But the researchers, led by Dong Xie, were suspicious of their own result. They knew that in population ecology, a long-standing warning called the Moran effect cautions observers that two populations can fluctuate in perfect synchrony without any interaction between them, simply because a shared external driver, such as weather, pushes both in the same direction at the same time. Could the same illusion be at work in water quality? The wetland’s stations all sit under the same sky, drained by the same seasonal rains, fertilized by the same agricultural calendar. When the summer production season opens, nutrient runoff surges everywhere at once; when winter closes it, the whole system settles down together. If every station rises and falls in lockstep with the seasons, then a spatial Markov chain trained on the raw records might mistake that shared seasonal rhythm for genuine spatial dependence, inflating the apparent influence of neighbours when in fact no spillover exists at all.

To test this, the team ran a carefully controlled comparison. They built both classic and spatial Markov chains from the full-year monthly WQI records, then repeated the analysis after deseasonalizing the data, stripping out the common seasonal signal so that each station’s residual fluctuations reflected only its own local dynamics. The results were striking. On the raw data, the spatial Markov chain found that a station’s transitions depended significantly on its neighbours’ state, with a p-value of 0.038, comfortably below the conventional threshold of 0.05. But when the researchers examined the own-versus-neighbour state table underlying that result, they found it was nearly degenerate: 22 percent of its cells were almost empty, because seasonal synchrony kept each station and its neighbours in the same state most of the time. The model had almost no opportunity to observe what happens when a station and its neighbours diverge, so its estimate of spatial dependence rested on a foundation of missing information.

After deseasonalization, the picture changed completely. The state table became fully populated, and the apparent spatial dependence vanished, with the p-value rising to 0.60, far above any threshold of significance. The reversal was not a fluke of one particular way of defining neighbours. The researchers tested six different spatial-weight specifications, ranging from simple binary adjacency to distance-decay formulations, and in every case the deseasonalized data produced p-values between 0.25 and 0.96, while the raw data had yielded values between 0.0002 and 0.077. The team also corroborated the finding with a Moran ladder, a sequence of spatial autocorrelation statistics computed at increasing lags. Pooled Moran’s I collapsed from 0.80 on the raw data to −0.11 after deseasonalization, and the spatial-autoregressive coefficient ρ fell from 0.70 to −0.16. In other words, once the shared seasonal driver was removed, the stations were essentially independent, and the strong positive autocorrelation in the raw records was attributable to the seasons, not to the hydrological connections between sites.

A two-way variance decomposition drove the point home. When the researchers partitioned the variance of the WQI into components attributable to time and to site, they found that 81.6 percent of the variance was explained by time, while only 7.9 percent was explained by site. The wetland’s water quality is overwhelmingly a temporal phenomenon, driven by the seasonal cycle of agricultural production, and the differences between stations are comparatively minor. The production season also opened a direct channel for collapse: during the active growing months, the WQI could plunge directly from the Good category to the Poor category, a transition that was absent in winter. This asymmetry suggests that the seasonal pulse of nutrient loading does not merely shift the mean water quality but changes the very dynamics of the system, creating windows of vulnerability in which a station’s condition can deteriorate abruptly without passing through intermediate stages.

The researchers were careful to acknowledge the limits of their own design. A power analysis showed that the study’s detection limit, the smallest deviation between conditional transition matrices that the design could reliably detect, was about 0.19, which is above the deseasonalized deviation of 0.09 that they actually observed. This means that while the strong apparent spatial dependence in the raw data can be confidently attributed to seasonal synchrony, the study cannot exclude the possibility of weak spatial interaction between stations. The hydrological connections in the wetland are real, and some subtle spillover may exist below the threshold of what 288 site-months can resolve. But the dominant signal, the one that inflated the spatial Markov chain’s estimate of dependence on the raw data, is clearly the Moran effect, the synchronizing influence of a shared seasonal driver rather than genuine spillover between sites.

The implications reach well beyond one wetland. Spatial Markov chains are increasingly popular in environmental monitoring, used to infer spatial dependence among air-quality stations, water-quality sites and pollution indices across entire river networks. Many of these applications train their models on raw records that retain the common seasonal signal, and the researchers’ results suggest that many published estimates of spatial dependence may be inflated by the same illusion. A model that finds significant neighbour effects on raw data may be detecting nothing more than the fact that all sites respond to the same seasonal driver, whether that driver is the agricultural calendar, the monsoon, the heating season or the annual cycle of temperature and light. The remedy is straightforward but rarely applied: deseasonalize the records before fitting the chain, and report the sparsity of the own-versus-neighbour state table along with the design’s detection limit, so that readers can judge whether the model had sufficient information to estimate spatial dependence in the first place.

The study also offers a broader lesson about the interpretation of spatial patterns in environmental data. Positive spatial autocorrelation, the statistical tendency of nearby sites to resemble each other, is often read as evidence of spatial process, of interaction, diffusion or spillover. But as ecologists have long recognized, and as this wetland study now demonstrates in a water-quality context, synchrony to a shared driver can produce exactly the same pattern without any interaction at all. Distinguishing between the two requires more than fitting a spatial model to raw records; it requires deliberately removing the shared signal and asking whether the pattern survives. In the Taihu wetland, it did not. The stations’ water quality rose and fell together because they share the same seasons, not because they influence each other, and the strong spatial dependence that the spatial Markov chain detected on the raw data was largely an artefact of that seasonal synchrony.

For the managers of agricultural wetlands, the findings carry a practical message. The dominance of the temporal component, with 81.6 percent of WQI variance attributable to time, means that water quality in this system is governed primarily by the seasonal cycle of agricultural production rather than by the spatial structure of the wetland itself. Interventions that target the seasonal pulse of nutrient runoff, such as adjusting fertilizer timing, establishing buffer zones during the production season or managing drainage during vulnerable months, are likely to be more effective than interventions aimed at the spatial connectivity between sites. The direct Good-to-Poor collapse channel opened by the production season highlights the importance of anticipating abrupt deterioration during the active growing months, when the system’s resilience is lowest and the consequences of nutrient loading are most severe.

As environmental monitoring networks grow denser and datasets grow longer, the temptation to fit increasingly sophisticated spatial models to raw records will only increase. This study is a reminder that the sophistication of the model is no substitute for the care of the analysis. A spatial Markov chain, however elegant, can only estimate spatial dependence from the information it is given, and if that information is dominated by a shared seasonal driver, the model will faithfully report the driver’s influence as if it were spatial process. The researchers’ prescription, deseasonalize first, check the state table for sparsity, and report the detection limit, is a simple workflow that could spare the field a generation of inflated estimates. In the fine-scale agricultural wetland of the Taihu basin, the seasons, not the neighbours, wrote the story of water quality, and the models that failed to listen to the seasons mistook the chorus for a conversation.

Subject of Research: Seasonal synchrony inflates spatial dependence in water-quality Markov chains: evidence from a fine-scale agricultural wetland

Article Title: Seasonal synchrony inflates spatial dependence in water-quality Markov chains: evidence from a fine-scale agricultural wetland

Article References: Xie, D., Xue, Y., Li, X., & Feng, Y. (2026). Seasonal synchrony inflates spatial dependence in water-quality Markov chains: evidence from a fine-scale agricultural wetland. Environmental Monitoring and Assessment, 198(9), Article 1027. https://doi.org/10.1007/s10661-026-15862-5

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15862-5

Keywords: Seasonal, synchrony, inflates, spatial, dependence, water-quality, Markov, chains, evidence, fine-scale, agricultural, wetland