A Novel Coral Reef Classification Method Combining Radiative Transfer Model With Deep Learning

Coral reef ecosystem is gradually being threatened, thus monitoring coral reefs using remote sensing is of great significance. There are difficulties in identifying and classifying coral reefs. We propose a radiative transfer model with deep learning (RTDL) to improve the accuracy of the coral reef classification. This model combines the radiative transfer theory with deep learning methods to consider the nonlinear fitting ability of the model and the constraints of the physical model, which improves the classification accuracy. This model utilizes ICESat-2 data to obtain the underwater topography photons (UTP) by active contours with variable convolution kernel model and retrieves water depth by combining the UTP with the Sentinel-2 images. An exponential model is used for reinforcing the reflectivity of the seabed features, which is executed by deep learning models. This study uses multidimensional and multitemporal images to illustrate the generalization and robustness of the RTDL model compared with other methods. The result indicates that the RTDL model obtains remarkable achievements in the classification of coral reef landforms. The method increases the accuracy by 5%, mean intersection over union by 13%, especially for areas, such as reef front slopes and lagoons, where the identification accuracy is increased by 26% for the reef front slopes, by 15% for sandbars, and by 9% for lagoons, compared with the traditional method across 14 areas. The model provides a stable classification framework with physically meaningful for identifying coral reef areas, providing robust method support for the conservation of their ecosystems.