Skip to main content

The Federated Learning Technique for Mitigating the Climate Change

Overall federated learning process in centralised setup

Lorenzo Carnevale, Assistant Professor at Università di Messina (UNIME)

Roberto Marino, Research Fellow at Università di Messina (UNIME)

Adrian Gambito, PhD Student at Università di Messina (UNIME)

Massimo Villari, Professor at Università di Messina (UNIME)

Standard and cutting-edge technologies often come to support socio-economic issues, e.g., climate change. In this regard, researchers from academia and industry, along with experts in the climate change field, are questioning how to prevent and mitigate the effects of such disasters. This is not a mystery that 2023 was the golden year for Artificial Intelligence (AI) due to the popularity of large language model (LLM) solutions, such as the OpenAI’s ChatGPT, which led the G7 leaders to a call for “AI guardrails”. As a result, many sectors have been interested in AI solutions and there was no exception for the one discussed in this blog. Indeed, a quick search on scientific paper databases (e.g., Scopus) with the keywords "Artificial Intelligence" and "climate change" shows how scientific production increased from 487 papers in 2022 to 746 papers in 2023. Although several AI solutions have been proposed to prevent and mitigate climate change (e.g., flood, wildfire, tornado, tsunami, earthquake), only a few are the solutions that involve distributed AI, such as Federated Learning (FL).

FL is a machine learning (ML) technique that allows training an algorithm through decentralised devices or servers that maintain the data, without the need to exchange it. This approach contrasts with traditional centralized ML techniques where data is uploaded to a server, or more traditional decentralised methods that assume local data is identically distributed. FL allows different participants to build a common and robust ML model, without exchanging any data. Using this technique allows us to address critical issues such as data protection, security and access rights and the use of heterogeneous data.

Federated learning process in centralised setup

Figure 1: Overall federated learning process in centralised setup (Source: Wikipedia). 

The scientific community is investing time and effort in proposing FL solutions for preventing and mitigating the effects of climate change. A FL framework for the Internet of Things (IoT) was proposed for wildfires [1]. The framework integrates with cameras strategically positioned in vulnerable areas. Indeed, prompt notification of fire detection increases the chance for the firefighters to reach the affected area in time, reducing the risk of damage. The motivation behind the use of FL is the poor abundance of datasets for this topic and the need to protect data privacy. In this regard, FL facilitates collaborative learning across a decentralised dataset. A novel FL methodology, called FedVIS [2], uses modern artificial intelligence models, such as the ones used in ChatGPT (e.g., Transformer) to improve the robustness of FL in the context of heterogeneous data. An application of FL for a network of Unmanned Aerial Vehicles (UAVs) is proposed in [5] to implement an effective solution for wildfire mitigation. Each UAV can hover at different locations and obtain images with distinctive features, applying the local dataset to a convolutional Deep Learning model, while a Main Ground Server aggregates the various models coming from different UAVs and delivers back a global model.


Figure 2: Landslide (Source: Wikipedia).

FL was proposed [3] also for improving the accuracy of single-point landslide displacement prediction while protecting the security of data. As each landslide displacement has unique characteristics, models trained on one landslide cannot be directly generalized. Instead, FL enables the collaborative training of individual landslide displacement prediction models resulting in a global model that can be generalized. The only constraint is selecting landslides of a similar type. 

The distributed AI methodology is also used [4] for optimising a routing protocol for ad hoc networks, really useful when network infrastructure is down due to a natural disaster. Moreover, Low-Earth-Orbit satellites are often involved in natural disaster management operations. FL is here used to eliminate the need to transmit row data, preserving bandwidth and energy. However, in this tough scenario, a generic FL scheme requires too many communication rounds between satellites and aggregators leading to substantial delays of up to several days in LEO constellations. The authors in [6] propose a one-shot FL framework for LEO satellite constellations able that needs only a single communication round to complete the entire learning process. This is possible by exploiting the knowledge distillation paradigm jointly with the technique of virtual model retraining.

LEO Sattelite

Figure 3: LEO Satellite (Source: Forbes).

In the context of the Horizon Europe 2020 TEMA project, FL technology can be used to distribute learning steps between devices for remote monitoring (Drone), between edge devices (Ground Stations) or between data centres located in different countries, thus avoiding expensive data transfers and lowering the need for bandwidth and computational time.

FL proves to be a promising solution for facing the issues concerning natural disaster management both for prevention and mitigation of the effects. Due to the scarcity of disaster events and the difficulty, data are often collected from public data sites. However, FL enables software architectures that can take advantage of fully decentralised training.



[1] Ali Akbar Siddique, Nada Alasbali, Maha Driss, Wadii Boulila, Mohammed S. Alshehri, Jawad Ahmad. Sustainable collaboration: Federated learning for environmentally conscious forest fire classification in Green Internet of Things (IoT). Internet of Things. Volume 25. 2024. 101013. ISSN 2542-6605.

[2] Hu, Y.; Fu, X.; Zeng, W. Distributed Fire Detection and Localization Model Using Federated Learning. Mathematics 2023, 11, 1647.

[3] Yuting Yang, Yue Lu, Gang Mei. A federated learning based approach for predicting landslide displacement considering data security. Future Generation Computer Systems. Volume 149. 2023. Pages 184-199. ISSN 0167-739X.

[4] Ricardo Pagoto Marinho , Luiz F.M. Vieira, Marcos A.M. Vieira, Antonio A.F. Loureiro. CAIN: An energy-aware and intelligent increasing coverage area routing protocol for future 6G networks. Computer Networks 228 (2023) 109733,

[5] Ahmed El Hoffy Sean (Seok-Chul) Kwon and Hen-Geul Yeh. Federated/Deep Learning in UAV Networks for Wildfire Surveillance. 2023 Wireless Telecommunications Symposium (WTS). DOI: 10.1109/WTS202356685.2023.10131685.

[6] Mohamed Elmahallawy, Tie Luo. One-Shot Federated Learning for LEO Constellations that Reduces Convergence Time from Days to 90 Minutes. The 24th IEEE International Conference on Mobile Data Management (MDM 2023), DOI: 10.1109/MDM58254.2023.00020