Tree-Based Sequence Models Reconstructed PM2.5 Gaps With the Highest Accuracy
Air pollution surveillance depends on continuity, and PM2.5 data rarely arrive in perfect order. Missing hours can distort exposure estimates, blunt seasonal analyses, and weaken any attempt to link particulate matter to health outcomes. In Pavlodar, Kazakhstan, investigators built a broad comparison of gap-filling strategies to see which approaches could most faithfully rebuild incomplete time series across short, medium, and extended interruptions.
The study tested 46 methods across five representative gap lengths ranging from 5 to 72 hours. A broad methodological sweep like this matters because a model that works for a few missing hours can collapse when the gap stretches across a full weather shift, a traffic pattern change, or a nighttime inversion. The central finding was clear: tree-based models combined with bidirectional sequence-to-sequence architectures performed best, with XGB Seq2Seq delivering a mean absolute error of 5.231 ± 0.292 μg/m3 for 12-hour gaps. Compared with basic statistical methods, that represented a 63% improvement.
Gap Length Changed the Value of Meteorological Data
The most practical message from the analysis concerns not only model choice, but also the length of the gap itself. For short 5-hour gaps, multivariate models that incorporated meteorological variables offered only modest gains of 2% to 3% over simpler approaches. As the missing interval widened, the value of those extra predictors increased sharply. For 48- to 72-hour gaps, the improvement climbed to 16% to 18%.
This pattern is biologically and physically plausible. Fine particulate pollution does not move in isolation; it follows wind, temperature structure, boundary-layer behavior, humidity, and other atmospheric drivers that shape dispersion and accumulation. When a gap spans only a few hours, local temporal autocorrelation can often carry the estimate. When the gap runs through multiple diurnal cycles or weather transitions, meteorology becomes indispensable because it helps the model infer how the atmosphere would have transported or trapped particles during the missing period.
Short gaps were one thing. Long gaps demanded weather-aware reconstruction.
Dynamic Models Handled Real-World Gaps Beyond Their Training Range
One of the most interesting operational findings came from dynamic multivariate models. Although they were trained on gaps up to 72 hours, they successfully processed real-world interruptions ranging from 1 to 191 hours. That flexibility is not trivial. In environmental monitoring, the useful method is rarely the one that performs beautifully only under tidy laboratory conditions; it is the one that continues to function when sensors fail for practical reasons, maintenance is delayed, or network interruptions extend beyond the expected window.
The ability to accept longer gaps than those used in training suggests a model with some resilience to distributional stress, though caution remains appropriate when extrapolating far beyond the training envelope. A 191-hour interruption is not a minor perturbation. At that scale, any imputation method is forced to infer prolonged pollution behavior from surrounding patterns, and confidence in the exact hourly trajectory should naturally be lower than in a short gap scenario. Even so, the fact that the model retained usable operational behavior across such a range strengthens its value for real surveillance systems.
Reconstructed Data Revealed Persistent Threshold Exceedance
Once the missing values were filled, the reconstructed PM2.5 series painted an unsettling exposure picture. 61.2% of monitored hours exceeded the World Health Organization daily threshold of 15 μg/m3. The time series also showed strong seasonal structure and pronounced diurnal cycles, reinforcing that ambient particulate pollution in the setting was neither random nor isolated. It was patterned, recurrent, and persistent enough to shape population exposure over large stretches of time.
This matters because incomplete surveillance can make a polluted environment look cleaner than it really is. Every missing hour risks biasing the exposure profile downward if the omitted period happened to occur during a pollution peak. A robust gap-filling method does not merely smooth a dataset; it affects how many people are classified as exposed, how episode frequency is counted, and how much weight policymakers assign to pollution mitigation.
Why This Approach Matters for Public Health Analysis
PM2.5 is one of the most consequential air pollutants in epidemiology because it penetrates deep into the lungs and is associated with cardiovascular and respiratory harm. Reliable hourly data are the backbone of exposure assessment, and exposure assessment is the backbone of any downstream inference about risk. When the data stream is fragmented, the scientific questions become harder to answer with confidence. Seasonal peaks can be underestimated, short spikes can disappear, and intervention studies can lose resolution.
The current evaluation offers a methodological answer to that problem. By comparing statistical approaches with neural-network-based and tree-based strategies, the work identifies a path toward more dependable reconstruction of incomplete environmental records. The best-performing methods did not rely on a single rigid template. They adapted to the shape of the gap, and in the multivariate setting they leveraged meteorological context when the missing stretch became longer. In a surveillance network where missingness is inevitable, adaptability is not a luxury; it is the difference between a usable exposure record and a compromised one.
Limits of Reconstruction Still Depend on the Missing Interval
No imputation system can recreate a missing pollution history with perfect certainty, especially when the gap spans several days. The study’s own results make that plain. Performance improved as models became more sophisticated, but the advantage of meteorological enrichment was not uniform across all gap lengths, and the longest interruptions necessarily pose the greatest inferential challenge. A model trained on gaps up to 72 hours may remain operational at 191 hours, yet the uncertainty around those far-end reconstructions should remain front of mind for analysts who use the completed series in health studies or regulatory review.
Still, the study delivers something practical: a ranked framework for choosing a filling method according to the length of the missing segment and the availability of auxiliary weather data. For short gaps, simpler strategies may suffice. For longer gaps, the evidence points toward multivariate tree-based sequence models, especially XGB Seq2Seq, as the strongest option among the approaches tested. In a field where missing data often force investigators to choose between discarding records and accepting crude interpolation, that is a meaningful advance in how environmental exposure data can be recovered and used.