Do Precipitation Correction Networks Lose the Mountains in Their Bottleneck?

Satellite rainfall products such as GPM IMERG, NASA's gridded multi-satellite precipitation estimate, are least accurate where rain gauges are scarcest: high, steep mountains. Machine-learning correction does not close this gap. Random Forest correction of IMERG over the Tianshan performs well on average but drops to the level of plain stepwise regression above 3000 m. The usual explanation is too few training samples. This note proposes a sharper, testable explanation. A deep correction network trained on gauges that cluster in the lowlands learns an internal feature space with few effective dimensions, shaped by lowland weather. When a high-altitude scene arrives, the features that make it distinctive fall outside that space and are discarded. What remains is the lowland "average" answer.

We propose to measure this loss directly. For each gauge, we compute how much of its input's internal representation lies outside the subspace spanned by the training data, and test whether that quantity predicts correction error better than elevation or sample count. If it does, the remedy is a training penalty on the network's narrow internal layer (its "bottleneck"). The penalty would stop that layer from encoding terrain, but only after conditioning on the true rainfall, so that real orographic rain is kept and only terrain-driven retrieval error is removed. This extends bottleneck debiasing, which so far needs discrete bias labels, to continuous terrain variables. It also makes the high-altitude failure a measurable property of the model rather than an anecdote.

Core Claim

The claim has two parts, and the second only matters if the first holds. The diagnostic part says that high-altitude degradation in deep satellite-precipitation correction is, to a measurable degree, dimensionality-induced information loss. Dimensionality-Induced Information Loss of Outliers in Deep Neural Networks showed that image classifiers have a transition layer where representations collapse to a low-rank subspace. At that layer, out-of-distribution inputs lose their distinguishing features and keep only the training set's bias. Projecting out the unaligned directions with a singular value decomposition raised outlier-detection AUROC to 0.978 on CIFAR-10 versus MNIST. That paper motivates our diagnostic, and we reuse its stable-rank and subspace-projection tools. It never looked at regression, at continuous covariate shift, or at geography.

The intervention part says that a bottleneck regularizer can shrink the elevation dependence of error without flattening orographic signal. It enforces that bottleneck features carry no terrain information once the gauge-observed rainfall is known.

Conditioning on rainfall is the design choice that separates this proposal from the obvious one. The obvious idea is to "disentangle terrain from precipitation" in the style of Disentangled Information Bottleneck. That work splits an input into a task-relevant code and a nuisance code with an adversarial independence penalty, on MNIST- and CIFAR-scale classification. Applied here without conditioning, it would be wrong on physical grounds: elevation genuinely causes rain through windward enhancement and rain shadows. Full invariance to elevation would delete exactly the signal the Lhasa and Nepal evaluations show IMERG already misses.

The nuisance is not elevation itself. It is the part of the satellite signal whose relation to true rain depends on elevation. EnD: Entangling and Disentangling Deep Representations for Bias Correction already encodes this conditional logic for discrete labels. It pulls together representations with the same target label from different bias groups, and pushes apart representations that share a bias group. This raised unbiased accuracy on Biased MNIST from 10.4% to 52.3%. EnD requires categorical bias labels and names continuous bias as its main limitation. We supply the continuous, regression version.

Research Program

The model is a compact convolutional network. Its input is a patch around each gauge pixel containing the IMERG Uncal, HQ (microwave-only) and IR (infrared-only) channels, ASTER GDEM derivatives (elevation, slope, aspect, curvature, wetness index, hillshade), and MODIS NDVI. It outputs a corrected rainfall value at the gauge. An explicit narrow layer of $d$ units sits between encoder and prediction head.

Diagnostic. We first train this network with ordinary loss on lowland-dominated data. We then compute three things at every layer: stable rank, the subspace spanned by the top singular directions of training-set features, and a per-sample "residual energy". Residual energy is the fraction of a sample's feature norm lying outside that subspace. We then ask whether residual energy predicts each gauge's absolute error after controlling for elevation, local training-sample density, and rainfall intensity.

