Monitoring Surface Water Using Satellite Image Data

Research Mentor: Yohn Jairo Parra Bautista, PhD, Assistant Professor at FAMU

Team: Keana Beaufort, Hampton University & Femi Adebisi, University of Maryland Eastern Shore

Climate and Data Science Research Internship

Profiles & Interests

Keana Beaufort, Undergraduate Student at Hampton University. Interests include studying the effects of atmospheric changes on the environment, weather prediction, human activities, and global warming.

Femi Adebisi, Undergraduate Student at University of Maryland Eastern Shore. Interests include Music, Arts, Black History and Afro Socialism.

Dr. Yohn Jairo Parra Bautista is an instructor in the computer and information sciences department at Florida A&M University. Interests include Machine learning, text mining, AI, and data behavior ethics.

Introduction

• Water is necessary for all life. Water is where life first began. In simple terms, life could not exist without this three-atom molecule.
• Surface Water is any body of water above ground, including streams, rivers, and lakes
• It is an important part of the hydrologic cycle and offers a range of environmental and society functions, such as irrigation for agriculture, drinking water, and habitat for aquatic life.
• Beyond our basic survival, clean water is vital for sanitation, hygiene and disease prevention
• Surface water usage has been increasingly impacted by population growth, climate change, and water management practices, with rising demand, climate-related challenges, and ongoing efforts to provide safe drinking water highlighting the need for efficient management and access.

Data Acquisition and Preprocessing Part 1

Dataset 1: A time-series of a earth observation system that is used to visualize over land and coastal water areas called cloudless Sentinel-2 imagery

1. Use https://s2maps.eu/ to get to know Sentinel-2 satellite imagery and illustrate how Earth’s lakes are changing.

Dataset 2: NWPU-RESISC45 is a dataset of land covers and uses.

• Write a Python program to download, unzip, split, and prepare the data.
• Load and display an example image and label in a Jupyter notebook.

3. Create our own label data.

• Create labels with https://www.makesense.ai/.
• Use the VGG JSON format for polygon annotations.

​

​

​

​

​

Satellite imagery is important for environmental monitoring, urban planning, and disaster management because it provides insight into the dynamic processes and changes that occur on our globe.

​

Data Acquisition and Preprocessing Part 2

​

1. https://s2maps.eu/

Better Post-Processing: Sharper look, more balanced colors - our improved post-processing provides much better results in various environments.

Lake expansion in the Tibetan Plateau

Dataset 2: NWPU-RESISC45 is a dataset of land covers and uses.

Enhancing and Segmenting Images-Workflow

1. Introduction to Google Colab
1. Navigate Google Colab.
2. Create a GPU instance: virtual machine that includes graphic processing units
2. Preparing datasets for a deep learning project
• Use image generator functions.
• Set up model training hyperparameters.
3. Image augmentation
• Learn about the principles of data augmentation and augmenting imagery with a high-level application programming interface and flexible machine learning framework making it easier to develop deep learning models called Keras/TensorFlow 2.0
• Create and merge image and mask generator functions.
• Create and visualize unpredictable transformations to input imagery.

U-Net Model

1. What is it
1. A convolutional neural network architecture designed for biomedical image segmentation. Introduced by Olaf Ronneberger et al. in 2015.
2. How it works
• Encoder (Contracting Path):
1. Consists of convolutional and max-pooling layers that reduce the spatial dimensions and capture features.
• Decoder (Expanding Path):
• Uses up-convolutions and concatenations with corresponding encoder features to reconstruct the image.
• Skip Connections:
• Connect encoder and decoder layers, allowing the network to retain high-resolution features.

Dataset: Sentinel-2 cloudless imagery

Evaluation Metric: Intersection over Union (IoU)

​

Model Training and Evaluation

1. Semantic segmentation of water using U-Net
1. Train a U-Net to detect water in Sentinel-2 cloudless imagery.
2. Use checkpoints and callback functions.
3. Train a U-Net model on NWPU-RESISC45 imagery.
4. Train with image augmentation.

Python code for the U net model

Model Evaluation 1

• In general terms, if the model does not perform well on the training data (high loss and/or low accuracy metrics), it’s best try adding more data and measuring the response before altering the model or training strategy.
• If the model does well on the training but not the testing data, that is a clue that the optimizer needs to change, or the model architecture (a bigger rocket, with more convolution layers).

Model Evaluation 2

​

Conclusions

• Could have used a number of alternative models to solve the problem of identifying water pixels in satellite imagery, some more complicated and some less.
• Made a good choice for a starting point of a baseline model consisting of a simple U-Net model architecture.
• The simplicity of U-Net makes it possible for us to modify the model for our own purposes, modifying it to our developing understanding of the structures and features in the data.
• Using U-Net, we were also able to focus more attention on developing an understanding of various training strategies.
• This research is essential for managing water resources, understanding variations in water supply, and determining how climate change is affecting ecosystems and communities.

Take away and Future work:

• Looking ahead, the knowledge we gained from this research will help us develop more advanced models and expand their applications.
• This progress will contribute to a more sustainable and resilient world. Our findings demonstrate the powerful combination of machine learning and environmental science in tackling some of today's most critical challenges.

Acknowledgements

This research was supported by the UMES Climate Science internship, which was funded by the National Science Foundation (NSF Award #2050874) and the NOAA Educational Partnership Program with Minority Serving Institutions Cooperative Agreement No: NA21SEC4810005. The student researchers acknowledge the NSF, the campus host-University of Maryland Eastern Shore, and UMES Faculty and Director, Dr. Paulinus Chigbu and Dr. Ligia DaSilva. We would also like to thank Ms. Alexia Jones and Rae Quadara, for facilitating the program and Dr. Yohn J. Parra Bautista, for being an outstanding mentor.