Environmental monitoring is essential to intervene promptly in environmental disasters or emergencies. Therefore, in the environmental field, as in other fields, it is essential to create reports quickly. In this article, we describe a possible system that uses Artificial Intelligence (AI) to automatically create environmental reports, always considering the supervision of the environmental expert. The system involves two phases, one for testing and the other for production. During the testing phase, the AI system is calibrated on a subset of the available data, while during the production phase, the AI system is fully operational and works on all the available data. We describe a practical case study using the temperature data provided by ARPA Lazio.
In recent years, Artificial Intelligence (AI) has made substantial progress thanks to the difusion of Large
Language Models (LLMs) [
We propose a system combining manual supervision with automatic analytics. There are two actors: the environmental expert, a human with deep knowledge of environmental data, and the AI model,
CEUR
ceur-ws.org Instructions
Manual Data
Analytics AI Automatic Data Analytics
Manual Evaluation Testing Phase
Production Phase Environmental
Data
Final Instructions
AI Automatic Data Analytics
Manual Control
Report
based on LLMs, which is tuned with environmental data through a Retrieval Augmented Generation
(RAG) process [
To test the efectiveness of our proposed approach, we analyzed the historical series of temperature
data recorded in the decade 2013-2023 from the AL007 station of the Micro Meteorological Network
(RMM) managed by ARPA Lazio and located in via Boncompagni 101 in Rome [
Through traditional data pre-processing techniques, the environmental expert (one of the authors) extracted average temperature values, maximum and minimum values, and mode for each month and considered each year. Next, they formulated the following questions, with an increasing level of complexity: 1) In which month and year was the highest temperature value recorded? 2) What was the hottest month for each year? 3) What is the percentage increase in temperature over the decade 2013-2023? We transformed the previous questions into instructions for the AI model. We defined the following basic structure for all the questions: Consider {context} where: - Year is the year - Month is the month - Mean is the average monthly temperature - Max is the maximum monthly temperature - Min is the minimum monthly temperature - Mode is the most frequent monthly temperature value Answer the following question: {question}
The context variable contains the input data, and the question variable is the specific question to be
answered. We implemented the AI model as a RAG-based system using LangChain [
In the first phase of the test, the AI model read and extracted data from only four years (2013, 2014, 2015, and 2020) compared to the entire available sample (2013-2023). Considering the entire data sample, the AI model answered the first two questions incorrectly. However, if we narrow it down to four years, the answers are correct. Regarding the last question, the AI model did not respond by providing a numerical value, but it ofered an interesting result: it precisely described the useful steps to answer the question, adding that it was not able to apply this procedure as it lacked a significant data, that of 2023, which does not fall within the data of the years known to it.
As a second step, we modified the structure of the questions to force the model to read all the years. We defined the following structure for all the questions: I have many datasets from 2013 to 2023.
Consider {context} where: - Year is the year - Month is the month - Mean is the average monthly temperature - Max is the maximum monthly temperature - Min is the minimum monthly temperature - Mode is the most frequent monthly temperature value Answer the following question considering all the data from 2013 to 2023: {question}
Compared to the previous version of the questions, we added more details related to the dataset range (highlighted in the previous code). The AI model answered the third question correctly, while it could not answer completely the first and second questions because it did not find data related to 2021, 2022, and 2023. The refining process throughout the testing phase should continue until the AI model can answer all the questions correctly. Our preliminary experiments demonstrated that although the AI model could not answer all the questions correctly, we can conclude that preliminary results are encouraging compared to the questions we answered manually. As a future work, we plan to improve the testing phase and also implement the production phase of the project.