Source-linked AI summary
Supraglacial Lake Fate Is Knowable Long Before the Season Ends
Emam Hossain, Md Osman Gani
TL;DR
The paper asks when supraglacial-lake outcomes become reliably identifiable rather than merely retrospectively labeled. It truncates inputs and retrains classifiers at successive cutoffs to measure each outcome’s earliest knowability date. Rapid and slow drainage resolve by 15 July and 1 August, while the ordering remains stable across learners and basin-held-out evaluation.
Problem
Retrospective labels identify what a lake did but not when observations become sufficient to distinguish its outcome.
Method
The study truncates inputs at 14 cutoffs, retrains the classifier at each cutoff, and measures the earliest per-class F1 target date using leakage-free features.
Results
Rapid drainage reaches the 80% per-class F1 target on 15 July and slow drainage on 1 August, leading the 15 October reference by 92 and 75 days.
Takeaways & Limitations
Monitoring should use outcome-specific release dates: rapid drainage can be flagged on 15 July rather than waiting for a uniform end-of-season release.
Takeaways & Limitations
The reported dates are specific to Greenland’s 2018 and 2019 melt seasons, and two years cannot establish shifts under warmer or cooler seasons.
Abstract
from arXiv · showhide
A supraglacial lake on the Greenland Ice Sheet ends its melt season in one of four ways: it drains rapidly through a hydrofracture, drains slowly across the surface, refreezes in place, or is buried by late-season snowfall. Which one occurs decides whether the meltwater reaches the ice bed. Satellite classifiers recover the outcome accurately but only after the season closes, and how much of a season each outcome actually requires has never been measured. We measure it directly: holding the representation and the classifier fixed, we truncate the input at $14$ cutoffs from 1 May to 31 December, retrain at each, and record the earliest cutoff at which each outcome's per-class $F_1$ reaches a fixed target. The outcomes resolve in a consistent order, two of them months early: rapid drainage by 15 July and slow drainage by 1 August, $92$ and $75$ days ahead of the earliest date a full-season pipeline can be computed at all, with buried and refreeze following at $44$ and $30$ days. Five further learners, from a majority-class floor and $54$ summary statistics to a trigger-based early classifier, leave the ordering intact: every learner that produces a per-class trajectory reproduces it despite end-of-season accuracies differing by up to $18$ percentage points, and it survives leave-one-basin-out evaluation, though not the substitution of machine labels for expert ones in an unseen season. Every feature we compute at day $t$ reads only days up to $t$, at a cost of at most $1.3$ percentage points. A monitoring system should therefore not have one release date: rapid drainage can be flagged on 15 July, three months before a full-season pipeline can be computed at all.
1 Introduction
The paper asks when satellite observations become sufficient to identify each supraglacial lake outcome, rather than merely labeling it retrospectively after the melt season. Using fixed representation and classifier with truncated, retrained inputs, it finds ordered outcome-specific knowability and supports different release schedules.
- Rapid drainage can deliver meltwater to the ice bed within hours, whereas slow drainage routes it across the surface without the same dynamic signature.
- Existing satellite systems classify lake outcomes from whole-season features and emit labels only after the season ends.
- The study truncates inputs at 14 cutoffs from 1 May to 31 December, retraining the fixed representation and classifier at each cutoff.
- The four outcomes reach a fixed per-class target in a consistent order, with drainage outcomes resolving before the storage outcomes.
2 Background and Related Work
Supraglacial lake outcomes differ in physical meaning and operational value, while prior classifiers generally consume the full season. The paper frames class-specific early knowability as distinct from applying one earliness policy to every outcome and requires leakage-free temporal evaluation.
- 2.1 Supraglacial lake evolution: The four classes are rapid drainage, slow drainage, refreeze, and burial by late-season snowfall, with rapid drainage most directly coupled to ice dynamics.
- 2.1 Supraglacial lake evolution: Because refreeze and burial are partly defined by drainage not occurring, they are physically harder to establish early than event-defined drainage outcomes.
