PaperlayerView exact revision on arXiv

Source-linked AI summary

Selective Transfer Learning for Groundwater Level Anomaly Forecasting: Accuracy-Risk Trade-offs

Yaxuan Wu, Ali Mohammed Mansoor Alsahag, Seyed Sahand Mohammadi Ziabari

arXiv:2610.37733v1cs.LGcs.AI

TL;DR

Groundwater transfer learning can improve average forecasts but may harm individual wells, and practical well-level evidence for selective transfer is limited. The study compares calibration-trained selective blending with fixed and non-contextual transfer policies; on the fixed Netherlands split, the policies trade pooled accuracy against well-level protection rather than one dominating.

  • Problem

    Groundwater transfer-learning studies offer limited evidence on practical well-level decisions for reducing negative transfer under heterogeneous conditions.

    Source: §2 · p. 4

  • Method

    The study compares no transfer, fixed transfer, global convex blending, and CASL, which contextually blends frozen parent predictions using a calibration-trained selector.

    Source: §1 · p. 2

  • Results

    CASL has 0.424–1.084% lower RMSE than No TL, while fixed transfer harms 18.38–51.86% of wells; neither CASL nor Global Convex Blend dominates.

    Source: §7 · pp. 18–19

  • Takeaways & Limitations

    On this fixed split, CASL more often has slightly lower pooled RMSE while Global Convex Blend more often protects harmed wells, supporting an accuracy–risk trade-off.

    Source: §7 · p. 19

    Five seeds use the same wells and fixed temporal split, so the evaluation does not establish generalization to other regions or periods.

    Source: §6.6 · p. 18

Abstract

from arXiv · show

This work studies context-aware selective transfer learning for multi-well groundwater anomaly forecasting in the Netherlands. Source-domain pretraining can raise average accuracy but is not uniformly reliable: it degrades performance on a substantial fraction of heterogeneous wells. Context-Aware Selective Learning (CASL) uses calibration-trained, constrained contextual blending of frozen no-transfer and fixed-transfer parents. Global Convex Blend uses the same parents with one calibration-fitted weight per seed. The corrected task requires nine consecutive observed monthly inputs and an exact three-calendar-month target, without interpolation. Evaluation covers 457 wells and 13,718 test windows under five predefined seeds, with the four-cell maxsim0.60 source pool frozen using Always-TL validation RMSE alone. Fixed transfer harms 18.38-51.86% of wells. CASL has 0.424-1.084% lower RMSE than No TL, partially rescuing 97.53-99.50% and fully rescuing 23.46-43.78% of fixed-harmed wells. Partial rescue can arise structurally from convex blending and does not establish contextual superiority. Global Convex Blend protects more wells in four seeds, while CASL has slightly lower pooled RMSE in four. Neither dominates: the results support an accuracy-risk trade-off on the fixed split, not rolling deployment validity.

1. Introduction

Transfer learning can help groundwater forecasting on average yet harm individual wells, motivating a well-level decision framework. On the corrected Netherlands dataset, CASL and Global Convex Blend show an accuracy–risk trade-off rather than a dominant policy.

1. Introduction

  • The corrected task forecasts standardized groundwater anomalies at exactly three calendar months using nine consecutive observed monthly inputs, without interpolation.It covers 457 eligible Dutch BRO wells and 13,718 site-level test windows under the 1996–2016 modeling constraints.Source: §1 · p. 1
  • Prior groundwater transfer studies emphasize average gains, leaving a practical well-level rule for reducing negative transfer under heterogeneous conditions insufficiently tested.Dutch wells span varied hydrogeology and management contexts, making well-specific transfer effects relevant to national forecasting.Source: §1 · p. 2
  • The reported evidence uses the source-compatible Netherlands target panel; Amsterdam motivated the design but is not the empirical scope.Recent-state and retrospective-shift descriptors show repeated associations, while source–target similarity associations with residual CASL harm are weak and sign-variable; these diagnostics establish neither causal mechanisms nor prospective prediction.Source: §1 · p. 2
  • The study compares No TL, fixed transfer, global convex blending, and selective transfer using a frozen maxsim0.60 source filter and five predefined seeds.CASL blends frozen parents with a bounded contextual weight; Global Convex Blend uses one calibration-fitted weight per seed.Source: §1 · p. 2
  • CASL achieves 0.424–1.084% lower test RMSE than No TL, while fixed transfer harms 18.38–51.86% of wells; neither CASL nor Global Convex Blend dominates.CASL partially rescues 97.53–99.50% and fully rescues 23.46–43.78% of fixed-harmed wells, but partial rescue can arise structurally from convex blending toward No TL.Source: §1 · p. 2

2. Related Work

Prior groundwater transfer-learning studies report average gains, while transfer can vary across locations and fail under domain mismatch. This study addresses the resulting gap with a matched comparison that evaluates pooled accuracy and well-level harm separately.

