Public Splits, Private Leaves: Tree Ensembles that Learn from Private Ground Sensors and Public Satellites

Many of the most useful environmental maps, such as fine-scale air pollution or soil moisture, come from combining two kinds of data. One is free, public satellite imagery that covers the whole planet. The other is a sparse set of ground sensors that measure the true value at specific points. Increasingly, the ground sensors belong to households, farms, or companies whose readings and locations are sensitive. A reading tied to a place reveals who lives where and how they live. Today's pipelines handle this by uploading everything in plaintext to one cloud platform. We propose a different split of labor for random forests (collections of decision trees, each fit to a resampled copy of the data, whose predictions are averaged). The shape of every tree, meaning which feature it splits on and where, is learned entirely from public satellite data, against a public coarse-resolution product that stands in for the private target. The private sensors are touched only to fill in the tree's leaf values. Those leaf values are just per-leaf sums and counts, and devices can contribute to them through cheap, non-interactive secret sharing. A differential privacy guarantee covers the result: a mathematical bound on how much any one sensor's data can change the output.

The opportunity comes from a known tension. Under differential privacy, choosing good split points eats the privacy budget, so the best private boosted trees give up and split at random. This is the "randomness paradox" of Federated Boosted Decision Trees with Differential Privacy. Spatial regression, however, depends on physically meaningful splits on lagged moisture, aerosol optical depth, or terrain, as the soil moisture downscaling study shows. Geospatial problems, unlike generic tabular ones, come with a free and well-correlated public signal: the coarse satellite product itself. Our claim is that growing structure against this public proxy gives splits that are meaningful and cost no privacy. This should close most of the gap between private and non-private forests at budgets where random splits fail.

Core Claim

Two conditions hold in a sensor-fusion task. A public coarse product correlates with the private fine-scale target. The tree structure is grown only from public rasters, using that proxy as the training target. Under those conditions, a forest whose leaves are estimated from secret-shared, noise-protected sensor sums should beat both totally random splits and impurity-weighted private splits at privacy budgets $\epsilon \le 1$. It should also beat a model trained on the proxy alone. Each device's cost stays linear in the number of leaves.

What is new is not the honest structure/estimation separation, which Multinomial Random Forest already uses. What is new is where the structure comes from. MRF picks splits from the private data through what it proves is an exponential mechanism, so it pays privacy for structure. The DP boosted-tree work picks splits randomly and pays nothing, but learns no structure. We pay nothing and still learn structure, because the information comes from a public source that exists in this domain and not in generic tabular benchmarks.

Research Program

Building the trees. Tree construction runs next to the satellite archive, in the compute-to-data style surveyed in From Copernicus Big Data to Extreme Earth Analytics. It reads public covariates with windowed reads and geographic samplers of the kind provided by TorchGeo, and the training target is the public coarse product. The covariates and coarse products depend on the task. For soil moisture, they are MODIS land surface temperature and vegetation indices, ERA5-Land precipitation, soil texture and elevation, with SMAP 36 km as the coarse product. For air quality, they are aerosol optical depth, NO2 and population, with a satellite-derived PM2.5 field as the coarse product. Fitting is ordinary greedy splitting on public pixels, so it is cheap and needs no cryptography. The finished trees are published.

Filling the leaves. Each sensor device fetches the public covariates for a coarse tile that covers its location. It routes itself through every tree locally. It then forms a vector of length $2L$, where $L$ is the total number of leaves, holding its clipped daily mean reading and a count of 1 in the slots of the leaves it falls into.

The tile is deliberately coarse. A precise windowed read would reveal the device's location to the raster server, much as One bad apple can spoil your IPv6 privacy shows that one leaked identifier can anchor a whole household.

The device masks and splits its vector across $k$ independently operated aggregators. This uses the non-interactive additive scheme of Non-Interactive and Secure Data Aggregation Scheme for Internet of Things, which costs each user $O(m)$ work and needs no user-to-user messages. That matters when devices drop offline. The aggregators add distributed Gaussian noise and release

$$\hat\mu_\ell = \frac{S_\ell + \eta^S_\ell}{\max(N_\ell + \eta^N_\ell,\ 1)},$$

where $S_\ell$ is the sum of clipped readings in leaf $\ell$, $N_\ell$ is the device count, and $\eta$ is the calibrated noise. Each device appears in exactly one leaf per tree. Sensitivity therefore grows with the number of trees $B$, not with $L$, and we account for it with Rényi differential privacy as in the DP boosted-tree work.

The central trade-off. Deeper trees fit the target better, but they spread devices thinner, so each noisy leaf average gets worse. The research question is how this trade-off behaves when the structure is informative. We will study it by sweeping depth, $B$, and $\epsilon$. A second structural lever is to merge sibling leaves that the public proxy cannot distinguish.

Physical consistency. After release, predictions are rescaled so that every coarse cell's average matches the public coarse product. This is the mass-conservation correction from the soil moisture paper. Because it uses only public data and released outputs, it costs no privacy.

Decentralization. We make a deliberate choice here. The aggregators are a small set of independent operators, for example municipal agencies, a research institute, and a citizen-science group. There is no ledger. Geospatial Blockchain documents that gas costs and state growth make on-chain geospatial computation impractical even for metadata. Since our protocol has a single aggregation round per forest, a ledger would add cost and no function.

The real trust assumption is that the aggregators do not collude. We state it plainly. Collusion would expose individual masked vectors, though released outputs would still carry differential privacy if the noise is added by each aggregator separately.

