Awesome

Sentinel-5p forecasting

This repository includes the PyTorch implementation of an encoder-decoder forecasting network. It was built for the prediction of air-pollution variables based on two ESA data products: Atmospheric measurements sensed by the Sentinel-5p satellite and an air-quality forecasting model provided by the Copernicus Atmospheric Monitoring Service that combines satellite observations with sophisticated chemistry and transport models.

Introduction

The Sentinel-5p mission is the first of the Sentinel series dedicated to atmospheric composition monitoring. At a spatial resolution of 5.5km and a daily temporal resolution worldwide it can retrieve the concentration of trace gases such as NO2, SO2 and CO.

In this project, it is investigated if a forecast superior to the numerically modeled one (CAMS) can be created using solely Sentinel-5p satellite data. We tackle this problem by employing a Convolutional LSTM for the prediction of sequential sentinel-5p images. Since the retrieval of trace gases is heavily affected by clouds and atmospheric noise, Sentinel-5p images contain a lot of No-Data-Values. To make the prediction model robust against these data inconsistencies we employ a masked loss. We also study the effect of using the corresponding CAMS images as an additional input to the Sentinel-5p prediction model.

Model architecture

The encoder-decoder network is composed of convolutional layers that are applied on the whole sequence and Convolutional LSTM layers that process each sequence element individually. In comparison to a regular LSTM the Convolutional LSTM preserves spatial structure of the input by using convolutions instead of element-wise multiplications in the cell state calculation.

Each convolutional layer is followed by Batch Normalization for better regularization. We use 3 layers in the encoder and the decoder. The model is trained using a masked MSE loss and the ADAM optimizer.

alt text

We employ two different versions of this architecture. One uses Sentinel-5 sequences solely (s5-fc), the other creates a prediction using the numerical forecast data as a conditional input to the decoder (Cond-S5-fc).

Results

The model's performance is evaluated on the Peak signal-to-noise ratio (PSNR) and Structural Similarity Index Measure (SSIM).

Model	PSNR	SSIM
S5-fc	21.08	0.52
Cond-S5-Fc	31.05	0.70

Both models were evaluated on a test set of 300 sequences with 5 frames being used as input and the following 5 frames being predicted by the network.

alt text

Setup to get started

Make sure you have Python3 installed. You can install the required python packages by running:

pip install -r requirements.txt

Before starting the model, place your configuration in the config.json file. You can configure the following parameters:

name: name of the folder in which logs and model checkpoints are saved.
n_gpu: number of gpus for training, multi-gpu training is supported.
arch: model parameters
data_loader: options for the loading of the dataset
optimizer: optimization type and learning rate
loss: loss function used for training
metrics: metrics used for evaluation
trainer: training specifications, such as number of epochs and early stopping

Usage:

Training:

python main.py -c config.json

Using multiple GPU:

python main.py --device 0,1,2,3 -c config.json

Evaluation

python test.py -c config.json -r /path/to/model_checkpoint

Future work:

use a more sophisticated loss, e.g. a weighted loss that assigns weights to each pixel according to its intensity (proposed in Deep Learning for Precipitation Nowcasting:A Benchmark and A New Model)
investigate if the use of optical imagery can improve the forecasting performance

License

This project is licensed under the MIT License. See LICENSE for more details. Except as contained in this notice, the name of the authors shall not be used in advertising or otherwise to promote the sale, use or other dealings in this Software without prior written authorization from the authors.

Acknowledgements

This work is the result of a challenge proposed by ESA as part of the Pi School of AI 2019 Q4 programme. We are grateful to all organizers, stakeholders and mentors for providing us this opportunity.