GeoTorch: A Spatiotemporal Deep Learning Framework

GeoTorch is a spatiotemporal deep learning framework on top of PyTorch and Apache Sedona. It enable spatiotemporal machine learning practitioners to easily and efficiently implement deep learning models targeting the applications of raster imagery datasets and spatiotemporal non-imagery datasets. Deep learning applications of raster imagery datasets include satellite imagery classification and satellite image segmentation. Applications of deep learning on spatiotemporal non-imagery datasets are mainly prediction tasks which include but are not limited to traffic volume and traffic flow prediction, taxi/bike flow/volume prediction, precipitation forecasting, and weather forecasting.

GeoTorch Modules

GeoTorch contains various modules for deep learning and data preprocessing in both raster imagery and spatiotemporal non-imagery categories. Deep learning module offers ready-to-use raster and grid datasets, transforms, and neural network models.

Datasets: This module conatins processed popular datasets for raster data models and grid based spatio-temporal models. Datasets are available as ready-to-use PyTorch datasets.
Models: These are PyTorch layers for popular raster data models and grid based spatio-temporal models.
Transforms: Various tranformations operations that can be applied to dataset samples during model training.
Preprocessing: Supports preprocessing of raster imagery and spatiotemporal non-imagery datasets in a scalable setting on top of Apache Spark and Apache Sedona. Users don't need to learn the coding concepts of Apache Sedona and Apache Spark. They only need to write their code on Python while PySpark and Apache Sedona implementations are hidden. The preprocessing module allows machine learning practitioners to prepare a trainable grid-based spatiotemporal tensor from large raw datasets along with performing various transformations on raster imagery datasets.

GeoTorch Design Principles

GeoTorch is designed in such a way that it has the necessary building blocks for developing raster and spatiotemporal DL applications within the PyTorch ecosystem. Various functionalities available in GeoTorch deep learning module are compatible with PyTorch core units such as neural network layers, datasets, and transformations. We make the deep learning module of GeoTorch GPU compatible so that PyTorch-provided scalability and parallelism on GPU can be achieved with GPU configured devices.

Although the data preprocessing module has dependencies on external big data processing libraries such as PySpark and Apache Sedona, the deep learning module only depends on PyTorch. Since the datasets component of the deep learning module provides preprocessed and trainable state-of-the-art benchmark datasets, designing applications with such benchmark datasets can be completed without requiring big data-related dependencies. Furthermore, to help machine learning practitioners build raster and spatiotemporal applications with their preferred raw datasets, our preprocessing module enables raster and spatiotemporal data processing in a pure Pythonic way without requiring the coding knowledge of Apache Spark, Apache Sedona, and other big data processing libraries while providing the scalability of Apache Spark at the same time.

Our preprocessing module is designed such that it minimizes the number of methods and classes in the API. Users can perform end-to-end spatiotemporal data preprocessing, which starts by loading raw datasets and ends by generating a trainable Tensor-shaped array, with a minimum number of method calls. It helps the users understand the API fast and reduces their confusion.

Documentation

Details documentation on installation, API, and programming guide is available on GeoTorch Website.

Dependency Set up

Following libraries need to be set up before using GeoTorch.

Dependencies for Deep Learning Module:

PyTorch >=1.10
Rasterio
Scikit-image

Dependencies for Preprocessing Module:

PySpark >=3.0.0
Apache Sedona >=1.2.0-incubating

Example

End-to-end coding examples for various applications including model training and data preprocessing are available in our binders and examples sections.

We show a very short example of satellite imagery classification using GeoTorch in a step-by-step manner below. Training a satellite imagery classification model consists of three steps: loading the dataset, initializing the model and parameters, and train the model. We pick the DeepSatV2 model to classify EuroSAT satellite images.

EuroSAT Image Classes

Annual Crop
Forest
Herbaceous Vegetation
Highway
Industrial
Pasture
Permanent Crop
Residential
River
SeaLake

Spectral Bands of a Highway Image

Spectral Bands of an Industry Image

Loading Training Dataset

Load the EuroSAT Dataset. Setting download=True will download the full data in the given directory. If data is already available, set download=False.

full_data = geotorch.datasets.raser.EuroSAT(root="data/eurosat", download=True, include_additional_features=True)

Split data into 80% train and 20% validation parts

dataset_size = len(full_data)
indices = list(range(dataset_size))
split = int(np.floor(0.2 * dataset_size))
np.random.seed(random_seed)
np.random.shuffle(indices)
train_indices, val_indices = indices[split:], indices[:split]

train_sampler = torch.utils.data.sampler.SubsetRandomSampler(train_indices)
valid_sampler = torch.utils.data.sampler.SubsetRandomSampler(val_indices)

train_loader = torch.utils.data.DataLoader(full_data, batch_size=16, sampler=train_sampler)
val_loader = torch.utils.data.DataLoader(full_data, batch_size=16, sampler=valid_sampler)

Initializing Model and Parameters

Model initialization parameters such as in_channel, in_width, in_height, and num_classes are based on the property of SAT6 dataset.

model = DeepSatV2(in_channels=13, in_height=64, in_width=64, num_classes=10, num_filtered_features=len(full_data.ADDITIONAL_FEATURES))
loss_fn = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.0002)

Train the Model for One Epoch

for i, sample in enumerate(train_loader):
    inputs, labels, features = sample
    # Forward pass
    outputs = model(inputs, features)
    loss = loss_fn(outputs, labels)
    # Backward pass and optimize
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

Evaluate the Model on Validation Dataset

model.eval()
total_sample = 0
correct = 0
for i, sample in enumerate(val_loader):
    inputs, labels, features = sample
    # Forward pass
    outputs = model(inputs, features)
    total_sample += len(labels)
    _, predicted = outputs.max(1)
    correct += predicted.eq(labels).sum().item()
val_accuracy = 100 * correct / total_sample
print("Validation Accuracy: ", val_accuracy, "%")

Other Contributions of this Project

We also contributed to Apache Sedona to add transformation and write supports for GeoTiff raster images. This contribution is also a part of this project. Contribution reference: Commits

RobertoTjesse/GeoTorch