Detecting Floods in Ungauged Basins by Coincidence Across Weak Sensors

Most machine-learning flood work treats the problem as regression: predict tomorrow's river discharge, score the prediction by squared error averaged over every day. In the basins that most need warnings there is often no river gauge at all. Those basins have only coarse satellite rainfall, perhaps one low-cost water-level probe, and long stretches of quiet baseflow broken by a few dangerous days. That setting resembles a rare-event search in particle physics more than it resembles regression. The question that matters is how many true floods you catch at a false-alarm rate a civil-protection office can tolerate. In this setting the "background" includes storms that never produce runoff, stuck or spiking sensors, and forged telemetry. Physicists facing this setup do not rely on one strong detector. They combine several weak, independent channels and ask whether these channels fire together in time and space more often than chance would allow.

We propose to build and test exactly that for floods. Each available channel produces a calibrated evidence score: a learned streamflow forecast transferred from gauged donor basins, a gridded rainfall field read by a small spatio-temporal network, and, where it exists, a single local sensor. A coincidence posterior then fuses these scores. It rewards agreement within a physically plausible rainfall-to-runoff lag and degrades gracefully when channels are missing. The claim we will test is narrow. At a fixed false-alarm rate, this posterior detects more flood events in held-out, zero-gauge or one-gauge basins than the best single channel thresholded on its own. The gain should be largest for "subthreshold" events that no single channel flags confidently.

Core Claim

The multimessenger method of Enhancing the significance of astrophysical events with multimessenger coincidences adds evidence from independent instruments. It weights each instrument by how well its timing and sky positions coincide, and in doing so raised a weak gravitational-wave candidate from a 7.7% to an 86% probability of being astrophysical. Its success there motivates this proposal. Its known weakness also constrains our design: the result depends strongly on the assumed emission model. In hydrology the analogue of the emission model is the distribution of delays between rain and river response. We therefore do not assume that distribution. We learn it from gauged donor basins as a function of static landscape attributes, and we report how sensitive detection is to errors in it.

Written out, for basin $b$ and day $t$, the log-odds that a flood is underway is

$$\log \frac{P(F_{b,t}\mid x)}{1-P(F_{b,t}\mid x)} = \log\frac{\pi_b}{1-\pi_b} + \sum_{c \in A_{b,t}} \max_{\tau} \Big[\log \Lambda_c(x_{c,t-\tau}) + \log k_b(\tau)\Big].$$

The symbols are defined as follows. $\pi_b$ is the basin's prior flood frequency, predicted from static attributes. $A_{b,t}$ is the set of channels actually reporting authenticated data at time $t$. $\Lambda_c$ is channel $c$'s calibrated likelihood ratio of "flood" versus "background." $k_b(\tau)$ is the learned lag kernel, which is the probability that channel $c$ leads the flood by $\tau$ days. A missing channel simply drops out of the sum rather than being imputed. This choice follows the finding in Short-Term Traffic State Prediction Based on the Spatiotemporal Features of Critical Road Sections that imputing structurally missing sensors propagates errors into downstream predictors. The sum assumes the channels are conditionally independent. That assumption is false for the two rainfall products, so testing it is a central experiment rather than a footnote.

Research Program

The work has three components.

The first is a rare-event benchmark built from existing large-sample hydrology. We use Caravan - A global community dataset for large-sample hydrology, which covers 6,830 catchments with ERA5-Land forcings, HydroATLAS attributes and daily streamflow, extracted through the kratzert/Caravan Google Earth Engine tooling. A flood event is a contiguous run of days above a per-basin high-flow quantile, and the detection unit is the event, not the day. Background comes in three kinds. The first is natural: heavy rain with no runoff response. The second is injected gauge artifacts, namely spikes, stuck values and slow drift added to the local-sensor channel at rates we control. The third is forged packets reporting rising water. We hold basins out by region. That way we can measure transfer into regions underrepresented in Caravan's North-America-and-Europe-heavy sample, an underrepresentation the Caravan authors document.

