摘要
针对广西地区作物种类多、蔗区调查复杂度高,以及因天气多变导致的卫星遥感图像获取困难等问题,该文提出了一种基于Sentienl-2影像的语义分割改进算法用于自动识别甘蔗种植区域,并在多时相的Sentinel-2和Landsat8影像数据基础上,提出了一种代表性光谱特征提取方法构建甘蔗产量预测模型。首先在轻量级网络BiseNetV2中加入了高效通道注意力模块(efficient channel attention,ECA),构建了ECA-BiseNetV2模型识别蔗田的种植区域;然后从识别到的甘蔗种植区域中提取不同时期的多种植被指数,利用线性回归模型将Landsat8植被指数转化为Sentinel-2植被指数,以减小Sentinel-2和Landsat8的数据差异;接着对各蔗区、各生长周期内的植被指数时间序列数据进行三次曲线拟合,提取最大值作为代表性光谱特征;最后使用了多种机器学习算法构建产量预测模型。结果表明,所提出模型总体精度达91.54%,甘蔗查准率达95.57%;基于植被指数拟合最大值构建的决策树模型的测试集R 2为0.792,比采用实际最大值构建的相应模型(R 2=0.759)提升了4.3%。该方法可有效解决因天气问题导致的甘蔗关键生长期遥感图像缺失而难以准确构建产量预测模型的问题,展示出较强的应用性。
This study aims to solve the challenges faced in the prediction of sugarcane yield in Guangxi,such as varied crops,complex investigations in the sugarcane planting areas,and difficult acquisition of remote-sensing images caused by the changeable weather.To this end,an improved semantic segmentation algorithm based on Sentinel-2 images was proposed to automatically identify sugarcane planting areas,and an extraction method for representative spectral features was developed to build a sugarcane yield prediction model based on multi-temporal Sentinel-2 and Landsat8 images.First,an ECA-BiseNetV2 identification model for sugarcane planting areas was constructed by introducing an efficient channel attention(ECA)module into the BiseNetV2 lightweight unstructured network.As a result,the overall pixel classification accuracy reached up to 91.54%,and the precision for sugarcane pixel identification was up to 95.57%.Then,multiple vegetation indices of different growth periods of the identified sugarcane planting areas were extracted,and the Landsat8 image-derived vegetation indices were converted into Sentinel-2 image-based ones using a linear regression model to reduce the differences of the indices derived using images from the two satellites.Subsequently,after the fitting of time-series data of the extracted vegetation indices using a cubic curve,the maximum indices were obtained as the representative spectral features.Finally,a yield prediction model was built using multiple machine learning algorithms.The results indicate that the test set of the decision tree model built using the fitted maximum values of the vegetation indices yielded R?of up to 0.759,4.3%,higher than that(0.792)of the model built using the available actual maximum values.Therefore,this method can effectively resolve the difficulty in developing an accurate sugarcane yield prediction model caused by changeable weather-induced lack of remote sensing images of sugarcane of the key growth periods.
作者
罗维
李修华
覃火娟
张木清
王泽平
蒋柱辉
LUO Wei;LI Xiuhua;QIN Huojuan;ZHANG Muqing;WANG Zeping;JIANG Zhuhui(School of Electrical Engineering,Guangxi University,Nanning 530004,China;Guangxi Key Laboratory of Sugarcane Biology,Guangxi University,Nanning 530004,China;Sugarcane Research Institute,Guangxi Academy of Agricultural Sciences,Nanning 530007,China;Guangxi Sugar Industry Group,Nanning 530022,China)
出处
《自然资源遥感》
CSCD
北大核心
2024年第3期248-258,共11页
Remote Sensing for Natural Resources
基金
广西重大科技专项项目“广西数字蔗田技术平台的构建与应用示范”(编号:桂科AA22117004)
广西重大科技创新基地建设项目“广西甘蔗生物学重点实验室”(编号:桂科2018-266-Z01)
国家自然科学基金项目“低空航拍图像融合田间环境及气象信息立体构建甘蔗长势、品质及产量预测模型”(编号:31760342)共同资助。
关键词
语义分割
植被指数
甘蔗产量预测
卫星遥感
时间序列
semantic segmentation
vegetation index
sugarcane yield prediction
satellite remote sensing
time-series
