Predicting Seagrass Habitats in Thailand with Machine Learning

Palika Wannawilai

Department of Geography (GEOG 350)

American River College, California, USA

conservation. Due to a tropical climate, Thailand is one of the countries that locates the most

seagrasses in the world, and more than 70% of seagrasses locate in the west coast along the

Andaman Sea of Thailand. To quantify seagrass abundance, this study will use the Ecological

Marine Unit point data (EMU), provided by USGS and Esri, with machine learning based on

ArcGIS Pro 2.6.3. The result found that suitable habitats of seagrasses were found along the

coastlines in Africa, Asia, Australia, North and South America, especially in high density around

Florida and the Gulf of Mexico. Although the accuracy of the model was about 95 % which was

high enough, the result may not be in high accurate because of using different dataset.

1. Introduction

To solve the reduction of seagrass habitats, Ecological Marine Units or EMUs would be

able to improve the seagrass conservation. By using known seagrass habitats around the United

States with machine learning, the result will be able to predict where seagrass grows worldwide

(Orhun). Moreover, the statistical data of EMUs are baseline 3D mapped ecosystems that will

support ocean sustainability for a framework to detect the change of the ocean (Esri, “Ecological

Marine Units”).

This study aims to create a predicting seagrass habitat map around Thailand in order to

raise concerns of the seagrass protection. However, the distribution of seagrasses around the U.S.

will be used to predict the correlation with seven variables of the ocean measurements from EMUs

with machine learning to achieve this: (i) to predict suitable seagrass habitats around the world and

and classification of physiographic and ecological information about features in the ocean. The

map was created from 3D data visualization and developed a statistical clustering to identify the

physiographic structure such as water column, temperature, salinity, and other factors that will

likely drive ecosystem responses. Users can navigate this marine ecology 3D map through the

ocean in a wide range of ocean parameters, and it is possible to observe other various

environments, such as mangroves and coral reefs. Moreover, to learn how to predict seagrass

habitats around the world, Esri provides a lesson of “Predict seagrass habitats with machine

learning” (Esri, “Predict seagrass habitats with machine learning”). Since the purpose of this study

is to create a suitable seagrass habitat map by using EMU dataset, applying the machine learning

method would help the result of this study to be more precise. According to the result from Esri

3.1 Create a dataset and metadata

The dataset of Global Distribution of Seagrasses derived from https://data.unep-

wcmc.org/datasets/7 by UNEP World Conservation Monitoring Centre (Figure 1) which is the

most recent data of seagrass distribution I have found. This study will use a different dataset from

Esri learning website to compare differences of two datasets. However, the dataset of EMUs and

others are downloaded from Esri learning.

Figure 1: Global Distribution of Seagrass dataset downloaded website.

To prepare input data, EMUs will be fixed its attribute table by using Fill Missing Values

tool, because it lost some information which need to be used for an analysis. Thus, there are seven

important variables; dissolve oxygen, nitrate, phosphate, salinity, silicate, stream, and temperature,

that may be influent in seagrass phenology. The result after filling the missing values in the

attribute table is shown in Figure 2.

layer and the World Seagrass layer will be calculated and given the categorical variable 1 and 0 to

show that there is known seagrass growth at that location for variable 1, and the variable 0 is for

unsuitable seagrass habitats.

3.2 Perform random forest classification

After preparing all datasets, I will use the machine learning libraries to create a prediction

model. First, the study will check the correlation of seven variables to make sure a random forest

classification is the best option. Then, the scripts will test dataset to predict seagrass occurrence

and the accuracy of the predictor. The Python scripts of every step are following in Figure 5, 6,

and 7.

Figure 5: The script for moving the data into Python.

The Kernel Density is the density of point features in a neighborhood around each output

raster cell (Esri, “How Kernel Density Works”). Therefore, creating a kernel density surface by

the Kernel Density tool will help to find locations where predicted seagrasses around the world

will have large concentrations.

The first result is shown in Figure 8 which is the correlation coefficient between seven

variables by using the Python script libraries. Moreover, the test accuracy of the model is about

Figure 9: The accuracy of the model is about 95%.

According to the result from the kernel density surface areas, the range of suitable seagrass

habitats around the world were calculated from 0 to 0.627. The areas of seagrasses are found

around the world, especially in the North Atlantic Ocean. Suitable habitats of seagrasses in

Thailand were found in wide areas along coastlines and the Gulf of Thailand, but in low density

(0.003). In contrast to the North America, suitable seagrass habitats were found in higher density

(0.208-0.549) along coastlines of Florida, and the Gulf of Mexico. The predicting of the

distribution of seagrass habitats around the world and Thailand are shown in Figure 10 and 11.

5. Analysis