Improving Peak River Discharge Predictions Using Urban Bioswale Soil Moisture

Draft research proposal for CyberWater Activity 4

Name: Omar Hegazy Affiliation: Marquette University Email: omar.hegazy@marquette.edu

Project Summary

This study will investigate whether antecedent soil moisture measured in urban bioswales can serve as a useful local proxy for broader antecedent wetness conditions and improve predictions of peak river discharge during rainfall events.

The study will use soil-moisture records from four previously monitored bioswale sites together with publicly-available soil moisture, rainfall and river-discharge observations.

Before modeling begins, a data-feasibility assessment will determine whether each site can be linked meaningfully to suitable rainfall and downstream discharge records and whether sufficient temporal overlap, sampling frequency, and storm events are available.

CyberWater will provide the conceptual framework for data retrieval, preparation, analysis, evaluation, and visualization. Custom scripts may be required for event identification, feature extraction, regression analysis, and model comparison.

Rationale and Objectives

Rainfall characteristics alone may not fully explain differences in river response between storms. The hydrologic response to similar rainfall can vary depending on antecedent wetness and storage conditions within a watershed.

Direct measurements of watershed-scale antecedent soil moisture may not always be available. Soil-moisture sensors installed within urban bioswales provide local observations that may respond to broader wet and dry conditions affecting the surrounding drainage area. This study therefore treats bioswale soil moisture primarily as a potential proxy for antecedent watershed wetness.

The central research question is:

Can antecedent soil-moisture measurements from urban bioswales provide useful additional information for predicting storm-event peak river discharge?

Testable hypothesis

Antecedent soil moisture measured within urban bioswales will provide predictive information about peak river discharge beyond rainfall characteristics alone and may provide additional predictive information beyond rainfall characteristics and pre-storm river discharge.

The study has the following objectives:

1. evaluate the relationship between antecedent bioswale soil moisture and observed peak river discharge; and
2. determine whether adding bioswale soil moisture reduces out-of-sample prediction errors relative to models based on rainfall and pre-storm discharge alone.

Predictive improvement would indicate that the bioswale sensor provides useful information about antecedent hydrologic conditions. It would not establish that the bioswale itself causes changes in river discharge.

Study Area and Data

The study will begin with soil-moisture records from four previously monitored bioswale sites around Milwaukee area. Site names, geographic locations, monitoring periods, sensor depths, sampling intervals, and drainage information remain to be documented.

Before constructing the predictive models, a data search and feasibility assessment will determine whether each bioswale site can be associated with suitable rainfall and downstream river-discharge observations.

The assessment will consider:

• the drainage relationship between each bioswale and a candidate discharge station;
• the size and characteristics of the contributing watershed;
• the hydrologic connectivity between the monitoring site and discharge station;
• temporal overlap among soil-moisture, rainfall, and discharge records;
• sampling frequency and data completeness; and
• the approximate number of usable storm events.

Sites that do not have sufficient spatial or temporal compatibility will not automatically be included in the modeling analysis.

If several bioswale locations share the same receiving watershed, their soil-moisture records may be evaluated individually and, if the number of observations permits, as a combined antecedent-moisture signal. If they drain to different watersheds, analyses will be conducted separately before comparing findings across locations.

Required input Intended use Likely source

Bioswale soil moisture Estimate antecedent Existing measurements local wetness before from the four monitored each rainfall event locations in Milwaukee.

Sensor metadata Assess data quality and Sensor locations, interpret measurements depths, calibration, maintenance, and sampling records

Rainfall observations Calculate storm depth, Local rain gauges or intensity, and duration suitable public rainfall datasets

River discharge Determine pre-storm Suitable downstream observations discharge and observed USGS or local event peak monitoring station

Site and drainage Establish whether the Site plans, watershed information sensor and discharge boundaries, storm-sewer records are maps, outlet correlated information

Analysis scripts and Execute and reproduce Custom scripts, event settings the analysis definitions, model settings, and evaluation configuration

Broader public soil-moisture datasets such as NRCS SCAN or NASA SMAP may be investigated as optional comparison datasets if suitable coverage exists. They will not be required for the primary workflow because their measurements occur at different spatial scales and should not be assumed to represent soil moisture inside the monitored bioswales directly.

Proposed Analysis

Data Feasibility Assessment

The first analytical step will determine whether the required observations can support the proposed event-based comparison.

For each candidate bioswale site, the assessment will identify compatible rainfall and discharge records and quantify their common period of record. Sampling frequency, missing observations, drainage relationships, and the approximate number of identifiable rainfall-response events will be evaluated.

This step will determine which sites proceed to the main analysis and will document why any locations or time periods are excluded.

Data Preparation and Event Identification

For feasible locations, soil-moisture, rainfall, and river-discharge observations will be checked for missing periods, inconsistent timestamps, sensor problems, and differences in sampling interval.

Records will then be aligned while preserving rainfall totals and sufficient discharge resolution to identify event peaks.

Storm-separation rules and river-response windows will be established before model fitting. Their exact definitions will depend on the temporal resolution and characteristics of the available records.

