Peng, Jianghai, Jiwei Li, Thomas C. Ingalls, Steven R. Schill, Hannah R. Kerner, and Gregory P. Asner. 2025. “A Novel Deep Learning Algorithm for Broad Scale Seagrass Extent Mapping in Shallow Coastal Environments.” ISPRS Journal of Photogrammetry and Remote Sensing 220:277–94. doi: 10.1016/j.isprsjprs.2024.12.008.
Abstract
Recently, the importance of seagrasses in the functioning of coastal ecosystems and their ability to mitigate climate change has gained increased recognition. However, there has been a rapid global deterioration of seagrass ecosystems due to climate change and human-mediated disturbances. Accurate broad-scale mapping of seagrass extent is necessary for seagrass conservation and management actions. Traditionally, these mapping methods have primarily relied on spectral information, along with additional data such as manually designed spatial/texture features (e.g., from the Gray Level Co-Occurrence Matrix) and satellite-derived bathymetry. Despite the widely reported success of prior methods in mapping seagrass across small geographic areas, two challenges remain in broad-scale seagrass extent mapping: 1) spectral overlap between seagrass and other benthic habitats that results in the misclassification of coral/macroalgae to seagrass; 2) seagrass ecosystems exhibit spatial and temporal variability, most current models trained on data from specific locations or time periods encounter difficulties in generalizing to diverse locations or time periods with varying seagrass characteristics, such as density and species. In this study, we developed a novel deep learning model (i.e., Seagrass DenseNet: SGDenseNet) based on the DenseNet architecture to overcome these difficulties. The model was trained and validated using surface reflectance from Sentinel-2 MSI and 9,369 field data samples from four diverse regional shallow coastal water areas. Our model achieves an overall accuracy of 90% for seagrass extent mapping. Furthermore, we evaluated our deep learning model using 1,067 seagrass field data samples worldwide, achieving a producer’s accuracy of 81%. Our new deep learning model could be applied to map seagrass extents at a very broad-scale with high accuracy.
Fig. 4. A conceptual illustration of the proposed SGDenseNet architecture. In this architecture, ’w’, ’h’ and ’c’ (’C’) represent the width, height, and channel number of the input dimension, respectively. The 10 attributes are comprised of 5 spectral reflectance, 4 spectral derivatives, and depth. ’BN’ and ’Conv’ denote batch normalization and convolution operations, respectively. Additionally, ’MaxPooling’ and ’AvgPooling’ refer to maximum and average pooling operations,
respectively.
