This repository contains the code used in my capstone project for the University of Wisconsin LaCrosse Masters of Data Science program titled “Predictive Dredging Models for the Mississippi River and Illinois Waterway”. Input data is stored in the Data Folder and includes historic river gage data, dredge data, Corps Shoaling Analysis Tool (CSAT) output and gage metadata. The code for this capstone is organized in an RMD document that can be easily run for analysis.
This capstone project uses 25 years of historic river gage observations and shoaling rates derived from hydrosurveys to develop machine learning models to predict shoaling rates as a proxy for dredging need. These models extend the forecasting window of dredging to support the U.S. Army Corps of Engineers Rock Island District dredging operations.
This script creates additional data plots and products that may not have been presented in capstone project. These products however; will be used in operational use of this modeling framework and in communications with key stakeholders.
Author: Barrie Chileen
Martinez, Geographer,
Rock Island District, U.S. Army Corps of Engineers
Program: Department of Data Science, University of Wisconsin – La
Crosse
Course: DS 785: Capstone
LSTM Model and suggestions for performance metrics visualizations were developed using Claude AI - Opus 4.5 on 12/01/2025. AI was used to build out the data preparation handling, performance checks, and workflow for LSTM. AI assisted in debugging and helped build model architecture to correctly handle temporal splits. Other reference materials include R Bloggers Forecasting Sunspots and Time Series Forecasting with LSTM RNN. LSTM structure is modeled after Asborno et al. 2024
├── Chileen_Martinez_Capstone_Code_Final.Rmd # Main analysis script
├── Data/
│ ├── UMR_IWW_1999_2024.csv # River gage observations (1999-2024)
│ ├── CSAT_DATA_Combined.csv # Shoaling rates from CSAT tool
│ ├── gage_metadata.csv # Gage locations and metadata
│ └── Dredge_Event_data.csv # Historical dredging events
├── Output/
│ ├── PCA/ # PCA biplots and variance tables
│ ├── xGBoost/ # xGBoost results and feature importance
│ ├── LSTM/ # LSTM predictions and horizon analysis
│ ├── Pool_Models/ # Pool-level model comparisons
│ ├── Comparisons/ # Model comparison tables
│ ├── Maps/ # Dredging urgency maps
│ └── EDA/ # Exploratory data analysis plots
└── README.Rmd
# Data manipulation and visualization
library(tidyverse)
library(lubridate)
library(zoo)
library(scales)
library(patchwork)
library(gridExtra)
library(viridis)
# Tables and reporting
library(knitr)
library(kableExtra)
library(gt)
# Machine learning
library(caret)
library(xgboost)
library(forecast)
# Deep learning
library(keras3)
library(tensorflow)
# Visualization
library(corrplot)
library(ggfortify)
library(webshot)
# Parallel processing
library(doParallel)
library(foreach)- R version 4.3.3
- Python 3.10
- TensorFlow v2.20.0
| Model | Purpose | Key Parameters |
|---|---|---|
| ARIMA | Baseline comparison | Seasonal, frequency=12 |
| xGBoost | Regression + Classification | 5-fold temporal CV, tuned hyperparameters |
| LSTM | Extended forecasting | 30-day lookback, 45-day horizon, 64→32 architecture |
# Open and knit the RMarkdown file
rmarkdown::render("Chileen_Martinez_Capstone_Code_Final.Rmd")The RMarkdown is organized into modular chunks and can be run in a step-wise approach:
- Data Loading & Cleaning - Load and preprocess gage/CSAT data
- EDA - Distribution plots, correlation matrix, seasonal analysis
- PCA - Principal component analysis by river
- Baseline Models - ARIMA, persistence, mean baselines
- xGBoost - River and pool-level gradient boosting
- LSTM - Deep learning time series forecasting
- Visualizations - Maps and comparison plots
- Results Summary - Export tables for paper
Asborno, M., et al. (2024). Forecasting sediment accumulation in the Southwest Pass with machine-learning models. Journal of Waterway, Port, Coastal, and Ocean Engineering, 150(2), 04023022.
Dunkin, L. M., Coe, L. A., & Ratcliff, J. J. (2018). Corps shoaling analysis tool: Predicting channel shoaling. U.S. Army Engineer Research and Development Center.
This project was developed for academic purpose and internal operational use within USACE Rock Island District. Code is available upon request.
Input data sources are publicly available:
- River gage data: rivergages.com
- Hydrosurvey data: eHydro
- Corps Shoaling Analysis Tool: CSAT