- 2.2 Classifying outcomes from satellite data: Prior systems use optical and radar observations to classify lake evolution, but they consume the complete season and do not report accuracy under truncated input.
- 2.3 Early classification of time series: Early time-series classification methods apply one stopping rule or cost trade-off across instances, whereas this study measures earliness separately by outcome class.
- 2.4 Leakage in temporal pipelines: Temporal interpolation, centered smoothing, and random spatial cross-validation can invalidate early claims, so the study restricts features to days ≤t and evaluates leave-one-basin-out splits.
3 Problem Setup
The problem setup represents each lake as a multivariate daily series, truncates it at specified calendar cutoffs, and trains a separate classifier for each cutoff. Knowability is defined by the earliest cutoff reaching a per-class F1 target, relative to the date when a full-season pipeline becomes computable.
- 3.1 Data and labels: The record contains 2018 and 2019 Greenland lake observations, six drainage-basin assignments, and expert labels for 1,000 balanced 2019 lakes.
- 3.1 Data and labels: Each lake is represented as a C-channel daily series on a 365-day grid, with nine input channels used by every reported model.
- 3.2 Prediction at a truncated cutoff: The truncation operator retains the first T days, and each cutoff model is trained and tested on inputs with that same truncated length.
- 3.2 Prediction at a truncated cutoff: The cutoff grid runs from 1 May through 15 October, with additional 31 October and day-365 reference cutoffs.
- 3.3 Knowability and lead time: Retraining at each cutoff avoids conflating missing evidence with the distribution mismatch caused by testing a full-season model on shortened inputs.
- 3.3 Knowability and lead time: Knowability is the earliest cutoff at which a class reaches the target per-class F1, reported throughout at τ = 0.80.
- 3.3 Knowability and lead time: Full-season features requiring October or complete-record observations become computable on 15 October, establishing the day-288 reference date for lead time.
4 Method
The method isolates how much season data is needed by retraining a fixed classifier at successive cutoffs while enforcing causal preprocessing and testing learner and spatial-split robustness.
- Leakage-free preprocessing: Every feature at day t uses only observations from days ≤t through forward filling, trailing smoothing, trailing outlier rejection, and expanding-window standardization.These substitutions prevent future observations from entering online features, though trailing smoothing introduces a conservative phase lag.
- Learner robustness: Five alternative learners isolate whether the ordering depends on majority-class prediction, summary statistics, catch22, ROCKET, or a trigger-based early classifier.All alternatives use identical folds, seeds, cutoffs, labels, and preprocessed inputs, so only the learner changes.
- Evaluation protocol: Evaluation combines random 5-fold cross-validation with leave-one-basin-out testing, the latter holding out each of six basins to assess transfer to new regions.The protocol is fixed before results are inspected, and reported uncertainty is summarized across five seeds or bootstrap resamples for the released full-season model.
- Evaluation protocol: Knowability dates use per-class F1, while aggregate comparisons use macro-recall to weight the four outcomes equally under class imbalance.F1 is chosen because recall alone can be achieved by over-predicting a class.
5 Results
The four lake outcomes become knowable in a consistent order, with drainage outcomes resolving months before storage outcomes. This ordering generally survives learner and basin changes, although transfer to an unseen season is limited by label and season shifts.
- 5.1 Outcomes become knowable in order: 15 July, 1 August, 1 September, and 15 September are the 80% knowability dates for rapid drainage, slow drainage, buried, and refreeze, respectively.These correspond to leads of 92, 75, 44, and 30 days relative to the 15 October full-season reference.
- 5.1 Outcomes become knowable in order: Rapid drainage resolves first because its sharp hydrofracture signature is dated and needs no later observations, whereas non-drainage evidence accumulates as the season closes.Slow drainage unfolds over weeks, while buried and refreeze are defined by an event not occurring; freeze-up confirms refreeze last.
- 5.1 Outcomes become knowable in order: 68.2 ± 0.8% F1 is achieved by buried lakes on 1 May, while macro-recall gains only 1.5 points after day 288 under random folds.Buried lakes have a distinctive location and radar signature early, and the post-season record contributes little to retrospective labeling.
