Imputation of urban environmental sensor data using gated attention bidirectional long short-term memory (GA-BiLSTM):...

Area of Science:

• Environmental Science
• Data Science
• Urban Planning

Background:

• Urban environmental monitoring networks face data gaps from sensor failures and communication issues.
• Ensuring data continuity and quality is crucial for effective environmental management.
• Existing data imputation methods may struggle with complex spatiotemporal dependencies in urban settings.

Purpose of the Study:

• To develop and evaluate an advanced model for imputing missing data in urban environmental monitoring networks.
• To assess the performance of the proposed gated attention bidirectional long short-term memory (GA-BiLSTM) model against established methods.
• To investigate the influence of different gap durations and the role of peripheral monitoring nodes.

Main Methods:

• Development of a gated attention bidirectional long short-term memory (GA-BiLSTM) model.
• Utilizing data from the CROCUS network in Chicago for model training and evaluation.
• Comparative analysis against XGBoost and K-nearest neighbors under various data gap scenarios (short-term and prolonged).

Main Results:

• The GA-BiLSTM model significantly outperformed XGBoost and K-nearest neighbors in data imputation accuracy.
• GA-BiLSTM demonstrated superior performance, especially in handling prolonged data outages (up to ten days).
• Analysis revealed the crucial predictive importance of peripheral rural sensor nodes for urban data imputation.

Conclusions:

• Advanced imputation techniques like GA-BiLSTM can substantially enhance the reliability of urban environmental monitoring data.
• The findings underscore the value of integrating peripheral data sources for robust urban monitoring systems.
• Improved data imputation supports more resilient data infrastructures essential for urban sustainability initiatives.