基于BSO-SVR的香蕉遥感时序估产模型研究

BSO-SVR-based Remote Sensing Time-series Yield Estimation Model for Banana

为了提高有限样本下遥感时序估产效果,本文提出一种基于BSO-SVR的香蕉遥感时序估产模型。该模型以广西壮族自治区扶绥县的71块香蕉田块为研究区,利用时间序列Sentinel-2遥感影像数据,结合实测产量数据,对2019—2020年香蕉产量进行预测与分析。融合阈值分割和形态学开操作方法,滤除香蕉关键生育期内遥感影像的厚云和云阴影区域;引入头脑风暴优化算法(Brain storming optimization algorithm,BSO)自动搜寻支持向量回归算法(Support vector regression,SVR)的最优惩罚因子和核函数参数,解决SVR模型的参数优化不足导致模型预测精度低的问题;搭建基于BSO-SVR的时间序列遥感估产模型,深入挖掘多时相遥感信息,以提升香蕉估产准确度。结果表明,相较于网格搜索算法(Grid search,GS)和灰狼优化算法(Grey wolf optimizer,GWO)搜寻SVR模型的最优参数,本文提出的头脑风暴优化算法具有更高的预测精度和更快的预测速度,在2019年和2020年BSO-SVR模型测试集的决定系数(Coefficient of determination,R;)分别为0.777和0.793,验证集R;分别为0.765和0.636,运行时间分别为0.320、0.331 s;与传统的岭回归模型(Ridge regression,RR)和偏最小二乘回归模型(Partial least squares regression,PLSR)相比,BSO-SVR模型的预测性能最佳,其次是RR模型,PLSR模型表现最差。本文提出的时序估产模型实现了香蕉田块产量的精准预估。

Timely,comprehensive and accurate estimation of banana yield can provide growers with decisions on variable fertilization,irrigation,harvest planning,marketing and forward sales.To improve the accuracy of banana remote sensing yield estimation,totally 71 banana fields in Fusui County,Guangxi were used as the study area,and a remote sensing prediction model for banana yield in 2019—2020 was conducted by using time-series Sentinel-2 remote sensing image data,combined with field measured yield data.The method firstly obtained Sentinel-2 images during the key banana phenological period of 2019—2020,then the threshold segmentation and morphological open operation methods were used to remove cloud and cloud shadow coverage areas,the average normalized difference vegetation index(NDVI)values of each plot were extracted,and finally the BSO-SVR model was used to predict and evaluate the banana yield in combination with the actual measured data of banana yield.The results showed that compared with the grid search(GS)and grey wolf optimizer(GWO)algorithms to optimize the penalty factor and kernel function parameters of the SVR model,the brain storming optimization algorithm proposed had higher prediction accuracy and faster prediction speed.The running times of the BSO-SVR model in 2019 and 2020 were 0.320 s and 0.331 s,respectively,and for the validation set,the R;of the BSO-SVR model was 0.777 and 0.793 in 2019 and 2020,respectively;for the test set,the R;of the BSO-SVR model was 0.765 and 0.636 in 2019 and 2020,respectively,except that the R;of the BSO-SVR model in 2019 is slightly lower than that of the GS-SVR model(R;=0.797)in 2019,except that the R;of the BSO-SVR model was higher than that of the GWO-SVR model and the GS-SVR model,and in addition,the overall performance of the RMSE and MAE of the BSO-SVR model was optimal in 2019—2020 compared with that of the GWO-SVR model and the GS-SVR model,indicating that the prediction results of the BSO-SVR model were closer to the actual values and with higher forecasting accuracy.Compared with the traditional ridge regression(RR)and partial least squares regression(PLSR)models,in 2019,the BSO-SVR model had the highest R;,followed by the RR model,and the PLSR model was the worst,where the BSO-SVR model had R;above 0.75 for both the validation and test sets,which was 0.113 and 0.174 higher than that of the RR model,and 0.192 and 0.184 higher than thta of the PLSR model,respectively.Meanwhile,the BSO-SVR model had the lowest RMSE and MAE compared with the RR model and PLSR model,indicating that the BSO-SVR model had good results in forecasting banana yield in 2019.In 2020,the BSO-SVR model had the best overall performance,with the average R;of 0.715 for the validation and test sets,and the R;of the validation and test sets were higher than that of the RR model by 0.035 and 0.014,respectively,and better than that of the PLSR model by 0.040 and 0.035,while the RMSE and MAE of the BSO-SVR model also had the best overall performance.The banana time-series yield estimation model proposed achieved accurate yield prediction of banana field plots,which can provide an effective way for field-scale crop yield estimation.