- 5.2 The ordering is model-independent: Seven of eight trajectory-producing learner-by-split cells preserve the ordering, with rapid drainage always earliest and refreeze latest or unreached.The 54-feature summary-statistic baseline recovers all four random-fold dates, while the majority-class floor reaches no class target.
- 5.2 The ordering is model-independent: MiniROCKET and ROCKET finish within 0.8 percentage points under both split schemes, but MiniROCKET leads early by 7.0 points under random folds and 8.6 under basin folds at 1 May.ROCKET’s sampled kernel lengths and dilations can exceed truncated series lengths, whereas MiniROCKET adapts its dilations to the available input.
- 5.3 Spatial and temporal transfer: 15 July and 1 August drainage dates remain unchanged under leave-one-basin-out, while buried moves two weeks later and spatial transfer reduces early macro-recall.At 15 July, held-out basin accuracy spans 17.4 percentage points, from 73.5% in Southwest Greenland to 56.1% in the Northwest.
- 5.3 Spatial and temporal transfer: Cross-year dates do not survive machine-label substitution: under 2019 machine labels drainage classes swap and refreeze never reaches target, while only buried reaches target on 2018.The label cost averages 10.2% macro-recall across cutoffs and the season cost adds 8.4%, yielding 47.4% at 15 July and 78.0% at year end.
6 Ablations
Ablations show that leakage-free preprocessing preserves the measured knowability dates while costing at most 1.3 percentage points, and atmospheric reanalysis adds no predictive value. The ordering also remains stable across target accuracies, while the dominant error pair shifts over the season.
- 6 Ablations: The leading error pair changes at the marked cutoff: drainage dominates earlier, storage later, and storage still carries two thirds of error on 15 September.The figure pools all remaining class pairs as a dotted line.
- 6.1 Cost of removing leakage: No knowability date changes between leakage-free and conventional preprocessing chains, while the conventional chain gains at most 1.3 percentage points.The leakage-free chain is ahead by 3.4 points in mid-May and differs by a mean absolute 0.9 points across cutoffs under random folds.
- 6.2 Atmospheric reanalysis channels: Adding five CARRA atmospheric fields never improves performance or knowability dates.Under basin folds, the mean macro-recall change is −1.1 points, with rapid drainage and refreeze dates moving later.
- 6.3 Sensitivity to the target: The four-class ordering survives every tested target, from τ = 0.75 to τ = 0.90.At τ = 0.75, both drainage classes retain 92 days of lead and refreeze retains 44; at τ = 0.90, rapid drainage retains 75 days and refreeze none.
7 Limitations
The reported dates are bounded by the two Greenland melt seasons and by the study’s label, sampling, resolution, and decision-rule choices. The transfer result is especially weakened by machine-generated labels rather than expert annotations.
- Seasonal and geographic scope: Results come from Greenland’s 2018 and 2019 melt seasons, so warmer or cooler seasons may shift the dates.Two years cannot establish how dates vary across seasons or under climate warming.
- Label provenance: Scoring 8,992 transfer lakes against machine-generated labels costs 10.2% macro-recall, making the cross-year conclusion weaker.The core measurement rests on 1,000 expert-labeled lakes; machine labels are used only for transfer experiments.
- Sampling and metrics: The balanced reference set contains 250 lakes per class and therefore does not represent deployment accuracies on the unbalanced ice-sheet population.Per-class F1 ordering and dates are unaffected, but precision would differ.
- Decision-rule dependence: Knowability dates are properties of a specified classifier family and accuracy target, not fixed constants.A stronger model or different target can move the dates, although the ordering survives the reported alternatives.
- Temporal resolution: Fourteen cutoffs resolve knowability dates to roughly two weeks; daily resolution would require 26 times the compute.The reported resolution is sufficient to establish the ordering but not exact daily dates.
8 Conclusion
The paper separates retrospective labeling from the time required for observations to support each outcome. It finds a stable, outcome-specific ordering and argues for release schedules that follow those dates rather than one end-of-season deadline.