Baselines

The naive reference point uses no private data beyond what is needed to fit a model. A random forest is trained on public covariates against the coarse product and applied directly at fine scale. Simple bilinear upsampling of the coarse product is a second, even weaker reference. If our method cannot beat the proxy-only forest, the private sensors are adding nothing that survives the noise.

The non-private ceiling is the plaintext model each benchmark paper reports. The soil moisture downscaling random forest reaches test R² of 0.848 and unbiased RMSE of 0.034 m³/m³ with lagged moisture features. The OpenAQ stacked ensemble, which combines random forest, XGBoost and LightGBM, reports a global median absolute error of 5.60 µg/m³ and overall R² of 0.49. We will reproduce both pipelines with our features before comparing against them, so the ceiling reflects our data rather than theirs.

The strongest private competitor is DP-TR Newton with Iterative Hessian splitting from Federated Boosted Decision Trees with Differential Privacy. That paper reports AUC within 1–2% of non-private XGBoost at $\epsilon = 0.5$ on Adult and Bank Marketing. No reported result was found on Lacuna for it on geospatial regression. It is the strongest competitor because it is the best documented way around the budget cost of split search. We will run it with regression loss and batched updates, on the same sensor data and the same $\epsilon$ grid, with matched Rényi accounting. Tuning will use the same number of configurations, chosen on public-proxy validation only, so that tuning leaks nothing.

The second private competitor is the Multinomial Random Forest. Its exponential-mechanism splits are run on the private data under the same total budget, with part of the budget given to structure. It reports average rank 1.5 against Breiman forest's 1.8 on 22 UCI datasets. No reported result was found on Lacuna for it under a stated $\epsilon$ on these geospatial tasks.

We also include a local differential privacy baseline, in which each device noises its own reading before upload. It needs no trusted aggregators and shows what the secret-sharing step buys.

Two ablations isolate the key ingredient. The first grows structure from public covariates without the proxy target, using unsupervised spatial partitioning such as k-d splits on covariates. This separates "public features" from "public proxy supervision." The second removes the mass-conservation step.

Our method counts as beating a competitor if, at $\epsilon \le 1$, its station-held-out RMSE is lower across at least five seeds with non-overlapping 95% intervals. We also want it to recover at least half of the gap between the proxy-only forest and the non-private ceiling.

Why not just…?

Why not just run the existing pipeline on Earth Engine with access controls and TLS? Access control protects data from outsiders, not from the platform operator or anyone with root access. That is the "hidden danger" identified in A Spatial Data Security Model Under the Cloud Environment. It also offers nothing against inference from the released map itself. Our design keeps the satellite processing in the cloud, where it belongs, and removes only the sensor plaintext. The comparison against the non-private ceiling measures exactly what this costs in accuracy.

Why not just use totally random splits, which already escape the randomness paradox? This is the most serious competitor, and we treat it as the main baseline rather than arguing it away. Random splits learn no feature selection. In spatial regression only a few features, such as lagged moisture or aerosol optical depth, carry the signal, so random splits should waste depth on irrelevant covariates. That is the prediction. If DP-TR matches us at $\epsilon \le 1$, the proxy is not adding information beyond what random structure plus private leaves already provide, and the claim fails.

Why not just train on the public coarse product and skip the private sensors? The proxy is biased at fine scale, which is why ground truth exists at all. The OpenAQ study's weak results in data-poor regions suggest that ground information matters most where proxies are least reliable. The threat is real, though. At very small $\epsilon$ or with very few sensors, the noise may swamp the correction. The proxy-only baseline shows where that crossover sits, and reporting the crossover is part of the result.

Why not run the whole forest inside an SGX enclave, or under Pivot-style cryptography? Secure IoT Data Analytics in Cloud via Intel SGX runs simple rule matching under a 128 MB page-cache limit and documents side-channel leakage from memory access patterns. Multi-band rasters overflow that limit, and enclaves give no protection for released outputs. Pivot is about 37.5× faster than pure MPC, but it still relies on Paillier encryption plus SPDZ, sized for organizations, not sensors. Our protocol needs only additions from devices. We will report per-device bytes and CPU time on a Raspberry Pi-class device so the gap is measured rather than asserted.

Proposed Experiments

The first testbed is SMAP downscaling from 36 km to 1 km. Its 120 CMA stations are treated as private devices, and we report station-held-out R² and unbiased RMSE. The second is the OpenAQ PM2.5 fusion task, with stations again treated as private and results reported as median absolute error and R², broken down by data-rich and data-poor regions. For both, we sweep $\epsilon \in \{0.1, 0.25, 0.5, 1, 2, \infty\}$, tree depth, and the number of trees $B$. We validate with spatially blocked cross-validation using grid samplers, because geographic autocorrelation inflates scores under random splits.

The decisive check is a curve of accuracy against privacy budget with all methods on the same axes. The proposal succeeds if public-proxy structure dominates DP-TR and MRF below $\epsilon = 1$. It also needs the unsupervised-structure ablation to sit between the two, which would show that proxy supervision, not just public features, does the work.

We will also test robustness to a weak proxy. To do this we degrade the coarse product's correlation with the target, by adding noise or shifting it in time, and find the correlation below which the advantage disappears. That threshold tells practitioners when the method applies.

Finally, the systems measurements cover per-device payload and computation, behavior when 30% of devices drop out mid-round, and coarse-tile size against location anonymity. They test whether the protocol is practical on real sensor hardware, not just correct on paper.