The research team, led by Baoni Xie and colleagues at Hebei GEO University, tracked ecological quality across northern Hebei from 2000 to 2022 using the remote sensing ecological index, or RSEI. This index integrates four satellite-derived components: greenness, wetness, dryness, and heat, into a single measure of environmental quality. Over the twenty-two-year record, mean RSEI across the study area rose by 29.79 percent, and 52.38 percent of the region showed a statistically substantial increase at a rate of 0.005 per year. On the surface, the picture looks like a clear ecological success story.

But the researchers went a step further. Rather than simply comparing areas where vegetation increased or decreased, they adjusted the RSEI responses against stable reference conditions, asking whether the change associated with a gain or a loss exceeded what would have happened had the land remained unchanged. This approach, borrowed from recent work on the asymmetric climate effects of forest cover change, allowed them to isolate the true ecological signature of each transition. Forest gain was consistently associated with positive adjusted RSEI responses, while forest loss was consistently negative, exactly as intuition would predict. Grasslands, however, told a very different story.

Grassland gain was associated with lower RSEI values, and grassland loss with higher ones, a reversal of what most observers would expect. The explanation lies in the land use pathways behind those changes. Areas where grassland expanded had mostly been converted from cropland, and newly established grassland on former farmland scored lower on the ecological index than stable grassland. Conversely, when grassland disappeared, the land typically transitioned into forest or back into cropland, and those destinations carried higher RSEI values than the grassland they replaced. In other words, the ecological quality of a pixel depends not just on whether grassland is present, but on what it was before and what it becomes after.

The asymmetry between gains and losses also varied over time. For forests, a significant mirror asymmetry appeared only during the 2004 to 2012 period, when the negative response associated with forest loss exceeded the positive response associated with forest gain, quantified by a mirror asymmetry coefficient of −0.049. For grasslands, the aggregate asymmetry across the full study period was not statistically supported, suggesting that the reversed responses, while real, did not accumulate into a consistent directional bias. These nuances matter because they show that the ecological meaning of vegetation change is contingent on both timing and context.

To understand what drives these divergent responses, the team turned to interpretable machine learning. They trained an eXtreme Gradient Boosting model, XGBoost, on the RSEI responses and then applied Shapley Additive Explanations, or SHAP, to decompose each prediction into contributions from individual environmental variables. This combination, increasingly popular in environmental science, provides both high predictive accuracy and transparent attribution, allowing researchers to see not just which factors matter, but how they matter and whether their influence is linear or nonlinear.

The results revealed striking differences between vegetation types and change directions. For forest gain, terrain elevation, captured by the digital elevation model, was the leading predictor at 28.04 percent of total importance, followed by evapotranspiration at 20.59 percent. Forest loss, by contrast, was dominated by nighttime light, a proxy for human activity and urbanization, at 22.66 percent, with precipitation second at 20.20 percent. Grassland gain was primarily associated with evapotranspiration at 34.50 percent and precipitation at 15.88 percent, while grassland loss was driven by evapotranspiration at 20.65 percent and elevation at 20.48 percent. These patterns suggest that forest expansion is largely a topographic and hydrological phenomenon, occurring where water and terrain favor tree growth, whereas forest loss is more closely tied to human pressure.

The SHAP dependence relationships were nonlinear and varied among vegetation types, change directions, and study periods, meaning that the same environmental variable could push ecological quality in different directions under different circumstances. Evapotranspiration, for example, emerged as a dominant factor for both grassland gain and grassland loss, but its influence likely operates through different mechanisms: high evapotranspiration may indicate vigorous, water-rich vegetation where grassland is establishing, while signaling water stress or degradation where grassland is being lost. This kind of context-dependent behavior is precisely what linear regression approaches miss and what interpretable machine learning is designed to expose.

The broader implication of the study is a caution against using vegetation area change as a proxy for ecological health. Northern Hebei sits at the intersection of China’s Grain for Green program, grassland conservation policies, and rapid urbanization, and the region has been the subject of extensive restoration efforts. The new findings show that a net increase in forest or grassland area can mask underlying asymmetries in ecological quality, and that gain and loss should be interpreted relative to stable conditions and specific land use pathways rather than in isolation.

For environmental managers and policymakers, the message is that monitoring programs need to account for transition pathways, not just land cover snapshots. A hectare of new grassland on former cropland is not ecologically equivalent to a hectare of stable grassland, and the loss of grassland to forest may represent an improvement rather than a degradation. As satellite records grow longer and machine learning tools become more accessible, studies like this one point toward a more nuanced accounting of ecological change, one that recognizes that the direction of vegetation change and the direction of ecological quality change are not always the same thing.

Subject of Research: Ecological quality responses to forest and grassland gain and loss in northern Hebei, China, assessed with RSEI and XGBoost-SHAP

Article Title: Assessing divergent ecological responses to forest and grassland dynamics using RSEI and interpretable machine learning in Northern Hebei, China

Article References: Xie, B., Li, H., Li, W., Qin, Z., Wang, Y., Zuo, H., Zhu, J., & Yu, S. (2026). Assessing divergent ecological responses to forest and grassland dynamics using RSEI and interpretable machine learning in Northern Hebei, China. Environmental Monitoring and Assessment, 198(11), Article 1209. https://doi.org/10.1007/s10661-026-16061-y

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16061-y

Keywords: RSEI, XGBoost-SHAP, forest gain, grassland loss, ecological quality, land use transition, northern Hebei, evapotranspiration, remote sensing, asymmetry, vegetation dynamics, interpretable machine learning