基于改进坐标注意力和U-Net神经网络的淡水养殖区提取

Extracting the images of freshwater aquaculture ponds using improvedcoordinate attention and U-Net neural network

针对淡水养殖区空间分布零碎以及样本数量不均衡等因素造成淡水养殖区提取不准确的问题,该研究提出了一种基于U-Net(U-shaped Network)的改进模型,制作了Landsat淡水养殖区动态监测的数据集,增加高、低维特征融合的坐标注意力机制提高模型的提取精度,构建多尺度特征学习更多位置信息,引入focal tversky loss损失函数提升零碎养殖区的识别率,实现1985—2021年研究区淡水养殖区的精确提取,分析近36年研究区淡水养殖区时空变化情况。结果表明:1)2021年淡水养殖区提取效果良好,改进后的模型总体分类精度为0.947,准确率为0.926、召回率0.966、F1分数0.946,均交并比0.899、Kappa系数为0.894,与其他模型相比,总体分类精度、Kappa系数大幅提升。2)1985—2021年,研究区淡水养殖区大致经历起步扩张、急速扩张、轻微萎缩3个阶段:1985—2000年研究区淡水养殖面积持续增加,总面积由1985年0.48km^(2)增长至2000年36.92km^(2),年度增加量大于1km^(2)且小于5km^(2);2000—2017年淡水养殖区面积急速增加至234.47km^(2),年度增加量大于5km^(2);2021养殖区面积209.58km^(2),2017—2021年养殖区面积减少了24.89 km^(2),转出的淡水养殖区多为建设用地所取代。综上,改进的模型具有较高的识别精度,该研究可以为淡水养殖区的提取提供参考,为水产养殖业的科学化管理提供信息依据。

Rapid extraction and precise identification of freshwater aquaculture ponds can play an important role in the management and decision-making of the aquaculture industry.Satellite remote sensing can be expected to provide the critical extraction of freshwater aquaculture ponds,due to its rapid,timely and effective way of Earth observation.However,it is still elusive to delineate the surface area and the variation in the aquaculture ponds from the satellite remote sensing images.In this study,an improved classification was introduced to delineate the changes in aquaculture ponds from Landsat remote sensing images using the coordinate attention with the U-Net neural network model.Firstly,the dataset was collected from the freshwater aquaculture ponds in the Gaochun district of Nanjing City,Jiangsu Province,China.Landsat remote sensing images were also captured from 1985 to 2021,and supplemented by GF-1,GF-2 satellite data and field survey data.Secondly,a coordinate attention model with U-Net as the backbone was created to fully extract the spatial information of features,where the information remained on the fragmented aquaculture ponds.Finally,the long-term Landsat images were selected to explore the temporal and spatial changes of freshwater aquaculture ponds.Six state-of-the-art models were also utilized to verify the improved model.The experimental results show that the performance of the improved model outperformed the rest,in terms of extraction accuracies.Furthermore,there was a dramatic variation in the surface area and distribution of freshwater aquaculture ponds in the study period.Specifically,the surface areas of aquaculture ponds increased from 0.48 to 36.92 km^(2)in the early stages from 1985 to 2000,with an annual increase of 2.43 km^(2).There was rapid growth in the second stage from 2000 to 2017.Among them,the surface areas of aquaculture ponds increased from 36.92 to 234.47 km^(2),with an annual increase of 11.62 km^(2).A gradual shrinkage was found in the surface areas of freshwater aquaculture ponds from 234.47 to 209.58 km^(2)in the third stage from 2017 to 2021,with an annual decrease of 6.22 km^(2).Moreover,the social and economic factors can be attributed to the main driving factors for the dramatic variation in the surface area of the aquaculture areas.In summary,the high overall accuracies of the improved model were achieved to rapidly extract the aquaculture ponds.Additionally,the food demand and economic value can be the significant driving factors for the rapid changes in the aquaculture ponds,while the policies and human activities can be the key factors for the area changes of the aquaculture ponds.The finding can also provide technical support for the scientific management and decision-making of the aquaculture industry.