- 分享主题：Spatiotemporal sequence prediction
- 论文标题：Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting
- 论文链接：https://arxiv.org/pdf/1506.04214.pdf
- 分享人：唐共勇

1. Summary

【必写】，推荐使用 grammarly 检查语法问题，尽量参考论文 introduction 的写作方式。需要写出

  1. 这篇文章解决了什么问题？
  2. 作者使用了什么方法（不用太细节）来解决了这个问题？
  3. 你觉得你需要继续去研究哪些概念才会加深你对这篇文章的理解？

In this paper, rainfall prediction is clearly defined, formalized as a machine learning task, and a model to solve the problem of Spatio-temporal prediction is proposed. According to the author, this is the first time to introduce the technology of deep learning into rainfall prediction. The author extends LSTM to obtain state information by convolution instead of matrix product. Moreover, the spatial information of the adjacent area is extracted with the help of CNN, which solves the problem that the traditional algorithm can not capture spatial information. The most valuable point of the whole article is that it combines CNN and LSTM to capture spatial information with CNN and time information with RNN.

2. 你对于论文的思考

需要写出你自己对于论文的思考，例如优缺点，你的takeaways

优点：
1.将CNN与RNN进行了融合，提供了解决时空预测的新方案
2.对降雨量预测进行了形式化定义，并达到了SOTA
缺点：
1.对空间信息的提取太过简陋，可以考虑增加注意力机制
2.针对短期降雨预报，即预测的视野太短，只能在特定领域取得较好效果

3. 其他

【可选】

image.png
image.png
ConvLSTM.pptx