial_lakes", min_positive_fraction=0.001)

# Create train / val / test splits
splits = dl4eo.splits.make_splits(
    "/data/glacial_lakes",
    ratios=(0.7, 0.15, 0.15),
    strategy="temporal",   # "random" | "temporal" | "spatial"
    valid_file="/data/glacial_lakes/valid_patches.txt",
)

# Global per-band statistics (training split only — no leakage)
stats = dl4eo.stats.compute("/data/glacial_lakes", split="train")
# → {"band_1": {"mean": 6032.7, "std": 3471.1, "p2": 540.0, "p98": 11752.0},
#    "band_2": {...}, ..., "_meta": {"n_files": 25, "split": "train"}}

3 — PyTorch dataset

from dl4eo.io import PatchDataset
from torch.utils.data import DataLoader

ds = PatchDataset(
    "/data/glacial_lakes",
    split="train",
    split_file="/data/glacial_lakes/splits.json",
    stats_file="/data/glacial_lakes/stats.json",
    norm="zscore",    # "zscore" | "minmax" | "percentile" | None
    bands=None,       # None = all bands; or e.g. [0, 1, 2, 6, 7]
)

sample = ds[0]
# sample["image"]  →  FloatTensor [C, H, W]
# sample["mask"]   →  LongTensor  [H, W]

loader = DataLoader(ds, batch_size=16, shuffle=True, num_workers=4)

PatchDataset inherits from torchgeo.datasets.NonGeoDataset when torchgeo is installed, and falls back to torch.utils.data.Dataset otherwise.

4 — Train a model (one-liner)

module = dl4eo.train(
    data_dir="/data/glacial_lakes",
    model="unet",            # see SUPPORTED_MODELS below
    backbone="resnet34",
    num_classes=2,
    split_strategy="temporal",
    norm="zscore",
    loss="dice_ce",          # "dice_ce" | "dice" | "ce" | "focal"
    batch_size=16,
    max_epochs=50,
    accelerator="gpu",
    devices=1,
)
# → auto-generates splits.json + stats.json if missing
# → saves best checkpoint (monitored on val/iou)
# → returns loaded SegmentationModule

5 — Build a model manually

from dl4eo.train import build_model, SegmentationModule, SegDataModule, SUPPORTED_MODELS
import lightning as L

print(SUPPORTED_MODELS)
# ['unet', 'unet++', 'deeplabv3+', 'fpn', 'pspnet', 'linknet', 'pan', 'manet',
#  'segformer', 'vit-tiny', 'vit-small', 'vit-base']

net    = build_model("segformer", in_channels=10, num_classes=2)
module = SegmentationModule(net, num_classes=2, lr=5e-4, loss="dice_ce")

dm = SegDataModule(
    data_dir   = "/data/glacial_lakes",
    split_file = "/data/glacial_lakes/splits.json",
    stats_file = "/data/glacial_lakes/stats.json",
    batch_size = 8,
)

trainer = L.Trainer(max_epochs=100, accelerator="gpu", devices=1)
trainer.fit(module, dm)

---

Pipeline stages

Stage	Description
1	Download Sentinel-2 L2A (STAC / Planetary Computer, cloud-filtered)
2	Preprocess S2: single-pass resample to 10 m + spectral index + stack
3	Generate patch AOIs: windowed reads, intersects user AOI polygon
4	Prepare DEM: one mosaic per scene, windowed reproject per patch
5	Prepare Sentinel-1 RTC: batched STAC search by date, VV+VH stack
6	Generate segmentation masks from label shapefile

Normalization is intentionally excluded from the pipeline. Use dl4eo.stats.compute() on the training split and PatchDataset(norm="zscore") at load time — this avoids per-patch scale inconsistency and data leakage.

---

Supported models

All models are trained from scratch on arbitrary input channels (no dataset-specific pretrained weights).

Model	Family	Default backbone	Constraints
unet	SMP	resnet34	—
unet++	SMP	resnet34	—
deeplabv3+	SMP	resnet34	batch_size ≥ 2 per GPU (BatchNorm)
fpn	SMP	resnet34	—
pspnet	SMP	resnet34	batch_size ≥ 2 per GPU (BatchNorm)
linknet	SMP	resnet34	—
pan	SMP	resnet34	input ≥ 128 px (pyramid pooling)
manet	SMP	resnet34	—
segformer	SegFormer	swin_tiny_patch4_window7_224	—
vit-tiny	ViT	vit_tiny_patch16_224	—
vit-small	ViT	vit_small_patch16_224	—
vit-base	ViT	vit_base_patch16_224	—

SMP models also support ImageNet-pretrained encoders for 3-channel input: weights="imagenet".

BatchNorm note: deeplabv3+ and pspnet will raise an error if a mini-batch contains only 1 sample. Ensure len(train_set) % batch_size != 1, or choose a batch_size that divides your training set evenly.

---

Output structure

base_dir/
├── stack/               # Scene-level S2 stacks (bands + spectral index)
├── images/              # Clipped S2 patches
├── DEM/                 # Per-scene DEM mosaics + per-patch stacks
├── GRD/                 # Downloaded SAR granules (VV, VH)
├── Clipped_SAR/         # SAR reprojected to patch grid
├── stacked/             # S2 + DEM patches  (10 bands)
├── stacked_with_sar/    # S2 + DEM + SAR patches  (primary output)
├── mask/                # Binary (or multi-class) segmentation masks
├── AOI_boxes/           # Per-scene patch grid shapefiles
├── splits.json          # Train / val / test split (after dl4eo.splits)
├── stats.json           # Per-band statistics   (after dl4eo.stats)
└── valid_patches.txt    # QC-passing patch list  (after dl4eo.qc)

---

Input requirements

Parameter	Description
aoi_shapefile_dir	Folder containing one or more AOI .shp files (study area polygon)
feature_shapefile	Label vector file (e.g. lake outlines) — used for mask generation and patch filtering
date_range	"YYYY-MM-DD/YYYY-MM-DD"

The AOI polygon controls which patches are generated. Only patches that intersect both the AOI and at least one label feature are kept.

---

Dependencies

Core (installed automatically): numpy, rasterio, geopandas, shapely, fiona, matplotlib, joblib, pystac-client, planetary-computer, requests, scipy

Training (pip install dl4eo[train]): torch>=2.0, lightning>=2.0, segmentation-models-pytorch>=0.3, timm>=0.9, torchmetrics>=1.0

Optional: torchgeo>=0.5 — enables NonGeoDataset base class for PatchDataset

---

Example use cases

• Glacial lake mapping and segmentation
• Flood extent extraction
• Multimodal image fusion (S2 + S1 + DEM)
• Patch-based dataset generation for semantic segmentation

---

Author

Developed by Saurabh Kaushik Postdoctoral Researcher · University of Wisconsin–Madison Earth Observation · Deep Learning · Geo-Foundational Models · Cryosphere

---

License

MIT License

---

Citation

If you use dl4eo in your research, please cite:

@misc{kaushik2026dl4eo,
  author       = {Saurabh Kaushik},
  title        = {{dl4eo: A Python package for multi-source Earth Observation dataset building and segmentation model training}},
  year         = {2026},
  howpublished = {\url{https://pypi.org/project/dl4eo/}},
}