摘要
归一化植被指数(normalized difference vegetation index,NDVI)是评估植被覆盖和生长状况的重要指数,提高NDVI预测精度对区域生态环境保护有重要意义。NDVI预测方法主要有传统回归模型、组合模型和深度学习模型,现阶段预测模型大多考虑了单因素特征,对多特征研究较少,而引入气候因子的多特征深度学习模型有助于提升NDVI的预测精度。因此,文章利用CMIP6气候数据,构建BiLSTM+Attention模型,对青藏高原2023—2024年NDVI进行多特征预测。结果表明:LSTM+Attention模型与传统平滑模型Holt-Winters相比,预测精度较高,RMSE、R 2、MAE分别为0.058、0.832、0.055和0.068、0.791、0.062;融合气候因子的多特征BiLSTM+Attention精度评价表现最好,预测精度RMSE、R 2、MAE分别为0.052、0.891、0.051;相关分析发现,气温和降水是影响NDVI变化的重要因素,融合气候因子能够提升NDVI时序预测精度,RMSE、R 2、MAE分别提升30、11、21个百分点。
The normalized difference vegetation index(NDVI)is a crucial indicator for assessing vegetation cover and growth conditions.Enhancing the accuracy of NDVI predictions is essential for regional ecological and environmental protection.Current methods for predicting NDVI primarily include traditional regression models,ensemble models,and deep learning approaches.However,most existing models tend to focus on single-factor features,with limited exploration of multi-feature methodologies.By incorporating climate factors into a multi-feature deep learning model,the accuracy of NDVI predictions can be significantly improved.Therefore,this study develops a BiLSTM+Attention model utilizing CMIP6 climate data to forecast NDVI in Qinghai-Tibet plateau for the years of 2023-2024.The results indicate that:the LSTM+Attention model demonstrates higher prediction accuracy compared to the traditional smoothing model,Holt-Winters,with RMSE,R 2,and MAE values of 0.058,0.832,and 0.055,respectively,versus 0.068,0.791,and 0.062 for Holt-Winters;the multi-feature BiLSTM+Attention model,which integrates climate factors,shows the best performance,achieving RMSE,R 2,and MAE values of 0.052,0.891,and 0.051,respectively;correlation analysis reveals that temperature and precipitation are significant factors affecting NDVI changes.Incorporating climate factors can improve the temporal prediction accuracy of NDVI,leading to enhancements of approximately 30%,11%,and 21%in RMSE,R 2,and MAE,respectively.
作者
刘宇航
庞国锦
王学佳
孟令震
舒梦瑶
LIU Yuhang;PANG Guojin;WANG Xuejia;MENG Lingzhen;SHU Mengyao(School of Surveying and Mapping and Geographic Information,Lanzhou Jiaotong University,Lanzhou 730070,China;National and Local Joint Engineering Research Center for Geographic National Conditions Monitoring Technology Application,Lanzhou 730070,China;Gansu Provincial Key Laboratory of Surveying and Mapping Science and Technology,Lanzhou 730070,China)
出处
《遥感信息》
北大核心
2025年第4期110-119,共10页
Remote Sensing Information
基金
国家自然科学基金(42161025)。
作者简介
刘宇航(2000-),男,硕士研究生,主要研究方向为植被变化预测。E-mail:12222005@stu.lzjtu.edu.cn;通信作者:庞国锦(1985-),女,硕士,博士,副教授,主要研究方向为生态环境遥感。E-mail:panggj@lzjtu.edu.cn。
