A Shallow-Water Substrate Classification Method Based on the Fusion of Multitemporal Remote Sensing Images Using a Random Forest Model

Substrate classification based on remote sensing images can provide fundamental data for offshore engineering construction and coastal ecological protection without in situ data. To address the inadequate consideration of features in existing substrate classification methods and to mitigate the low classification accuracy caused by noise in single-temporal images, this paper improves the existing remote sensing classification methods and proposes a shallow-water substrate classification method based on the fusion of multitemporal remote sensing images using a random forest model. The proposed method proposes and applies adaptive weighted fusion to bathymetric inversion results from single-temporal images for optimal water depth and terrain features. It derives and constructs a depth-independent bottom reflectance equation, uses median fusion to generate optimal images for accurate bottom reflectance and spectral features, then feeds selected features into the random forest model for training to produce final classification results. Three different types of shallow-water areas were selected for the experiments. The results show that the proposed method effectively leverages multiple features and capitalizes on multitemporal advantages to circumvent the limitations of single-temporal classification, the classification accuracy reaches over 90.34% on different substrates, the maximum misclassification rate is below 8.13% and the number of misclassified pixels is significantly less than that of single-temporal images. Compared with the four mainstream classification algorithms, the proposed method not only achieves overall accuracy (OA) and F1-score exceeding 0.94 and a kappa coefficient greater than 0.90, but demonstrates excellent detail preservation capability and strong adaptability to complex shallow-water environments, and delivers stable classification results.