DynamicSeq2SeqXGB for PM2.5 imputation in extremely sparse environmental monitoring networks

Abstract

Environmental monitoring networks face critical data gaps that compromise public health protection and regulatory compliance, with missing data rates often exceeding 40% in operational settings. This study validates DynamicSeq2SeqXGB, a novel hybrid model that integrates a sequence-to-sequence encoder–decoder for temporal pattern extraction with an XGBoost regressor for robust gap reconstruction under extreme sparsity. Data from five monitoring stations in Pavlodar, Kazakhstan, collected over a 15-month period from May 23, 2024 to July 19, 2025, were analyzed representing severely compromised infrastructure (completeness rates 23.3–57.5%). The methodology employs adaptive context processing and implements hierarchical decomposition for extended outages. Two data preparation strategies were evaluated: selective compression applying quality thresholds versus full compression retaining all available observations. Benchmarking against classical methods using synthetic gaps of 5–72 hours demonstrated DynamicSeq2SeqXGB’s superiority in 96% of cases under full compression and 100% under selective compression (average 48.8% improvement for both strategies) with corresponding MAE values of 3.7–8.5 μg/m 3 across the Pavlodar stations. Notably, full and selective compression showed equal overall effectiveness (50% win rate each), with optimal strategy depending on station-specific characteristics. External validation on the Beijing dataset (Guanyuan station, 2016) with controlled degradation confirmed cross-regional transferability, achieving MAE of 8.50 μg/m 3 and coefficient of determination (R 2 ) of 0.944 (68–79% improvement over baselines). The method successfully reconstructed PM 2.5 time series even at 23.3% completeness, demonstrating robust performance for operational deployment in severely degraded monitoring networks.