# Water detection on SWOT images

Introduction

This page presents a work that was done during my PhD in the context of water detection in SWOT amplitude images. After an introduction about the SWOT mission, the first two parts are aimed at people having no background in image processing, and introduce the basic principles of Markov Random Fields (MRF) and their optimization (including using graph-cuts). The third and the last parts allow you to tune the parameters of our algorithms and see the results on the SWOT test image. Finally, part 4 explains a new model developed in the framework of this PhD. I would like to acknowledge the CNES and the program Futur & Ruptures from the Institut Mines-Telecom for the funding. I would also like to thank the co-authors of the work demonstrated here: Loic Denis, Roger Fjortoft and Florence Tupin

This PhD is conducted in the context of SWOT mission, led by CNES and NASA's JPL. The main objective of this mission is to recover water elevation with interferometric images using a Ka-band Radar INterferometer (KaRIn) operating at 35.75GHz. The satellite using this instrument will be launch in 2020. It will allow for repetitive measurements (every 22 days) independently of the weather conditions or solar illumination. One of the first step in the processing of the data acquired by SWOT is to detect water bodies in the images. Due to the fact that we want to have a signal in water bodies (which is usually not the case in SAR (Synthetic aperture radar) images), it uses uncommon acquisition parameters: in addition to the high frequency we mentioned (35.75GHz), it uses a near-nadir incidence angle (between 1 and 4 degrees). These unusual acquisition parameters call for the development of new methods. In this demo, we work our way through one of them, dedicated to large areas such as rivers.

Here is the outline of this demo:

If you are not familiar with image processing and MRF, you should start with part 1. Otherwise, you might want to jump to part 3.

Part 1: Markov Random Fields, why ?

Input image

Classify me

Classes

Background

Water

Part 2: Markov Random Fields, how ?

Input image ($v$)

Class image ($u$)

Part 3: [Demo] Influence of $\beta$

In this part, you can play with the parameter $\beta$ defined in part 2. Note that the higher $\beta$ is, the more regularized is the output. Also, note that $\beta = 0$ is equivalent to the pixellic classification shown in part 1. Green pixels are true positives (water classified as water), black are true negatives (background classified as background), red are false positive (background classified as water) and blue are false negatives (water classified as background).
Can we do better ? Let's jump to the next part to find out!

Input image

Output

Options

Beta

4

Part 4: Adapting the algorithm to SWOT

Input image

Output

Part 5: [Demo] Influence of $\beta$, $\beta_{az}$, $\beta_{rg}$ and $\beta_{th}$

In this part, you can play with the parameters defined in part 4 and see the results of the algorithm. As a reminder, $\beta$ tunes the smoothness of the solution, whereas $\beta_{az}$, $\beta_{rg}$ and $\beta_{th}$ tune the priors on the parameter image (in azimuth/vertical direction, in range/horizontal direction, and with respect to the theoretical antenna pattern respectively).
Green pixels are true positives (water classified as water), black are true negatives (background classified as background), red are false positive (background classified as water) and blue are false negatives (water classified as background).
Note that when you set $\beta_{az}$, $\beta_{rg}$ and $\beta_{th}$ to 0, you have the same results as in part 3.

Input image

Output

Options

$\beta$

4

$\beta_{az}$

120

$\beta_{rg}$

200

$\beta_{th}$

2

Metrics

TPR:

Test

FPR:

Test

Error Rate:

Test

MCC:

Test