Groundwater forecasting and transfer learning

  • Ma et al. (2021b,a, 2022) and Profeta et al. (2025) report groundwater anomaly forecasting gains from transfer, but do not test well-level conditional decisions.Patra, Chu and Tatas (2023) find that cross-site benefits vary and performance can deteriorate in hydrogeologically distinct areas such as coastal wells.Source: §2 · p. 3

Conditional transfer and negative transfer

  • Pan and Yang (2010) frame transfer learning as relaxing same-distribution assumptions, while Zhang et al. (2023) describe negative transfer when source and target domains are insufficiently aligned.Transferability estimation and time-series domain adaptation motivate conditional strategies, but the proposed selector is a simpler calibration-stage blending rule.Source: §2 · p. 3

Multi-site heterogeneity and architectural choices

  • CASL is post-hoc bounded blending over frozen no-transfer and fixed-transfer parents, while Global Convex Blend provides a non-contextual comparator.The comparison does not attribute reduced harm uniquely to contextual adaptation; a single transfer rule may not suit all wells.Source: §2 · p. 3

Drought-relevant evaluation

  • The study uses standardized groundwater anomalies for drought-relevant monitoring and does not benchmark stronger backbones beyond its controlled LSTM family.The authors identify stronger-backbone testing as future work.Source: §2 · p. 4

Research gap

  • The missing evidence is a practical well-level decision rule for reducing negative transfer under heterogeneous groundwater conditions within a fair protocol.The study compares no transfer, fixed transfer, global convex blending, and similarity-filtered selective transfer on the Netherlands target dataset.Source: §2 · p. 4

3. Methodology

The methodology compares transfer policies on matched data and frozen parent predictions, then fits calibration-stage selectors to blend the parents. The design separates pooled forecast accuracy from well-level transfer harm.

3.1. Research Design

  • The matched design compares No TL, Always TL, Global Convex Blend, and CASL under common preprocessing, chronological splits, input windows, and parent-training conditions.Site-Hard is a secondary routing comparison; reusing frozen parent predictions separates policy comparisons from changes in data construction or parent architecture.Source: §3.1 · p. 4

3.2. Context-Aware Selective Learning Framework

  • CASL trains two parent experts first, then learns a selector over their fixed predictions during a separate calibration stage.The parents are a target-only model and a source-pretrained, target-fine-tuned transfer model.Source: §3.2 · p. 5

3.2.1. Architecture and Feature Split

  • The architecture separates transferable hydroclimatic inputs from local site context, with distinct shared and local sequence encoders.The shared branch is pretrained; local groundwater history, pumping proxy, and normalized coordinates provide site context. Inputs require nine consecutive observed months, while explicit climate lags extend predictor provenance to t−11.Source: §3.2.1 · p. 5

3.2.2. Transfer Learning and Source Filtering

  • No TL and Always TL share an architecture, differing in whether the transferable encoder is source-pretrained before target fine-tuning.The source pool is selected once using corrected Always-TL validation RMSE alone: maxsim0.60 retains four of 512 cells and is frozen before gate fitting and across seeds.Source: §3.2.2 · p. 5

3.2.3. Selector Design

  • CASL’s calibration-trained gate uses 47 contextual inputs to produce a bounded per-observation blend weight between the frozen parents.Calibration-only clipped blend targets are fitted by weighted regression; histogram gradient boosting is selected in all five seeds. Site-Hard is a secondary discrete choice between the same parents.Source: §3.2.3 · pp. 5–6

3.3. Baselines and Comparison Models

  • Global Convex Blend uses the same frozen parents with one clipped global calibration weight per seed, while Site-Hard is a secondary selector.The benchmark also includes No TL and Always TL; the Global Convex Blend and CASL calibration objectives differ, so their comparison does not isolate a causal contribution of contextual inputs.Source: §3.3 · p. 6, §3.2.3 · p. 6

4. Experimental Setup

The experiments use a corrected Netherlands groundwater panel, fixed chronological splits, and a matched comparison of transfer policies. Evaluation combines pooled forecasting metrics with per-well harm and rescue measures, while reliability is bounded by the fixed-split design and noted data limitations.

