现代农业研究2025，Vol.31Issue(9)：46-52,57,8.
基于G-SMOTE-CNN的高光谱遥感土壤分类研究
Research on Hyperspectral Remote Sensing Soil Classification Based on G-SMOTE-CNN
摘要
Abstract
Hyperspectral remote sensing technology,with its ability to obtain high-dimensional and continuous spectral information,provides significant support for achieving high-precision soil classification and digital mapping,which is of great significance for scientifically formulating land use policies and ecological environment protection strategies.However,the problem of imbalanced distribution of known samples in the automatic identification of soil types has always affected the accuracy and stability of classification models.Therefore,this paper combines the optimized Geometric-SMOTE(G-SMOTE)with the 3D-CNN to conduct hyperspectral classification research on four typical soil types:black soil,black soil with albic horizon,albic soil,and meadow soil.The specific methods include:first,conducting statistical analysis on the collected soil sample data and using spatial analysis methods for data expansion.Then,using the default strategy and numerical directional optimization strategy,the G-SMOTE method is used to oversample the expanded soil sample data set to generate a balanced data set.Finally,the 3D-CNN classification model is used for fine classification of different soil types,and some machine learning algorithms are selected for comparative analysis.The results show that G-SMOTE effectively enhances the model's recognition ability for minority soil types,and the 3D-CNN model based on the numerical directional optimization strategy achieves the highest total classification accuracy(91.57%)and Kappa coefficient(0.7270),confirming that the combined application of G-SMOTE and 3D-CNN can effectively address the problem of imbalanced sample distribution in hyperspectral soil classification,providing an efficient and feasible new path for the fine identifica-tion of complex soil types and high-precision soil mapping.关键词
高光谱/土壤分类/几何合成过采样/三维卷积神经网络Key words
hyperspectral/soil classification/G-SMOTE/3D-CNN分类
管理科学引用本文复制引用
林楠,张中帅,姜然哲,彭少聪,麻玮玮..基于G-SMOTE-CNN的高光谱遥感土壤分类研究[J].现代农业研究,2025,31(9):46-52,57,8.基金项目
吉林省自然科学学基金优秀青年基金项目"多源遥感数据与环境变量协同作用的土壤盐分估测及精细化制图方法研究"(项目编号:20230101373JC). （项目编号:20230101373JC）
