基于改进DeepLab v3+模型和迁移学习的高分遥感 耕地 信息提取方法

High-resolution remote sensing arable land information extraction method based on improved DeepLab v3+ model and transfer learning

针对实际遥感耕地信息提取工作中,多源数据特征复杂、样本标注工作繁重等导致高空间分辨率影像解译精度不高、自动化程度不够的问题,本研究基于DeepLab v3+模型,提出一种融合邻域边缘加权模块(NEWM)和轴向注意力机制模块(CBAM-s)的卷积网络模型DEA-Net,结合迁移学习方法进行高分辨率遥感影像耕地信息提取。首先,在浅层网络结构中加入邻域边缘加权模块,提升高分辨率下地物的连续性,细化边缘分割粒度;其次,在深层网络结构中添加轴向注意力机制模块,增加细小地物的关注权重,减少深度卷积导致地物丢失的情况;最后,采用迁移学习的思想,降低样本标注工作量,提高模型学习能力。利用高分卫星土地覆盖数据集(GID)数据构建源域数据集进行模型预训练,将获取的模型参数及权重信息迁移至大数据与计算智能大赛(BDCI)遥感影像地块分割竞赛数据集和全国人工智能大赛(NAIC)遥感影像数据集制作的2种不同目标域数据集中,微调训练后应用于耕地信息提取研究。结果表明,本研究构建方法能够增强模型的空间细节学习能力,提高耕地语义分割精度的同时,降低2/3以上的训练样本数量,为遥感耕地信息提取及农业数据智能化利用提供新的思路和方法。

The complex multi-source data features and heavy sample annotation work in the practical remote sensing arable land information extraction work will lead to low accuracy and insufficient automation of high spatial resolution image interpretation.In view of the above problems,based on DeepLab v3+,we proposed a convolutional network model DEA-Net that incorporated the neighborhood edge weighting module(NEWM)and the axial attention mechanism(CBAM-s),and combined the transfer learning method to extract arable land information of high-resolution remote sensing images.First,the NEWM was added to the shallow network structure to improve the continuity of features under high resolution and refine the granularity of edge segmentation.Then,the CBAM-s was added to the deep network structure to increase the attention weight of fine features and reduce the loss of features due to deep convolution.Finally,the idea of transfer learning was adopted to reduce the sample annotation workload and improve the learning ability of the model.The source domain dataset was constructed using the Gaofen image dataset(GID)for model pre-training,and the acquired model parameters and weight information were migrated to two different target domain datasets produced by big data&computing intelligence contest(BDCI)and national artificial intelligence challenge(NAIC),and fine-tuned and trained for arable land information extraction.The results showed that the method constructed in this study could enhance the spatial detail learning ability of the model,improve the semantic segmentation accuracy of arable land,and reduce the number of training samples by more than 2/3.It can provide new ideas and methods for remote sensing arable land information extraction and intelligent utilization of agricultural data.