4.1. Data Sources

  • The corrected target panel retains 457 wells and 72,801 eligible model windows, including 13,718 test windows, while the source domain has 512 gridded cells.The pumping feature is a coarse proxy based on licensed groundwater-use volumes within a 5 km buffer, not measured per-well abstraction.Source: §4.1 · pp. 6–7

4.2. Data Processing

  • Forecasting requires complete observed inputs at months t−8 through t and an observed response exactly at t+3; missing groundwater months are not interpolated.The fixed target-date splits cover training from 1996-10 to 2008-07, validation from 2008-08 to 2011-06, calibration from 2011-07 to 2013-05, and testing from 2013-06 to 2016-06.Source: §4.2 · p. 7

4.3. Training Configuration

  • The selected CASL gate is histogram gradient boosting, fitted after both parent models are frozen; Table 1 summarizes the corrected training setup and settings.Source: Table 1 · p. 8, §4.3 · p. 8

4.4. Evaluation Strategy

  • The evaluation defines RMSE as its main aggregate metric and distinguishes pooled metrics from per-well harm and rescue rates.Partial rescue is descriptive: convex blending toward No TL with α < 1 guarantees partial rescue of fixed-harmed wells and introduces no new harm.Source: §4.4 · p. 8

4.5. Reliability

  • The five predefined seeds characterize fixed-split seed sensitivity, not population uncertainty or temporal robustness; rolling or blocked temporal validation was not performed.Reliability is supported by saved corrected inputs, manifests, hashes, predictions, and well-level outcomes, but public access to the corrected release remains unconfirmed.Source: §4.4 · p. 8, §4.5 · pp. 8–9

5. Results

Under the corrected fixed-split evaluation, CASL modestly improves pooled RMSE over No TL while transfer harms a substantial, seed-variable subset of wells. Harm reduction and accuracy results are mixed across CASL and Global Convex Blend, and the diagnostics do not establish temporal robustness or causal explanations.

5.1. Main Model Comparison

  • CASL’s test RMSE is 0.424–1.084% lower than No TL across seeds and lower than Global Convex Blend in four of five seeds.The evidence does not show universal superiority: CASL has higher RMSE than Always TL in one seed, and the reported mean MAE is slightly lower for Always TL.Source: §5.1 · p. 11, Table 2 · p. 12

5.2. Negative Transfer and Selective Recovery

  • Always TL harms 18.38–51.86% of wells; CASL partially rescues 97.53–99.50% and fully rescues 23.46–43.78% of fixed-harmed wells.Partial rescue is a reduction in error relative to Always TL, not necessarily restoration of the No-TL baseline; convex shrinkage toward No TL can produce it. CASL reduces the net harmed-well rate by 7.22–17.29 percentage points, while Global Convex Blend has lower harmed-well prevalence in four seeds.Source: §5.2 · p. 11, Table 3 · p. 12

5.3. Robustness

  • Seed variation reflects training randomness on a fixed split, not population uncertainty or temporal robustness; affected-well identities vary across seeds.Only three wells are harmed by Always TL in all five seeds, and no well is fully rescued or persistently harmed in all five; eligibility depends on whether Always TL harms that well in a seed.Source: §5.3 · pp. 12–13

5.4. Source Filtering Sensitivity

  • The four-cell maxsim0.60 source pool was frozen using seed-42 Always-TL validation RMSE alone; maxsim0.50 and maxsim0.60 differed by only 0.001150 RMSE.This operational choice does not establish a scientifically optimal threshold.Source: §5.4 · p. 13

5.5. Selector Interpretation

  • Recent anomaly, endpoint-change, and retrospective train–test-shift descriptors associate positively with CASL-minus-No-TL RMSE, while source-similarity associations are weak and sign-variable.These associations do not establish causal mechanisms or prospective failure prediction; contextual inputs and nonconstant weights alone do not establish added value over non-contextual blending.Source: §5.5 · p. 13

5.6. Gate Behavior and Functional Sensitivity

  • The calibration-only procedure selected hgb_env in every seed, and no seed showed the checked forms of gate collapse.The gate’s mean test weight ranged from 0.425 to 0.605; inference-only permutations showed sensitivity to recent groundwater and hydroclimatic state, not causal importance or accuracy gains.Source: §5.6 · p. 13

5.7. Error Analysis

  • Across the fixed evaluation, CASL reduced net harmed-well rates and improved RMSE versus No TL in every seed, but had worse aggregate RMSE than Always TL in one.Figure 2 compares held-out aggregate metrics across the same wells and observations; Figures 3–5 distinguish all-well harm, conditional rescue, paired effects, and outcome recurrence across seeds. These retrospective contrasts do not establish temporal robustness or superior protection over Global Convex Blend.Source: §5.7 · p. 14, Figure 2 · p. 15, Figure 3 · p. 15

6. Discussion

Across the corrected Netherlands panel, selective transfer yields a small pooled-accuracy gain but does not dominate global blending on well-level harm. The discussion emphasizes metric-dependent preferences, limits on source-filter and selector interpretation, and the fixed-split scope of the findings.

