GEOSPATIAL MACHINE LEARNING MODELING FOR BLUE CARBON ESTIMATION IN MANGROVE ECOSYSTEMS, TIDAL MARSHES AND SEAGRASS

• Main Author: Simarmata, Nirmawana
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/84226
• Institution: Institut Teknologi Bandung

Blue carbon is carbon sequestered and stored in oceans and coastal ecosystems, including mangroves, tidal marshes, and seagrass beds. Blue carbon is stored above, below, and in the water Blue carbon has the potential to mitigate and reduce climate change. In addition, blue carbon also provides other benefits, such as coastal protection and conservation of marine areas. Despite its potential and benefits, coastal ecosystems are experiencing degradation due to development and land use changes that cause the release of carbon stocks. One of the efforts that can be made for blue carbon management is identifying coastal ecosystems that include mangroves, tidal marshes, and seagrass beds and measuring the carbon stocks stored in them. One of the identification technologies is remote sensing technology, that can analyze a large area efficiently. Remote sensing technology is evolving with the integration of machine learning, which can process data quickly and accurately to identify and map blue carbon ecosystems in greater detail. This study applied remote sensing technology to a portion of Lampung covering the east and south coasts. This research aims to determine the classification model of mangrove, tidal marsh, and seagrass areas using machine learning methods and geospatial modeling of blue carbon estimation in coastal ecosystems in an integrated manner with high accuracy. The data used are Sentinel 2A images and measurements using a field spectroradiometer. This research uses a machine learning method that is divided into two stages, namely the classification and estimation stages. The number of samples used in this study was 986 sample . The process for classification used is random forest classification, while estimation modelling uses random forest regression (RFR). The parameters used in blue carbon estimation are image spectral reflectance, field measurement spectral reflectance, vegetation density, land cover, moisture, and open soil index, with 26 variables. The results obtained from this study are classification maps from the best model, namely model 3 with 26 variables where the mangrove ecosystem has RMSE = 0.61 and AUC = 0.85, tidal marsh ecosystem has RMSE = 0.22 and AUC = 0.87 and seagrass ecosystem has RMSE = 0.18 and AUC = 0.85. The accuracy value of the classification results for the mangrove ecosystem obtained OA = 95%, kappa value = 0.92, UA = 96% and PA = 100%, tidal marsh ecosystem OA = 95%, kappa value = 0.92, UA = 100% and PA = 100%, and seagrass ecosystem has OA = 95%, kappa value = 0.92, UA = 95% and PA = 100%. Furthermore, the results of the geospatial model of mangrove blue carbon estimation show that the best model is Model 26 with 26 variables, with a value of RMSE = 81.44, MAE = 39.05, R² = 0.81. The tidal marsh blue carbon estimation model obtained the best model is Model 30 with 18 variables, which has RMSE = 44.59, MAE = 36.23, R² = 0.85. The seagrass blue carbon estimation model is Model 11 with 11 variables with RMSE = 42.1, MAE = 10.01, R² = 0.88. The final results of the integrated blue carbon estimation model in coastal ecosystems obtained RMSE = 46.2, MAE = 42.6, and R2 = 0.80. The modelling results above show that the RFR model's performance ability is proven to be efficient in estimating blue carbon This allows the model to be used for more accurate and sustainable monitoring and management of blue carbon ecosystems in mangrove, tidal marsh, and seagrass ecosystems with similar species characteristics. This study shows that machine learning algorithms can provide a robust and effective solution to the challenges of estimating blue carbon in coastal ecosystems.