Introduction
Transboundary haze is a recurring environmental and public health crisis in Southeast Asia, primarily driven by large-scale forest fires, agricultural "slash-and-burn" practices, and industrial emissions. Because pollutants are carried by atmospheric currents across national borders, localized air quality management is often insufficient. Predicting these events requires a holistic approach that integrates data from multiple countries and sources. The MediaEval 2021 "Insight for Wellbeing" task was established to address this challenge, focusing on the prediction of PM10 levels in Brunei, Singapore, and Thailand through cross-data analytics.
Problem Statement
Traditional air quality forecasting models frequently rely on local sensor data, which fails to account for the regional transport of pollutants. In Southeast Asia, a spike in particulate matter in one nation is often the result of activities in another. This "transboundary" nature necessitates models that can handle varying data granularities, regional atmospheric transport patterns, and the complex interplay between meteorological conditions and pollutant diffusion. The primary objective of the task is to improve 3-day air pollution forecasts, providing a window for actionable policy interventions and public health alerts.
Task Framework and Methodology
The MediaEval 2021 task organizers structured the challenge into three distinct subtasks to evaluate different aspects of predictive modeling:
- Localized Prediction: Participants develop models using only the historical weather and air quality data of a specific country. This serves as a baseline to determine the predictive power of local datasets.
- Transboundary Prediction: This core subtask requires the integration of "cross-data" sources. Participants are encouraged to use meteorological and air quality data from neighboring countries (e.g., using Indonesian weather patterns to predict haze in Singapore), alongside remote sensing data, terrain information, and non-traditional sources like social media or news feeds.
- Transfer Learning: This subtask investigates the feasibility of applying models trained on data-rich regions (such as Singapore) to regions with scarcer or less granular data (such as Brunei), testing the robustness and generalizability of the proposed architectures.
The task overview suggests several advanced architectural approaches for these problems, including Attentive Temporal Graph Convolutional Networks (ATGCN) to capture inter-station spatial-temporal relationships and CRNN-LSTM models for processing atmospheric sensing sequences.
Dataset and Evaluation
The provided dataset consists of multi-national time-series data in CSV format, covering the period from 2010 to 2019.
- Target Regions: Brunei, Singapore, and Thailand.
- Parameters: PM10 and PM2.5 concentrations, supplemented by meteorological data such as wind speed and direction.
- Temporal Structure: Training data spans 2010–2017, while testing data covers 2018–2019. The testing set includes "temporal gaps" of eight consecutive days; participants are tasked with predicting PM10 values for the first three days of these gaps.
- Supplemental Data: Additional data from Indonesian stations is provided to facilitate the transboundary modeling subtask.
Performance is evaluated using Mean Absolute Error (MAE), with submissions ranked based on the average MAE across all monitoring stations within a target country. The focus on a 3-day lead time reflects the practical requirements for health-related decision-making.
Key Insights
The task highlights several critical considerations for regional air quality modeling:
- The Synchronization Challenge: Disparities in recording frequencies (hourly vs. daily) and available parameters across ASEAN countries necessitate robust temporal interpolation and summarization techniques.
- Spatial Interpolation: Given that sensor networks can be sparse or prone to failure, GIS-based spatial interpolation is identified as a vital tool for increasing data density and filling missing values.
- Haze as a Diffusion Process: Framing haze as a diffusion process rather than a simple time-series problem suggests that graph-based neural networks are uniquely suited to model the relationships between cities as nodes in a dynamic network.
Conclusion and Future Directions
While the MediaEval 2021 task provides a rigorous framework for haze prediction, several challenges remain. Data heterogeneity across borders continues to hinder the development of "plug-and-play" models. Furthermore, the integration of unstructured data—such as satellite imagery and social media—with structured sensor data remains an open research question. Future work will likely focus on improving the modeling of long-range biomass fire transport, which is highly sensitive to shifting seasonal weather conditions and regional land-use changes.
Summary updated