MODELING MANGROVE AREA CHANGES AND THE INFLUENTIAL LOCAL FACTORS USING A GEOSPATIAL MACHINE LEARNING (GML) APPROACH
The condition of mangrove changes is increasing and tends to be uncontrolled. The national mangrove rehabilitation program is also being carried out more intensively and should be accompanied by efforts to prevent potential greater mangrove changes, one of which is using a machine learning approach....
Saved in:
| Main Author: | |
|---|---|
| Format: | Dissertations |
| Language: | Indonesia |
| Online Access: | https://digilib.itb.ac.id/gdl/view/83822 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | Institut Teknologi Bandung |
| Language: | Indonesia |
| Summary: | The condition of mangrove changes is increasing and tends to be uncontrolled. The national mangrove rehabilitation program is also being carried out more intensively and should be accompanied by efforts to prevent potential greater mangrove changes, one of which is using a machine learning approach. The use of machine learning can minimize misclassification and provide high accuracy in identifying mangrove changes. The factors that cause mangrove changes in one location and another tend to be different. If mangrove changes are modeled using the global regression method, the resulting model is not optimal because it does not represent the actual phenomenon.
Based on the above, the study aims to develop a hybrid classification algorithm based on a modified machine learning to separate mangrove and non-mangrove land cover and determine local factors that influence mangrove changes. The study location was carried out in the Segara Anakan area, Cilacap Regency, Central Java Province with a total of 325 field sampling points. The hybrid classification technique approach applied is by optimizing the random forest (RF) and decision tree (DT) classification algorithms and also combining the use of spectral and non-spectral data. The selection of the feature importances in building a hybrid classification use the modified recursive feature elimination (MRFE). The geographically weighted logistic regression (GWLR) method is used to estimate the local factors causing mangrove changes so that the relationship between the dependent and independent variables varies greatly in all locations.
The proposed novelty of this study includes finding the important features that can be used to build a hybrid classification algorithm based on modified machine learning in identifying mangrove changes with high accuracy. In addition, the results of analysis of the factors causing mangrove changes which vary have implications for making and formulating policies that also differ from one mangrove location to another so that they can support sustainable mangrove management.
The results show that the hybrid classification algorithm is built based on five important features, namely digital elevation model (DEM), near-infrared (NIR), normalized difference moisture index (NDMI), normalized difference water index (NDWI) and distances from brackish water rivers (DBWR) which are able to map mangrove and non-mangrove objects in a time series and are able to increase overall accuracy by 0.48% and kappa accuracy of 0.01. The local factors that influence mangrove changes are dominated by population density and a fraction of salinity in the western area, distance from ponds in the central area, and distance from settlements in the eastern area. The results and novelty contribute to providing an understanding of the mangrove classification process using a machine learning approach and features importance that are closely related to mangrove ecosystem habitat and supports appropriate mitigation steps for mangrove damage. |
|---|