The second component is per-channel evidence models. The forecast channel is an Entity-Aware LSTM (a recurrent network conditioned on static basin attributes). It is trained on donor basins and applied to target basins with their streamflow hidden, following the training regime of The proper care and feeding of CAMELS: How limited training data affects streamflow prediction. That paper found that training on many diverse basins beats training on long records from one basin, which is why transfer into ungauged basins is plausible at all. The rainfall channel replaces the lumped basin-average forcing used in that work with a gridded patch over the catchment from ERA5-Land and GPM IMERG. We read the patch with a CNN-then-LSTM cascade, in which a convolutional network summarizes each day's rainfall map into a vector and a recurrent network tracks those vectors over time. The architectural pattern is the one used across the spatio-temporal literature, for example in Electrical Energy Prediction in Residential Buildings for Short-Term Horizons Using Hybrid Deep Learning Strategy. The reason to keep the map is physical: storms concentrated on steep headwaters behave differently from the same rainfall total spread across a floodplain. Each channel's raw score is turned into a likelihood ratio by calibrating it on donor basins. We borrow from PandaX-III (Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment) the practice of reporting signal efficiency against background rejection, rather than a single error number. That paper's seven-orders-of-magnitude rejection at above 20% signal efficiency came from several shape features rather than from energy alone.

The third component is a telemetry constraint. The local-sensor channel is only useful if its packets are trustworthy and arrive within the warning window. Integrated signcryption produces a single combined signature-and-encryption operation, which saves bytes and compute relative to signing and then encrypting separately. An Efficient Deniable Authenticated Encryption Scheme for Privacy Protection reports up to 30% cheaper encryption and 50% cheaper decryption, and Signcryption in Hierarchical Identity Based Cryptosystem reports saving at least one group element per message. Prior evaluations ran on desktops, however. That includes PQ-CLSC in Efficient and Secure Post-Quantum Certificateless Signcryption for Internet of Medical Things, measured at about 31.634 ms on an AMD Ryzen 7 machine. It also includes GCD-RSD in Granular Content Distribution for IoT Remote Sensing Data Supporting Privacy Preservation, which offloaded unsigncryption to desktop computers. We treat security as a measured budget rather than a contribution. Packets that fail verification leave the active channel set $A_{b,t}$. We then check that the detector does not mistake that dropout for evidence of a flood.

Baselines

The naive reference point is per-basin climatology. It raises an alarm whenever ERA5-Land basin-average precipitation exceeds a high quantile of that basin's own history. It needs no learning and represents what a district with no modeling capacity could run. Every method is scored by event recall at false-alarm rates of one and four per basin-year, on the same held-out basins.

The strongest existing competitor is a thresholded transferred streamflow forecast. We train an EA-LSTM on donor basins, predict discharge in the ungauged target, and alarm when predicted discharge crosses the basin's flood quantile. This is what an operational group would build today. On regression it is well characterized: The proper care and feeding of CAMELS: How limited training data affects streamflow prediction reports a median Nash-Sutcliffe efficiency of 0.74 on 531 CAMELS basins with 9 years of data, against about 0.66 for XGBoost and the calibrated mHM model. We will reproduce that number on the CAMELS subset of Caravan before running anything else, to confirm our pipeline is sound. For event detection at fixed false-alarm rate, there is no reported result found on Lacuna for EA-LSTM or any other model. The baseline therefore gets the same donor basins, the same inputs and the same hyperparameter search budget as our channel models, and its alarm threshold is tuned on donor basins exactly as ours is. Beating it means a recall gain at one false alarm per basin-year whose bootstrap confidence interval over held-out basins excludes zero. We expect the gain to concentrate in events where the forecast alone peaks below threshold.

The physics-guided transfer method of Transfer Learning Using Inaccurate Physics Rule for Streamflow Prediction is included as a second forecast-channel variant. It corrects an imperfect mass-balance rule with a residual model predicted from basin attributes, and reports test RMSE falling from 11.871 to 2.116 with 60 training basins. If its forecasts are sharper, it replaces the EA-LSTM as the baseline to beat.

