基于ConvLSTM多特征融合的农田杂草群落扩散模型研究  

Study on weed community dispersal model based on ConvLSTM multi-feature fusion in agricultural fields

作　　者：权龙哲[1,2] 姜浩 夏福霖 QUAN Longzhe;JIANG Hao;XIA Fulin(School of Engineering,Northeast Agricultural University,Harbin 150030,China;School of Engineering,Anhui Agricultural University,Hefei 230036,China)

机构地区：[1]东北农业大学工程学院,哈尔滨150030 [2]安徽农业大学工学院,合肥230036

出　　处：《东北农业大学学报》2023年第9期53-66,共14页Journal of Northeast Agricultural University

基　　金：安徽省高校自然科学重大项目(2023AH040138);国家自然科学基金项目(32271998,52075092)。

摘　　要：杂草在作物生长初期受环境变化影响快速扩散,严重压缩作物生长环境。为有效管理农田并准确获取杂草群落扩散位置和生长状况,采用深度学习技术,基于卷积长短期记忆网络(Convolutional long short term memory,ConvLSTM)模型及多特征融合对农田杂草群落扩散准确预测,通过无人机获取具有时间序列的数字正射图像(Digital orthophoto map,DOM)数据,数据预处理后,优化土壤调节植被指数(Optimizing soil adjustment vegetation index,OSAVI)阈值法,构建多种输入特征制作数据集。将ConvLSTM模型与多种输入特征融合并对模型进行堆叠优化,构建多特征融合卷积长短期记忆网络(Multi-feature convolutional long short term memory networks,MF-ConvLSTM)模型,实现多步预测,使用制作数据集进行网络训练,综合对比MF-ConvLSTM、ConvLSTM、深度神经网络(Deep neural network,DNN)、全连接长短期记忆网络(Fully-connected long short term memory networks,FC-LSTM)4个模型。结果表明,构建的MF-ConvLSTM模型预测效果较好,其综合性能优于ConvLSTM、DNN和FC-LSTM,均方误差(Mean square error,MSE)值为0.0191,较传统FC-LSTM模型下降0.0087、POD提高0.0702、CSI提高0.0583、FAR降低0.0727。在不同覆盖度和降雨量条件下,MF-ConvLSTM模型杂草群落扩散预测结果较为平均,拥有较稳定MSE值及预测精度,体现模型较好的鲁棒性。此外,根据试验可知特征输入和预测步长对MF-ConvLSTM模型有不同程度影响。研究提出MF-ConvLSTM模型能自适应学习短期时空依赖关系,在多特征共同输入和短期预测步长情况下达到最佳性能。研究为准确获取农田杂草群落扩散位置和生长状况提供思路和方法,也可为后续农田精准除草和制作杂草处方图提供参考。Weeds will spread rapidly in the early stage of crop growth under the influence of environmental changes,seriously compressing the crop growth environment.In order to effectively manage the farmland and accurately obtain the location of weed community proliferation and growth conditions,deep learning technology was used to accurately predict the proliferation of weed communities in farmland based on the ConvLSTM model and multi-feature fusion,DOM data with time series were acquired by UAV,data preprocessing,OSAVI vegetation index thresholding,and construction of a variety of input features to produce a dataset.The MF-ConvLSTM model was constructed by fusing the ConvLSTM model with multiple input features as well as stacking optimization of the model to achieve multi-step prediction,network training using the production dataset,and comprehensively comparing the four models,MF-ConvLSTM,ConvLSTM,DNN,and FC-LSTM.The results showed that the constructed MF-ConvLSTM model predicted better and its comprehensive performance was better than ConvLSTM,DNN,and FC-LSTM,with an MSE value of 0.0191,which was a decrease of 0.0087,an increase of 0.0702 in POD,an increase of 0.0583 in CSI,and a decrease of 0.0727 in FAR compared to the traditional FC-LSTM model.Under different conditions of coverage and rainfall,the MF-ConvLSTM model predicts the spread of weed communities more evenly,with more stable MSE values and prediction accuracy,reflecting the model's better robustness.In addition,according to the experiments,it could be seen that the feature input and the prediction step size had different degrees of influence on the MF-ConvLSTM model.It was proposed that the MF-ConvLSTM model could adaptively learn the short-term spatial and temporal dependencies,and could achieve the best performance under the condition of multi-feature common input and short-term prediction step.The study could provide ideas and methods for accurately obtaining the dispersal location and growth status of weed communities in farmland,and could also provid

关 键 词：ConvLSTM 无人机 杂草群落扩散 时空预测 农业信息 精准农业 

分 类 号：S126[农业科学—农业基础科学] S127S451