Intervention. Let $h$ be the bottleneck features, $z$ the vector of terrain covariates, and $y$ the gauge rainfall. The intervention adds a kernel dependence penalty,

$$\mathcal{L} = \ell(\hat y, y) + \lambda \sum_{b} \widehat{\mathrm{HSIC}}\big(h, z \mid y \in b\big),$$

where $\ell$ is the regression loss, $\widehat{\mathrm{HSIC}}$ is a kernel measure of statistical dependence (zero when independent) computed within each minibatch, $b$ indexes quantile bins of observed rainfall, and $\lambda$ sets the penalty strength.

Binning on $y$ plays the role of EnD's "same target label". Using a kernel on continuous $z$ replaces EnD's discrete bias groups: samples are compared by terrain similarity instead of by group membership. The network still receives $z$ as input, so it can use terrain to correct the retrieval. The penalty only forbids terrain from leaving a trace in the features beyond what the rain itself implies. We then repeat the diagnostic on the regularized model. The prediction is that high-altitude residual energy falls, and that the fall tracks the error reduction.

Datasets. The Tianshan network of about 1000 stations, used by Quantitative Precipitation Estimation in the Tianshan Mountains Based on Machine Learning, gives monthly warm-season gauges spanning low plains to above 3000 m. It is the primary testbed for elevation-band splits. The 67 high-mountain gauges of Evaluation of GPM IMERG V05B and TRMM 3B42V7 Precipitation Products over High Mountainous Tributaries in Lhasa with Dense Rain Gauges supply 6-hourly and daily targets for testing extremes. So do the 25 Nepal stations of Evaluation of the Spatiotemporal Distribution of Precipitation Using 28 Precipitation Indices and 4 IMERG Datasets over Nepal.

Risks. Two data risks are real. First, access to the regional Chinese networks may require collaboration with their authors. Second, the Tianshan data are monthly, which limits the diagnostic there to seasonal-scale representations. We treat Lhasa and Nepal as the daily-scale check, not as optional extras.

Baselines

Do-nothing reference. Uncorrected IMERG (the gauge-calibrated Final Run) evaluated at the gauges. In Lhasa, Evaluation of GPM IMERG V05B and TRMM 3B42V7 Precipitation Products over High Mountainous Tributaries in Lhasa with Dense Rain Gauges reports a daily mean correlation coefficient (CC) of 0.50 and a daily relative bias of 10% for IMERG. That work also documents that IMERG overestimates trace rain and underestimates rain above 50 mm per day. Any correction must beat this at daily scale above 3000 m.

Strongest existing method. Random Forest on DEM and NDVI covariates, from Quantitative Precipitation Estimation in the Tianshan Mountains Based on Machine Learning. It reports CC 0.81, Nash–Sutcliffe efficiency (NSE) 0.66, RMSE 19.81 mm and relative bias 1.06% over the Tianshan. Geographically weighted regression reached CC 0.79 and NSE 0.62. That paper gives no separate number for the above-3000 m band; there is no reported result on Lacuna for RF skill restricted to high elevations, so we will compute it ourselves. We will rerun RF and geographically weighted regression on exactly our splits and covariates, tuned by the same cross-validation budget as the network. Beating RF means a statistically significant gain in NSE and RMSE in the held-out high-elevation band, with no loss on lowland gauges. Matching RF's all-elevation CC of 0.81 is a sanity check, not a win.

Unconstrained deep network. The same architecture and inputs with no penalty and no narrow layer. This follows the "concatenate terrain into a CNN" design shared by STAS and IC-MLNet. STAS: Adaptive Selecting Spatio-Temporal Deep Features for Improving Bias Correction on Precipitation reports a heavy-rain Threat Score of 0.38 on its Eastern China benchmark. Multi-Layer Networks for Ensemble Precipitation Forecasts Postprocessing reports a 17.8% MAE improvement over its ensemble baseline in Southern China. Both are forecast post-processing in flat or coastal regions, so we use them as architectural references rather than numbers to beat. IC-MLNet's intensity stratification will be included as an optional head, because the question of whether our penalty hurts extremes should be tested against a head designed for extremes.

