Abstract:
To improve the lack of deep learning dataset of coastal ecological
scenes and low accuracy of multi-scale objectives semantic segmentation
for remote sensing image classification, we take three types of coastal typical
ecological supervision multi-scale objectives of mangrove, raft cultivation
and pond aquaculture as research objects, constructs a benchmark dataset
for coastal ecological supervision, improves the UNet feature fusion by
integrating batch normalization and spatial dropout modules, and proposes
a multi-scale deep convolutional semantic segmentation model. The model
has an overall accuracy of 92% on the test set, a kappa coefficient of 0.87, and a
mIoU of 82%. The experimental results show that the coupled stacking of batch
normalization and feature fusion spatial dropout can effectively suppress multiscale
objectives semantic segmentation overfitting and improve the model accuracy
and generalization performance. The proposed model and the constructed semantic
segmentation dataset for coastal ecological supervision can provide decision support for
ecological restoration, mapping and comprehensive management in coastal areas.
针对缺少滨海生态场景深度学习数据集,面向遥感影像分类的多尺度目标语义分割精度不高等问题,研究以红树林、浮筏养殖和围塘养殖三类滨海典型生态监管多尺度目标为研究对象,构建了面向滨海生态监管的多目标语义分割数据集,通过集成批归一化和空间置弃算法,改进UNet 特征融合策略,提出了一种多尺
度深度卷积语义分割模型。模型在测试集上总体精度92%,Kappa 系数0.87,平均交并比82%。实验结果表明批归一化与特征融合空间置弃的耦合堆叠可有效抑制多尺度目标语义分割过拟合,提高模型精度和泛化性能。研究提出的模型及构建的面向滨海生态环境监管的多目标语义分割数据集可为滨海区域生态修复、测绘和综合治理提供决策支持。