For the dense-station sub-study we use APS-LSTM (APS-LSTM: Exploiting Multi-Periodicity and Diverse Spatial Dependencies for Flood Forecasting). It finds dominant periods in the series by Fourier transform and applies attention across both periods and stations. On TunXi it reports a 4.1% RMSE and 9.8% MAE reduction over the second-best model. The file gives only these relative figures, and for ChangHua there is no reported number found on Lacuna.

Two ablations isolate the key ingredient. The first is a learned joint classifier: gradient-boosted trees given every channel's features concatenated, with missing values marked. It tests whether the factorized coincidence structure matters or only the extra inputs do. The second removes the lag kernel by setting $k_b(\tau)$ uniform. It tests whether temporal coincidence, rather than mere score addition, carries the gain.

Why not just…?

Why not just threshold the best streamflow forecast? This is the real threat. If the transferred LSTM already captures most of the information in rainfall, the coincidence posterior adds little, because its rainfall channel is partly redundant with the forecast's own inputs. The main experiment settles this directly. If recall at fixed false-alarm rate does not improve over the thresholded forecast, we report the negative result. The remaining contribution is then the event-level benchmark and the measured robustness to sensor glitches, which a thresholded forecast that consumes a local gauge does not have by construction. We specifically expect gains only where an independent channel exists: gridded satellite rainfall the forecast did not see, or a local probe. The evaluation is stratified by which channels are present so that the gain, or its absence, is localized.

Why not just train one end-to-end classifier on all inputs? With enough labeled floods this would probably win, since it can learn correlations between channels that the independence assumption ignores. But floods are rare per basin. The joint model must also learn what missingness patterns mean, whereas the factorized posterior simply drops absent channels. The gradient-boosted joint classifier ablation tests this. Gauch et al. found tree models competitive on small data, so it is a fair stand-in. If it matches the posterior in full-channel basins but falls behind as channels are removed, that confirms the structural argument. If it matches everywhere, the posterior's remaining value is interpretability and calibration, and we would say so.

Why not just sign then encrypt with standard primitives? For the detection question this probably works just as well. We do not claim otherwise, which is why signcryption is a measured constraint rather than the contribution. On a microcontroller-class field node we measure per-packet bytes, time and energy for an elliptic-curve signcryption scheme against Ed25519 signing followed by AES-GCM encryption. We then convert both into the warning lead time lost at a fixed battery-limited reporting cadence. The prior savings figures for signcryption, such as BASUV's 632 bytes against 860 to 1,332 bytes for competing protocols in BASUV: A Blockchain-Enabled UAV Authentication Scheme for Internet of Vehicles, suggest the difference may matter only on very narrow links. If the standard composition fits the budget, we recommend it.

Proposed Experiments

The decisive experiment is the regional hold-out on Caravan. We compare event recall at one and four false alarms per basin-year for five methods: the climatology alarm, the thresholded EA-LSTM, the physics-corrected forecast, the joint classifier, and the coincidence posterior. We report results separately for target basins with zero local sensors and with one simulated local sensor, the latter built from the basin's own gauge degraded by injected artifacts. We report recall separately for events whose best single channel scores below its own alarm threshold, since that is where coincidence should matter.

Two further checks probe the method's assumptions. The first is calibration of the posterior on held-out basins. In particular, we test whether the ERA5-Land and IMERG rainfall channels double-count evidence. If they do, we merge them into one channel and re-run the comparison. The second is lag-kernel sensitivity. We perturb $k_b(\tau)$ by the error we observe when predicting it from attributes, mirroring the emission-model sensitivity Veske et al. report.

On TunXi and ChangHua, where data are hourly and stations are dense, we progressively remove rainfall stations down to one. We measure whether APS-LSTM's spatial attention degrades faster than the coincidence approach. The station-dependence of that attention, noted in the file's gap analysis, predicts it will.

Finally, we inject forged rising-water packets and stuck-sensor episodes at controlled rates. We report false alarms with and without authenticated telemetry, along with the measured lead-time cost of each security scheme on field hardware.