Hydrologic Prediction with UNet Model and Advanced Error Detection

Machine Learning Internship Project at UC Irvine's Center for Hydrometeorology and Remote Sensing (CHRS)

Project Overview

This repository includes my internship work focused on enhancing hydrologic prediction accuracy using a UNet model alongside an advanced error detection tool developed in MATLAB. Leveraging satellite datasets—microwave (MW128), infrared (IR128), and rainfall (RAIN128) imagery—this project optimizes data quality control and improves water resource management predictions.

Key Contributions

• Data Pipeline Optimization: Efficient loading and preprocessing of satellite imagery datasets, reducing preparation time and ensuring high-quality inputs for model training.
• UNet Model for Hydrologic Prediction: Custom UNet architecture designed for satellite image segmentation, tailored to predict hydrologic patterns and trends.
• Advanced Error Detection Tool: MATLAB-based tool for satellite image classification quality control, achieving a 35% improvement in prediction accuracy by reducing data noise.
• Performance Improvements: Incorporates optimization techniques like learning rate scheduling and early stopping to boost model accuracy.
• Visual Analysis: Visualizes model predictions, spatial data mappings, and performance metrics to provide a comprehensive view of predictive capabilities.

Project Structure

• UNET_Global01.ipynb: Main notebook with data loading, UNet model setup, training routines, and visual analysis.
• processImages.m: MATLAB script implementing the error detection tool for satellite image classification, enhancing data quality and improving model prediction accuracy by 35%.
• data/: Placeholder for satellite datasets (MW128, IR128, RAIN128).
  Note: Datasets are not included in this repository.
• results/: Placeholder for visual outputs and model results.
  Note: Results are not included in this repository.

Datasets

• MW128: Microwave images, highlighting soil moisture and water content.
• IR128: Infrared images, capturing temperature variations and surface details.
• RAIN128: Rainfall data, supporting the study of precipitation patterns.

Model and Tool Details

• UNet Model

  • Input: Multi-channel satellite images (MW, IR, and RAIN) at 128x128 resolution.
  • Layers: Multiple convolutional and pooling layers for feature extraction, followed by upsampling layers for segmentation.
  • Output: Predictions mapped to geospatial data for visual and practical applications.

• Error Detection Tool

  • Developed in MATLAB for identifying and correcting errors in satellite classification data.
  • Processes images to improve data quality, achieving a 35% increase in hydrologic prediction accuracy by minimizing data inconsistencies.

Visualizations and Results

• Model Predictions: Comparison of input images with predicted segmentation, showcasing the model’s hydrologic feature capture.
• Geospatial Mapping: Visual representations overlaid on spatial maps to interpret predictions in geographic contexts.
• Performance Metrics: Accuracy and loss curves show model performance over training epochs.

Technologies Used

• Python & TensorFlow: Core for model training and data manipulation.
• MATLAB: Advanced error detection and data quality control.
• NumPy & Matplotlib: Data manipulation and visualization.
• Basemap: Geographic data visualization.

Acknowledgments

This project was completed during my Machine Learning Engineering Internship at CHRS, UC Irvine. Thanks to my mentors for their support and guidance.

Contact

For questions or collaboration, please reach out to linhppham0104@gmail.com or connect on https://www.linkedin.com/in/evelynpham04/.