Spatial downscaling turns a coarse satellite or climate-model field, such as 25-km soil moisture, into a 1-km map by learning how fine-scale covariates like elevation and surface temperature relate to the coarse values. A random forest is the usual tool. Random forests offer a well-known free source of uncertainty: each tree is trained on a bootstrap resample, so about a third of the data is "out-of-bag" for that tree, meaning that tree never saw it. Residuals on those held-out points can be turned into prediction intervals with distribution-free coverage guarantees, and no data has to be set aside. This note argues that in downscaling these intervals answer the wrong question. The forest's training labels are the coarse satellite values themselves. Its out-of-bag residuals therefore measure how consistently the covariates reproduce the satellite field. They do not measure how far the map sits from the true ground state. When the source is biased, as with AMSR-E soil moisture over North China, which correlates with in-situ probes at only R = 0.02 (Zhang et al., 2022), an interval can be perfectly calibrated to the satellite and still badly miss the ground.
We propose to measure this gap and close it with a "two-ledger" interval. The first ledger is a spatially blocked out-of-bag estimate of allocation error, the error in spreading a coarse value across sub-pixels. It is dense, local, and free. The second ledger is a small conformal step on the scarce ground stations that rescales the first to cover inherited source error. Validity comes from the stations. Spatial adaptivity comes from the forest. The central prediction can be checked directly: as source quality degrades, naive out-of-bag coverage of ground truth should fall in proportion to source error, while the two-ledger interval should hold near nominal coverage with widths that still vary across terrain.
Core Claim
Write the fine-scale truth at location $x$ as $y(x)$, the downscaled prediction as $\hat y(x)$, and the error as $y(x) - \hat y(x) = a(x) + s(x)$. Here $a(x)$ is allocation error, the gap between the prediction and the fine field that a perfect model would produce from the given coarse input. The term $s(x)$ is inherited source error, the part of the coarse input's deviation from truth that passes through to the fine grid. The literature on source dependence shows that $s$ often dominates. Super-resolved 1-km microwave land surface temperature had nearly the same error as the 25-km and 10-km inputs (Li et al., 2024). Downscaling uncorrected satellite precipitation passes its errors straight to 1 km (Lu et al., 2020). Climate-model biases flow into diffusion-generated local weather (Schmidt et al., 2025). These papers motivate the proposal: they establish that $s$ is large and structured.
The out-of-bag methods we build on are Jackknife+-after-bootstrap, which forms leave-one-out residuals from the trees that excluded each point (Kim et al., 2020); QOOB, which adds quantile forests to that idea and gives tighter intervals in small samples (Gupta et al., 2022); and stability results showing that out-of-bag residuals track true test residuals for forests (Wang et al., 2023). All three see only $a$, because their residuals are computed against the labels the forest was trained on. A further problem specific to space is that $a$ itself is underestimated. Under uniform bootstrap draws on a grid, a cell's neighbors are almost always in-bag, so out-of-bag residuals are optimistic even for allocation error. Data-OOB valuation (Kwon & Zou, 2023; 2D-OOB, 2024) faces a matching limit as a source-quality auditor. It flags coarse cells that disagree with their covariates, such as isolated radio-interference spikes, but a smooth regional bias agrees with itself and is invisible to it. The claim is that no amount of out-of-bag cleverness recovers $s$. The ground stations are needed, and they should be used only for the one thing out-of-bag residuals cannot supply.
Research Program
The construction has three steps. First, we train the downscaling forest at coarse scale, following the practice of Zhang et al. (2022) and Li et al. (2022). The one change is that bootstrap draws are taken over spatial blocks with a buffer: a cell counts as out-of-bag for a tree only if no cell within a chosen distance was in that tree's bag. The distance is set from the empirical range of residual autocorrelation. From these blocked out-of-bag residuals we fit a local quantile model, in the spirit of the localized residual quantile forest of Jiang & Xie (2026) but with no held-out split. This yields an allocation scale $\sigma_a(x)$ at every fine pixel.
