【摘要】 针对复杂山地空间异质性显著、耕地信息破碎化严重、提取困难等问题,对DeepLabV3+模型进行改进,基于恐龙谷高分二号卫星影像,构建一种用于复杂山地耕地信息自动提取的MDS-DeepLabV3+模型.使用MobileNetV2作为特征提取器,引入其在ImageNet数据集上的预训练权重,降低复杂度,加速模型拟合;提出密集连接的空间空洞金字塔池化模块与scSE注意力模块结合的DscASPP模块,获取多尺度图像特征,整合空间通道信息.采用CARAFE算子替代原始上采样方法,在较大的感受野范围内聚合上下文信息,实现更准确和高效的特征重建.结果表明,MDS-DeepLabV3+模型平均交并比DeepLabV3+提升6.5%,平均像素准确率增加4.08%,F₁上升4.04%,模型参数量仅有3.97 MB.在禄丰数据集上对各种耕地类型的提取效果均优于其他分割网络,有效降低耕地漏提率和误提率,提取效率及准确性较高.更多还原

【Abstract】 In order to solve the problems of significant spatial heterogeneity,serious fragmentation of cultivated land information,and difficulty in extraction of complex mountainous land,the DeepLabV3+ model was improved,and based on the satellite images of Dinosaur Valley Gaofen-2,an MDS-DeepLabV3+ model for automatic information extraction of complex mountainous cultivated land was constructed.MobileNetV2 was used as the feature extractor,and its pre-trained weights on the ImageNet dataset were introduced to reduce the complexity and accelerate the model fitting.A DscASPP module combining the densely connected atrous spatial pyramid pooling module and the scSE attention module is proposed to obtain multi-scale image features and integrate spatial channel information.The contentaware reassembly feature extraction operator is used to replace the original upsampling method to aggregate the context information in the large receptive field to achieve more accurate and efficient feature reconstruction.The results show that the average intersection ratio of the MDS-DeepLabV3+ is increased by 6.5%,the average pixel accuracy is increased by 4.08%,and the F₁ is increased by 4.04%,and the number of model parameters is only 3.97 MB.The extraction effect of various cultivated land types on the Lufeng dataset is better than that of other segmentation networks,which effectively reduces the rate of under-extraction and mis-extraction of cultivated land,and the extraction efficiency and accuracy are high.更多还原