Environmental studies often want to know which driver controls a response such as evaporation, surface cooling, or bird song timing, when the candidate drivers move together: temperature, humidity deficit, radiation, soil moisture, leaf area. The usual tool is hierarchical partitioning. It fits every subset of the candidate predictors and averages how much each one adds to explained variance. That separates a driver's own contribution from the part it shares with correlated drivers. Every environmental application of it we found runs on pooled, static data: one cross-section, monthly means, or decades of records lumped together. Continuous sensor streams tell a different story. Thermal and ecological behavior changes sharply between the noon radiation peak and the night, and between one weather front and the next.
We propose to split a continuous half-hourly record into nested time scales and then run hierarchical partitioning separately at each scale. The scales run from within-day through synoptic (multi-day weather systems), seasonal, interannual, and between-site. The result is a table that says both where in the time hierarchy the variance lives and why, meaning which driver independently explains it at that scale. Our central claim is that a pooled partitioning is close to a variance-weighted average of these scale-specific answers. A pooled partitioning therefore systematically hides drivers that dominate low-variance scales, and the scale-resolved table is better than the pooled ranking at predicting which drivers help forecasting at a given horizon.
Core Claim
The closest existing work comes in three forms. The global synthesis Phenology acts as a primary control of urban vegetation cooling and warming ran hierarchical partitioning on 3,634 pooled observations of the air temperature difference between vegetated and background sites. Plant growth stage explained 37.6% of the variance, seasonality 23.0%, and latitude 11.4%, while time of day came out minor. The FLUXNET study Ecosystem Evapotranspiration Partitioning and Its Spatial-Temporal Variation attributed interannual variation in the transpiration fraction to leaf area index (22%) and vapor pressure deficit (5%), but only at one scale and without formal partitioning across nested strata. The soil nutrient study Evaluation of Spatial Variability of Soil Nutrients in Saline-Alkali Farmland used partitioning on monthly climate means and a single sampling campaign.
All three motivate the gap. Each attributes variance with collinearity handled, but at one pooled scale. Our hypothesis is that "time of day is minor" and "VPD explains 5%" are partly artifacts of pooling. The diurnal and synoptic components of a response carry less total variance than the seasonal component, so a pooled decomposition gives them little weight even when one driver almost completely explains them.
We borrow the nesting idea from the analysis of molecular variance used in population genetics. That method splits total squared deviation into within-population, among-population, and among-region strata, as in the Qinling primate study by Kang Huang et al.. We reuse that variance bookkeeping on time scales instead of geographic strata. We then add, at each stratum, the driver-level partitioning that the genetics method does not have.
Research Program
Let $y_t$ be a half-hourly response at one site, for example latent heat flux. We decompose it into nested block-mean components: within-day deviations from the daily mean, daily means minus weekly means (synoptic), weekly means minus seasonal means, seasonal means minus the annual mean, and annual means minus the site mean. Above the site level, the strata are site-within-biome and biome. For balanced blocks these components are exactly orthogonal, so
$$\mathrm{Var}(y) = \sum_{s} \mathrm{Var}(y^{(s)}),$$
where $y^{(s)}$ is the component of the response at scale $s$. Each candidate driver $x_j$ (vapor pressure deficit, net radiation, air temperature, soil moisture, leaf area index, wind speed, and a phenology stage) is decomposed the same way.
At each scale we regress $y^{(s)}$ on the drivers' same-scale components. We run hierarchical partitioning over all $2^K$ subsets, which is cheap at $K\approx 8$. This gives an independent contribution $I_{j,s}$ and a joint contribution $J_{j,s}$. The variance share of driver $j$ at scale $s$ is then
$$A_{j,s} = \frac{\mathrm{Var}(y^{(s)})\, I_{j,s}}{\mathrm{Var}(y)},$$
which reads directly as "the fraction of all variance in the response that sits at scale $s$ and is independently explained by driver $j$."
The analytic part of the project is to characterize when a pooled partitioning of $y$ on $x$ approximates $\sum_s A_{j,s}$. It should do so when the scale components are near-orthogonal and the driver effects are scale-separated. It should fail when cross-scale terms are large, for example when a seasonal driver modulates a diurnal response. We will measure that failure rather than assume it away.
Three practical problems need explicit handling. The first is missing data and sensor dropouts, which unbalance the blocks. We will use gap-weighted block means and report how attributions change with an imposed gap fraction. The second is negative joint contributions under suppression, which we report separately rather than truncate. The third is lagged dependence: soil moisture acts at synoptic scale through accumulated rain. We will test same-scale lagged driver terms as an extension.
Baselines
The naive reference is pooled hierarchical partitioning, run on the same half-hourly FLUXNET2015 records exactly as the global cooling synthesis did on its pooled data (Phenology acts as a primary control). It uses the same driver set and the same subset enumeration. It is the default practice, so it is what our claim has to beat. "Beating it" means three things: the scale-resolved table recovers known drivers on semi-synthetic data where the pooled ranking does not, and it does better on the forecasting check described below. A second naive point is ranking drivers by their marginal correlation with the response, ignoring collinearity entirely.
