This repository contains the data and software used to conduct the analysis for the research paper: 'Improving localized weather predictions for precision agriculture: a Time-Series Mixer approach for hazardous events detection', currently under submission in Environmental Modelling & Software.

This study leverages the Time-Series Mixer (TSMixer) neural network to improve weather forecasts for precision agriculture. The TSMixer model predicts temperature, wind speed, relative humidity, and precipitation over a 45-hour horizon. Trained using predictions from the MOLOCH model and observational data from ARPA stations near six agricultural sites in Northern Italy, TSMixer demonstrates:

• Greater Accuracy: Outperforming the MOLOCH model in forecasting key weather variables.
• Hazardous Event Detection: Excelling in identifying critical events for precision agriculture, such as:
  • Frost damage
  • Heat stress
  • Germination block The findings underscore the model value for environmental risk management in agriculture.

The research was conducted primarily on Google Colab Notebooks, utilizing GPUs for efficient neural network training. To reproduce the results, we recommend accessing the project folder in Google Drive and opening the notebooks in the notebooks directory using the Google Colab application.

The folder LocalWeatherPredictionsDL is structured as follows:

• data: Contains the ARPA and MOLOCH data.
• notebooks: Includes Colab Python notebooks for reproducing the main results.
• lib: Contains Python files with auxiliary methods used in the notebooks.
• results: Stores various results generated by the notebooks, as model checkpoints.

The ARPA data can be accessed from the ARPA Lomabardia and Piemonte websites (https://www.arpalombardia.it; https://www.arpa.piemonte.it). MOLOCH data are available in grib2 format upon request from the server https://tds.bo.isac.cnr.it/ Please send your request to dinamica@isac.cnr.it or to the corresponding author. The data should be organized in Pandas dataframes where each column contains the hourly value of the weather variables for each study locations and located in the 'data' folder with the names 'df_moloch_agro.p' and 'df_arpa_agro.p' saved as pickle files.

Contents of the notebooks folder:

1. agro_tsmixer_double_reinforced_moloch_predictions.ipynb: the notebook contains the procedure to optimize, train and test the TSMixer neural networks with the ARPA and Moloch data. Inside the notebook there are the analysis of the extreme events and the of the interpretation of the TSMixer neural network.
2. agro_tsmixer_past_local_data.ipynb: the notebook contains the procedure to optimize, train and test the TSMixer neural networks trained only with the past ARPA data.