DeepLearning for Short-Range Weather Forecasts

Weather forecasting is a long-studied field. Numerical Weather Prediction (NWP) models have been around since revolution in high performance computing. NWP, refers to forecasting the future state of the weather through large computer simulations that take into account the current states of variables. This method uses Partial Differential Equations to predict the temperature, wind, precipitation, and clouds.

The motivation behind this project is to propose a different Deep Learning model to not only forecast precipitation, but also temperature as well as additional land, oceanic and atmospheric climate variables, so that less computing power is used, and the predictions are more accurate.

Dataset Construction

To train our model, we went ahead and created our own dataset. We used the Google Earth Engine (GEE) to get the precipitation and temperature hourly aggregates as images, in which each pixel covers ~11KM on ground. Following are some samples of temperature and precipitation dataset:

Region Coordinates

23.155933,59.545898,37.439499,79.541016

Preprocessing

We have resized the dataset images into two resolutions for different models

74 x 104
54 x 74

Our GEE script downloads the images with a matplotlib axis padding, So we have written a script to crop the images as well.

Lastly, We have concatenated the images into numpy arrays such that 24 hours correspond to 24 images. For example, a sample of one month will have a 4 dimensional tensor of 30 x 24 x width x height. If we were to concatenate daily aggregates of precipitation and temperature, then our samples would like the following gifs:

Architecture

Inspired by Xingjian Shi et al., 2015, We have used ConvLSTMs for our model.

Michael R Lin, Predicting Fist Movement with Convolutional LSTMs. Available online: https://m-lin-dm.github.io/ConvLSTM_Fish_1/ (accessed on June 11, 2021)

Dependencies

tensorflow
ee
geemap
numpy
matplotlib
opencv
moviepy
imageio
pandas

Training - Loss

PSNR, SSIM and MS-SSIM metrics over lead times

Results

We have implemented our models in TensorFlow. We have trained four different models and on different configurations, using Adam optimizer with the loss suggested from paper.

Model	Type	Comments
1.	Precipitation	Model 1 captures the dynamics of the motion but the amount of prediction precipitation is high.
2.	Temperature	Model 2 is trained on the temperature data, and it captures both the motion and predictions are aligned with the ground truth.
3.	Precipitation	Model 3 improves over Model 1 and performs better.
4.	Precipitation	Model 4 has successfully captured the motion as well as the amount of precipitation occuring in different regions.

Wind Speed

For the wind speed we have made some changes to the model by adding dropout and l2 regularization, as the model was overfitting. These changes improved model's performance. We trained our model on U and V components of the Wind, and also on hourly data as well as daily averages since 1940. We have kept the region coordinates same as above.

Results for the Wind Speed and Direction

Predictions - Autoregressive

Precipitation

Temperature 2m

U component of Wind

V component of Wind

iamiqbal / deep-learning-for-short-range-weather-forecasts Goto Github PK