Giovanni Bricconi

This week I have found a nice article:

Fu Jie Tey, Tin-Yu Wu, and Jiann-Liang Chen. 2022. Machine Learning-based Short-term Rainfall Prediction from Sky Data. ACM Trans. Knowl. Discov. Data 16, 6, Article 102 (September 2022), 18 pages. https://doi.org/10.1145/3502731

I enjoyed reading it for various reasons. First of all, it gently introduces some terms and provides many pictures that let you familiarize with the needed concepts. We can guess the meaning of high accuracy, high precision, and recall rate, but it is much better to see the formulas that define them. This is concentrated in a couple of pages, it is a good reference. For instance, accuracy is defined as (# of true positive+true negative predictions)/(all predictions), and precision is just (# of true positive prediction)/(# of true or false positive predictions). Having these 2 pages greatly simplifies interpreting the results presented.

Then it describes how sky data has been acquired. They have deployed some Raspberry-pi modules, taking sky pictures, and also some Arduino-based weather sensors collecting temperature, UV rays, humidity, etc. They describe some difficulties encountered, and some workarounds used, for instance why they have chosen a particular angle when taking sky pictures.

For short-term prediction they mean answering the question: will it rain within 30 minutes? The purpose of the study was to see if a machine learning model, based on sky pictures and some other measures can be accurate enough. Usually, weather forecasts are made using complex mathematical models, on big computers, and their approach requires much less resources.

Their machine learning model is composed of a ResNet-152 convolutional neural network followed by a Long short-term memory recurrent neural network. It has been trained with about 60’000 images and has a recall of 87% on predicting rain absence and 81% on predicting rain.