Second, we form a normalized score at each of the $n$ ground stations,
$$r_i = \frac{|y_i - \hat y(x_i)|}{\sqrt{\sigma_a(x_i)^2 + \tau(x_i)^2}},$$
where $y_i$ is the station measurement and $\tau(x)$ is a source-error scale. In the simplest version $\tau$ is a single constant fit on the stations. In the richer version $\tau$ is a small function of source-quality indicators: the sensor's quality flags, the disagreement between sensors where several products exist, and the Data-OOB value of the enclosing coarse cell. That last indicator is the one place data valuation enters, and it is ablated.
Third, we apply a jackknife+ step over stations to the scores $r_i$. Because only a scalar rescaling is refit, this costs almost nothing. To respect spatial clustering, stations are grouped into geographic clusters and held out a cluster at a time. The final interval is $\hat y(x) \pm \hat q \sqrt{\sigma_a(x)^2 + \tau(x)^2}$, where $\hat q$ is the conformal quantile of the station scores. The coverage guarantee holds at locations exchangeable with the station clusters. We will state this limit plainly, together with the gap between a point probe and a 1-km pixel, which the source ledger absorbs as part of $\tau$.
The analysis has two parts. We will show, on controlled data, that the ground-truth coverage of naive out-of-bag intervals falls roughly as $\mathbb{P}(|a+s| \le w_a)$. Here $w_a$ is the out-of-bag half-width, so coverage is predictable from the ratio of source to allocation error. We will also verify that block-buffered out-of-bag recovers the allocation coverage that uniform out-of-bag overstates.
Baselines
The naive reference point is the deterministic forest of Zhang et al. (2022), which reports RF R = 0.261 and RMSE = 0.181 m³/m³ against 7 stations in North China, and raw AMSR-E RMSE = 0.20 m³/m³. A bare point map has no interval, so its coverage is undefined. We include it so that point accuracy is checked and shown not to degrade.
The strongest internal competitor is Jackknife+-after-bootstrap or QOOB applied directly to the downscaling forest, using standard uniform bootstrap and residuals against the coarse labels. Kim et al. (2020) report empirical coverage near the 90% target on Communities and Crime, BlogFeedback and MEPS. These are tabular results; no reported result found on Lacuna for these methods on spatial downscaling. It is the strongest competitor because it is exactly what a practitioner would bolt onto an existing random-forest pipeline.
The strongest external competitor is Localized Spatial Conformal Prediction (Jiang & Xie, 2026) calibrated on held-out stations. It reports reaching 90% coverage with intervals about 30% tighter than the next competitor in one simulation, and coming closest to nominal on New Mexico signal-strength data. No numeric coverage on downscaling was found on Lacuna.
We also run EnbPI (Xu & Xie, 2021) as a further out-of-bag competitor, and kriging variance as the classical one. All methods share the same forest (same covariates, number of trees and depth, tuned once on blocked out-of-bag error), the same stations, and the same cluster-level evaluation folds. The only differences are how intervals are formed and which data calibrate them.
Two ablations isolate the ingredients. One uses blocked out-of-bag with no station step, testing whether fixing leakage alone suffices. The other uses a station step with constant width and $\sigma_a$ set to one, testing whether the forest's local signal adds anything. Beating the baselines means achieving ground-truth coverage within two points of 90% across source-quality levels, where the out-of-bag baselines fall well short. It also means achieving narrower intervals than the station-only ablation at matched coverage, and better coverage than it within terrain bins.
Why not just…?
Why not just conformalize the station residuals directly and skip out-of-bag entirely? This is the real threat. With 56 Tibetan stations, a station-level jackknife+ gives valid marginal coverage at station-like sites by itself. Its weakness is that it yields one global width, or a handful of widths for a few bins. It cannot tell a homogeneous plain from a heterogeneous ridge, where allocation error is large. The constant-width ablation is exactly this method. If it matches the two-ledger interval on coverage within terrain-heterogeneity bins and on mean width, the forest's local signal adds nothing and the proposal reduces to "use your stations." We would report that as the finding.
