摘要
针对云检测任务中云和背景样本不均衡易造成模型泛化能力差的问题,应用代价敏感学习方法,在卷积网络的损失函数中引入代价系数,同时使用F1分数代替总体精度指标进行模型选择,可有效克服样本不均衡问题。以高分一号影像为实验数据,提取了不同下垫面的云,验证了本方法的有效性。
In order to overcome the generalization problem of the prediction model caused by the imbalance of cloud and background samples in the cloud detection task,we used the cost sensitive learning method,introduced cost coefficients into the loss function of convolution network,and used F1 score instead of overall accuracy to select model,which could effectively overcome the sample imbalance problem.Then,taking GF-1 satellite images as the test data,we extracted cloud areas with different underlying surfaces,which could verify the effectiveness of this method.
出处
《地理空间信息》
2021年第8期54-57,I0006,共5页
Geospatial Information
基金
湖南省自然科学基金资助项目(2019JJ50177)
河南省国土资源厅基于云架构的河南省国土资源遥感监测“一张图”项目。
