An Application of Machine Learning for Mangrove Heath Assessment with Unmanned Aerial Vehicle

Abstract

Background and Objectives: Mangrove forests are important tropical ecosystems that play significant ecological and economic roles. They function as carbon sinks, provide habitats for diverse species, reduce coastal erosion, and treat wastewater before discharge into the sea. Therefore, monitoring and assessing mangrove forest health is essential for biodiversity conservation and sustainable coastal management. However, traditional approaches for assessing mangrove health mainly rely on field surveys, which are time-consuming, resource-intensive, and unable to comprehensively cover large areas. The application of remote sensing technologies, such as satellite imagery and Unmanned Aerial Vehicle (UAV) imagery, has therefore become an important tool. Although satellite imagery can cover extensive areas, cloud cover and atmospheric disturbances pose major limitations to image quality, particularly in tropical humid regions. High-resolution UAV imagery thus represents an alternative for mangrove health assessment. Assessments based on UAV imagery commonly employ vegetation indices in combination with machine learning (ML) techniques; however, studies focusing on mangrove health assessment using UAV data remain limited. Accordingly, this study aims to select appropriate vegetation indices for classifying mangrove health from multispectral UAV imagery and to develop machine learning–based classification models to enable efficient and accurate monitoring of mangrove health changes.

Methodology: This study was conducted in the mangrove forest of Khlong Koh Phi, Phuket Province. Two types of data were collected: field data obtained by measuring leaf spectral reflectance using a spectroradiometer, and multispectral UAV imagery acquired during the summer and rainy seasons. Sixteen vegetation indices were calculated, including NDVI (Normalized Difference Vegetation Index), GNDVI (Green Normalized Difference Vegetation Index), NDWI (Normalized Difference Water Index), SAVI (Soil Adjusted Vegetation Index), EVI (Enhanced Vegetation Index), NDRE (Normalized Difference Red Edge Index), NDI (Normalized Difference Index), TCARI (Transformed Chlorophyll Absorption Ratio Index), RECI (Red Edge Chlorophyll Index), CCCI (Canopy Chlorophyll Content Index), GCC (Green Chromatic Coordinate), VDVI (Visible Difference Vegetation Index), MTCI (MERIS Terrestrial Chlorophyll Index), ExG (Excess Green Index), PSRI (Plant Senescence Reflectance Index), and VARI (Visible Atmospherically Resistant Index). Mangrove canopy images were collected using UAV at predefined locations and classified into healthy and unhealthy categories based on visual interpretation of leaf color, canopy density, and structural abnormalities. The visual classification results were validated by comparison with plant health analysis obtained from the spectroradiometer. The health boundaries were then extended to other tree canopies to serve as reference data for ML model development. Averaged vegetation index values were calculated within the defined health boundaries, and indices capable of distinguishing between healthy and unhealthy canopies were selected using boxplot analysis as input variables for three ML models: Random Forest (RF), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost). The classification results were divided into three classes: healthy areas, unhealthy areas, and other areas (water, mud, and built-up areas). Each model was trained 50 times, with data split into training and testing sets at a ratio of 70:30 in each iteration. The model achieving the highest Overall Accuracy (OA) was selected as the most suitable model. All analytical processes were conducted using Python.

Main Results: The results indicated that 6 out of the 16 vegetation indices could distinguish between healthy and unhealthy mangrove canopies: GNDVI, NDWI, NDRE, CCCI, ExG, and MTCI. When these six indices were used as input variables for the three ML models (RF, SVM, and XGBoost), RF achieved the highest classification performance, with an Overall Accuracy of 77% and a Kappa coefficient of 0.62. The developed model was further tested on areas outside the training dataset. For the healthy class, Producer’s Accuracy was 99% and User’s Accuracy was 95%. For the unhealthy class, although Producer’s Accuracy was 61%, User’s Accuracy reached 88%, indicating the model’s potential for effective plant health analysis.

Conclusions: This study applied three machine learning algorithms RF, SVM, and XGBoost in combination with multispectral UAV imagery to assess mangrove forest health. The findings confirm that UAV imagery has high potential for plant health classification. Among the tested models, RF combined with six vegetation indices (GNDVI, NDWI, NDRE, CCCI, ExG, and MTCI) produced the best results. The developed model can be used to monitor mangrove health in both dry and rainy seasons and to support restoration, management, and long-term systematic monitoring of mangrove forest resources.