6.1. Negative Transfer and the Accuracy–Risk Trade-off

  • CASL lowers pooled RMSE versus No TL by 0.424–1.084%, but GCB protects more wells in four seeds while CASL has slightly lower pooled RMSE in four; neither dominates.Partial rescue can arise structurally from convex blending toward No TL, so it does not establish contextual added value.Source: §6.1 · p. 17

6.2. Why the Aggregate Gain is Small but Still Meaningful

  • Pooled RMSE can hide well-level harm, and GCB’s zero weights in three seeds restore No TL exactly without strict full rescues.Among the two seeds with positive GCB weights, full-rescue comparisons are mixed.Source: §6.2 · p. 17

6.3. Metric-Dependent Model Preference

  • Always TL has lower mean pooled MAE than CASL (0.90227 versus 0.90420), while GCB has lower mean harmed-well prevalence than CASL (9.85% versus 26.13%).The discussion recommends matching model comparison to the loss or risk measure of interest.Source: §6.3 · p. 17

6.4. Source–Target Mismatch and Source Filtering

  • The maxsim0.60 source filter retains four cells and is frozen before calibration, but its 0.001150-RMSE validation advantage over maxsim0.50 does not establish a generally optimal threshold.The source consists of gridded hydroclimatic model output rather than observed groundwater wells; this mismatch may contribute to heterogeneous outcomes but is not established as the mechanism.Source: §6.4 · p. 17
  • Recent-state and retrospective-shift descriptors repeatedly associate with residual CASL harm, whereas similarity associations are weak and sign-variable; these comparisons establish neither input importance nor prospective failure detection.They are descriptive associations, not evidence of a physical causal mechanism.Source: §6.4 · pp. 17–18

6.5. Operational Implications

  • CASL’s blending weight indicates each frozen parent’s contribution, not transfer-safety probability or physical causality; the Site-Hard variant offers an explicit no-transfer fallback.Reporting the weight could let downstream users audit which parent contributes more.Source: §6.5 · p. 18
  • No drought-event evaluation was performed, and the fixed target-date splits lack a forecast-origin embargo, so results support fixed-split robustness rather than strict rolling-origin deployment validity.The study is a forecasting and transfer-policy comparison, not a complete drought early-warning system.Source: §6.5 · p. 18

6.6. Limitations and Future Work

  • The findings apply primarily to comparably monitored Dutch wells; five seeds on the same wells and fixed split do not establish generalization to other regions or periods.CASL and GCB use different calibration objectives, so their comparison does not isolate the contribution of contextual inputs. Future work prioritizes rolling-origin evaluation and selector ablations under matched calibration objectives.Source: §6.6 · p. 18

7. Conclusion

On the corrected Netherlands panel, transfer is conditionally useful: fixed transfer harms a substantial share of wells, while CASL gives small aggregate gains and partial rescue. The comparison supports an accuracy–risk trade-off, not strict rolling-origin deployment validity.

7. Conclusion

  • Across five seeds, CASL lowers RMSE versus No TL by 0.424–1.084%, while fixed transfer harms 18.38–51.86% of target wells; neither CASL nor GCB dominates.CASL has slightly lower pooled RMSE than GCB in four seeds, while GCB protects more wells from harm in four.Source: §7 · p. 18
  • CASL partially rescues 97.53–99.50% and fully rescues 23.46–43.78% of fixed-harmed wells, but convex blending toward No TL can produce partial rescue structurally.The conclusion describes the main hypothesis as receiving partial support.Source: §7 · p. 18
  • The five-seed check uses the same fixed split and does not establish strict rolling-origin deployment validity or constitute a complete drought early-warning system.Source-compatible panel, gridded source data, and monthly aggregation bound the conclusions.Source: §7 · pp. 18–19

A. Additional Methodology Details

The appendix gathers methodology support for the target-data filtering funnel and source-filter comparison. Its filter comparison uses seed-42 Always-TL validation RMSE, with the selected filter frozen before gate fitting.