- 8 Conclusion: Rapid drainage reaches the target by 15 July and slow drainage by 1 August, 92 and 75 days before a full-season pipeline can be computed.The ordering remains across alternative learners, basin-held-out evaluation, and tested targets.
- 8 Conclusion: A uniform end-of-season release discards roughly three months of lead for rapid drainage, the outcome most relevant to ice dynamics.Refreeze does not reach the target before mid-September.
- 8 Conclusion: More expert-labeled seasons are needed to test date stability, climate drift, and basin variation, with a 17.4-point basin spread worth testing first.The paper also identifies per-outcome trigger functions as the natural instrument for operational deployment.
A Reproducibility
The study provides access to its inputs, code, preprocessing chains, tensors, and complete result records. These deposits support reproduction of the splits, seeds, fits, and reported per-class metrics.
- Inputs: The satellite record, expert labels, released full-season models, and CARRA reanalysis are identified as externally deposited inputs fetched by DOI.Neither source is re-hosted in the study.
- Code: The preprocessing chains, truncated sweep, baseline arms, and figures are released in the project repository.The repository URL is provided in the paper.
- Data and results: The deposited tensors include both preprocessing chains and seasons, fold and basin assignments, expert and machine labels, and each lake’s drainage date.This allows exact splits and seeds to be reused without regeneration.
- Data and results: The results archive records all 3,080 main-sweep and baseline fits, including per-fold, per-seed confusion matrices and per-class precision, recall, and F1.The records cover both preprocessing chains and variable sets.
- Data and results: The deposited tensors contain modified Copernicus Climate Change Service information, for which the European Commission and ECMWF disclaim responsibility.This is a data-use disclaimer rather than a methodological limitation.
B Dataset details
The study uses expert- and machine-labeled Greenland lake records, with basin imbalance and differing machine-label populations shaping evaluation. Cutoffs span the melt season from 1 May through year-end under specified preprocessing assumptions.
- Dataset composition: 1,000 expert-labeled lakes are evenly divided across four classes, but their basin composition is uneven across leave-one-basin-out folds.Southwest Greenland contributes 245 lakes, including 89 rapid-drainage lakes, while Southeast Greenland contributes 92, including 10.
- Cutoff design: 9, -
C Implementation and compute
The implementation uses established time-series learners and records computational settings, fit counts, preprocessing comparisons, and transfer-label coverage.
- MiniROCKET, ROCKET, catch22, and TEASER use sktime implementations, while ridge classification and bootstrap procedures use scikit-learn.
- 3,080 fits form the main sweep, with 770 additional ROCKET fits, 154 catch22 fits, and 154 summary-statistic fits.TEASER contributes 55 runs and 10,000 decisions; the majority-class floor fits nothing.
- TEASER uses 2,000 kernels instead of the full kernel count because matched settings would multiply its cost by roughly five without changing the comparison.
- The conventional preprocessing chain is compared with the leakage-free chain at every cutoff, and no class’s knowability date differs between them.
- Transfer-target tables exclude the 1,000 expert-labeled lakes, and the 2018 season is machine-labeled throughout because no expert labels exist.
D Additional results
Additional analyses examine basin variation, channel variance structure, and early performance of the full-season reference pipeline. They show substantial spatial differences and limited early discrimination relative to chance.
- Table 7 repeats the per-class F1 analysis under conventional preprocessing so results can be compared cutoff by cutoff rather than only at summarized knowability dates.
- 17.4 percentage points separate held-out basin accuracies on 15 July, narrowing to 7.4 points by 31 December.Northwest Greenland is weakest at every cutoff, while Southwest Greenland is strongest at most.
- Between-lake variance accounts for 92.3% of sdeswe variance and 97.5% of sp variance, compared with 13.4% for runoff, 19.9% for rh2m, and 22.6% for albedo.
- On 1 May, the Sentinel-1 component reaches 21.0% accuracy against a 25.0% chance level, while Sentinel-2 reaches 32.9% against 33.3%.Neither component is distinguishable from chance before mid-June because both bootstrap intervals contain the corresponding chance levels.