Multi-model convolutional neural network architectures for coastal forest extent and aboveground biomass estimation

Abstract

Mapping coastal forests through large-scale remote sensing remains challenging, despite extensive local, national, and global efforts. In particular, Sarangani Bay Protected Seascape (SBPS) in the Philippines has been largely overlooked in both national and global coastal forest mapping initiatives. To address this gap, we evaluated the performance of three convolutional neural network (CNN) models, U-Net, DeepLabV3, and PSPNet, in identifying coastal forests within SBPS. These forested areas detected were subsequently analyzed for leaf area index (LAI), which was then used to estimate aboveground biomass (AGB). Among the models tested, U-Net demonstrated the highest accuracy, achieving an overall accuracy of 92.66 %. In contrast, DeepLabV3, while the fastest to train, yielded lower accuracy. AGB estimates revealed that the municipalities of Glan and Maasim had the highest AGB, with 2582.43 Mg ha⁻¹ and 1260.57 Mg ha⁻¹, respectively, while Alabel recorded the lowest at 27.27 Mg ha⁻¹. Although distinguishing true mangroves from non-true mangrove classes in coastal forests remains a limitation, the integration of remote sensing and deep learning offers strong potential for enhancing the accuracy and efficiency for land use and land cover classification, as well as AGB estimation.