Ablations. Four ablations isolate the key ingredient. The first uses a narrow bottleneck only, testing whether capacity alone helps, as the VAE-DG result would suggest. The second uses unconditional terrain invariance, which should hurt if the orographic argument is right. The third uses EnD with elevation discretized into bands, testing whether continuity matters. The fourth reweights samples by inverse gauge density. All models share data splits, input channels, epochs, and a fixed hyperparameter-search budget.

Why not just…?

Why not just use Random Forest with DEM covariates, which already works? It is the strongest competitor, and on lowland and all-elevation averages we expect it to be hard to beat. Its documented failure is specifically above 3000 m, where Quantitative Precipitation Estimation in the Tianshan Mountains Based on Machine Learning finds it no better than stepwise regression. RF also cannot use the spatial pattern of the surrounding satellite field. If RF matches our model in the held-out high band, the intervention has failed. The diagnostic would still tell us whether deep models fail for the same reason.

Why not just oversample or reweight high-altitude gauges? This is the most serious threat. If the failure is sample starvation, reweighting should fix it. The two explanations make different predictions, and the diagnostic separates them. If residual energy predicts error after controlling for local sample density, collapse is doing work that reweighting alone does not address. We run reweighting as a full baseline, including in combination with the penalty. If reweighting alone closes the gap, we will report that the bottleneck story is unnecessary.

Why not just bin elevation and apply EnD directly? This is cheap, and it may be enough. Binning, however, forces arbitrary thresholds on a variable that interacts with slope and aspect. Samples near a band edge would be treated as maximally different despite being physically similar. The discretized-EnD ablation tests this directly. The continuous kernel penalty must beat it in the high-elevation band to justify itself.

Why not just aggregate to monthly scale? The Lhasa study shows IMERG's CC rising from 0.50 daily to 0.92 monthly. The hazards that matter in mountains, flash floods and debris flows, live at daily and sub-daily scale, and aggregation does not remove the heavy-rain underestimation documented in Nepal. We therefore report daily-scale results as primary wherever the gauges allow.

Proposed Experiments

The central split is leave-elevation-band-out on the Tianshan. Models train on gauges below 2500 m and are evaluated separately on the 2500–3000 m band and the above-3000 m band. We also run spatial block cross-validation, so that neighboring gauges never straddle the train–test split. Metrics are CC, NSE, RMSE and relative bias per band. We also report an "elevation slope of error", the regression of absolute error on elevation. It summarizes regional variation in one number, and we want it driven toward zero.

For extremes, we train on Nepal and Lhasa daily data. We report probability of detection, false alarm ratio and equitable threat score at 10 and 50 mm per day. This directly tests the open worry that compression erases localized convective peaks.

The decisive checks are three.

Check 1: does collapse explain the mountain error? In the unregularized network, the partial correlation between per-gauge residual energy and absolute error, controlling for elevation, sample density and intensity, should be positive and robust across random seeds. If it is near zero, the collapse hypothesis is refuted and the intervention loses its motivation.

Check 2: does the conditional penalty help without hurting orographic rain? It should reduce high-band RMSE and the elevation slope of error relative to the unconstrained network, reweighting, and discretized EnD. It should do so without lowering heavy-rain probability of detection. The unconditional-invariance ablation should do visibly worse on windward gauges, confirming that conditioning on rainfall protects real signal.

Check 3: is the gain mediated by the representation? Across the $\lambda$ sweep, reductions in high-altitude residual energy should track reductions in high-altitude error. If error improves while residual energy does not move, the penalty is working through some other route, and we will say so.