Multi-Sensor Decision Tree Machine Learning Algorithm For Identifying Mangrove Ecosystem

Abstract

Abstract. Mangrove forests are the forest type with the highest carbon stocks (above all soil organic carbon) and fulfil numerous other important ecological and environmental functions. Monitoring of mangrove forests and providing reliable data about their development does, therefore, support the informed management of mangrove forests. This paper compares the reliability of multi-spatial resolution PlanetScope in Tapanuli Tengah, Sentinel-2A, and Landsat-8 imagery in Asahan Regency, for detecting mangroves and their vegetation density using a decision tree machine learning (DT-ML) approach. Landsat 8 and Sentinel-2A were used to identify mangrove and non-mangrove classes, while high-resolution PlanetScope was specifically examined to detect mangrove density, i.e., dense, medium, and sparse. The raw satellite data were initially transformed into synthetic images, such as NDVI, NDWI, SAVI, VARI, and ARVI, and then combined with socio-geo-bio-physical variables, including substrate type, elevation, slope, and proximity to the coastline.  In general, this study demonstrates that DT-ML achieves overall accuracies of 92.4% (Landsat 8), 93.0% (Sentinel-2A), 92.1% (PlanetScope Tapanuli Site), and 94.5% (PlanetScope Langkat Site).  It was identified that NDVI and substrate are the most influential variables across all datasets, particularly in distinguishing between mangrove and non-mangrove ecosystems, thereby reducing misclassification of non-mangrove classes. PlanetScope's high spatial resolution (3 m) produces superior detail, enabling better detection of canopy density. The presence of the Sentinel-2A Red Edge channel is crucial for distinguishing mangroves from other types of vegetation. The Decision Tree algorithm has been successfully adapted for multi-resolution imagery, highlighting a model with good interpretability for straightforward generalization. It is concluded that Decision Tree Machine Learning may provide better accuracy and the ability to integrate spectral variables from satellite imagery and socio-geo-biophysical non-spectral variables.