For each eligible event, the analysis will calculate:

• antecedent bioswale soil moisture;
• local minimum soil moisture level before the storm event;
• total storm rainfall depth;
• maximum rainfall intensity;
• storm duration;
• local minimum river discharge immediately before the event; and
• observed peak river discharge associated with the event.

Events without adequate coverage will be excluded, with the reasons for exclusion recorded.

The definition of antecedent soil moisture and the rainfall-intensity interval will be selected according to the available sampling frequency and will be applied consistently within each analysis.

Exploratory Analysis

Scatterplots and correlation coefficients will first be used to examine associations among antecedent soil moisture, rainfall characteristics, pre-storm river discharge, and observed peak discharge.

Pearson or Spearman correlation may be used depending on the form and distribution of the observed relationships.

These analyses will be exploratory. Correlation between bioswale soil moisture and peak discharge alone will not be interpreted as evidence that the soil-moisture measurement improves prediction independently of rainfall or pre-storm discharge.

Regression Modeling and Evaluation

The primary analysis will compare three nested predictive models.

Model 1 --- Rainfall model

Peak discharge predicted from:

• storm rainfall depth;
• maximum rainfall intensity; and
• storm duration.

Model 2 --- Rainfall + initial-flow model

The same predictors as Model 1, plus:

• river discharge immediately before the storm.

This model tests whether information about antecedent watershed state improves prediction beyond rainfall characteristics alone.

Model 3 --- Soil-moisture model

The same predictors as Model 2, plus:

• antecedent bioswale soil moisture.

This model provides the main test of whether the bioswale soil-moisture measurement contains useful predictive information beyond rainfall and pre-storm river discharge.

Multiple regression will be used initially. Ridge regression may be considered if predictor correlation causes unstable coefficient estimates.

All models will use the same eligible storm events and the same training and evaluation sets.

Where the length of the record permits, a later block of storms will be reserved for testing so that predictive performance is evaluated on events not used to fit the models. Entire storm events will remain together, and preprocessing, model fitting, predictor decisions, and any tuning will be performed using the training data only.

Predictive performance will be evaluated using paired errors on the held-out storms. Mean absolute error and root mean squared error in predicted peak discharge will be reported.

The comparison between Models 1 and 2 will indicate the value of antecedent river conditions, while the comparison between Models 2 and 3 will determine whether the bioswale soil-moisture measurement provides additional predictive information.

A finding of little or no improvement will also directly address the research question.

This analysis uses observed storm rainfall to predict event peak discharge retrospectively. Forecasting peak discharge before a future storm would require forecast rainfall or a separately defined procedure for updating predictions as rainfall observations become available.

Desired CyberWater Workflow

The desired user experience is to select or provide:

• bioswale soil-moisture records;
• site and monitoring information;
• a rainfall source;
• a downstream discharge station; and
• a study period.

The workflow would first determine whether sufficient compatible observations exist for analysis. For feasible data combinations, it would retrieve and align the required time series, identify storm events, calculate event-level predictor variables, execute the model comparisons, and return prediction and evaluation results.

Conceptually, the workflow is:

Data selection and retrieval → Data preparation → Regression modeling → Evaluation → Visualization and export

The completed workflow should produce:

• a record of data availability and exclusions;
• an event-level analysis dataset;
• exploratory correlation plots and coefficients;
• observed and predicted peak discharge charts;
• error summaries for the three models; and
• figures showing whether predictive performance changes when antecedent discharge and bioswale soil moisture are added.

CyberWater Capabilities and Development Needs

Several existing CyberWater capabilities may support parts of the proposed workflow.

Documented capability Proposed application

UsgsAgent Retrieve discharge observations where an appropriate USGS gauge exists

PythonSource Integrate custom data-preparation or analysis scripts

msmShowChart Display compatible analysis results

Built-in components for sensor-file import, rainfall retrieval, storm-event identification, regression modeling, and paired predictive evaluation have not yet been established for this workflow.

Where existing components are unavailable, these operations could initially be implemented through custom scripts.

Expected Outcomes and Significance

The completed study will determine whether local soil-moisture observations collected within urban bioswales provide information useful for predicting storm-event peak river discharge.

Another outcome would be the use of bioswale soil-moisture measurements as a useful local proxy for hydrologic conditions relevant to river response.

The study will also demonstrate how local sensor datasets can be combined with public hydrologic observations in a reproducible CyberWater workflow.

Sources of Uncertainty

Temporal overlap may limit the analysis. The soil-moisture, rainfall, and discharge records must overlap for enough rainfall events to support separate model fitting and evaluation.

Additional uncertainties include:

• missing sensor observations;
• sensor drift or calibration problems;
• limited storm-event counts;
• differences in sampling frequency;
• spatially variable rainfall;
• upstream flow regulation;

The number of available events will also determine how complex the model comparison can reasonably be. Additional predictors or separate wet/dry-condition analyses will only be considered if the available sample size supports them.

Implementation feasibility depends on successfully importing the sensor records, retrieving suitable rainfall and discharge observations, executing the required analysis scripts, and passing results through the CyberWater workflow.

Team Feedback and Proposal Revisions

• CyberWater capabilities cover most of the required analysis. Certain visualization requirements can be added using Python-source scripts.