Why not just correct the source first, by merging with gauges, and then use ordinary out-of-bag intervals? Merging helps. In the Tianshan, optimum-interpolation correction lifted IMERG correlation with gauges from 0.516 to 0.616 before downscaling (Lu et al., 2020). But the corrected field still has residual source error that out-of-bag residuals cannot see, and merging spends the same stations without producing an interval. We will run merge-then-out-of-bag on the Tianshan data and expect it to under-cover by the residual source error. Merging can also be stacked under our method, and we will test that combination too.
Why not just use kriging variance? Kriging variance assumes a stationary Gaussian field with a fixed semivariogram. Jiang & Xie (2026) argue that these assumptions are brittle in non-stationary terrain, and kriging trailed learned models across 62,400 precipitation pairs in RainNet (Chen et al., 2020). Kriging is run as a baseline, so its coverage is measured rather than assumed.
Proposed Experiments
The mechanism study uses RainNet (Chen et al., 2020). We take the 4-km StageIV radar-gauge analysis as fine truth and aggregate it to 13 km. We then inject coarse source errors of controlled type and size: smooth regional bias, multiplicative drift, and isolated spikes resembling radio interference. Truth is known everywhere, so we can trace coverage as a function of the ratio of source to allocation error and test the predicted decline of naive out-of-bag coverage. We can also check that Data-OOB catches the spikes but not the smooth bias. In this setting stations are simulated by sampling 7, 56 or 120 pixels, to measure how few stations the second ledger needs.
The real-data test is the Central Tibetan Plateau (Jiang et al., 2017), where the same 56 stations validate products of very different quality: NASA, LPRM and JAXA AMSR-E retrievals, AMSR2, and SMOS. The reported best product is descending-node SMOS downscaled with a neural network, at r = 0.859 and RMSE = 0.052 m³/m³. This gives a natural ladder of source quality with fixed ground truth. The decisive check is the slope of ground-truth coverage against product RMSE, which should be steep for out-of-bag methods and flat for the two-ledger interval.
We replicate on the SMAP 36-km product with 120 stations (Li et al., 2022), where the reported RF fit is R² = 0.848 and ubRMSE = 0.034 m³/m³. This setting adds a stress test: whether intervals around mass-conserving predictions retain coverage. North China, with 7 stations, serves as the extreme-scarcity case. There we report coverage with honest uncertainty bars rather than claiming significance. The metrics are empirical coverage at held-out station clusters (overall, by terrain bin, and by product), mean interval width, and interval score, alongside RMSE and correlation to confirm the point map is unchanged.
References
- Downscaling of AMSR-E Soil Moisture over North China Using Random Forest Regression
- Downscaling Land Surface Temperature Derived from Microwave Observations with the Super-Resolution Reconstruction Method: A Case Study in the CONUS
- The Development of a Two-Step Merging and Downscaling Method for Satellite Precipitation Products
- A Generative Framework for Probabilistic, Spatiotemporally Coherent Downscaling of Climate Simulation
- Predictive inference is free with the jackknife+-after-bootstrap
- Nested conformal prediction and quantile out-of-bag ensemble methods
- Stability of Random Forests and Coverage of Random-Forest Prediction Intervals
- Data-OOB: Out-of-bag Estimate as a Simple and Efficient Data Value
- 2D-OOB: Attributing Data Contribution Through Joint Valuation Framework
- A Spatial Downscaling Method for Remote Sensing Soil Moisture Based on Random Forest Considering Soil Moisture Memory and Mass Conservation
- Spatial Conformal Inference through Localized Quantile Regression
- Conformal prediction interval for dynamic time-series
- RainNet: A Large-Scale Dataset for Spatial Precipitation Downscaling
- Evaluation of Multiple Downscaled Microwave Soil Moisture Products over the Central Tibetan Plateau