A. Additional Methodology Details

  • The appendix collects supporting methodology details, including target-data filtering, training configuration, source-filter comparison, and a CASL pipeline schematic.GCB and Site-Hard use the same frozen parents; their calibration rules are specified in the main Methodology.Source: A. Additional Methodology Details · p. 22
  • Table 4 presents the Netherlands target-data filtering funnel, with final eligibility and test counts based on the corrected calendar protocol..Source: Table 4 · p. 23
  • Table 5 compares corrected source filters using seed-42 Always-TL validation RMSE; the chosen filter is frozen before gate fitting and not reselected by seed.The leading candidates are numerically near-equivalent.Source: Table 5 · p. 23

B. Robustness and Sensitivity Figures

The appendix supplies per-seed performance and seed-42 diagnostics for harm, rescue, and blending. Seed 42 is illustrative rather than representative, and its outcomes are descriptive comparisons among policies.

B. Robustness and Sensitivity Figures

  • The appendix reports complete per-seed RMSE and calibration α plus corrected seed-42 diagnostics; seed 42 has the highest CASL partial/full rescue and smallest RMSE improvement versus No TL.No model is fitted or selected from test labels.Source: B. Robustness and Sensitivity Figures · p. 23
  • In seed 42, CASL has 88 full rescues, 112 partial-only rescues, one case without partial rescue, 247 preserved benefits, and nine new harms.CASL has 113 persistent harms; GCB has 124 persistent harms and zero new harms in this seed.Source: B. Robustness and Sensitivity Figures · p. 23
  • Table 6 compares seed-42 aggregate performance and transfer risk on 13,718 observations from 457 wells, with harm and rescue denominators specified separately.Equality with No TL is not strict full rescue.Source: Table 6 · p. 24
  • Table 7 reports per-seed corrected test RMSE and calibration-only GCB coefficients for identical wells, test windows, and frozen source filters; zero α reproduces No TL exactly.Seed 42 is retained as a reference, not a representative realization.Source: Table 7 · p. 24
  • Figure 6 gives the seed-42 reference comparison of No TL, Always TL, CASL, and Site-Hard; its close point estimates are not the five-seed GCB comparison.The main comparison tables report GCB and five-seed accuracy–risk results.Source: Figure 6 · p. 24
  • Figure 7 separates strict harm among 457 wells from partial and full rescue among 201 fixed-harmed wells; full rescue is nested within partial rescue.Its partial-rescue comparison is descriptive and does not establish contextual added value over global blending.Source: Figure 7 · p. 25
  • Figure 8 partitions 457 seed-42 wells into five mutually exclusive outcomes; partial mitigation is distinct from eliminating harm.GCB baseline ties have separate equality handling.Source: Figure 8 · p. 25
  • Figure 9 compares CASL with Always TL on the 201 fixed-harmed wells: negative histogram values indicate reduced error, and below-diagonal points indicate partial mitigation.These changes do not show stronger protection than GCB.Source: Figure 9 · p. 25

C. Error Analysis Details

Descriptive cross-seed and seed-42 diagnostics compare rescue outcomes with well descriptors, support, and CASL blend weights. The reported comparisons do not establish causal mechanisms or prospective failure prediction.

C. Error Analysis Details

  • The seed-42 support distribution has median 33 observations per well (IQR 24–37; range 1–37); all four fixed-harmed wells with fewer than ten observations receive partial rescue.Three of those four remain harmed by CASL, and one receives full rescue; all nine newly harmed wells have 15–37 observations.Source: C. Error Analysis Details · p. 26
  • Seed-42 full-rescue versus persistent-harm median origin anomaly magnitude is 0.6733 versus 0.8352, and absolute endpoint change per month is 0.0866 versus 0.1097.These reference contrasts do not establish a stable similarity mechanism; descriptor comparisons are descriptive, not causal or prospective failure detectors.Source: C. Error Analysis Details · p. 26
  • Recent-change full-rescue-versus-persistent contrasts repeat direction in five seeds, while magnitude and shift contrasts repeat in four; reused wells mean seeds are not independent ecological replications.The seed-23 new-harm contrast is undefined because that seed has no newly harmed wells.Source: C. Error Analysis Details · p. 26, Figure 10 · p. 27
  • Seed-42 test γ has mean 0.46717, median 0.44791, and standard deviation 0.21957; under the γ≥0.5 convention, 41.17% favor Always TL and 58.83% favor No TL.The convention describes dominant blend weight, not hard routing or event probability; calibration used 6,562 observations, and nonconstant weights do not establish per-well protection.Source: C. Error Analysis Details · p. 26, Figure 12 · p. 28
  • Inference-only sensitivity averages absolute changes in clipped γ across eight joint permutations of nine disjoint groups covering 47 inputs, without refitting or using test truth.Group size, correlated inputs, and unrealistic combinations constrain interpretation; permutation spread is not predictive uncertainty or evidence of causal mechanism or stronger harm protection than GCB.Source: C. Error Analysis Details · p. 26, Figure 13 · p. 29
Loading 2610.37733v1…