He Chunlong, Jiang Qigang, Tao Guofang, Zhang Zhenchao
College of Geoexploration Science and Technology, Jilin University, Changchun 130026, China.
Sensors (Basel). 2023 Oct 16;23(20):8493. doi: 10.3390/s23208493.
Nearshore water depth plays a crucial role in scientific research, navigation management, coastal zone protection, and coastal disaster mitigation. This study aims to address the challenge of insufficient feature extraction from remote sensing data in nearshore water depth inversion. To achieve this, a convolutional neural network with spatial location integration (CNN-SLI) is proposed. The CNN-SLI is designed to extract deep features from remote sensing data by considering the spatial dimension. In this approach, the spatial location information of pixels is utilized as two additional channels, which are concatenated with the input feature image. The resulting concatenated image data are then used as the input for the convolutional neural network. Using GF-6 remote sensing images and measured water depth data from electronic nautical charts, a nearshore water depth inversion experiment was conducted in the waters near Nanshan Port. The results of the proposed method were compared with those of the Lyzenga, MLP, and CNN models. The CNN-SLI model demonstrated outstanding performance in water depth inversion, with impressive metrics: an RMSE of 1.34 m, MAE of 0.94 m, and R of 0.97. It outperformed all other models in terms of overall inversion accuracy and regression fit. Regardless of the water depth intervals, CNN-SLI consistently achieved the lowest RMSE and MAE values, indicating excellent performance in both shallow and deep waters. Comparative analysis with Kriging confirmed that the CNN-SLI model best matched the interpolated water depth, further establishing its superiority over the Lyzenga, MLP, and CNN models. Notably, in this study area, the CNN-SLI model exhibited significant performance advantages when trained with at least 250 samples, resulting in optimal inversion results. Accuracy evaluation on an independent dataset shows that the CNN-SLI model has better generalization ability than the Lyzenga, MLP, and CNN models under different conditions. These results demonstrate the superiority of CNN-SLI for nearshore water depth inversion and highlight the importance of integrating spatial location information into convolutional neural networks for improved performance.
近岸水深在科学研究、航海管理、海岸带保护和海岸减灾中起着至关重要的作用。本研究旨在应对近岸水深反演中遥感数据特征提取不足的挑战。为此，提出了一种具有空间位置整合的卷积神经网络（CNN-SLI）。CNN-SLI旨在通过考虑空间维度从遥感数据中提取深度特征。在这种方法中，像素的空间位置信息被用作两个额外的通道，与输入特征图像连接。然后将得到的连接图像数据用作卷积神经网络的输入。利用GF-6遥感影像和电子海图实测水深数据，在南山港附近海域进行了近岸水深反演实验。将该方法的结果与Lyzenga、MLP和CNN模型的结果进行了比较。CNN-SLI模型在水深反演中表现出卓越的性能，指标令人印象深刻：均方根误差（RMSE）为1.34米，平均绝对误差（MAE）为0.94米，相关系数（R）为0.97。在整体反演精度和回归拟合方面，它优于所有其他模型。无论水深区间如何，CNN-SLI始终实现最低的RMSE和MAE值，表明在浅水和深水中均具有出色的性能。与克里金法的对比分析证实，CNN-SLI模型与插值水深的匹配度最佳，进一步确立了其相对于Lyzenga、MLP和CNN模型的优越性。值得注意的是，在本研究区域，当使用至少250个样本进行训练时，CNN-SLI模型表现出显著的性能优势，从而得到最优的反演结果。在独立数据集上的精度评估表明，在不同条件下，CNN-SLI模型比Lyzenga、MLP和CNN模型具有更好的泛化能力。这些结果证明了CNN-SLI在近岸水深反演方面的优越性，并突出了将空间位置信息整合到卷积神经网络中以提高性能的重要性。