The strongest existing approach is SHAP attribution on a pooled gradient-boosted model. This is the attribution route used for soil nutrients (Evaluation of Spatial Variability of Soil Nutrients), applied to an XGBoost model like the one in the FLUXNET evapotranspiration study. That study reports soil-evaporation skill of NSE = 0.657, R = 0.806, and RMSE = 11.344 W/m² across ten biomes (Ecosystem Evapotranspiration Partitioning). We will retrain XGBoost on the same sites with the same driver inputs plus time-of-day and day-of-year features. We will use the same hyperparameter search budget as any learned component of our pipeline, and check that we reproduce comparable skill before comparing attributions. We will also reproduce its interannual numbers (LAI 22%, VPD 5%) to check our pipeline. No reported result was found on Lacuna for scale-resolved attribution on FLUXNET2015, so we have no prior number to target there.
The ablations isolate the key ingredient, which is partitioning after scale separation. "Scale decomposition without partitioning" reads attribution from standardized coefficients of a multiple regression at each scale. "Partitioning without scale decomposition, but stratified" runs separate partitioning for day versus night and for each season, using bins rather than nested components.
Why not just…?
Why not just stratify the data by day/night or season and run partitioning in each bin? This is cheap and partly answers the question. However, bins still mix scales: a daytime bin contains diurnal, synoptic, and seasonal variance together. Bins also give no variance-weighted account of how much of the total each stratum carries. We include stratified partitioning as an ablation. If its rankings match the nested decomposition on the semi-synthetic benchmark and on the forecasting check, then the nested machinery is unnecessary, and we would report that.
Why not just use wavelet coherence or spectral regression? This is the most serious threat, because wavelet methods already give scale-dependent covariance. Standard coherence is pairwise, though, and collinear drivers such as VPD, temperature, and radiation show high coherence at the same scales. That is exactly the confusion partitioning exists to resolve, as the blackbird study showed: traffic noise, light at night, and building density could only be separated by partitioning (Temporal shift in Eurasian Blackbird and European Robin song). We treat the scale decomposition as replaceable: nested block means and a maximal-overlap Haar wavelet both run in the pipeline. The claim rests on partitioning within scales, not on the choice of filter. Per-driver wavelet coherence is a baseline, and on semi-synthetic data with planted collinear drivers it should misattribute where our method does not.
Why not just use SHAP on a flexible model with time features? SHAP values for a fitted model answer how the model uses each feature, averaged across all scales. A tree ensemble can route diurnal variation through a time-of-day feature rather than through radiation. In that case SHAP credits "time" and hides the physical driver, which is the spurious-shortcut failure that Hierarchical interpretations for neural network predictions documents for deep classifiers. If SHAP restricted to physical drivers already matches our scale-resolved ranking, our contribution shrinks to a bookkeeping device. The semi-synthetic recovery test settles this directly.
Why not a mixed-effects model with variance components? Random-effects models split variance across nested strata well, which is the "where." They do not split collinear fixed-effect drivers into independent shares within each stratum, which is the "why." The neurodevelopment study José Villar et al. illustrates the "where" half only. We report variance components alongside our table, and they should agree with our $\mathrm{Var}(y^{(s)})$ column.
Proposed Experiments
The main testbed is FLUXNET2015, with half-hourly latent heat, VPD, net radiation, and air temperature across 55 sites and 10 biomes, plus ERA5-Land soil moisture and MODIS leaf area index, as assembled in Ecosystem Evapotranspiration Partitioning. The responses are latent heat flux and the transpiration fraction.
The first decisive check is recovery on semi-synthetic data. We keep the real, collinear driver series, generate responses from planted rules such as "VPD drives diurnal variance, LAI drives seasonal variance, soil moisture drives synoptic variance," and add realistic noise and gaps. We then score each method by rank correlation and absolute error between recovered and planted shares, scale by scale.
The second check is pooled versus resolved on real data. We predict that VPD and radiation carry large independent shares of within-day and synoptic variance even where the pooled analysis assigns them single-digit shares, as in the 5% interannual VPD figure. If pooled and resolved rankings agree across most sites, the central claim fails.
The third check is forecast relevance. We build forecasters at 1-hour, 1-day, and 1-week horizons, restricting inputs to the top-ranked drivers at the matching scale under each attribution method. We compare out-of-site skill (NSE) against all-driver and persistence models. A useful attribution should pick feature sets that lose little skill relative to all drivers and beat feature sets picked by pooled rankings.
A secondary check reuses the daily station records from AgroML and Evaluation of Eight Decomposition-Hybrid Models. The latter found that high-frequency decomposition components inflate test scores but destabilize forward forecasts. We ask whether bands with near-zero driver-explained share are the ones whose removal stabilizes forward rollouts.
Bootstrap resampling by site and by year gives uncertainty for every entry in the table.
References
- Phenology acts as a primary control of urban vegetation cooling and warming: A synthetic analysis of global site observations
- Ecosystem Evapotranspiration Partitioning and Its Spatial-Temporal Variation Based on Eddy Covariance Observation and Machine Learning Method
- Evaluation of Spatial Variability of Soil Nutrients in Saline-Alkali Farmland Using Automatic Machine Learning Model and Hyperspectral Data
- Population structure of the golden snub‐nosed monkey Rhinopithecus roxellana in the Qinling Mountains, central China
- Temporal shift in Eurasian Blackbird Turdus merula and European Robin Erithacus rubecula song under urban pressures
- Hierarchical interpretations for neural network predictions
- Neurodevelopmental milestones and associated behaviours are similar among healthy children across diverse geographical locations
- AgroML: An Open-Source Repository to Forecast Reference Evapotranspiration in Different Geo-Climatic Conditions Using Machine Learning and Transformer-Based Models
- Evaluation of Eight Decomposition-Hybrid Models for Short-Term Daily Reference Evapotranspiration Prediction