基于卷积长短期记忆网络的NDVI预测方法研究

An Approach of the Convolutional Long Short-term Memory Network for NDVI Prediction

植被作为生态系统的重要组成部分,分析、预测其分布特征对目标地物监测、估产、自然灾害预测等应用具有重要意义。选取内蒙古自治区锡林郭勒盟为研究区,基于卷积长短期记忆网络(ConvLSTM),利用2000年2月至2012年12月500 m分辨率的MODIS逐月NDVI时序特征影像产品(MODND1M),通过对不同参数进行优化与调整,建立预测模型,实现了2013年逐月NDVI值的预测,并与同期NDVI产品进行对比分析。结果表明,植被生长周期4~10月预测数据整体上与产品数据趋势一致,决定系数R 2平均值为0.678,精度最高可达0.835。该预测算法能充分利用空间时序特征实现对NDVI的预测,并为农作物估产、地上植被生物量估算等研究提供有力的技术支撑。

Vegetation is one of the important parts of an ecosystem.The analysis of vegetation distribution is of great significance to vegetation monitoring,crop yield estimation and natural disaster prediction.Taking the Xilingol League of Inner Mongolia Autonomous Region as an example,we propose an approach of the Convolutional Long Short-Term Memory Network(ConvLSTM).The monthly NDVI timeseries MODIS images(MODND1M)with 500 m spatial resolution from February 2000 to December 2012 are involved in this study.Through parameters optimization and adjustment,a prediction model is established to predict monthly NDVI values in 2013.We find that the prediction results and NDVI products in the same period show a consistent trend from April to October,which is the growth cycle of vegetation.The average value of the coefficients of determination R2 is as high as 0.678.And the maximum value of accuracy is up to 0.835.In other words,the approach proposed in this article makes full use of the spatio-temporal characteristics of images to predict NDVI,and can be used for crop yield estimation and biomass estimation of above-ground vegetation.