郁闭度超过60%的林地具有树种种类复杂、种内特征差距小及种间位置间距小等特点。虽然传统卷积神经网络(CNN)较其他分类方法具有精度高与自动化水平高等优势,但其存在学习效率低、识别精度提升困难及可解释性差等缺点。此外,传统规则分割绘制树种图的方法忽视了树种及遥感地物边界特征的变化,易在高郁闭度的林区产生椒盐现象。为解决上述问题,提出基于类激活映射及自注意力模型(ST)的新的树种分类方法(G-ST),它集成了迁移学习、ST分类模型与梯度下降的类激活映射,通过综合长距离特征、数据增强、其余领域的特征知识及预测训练关注度,提升G-ST分类精度、模型泛化能力及可解释性,结合简单线性迭代聚类方法生成树种专题地图。结果表明,该方法得到的树种图精度较传统CNN结合规则分割的制图方法更高,林木及遥感地物边界更趋近于矢量化结果,能有效为树种影像分割、制图及分布统计工作提供参考。

The characteristics of the forest land with a canopy density of more than 60% include complex tree species,a narrow interspecific character gap,and modest interspecific distances.Traditional convolutional neural networks (CNN) are superior to other classification approaches in terms of accuracy and automation,but they also have a low learning efficiency,a tough time enhancing recognition accuracy,and poor interpretability.Additionally,the typical approach of creating maps of tree species through regular segmentation overlooks changes in tree species and boundary characteristics of remote sensing objects,making it simple to create the salt and pepper phenomenon in forest areas with dense canopies.Therefore,a Grad-Swin transformer (G-ST) based on class activation mapping and Swin transformer (ST) classification model was suggested as the solution to the aforementioned issues.It incorporates transfer learning,the ST classification model,and gradient descent class activation mapping,and by integratinglong-distance features,data enhancement,feature knowledge from other fields,and prediction training attention,it increases the G-ST classification accuracy as well as the model's generalization and interpretability abilities.Thematic maps of tree species are created by using a straightforward linear iterative clustering algorithm.The results demonstrated that the accuracy of the tree species map produced by this method was greater than that of the conventional CNN combined with regular segmentation,and the boundary between trees and remote sensing objects was closer to the vectorization result.This method can therefore serve as a useful reference for tree species image segmentation,mapping,and distribution statistics.