This paper describes our experiments of UNet based deep convolutional neural network model applied on Weather4cast challenge 2021 stage 1. This challenge's task is to predict future weather patterns, which is composed of four target variables per pixel: temperature, convective rainfall rate, probability of occurrence of tropopause folding, and cloud mask. Given equal size of input and output weather image, we trained UNet shaped neural network model to make predictions for each target variable. Evaluation results show competitive performance compared to the baseline method.
In this paper, we describe our methods on Weather4cast
2021’ stage 1 challenge. This challenge’s task is to
predict future 8 hours weather image given prior 1 hour’s
weather data [
Training data contains weather images acquired from 3
distinct regions worldwide, named R1, R2 and R3 (which
corresponds to Nile region, Eastern Europe, and South
West Europe, respectively) in this challenge. In the core
challenge task, both training data and held out test data
came from R1, R2 and R3 regions. In the transfer learning
task, training data is same as core challenge task, but held
out test data is coming from diferent regions, named R4,
shown efective performance in similar tasks [
Evaluation results show competitive performance compared to the baseline methods for this task. 2.
images. vector.
Prediction takes prior 1 hour’s weather data as input, and should produce prediction ranging to next 8 hours. Each weather image has 15 minutes interval, so input has 4 weather images, and output should have 32 weather
Input weather image contains various weather-related variables data per pixel. There are 9 continuous variables, and 16 discrete variables. Continuous variables are normalized to minimum 0.0, maximum 1.0 range, and discrete variables are converted into one-hot encoding
In this study, recurrent model such as LSTM (long
short term memory network) [9] was not utilized. So,
time dimension is simply merged into feature channel
dimension.
2.2. Model
We implemented UNet [
Each blue box represents dense convolution layers block with average pooling layer. Each orange box represents deconvolution layers. Green arrow represents skip connections between downsampling path and upsampling path.
We trained model for each of 4 target variables separately, but all of them have identical UNet with DenseNet model structure described above. For temperature, convective rainfall rate, probability of occurrence of tropopause folding, mean squared error is used as loss function. For cloud mask, since it can have only binary output (0 or 1), sigmoid loss is used instead.
Training data has 316 days weather images from R1, R2 and R3 region, time period ranging 2019-2020. Held-out test data is captured from separate time period (20202021).
In this study, we haven’t tried to train region-specific model. All training weather images from diferent regions were merged and used together to train general model that can be used to apply on all regions.
We trained 1 model for the temperature prediction, 2 models for the convective rainfall rate prediction, 2 models for the probability of occurrence of tropopause folding prediction, and 3 models for the cloud mask prediction. When multiple models are trained for the same target variable, output prediction values are simply averaged to produce final output. Our experimentation code is publicly available at https://github.com/sungbinchoi/ w4c_st1.
Oficial evaluation metric for this challenge was mean squared error divided by mean squared error of the baseline prediction, which is taking last previous weather image as prediction output on all out timeframe, so baseline prediction output is scored as 1.0.
In this study, we trained only region-agnostic model. Same models used to produce prediction for the core task is used for the transfer learning task without any change. Our best evaluation results from held out test set was 0.507325 in the core challenge task and 0.465760 in the transfer learning task, which means roughly loss is halved compared to the baseline.
In this experiment, we used UNet for the weather forecast task. It showed competitive performance compared to the baseline.
In our study, we trained only region-agnostic model. In our preliminary experiment, it was not clear whether a region-specific model is more efective than the general model for the core challenge task. As a layperson, this was quite counterintuitive, because weather from R1 (Nile region) seems very diferent from R3 (South West Europe). We will explore more various research ideas and hope to find more efective methods in the next stage of the challenge. [7] H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y. Iwamoto, X. Han, Y. Chen, J. Wu, Unet 3+: A full-scale connected unet for medical image segmentation, in: 2020 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2020, Barcelona, Spain, May 4-8, 2020, IEEE, 2020, pp. 1055–1059. [8] G. Huang, Z. Liu, L. Van Der Maaten, K. Q. Weinberger, Densely connected convolutional networks, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 4700–4708. [9] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural Computation 9 (1997) 1735–1780.