WO2020047739A1 - 基于多时序属性元素深度特征的小麦重度病害预测方法 - Google Patents

基于多时序属性元素深度特征的小麦重度病害预测方法 Download PDF

Info

Publication number
WO2020047739A1
WO2020047739A1 PCT/CN2018/103965 CN2018103965W WO2020047739A1 WO 2020047739 A1 WO2020047739 A1 WO 2020047739A1 CN 2018103965 W CN2018103965 W CN 2018103965W WO 2020047739 A1 WO2020047739 A1 WO 2020047739A1
Authority
WO
WIPO (PCT)
Prior art keywords
time
disease
wheat
image
information storage
Prior art date
Application number
PCT/CN2018/103965
Other languages
English (en)
French (fr)
Inventor
陈天娇
王儒敬
谢成军
张洁
李�瑞
陈红波
胡海瀛
吴晓伟
Original Assignee
安徽中科智能感知大数据产业技术研究院有限责任公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 安徽中科智能感知大数据产业技术研究院有限责任公司 filed Critical 安徽中科智能感知大数据产业技术研究院有限责任公司
Priority to CN201880024483.5A priority Critical patent/CN110622182A/zh
Priority to PCT/CN2018/103965 priority patent/WO2020047739A1/zh
Publication of WO2020047739A1 publication Critical patent/WO2020047739A1/zh

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/084Backpropagation, e.g. using gradient descent
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q10/00Administration; Management
    • G06Q10/04Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06QINFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
    • G06Q50/00Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
    • G06Q50/02Agriculture; Fishing; Forestry; Mining
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis

Definitions

  • the invention relates to the technical field of agricultural plant protection prediction, in particular to a method for predicting severe wheat diseases based on the depth characteristics of multiple time-series attribute elements.
  • wheat disease goes through multiple states, corresponding to many time periods. Diseases also show different characteristic states at each time period. Severe diseases are based on evolution from scratch to mild to moderate, and there is a high correlation between different time points.
  • the purpose of the present invention is to provide a method for predicting severe wheat diseases based on the depth characteristics of multiple time-series attribute elements, and to solve the dependency relationship between time observations of wheat disease data that cannot be analyzed in the prior art. Information predicts severe disease defects.
  • the first step is to obtain basic data: the basic data includes captured image data sets and environmental information data;
  • the second step is to construct a wheat severe disease prediction model: a deep convolutional neural network and a time series information storage network are used to fuse wheat disease environmental information, image semantics and location environment attributes to construct a wheat disease severe prediction model;
  • the third step is the joint training of the temporal information storage network and the deep convolutional neural network.
  • the multi-day image data set is used as the training sample of the deep convolutional neural network
  • the environmental information data is used as the training sample of the temporal information storage network.
  • the fourth step is to obtain the to-be-predicted image and the to-be-predicted environmental information data
  • the fifth step is the prediction of severe wheat disease: the image to be predicted and the environmental information data to be predicted are input into the model to obtain the prediction result of severe wheat disease.
  • the second step includes the following steps:
  • Time-domain joint learning of environmental information data and image information data Model the characteristics of wheat severe disease data, select several types of environmental information data that affect wheat disease occurrence, and captured image information data for time-domain joint learning;
  • W and b are weight and offset terms, respectively.
  • the step S2.1 specifically includes the following steps:
  • K, L, and M respectively represent species, soil type, and topographical characteristics
  • time series information storage network update rules are set as follows:
  • a 1 is a constant environmental attribute element ⁇ K, L, M ⁇ in the time domain
  • a 2 is a variable attribute element ⁇ I t , C t , S t ⁇ in the time domain
  • T 1 and T 2 are A transformation matrix of time-invariant attributes and time-varying attributes.
  • the time-invariant environment attribute factor feature set of the time-series information storage network is (x 0 , x 1 , x 2 , ..., x T );
  • S2.1.4 Set the number of layers of the time series information storage network model to be consistent with the time point, and each layer is provided with inputs and outputs.
  • the first layer of the time series information storage network model has the input time constant environment attribute factor characteristics x 0 , Its impact results are continuously transmitted to the prediction of each moment;
  • the image passes through the deep feature extraction network and is fused with the time-varying environment attribute factors to be used as the input of the n + 1 layer of the time-series information storage network model, and the time-series information storage network model levels are sequentially input according to the sequence of different time-series shooting;
  • the step S2.1.6 specifically includes the following steps:
  • the long-term storage unit c t-1 forgets and discards information through the forgetting unit f t ;
  • the setting forgetting unit f t is controlled by the external input x t at the current time, the short-term storage output h t-1 at the previous time, and the long-term storage c t-1 at the previous time.
  • the formula is as follows:
  • W xf , W hf , W cf , b f respectively represent the weights and offsets of external input, short-term storage and long-term storage;
  • i t is controlled by x t , h t-1 , and c t-1 .
  • It t , c t , f t , and o t are input units, storage units, forgetting units, and output units on the timing information storage network unit, respectively.
  • ⁇ () represents an S-shaped activation function
  • represents a component component multiplication
  • W is a weight matrix connecting different units
  • the output unit o t is controlled by x t , h t-1 and the long-term storage c t at the current time.
  • the step 3 specifically includes the following steps:
  • the time series information storage network calculates i t , c t , f t , o t , h t values of the five vectors;
  • the back propagation of the error term of the time series information storage network includes two directions: one is the back propagation along time, starting from the current time t, calculating the error term at each time; one is propagating the error term to the upper layer of the feature extraction network, According to the corresponding error term, the gradient of each weight is calculated.
  • the image data set includes image information of a mild disease image, a moderate disease image, and a severe disease image
  • the environmental information data includes temperature, humidity, soil moisture parameters, and historical disease forecast data.
  • the image data set includes approximate disease incidence, mild disease, mild to moderate disease, moderate disease image, and severe disease image information
  • the environmental information data includes temperature, humidity, soil moisture parameters, and historical disease reports. data.
  • the time-series information storage network model includes a storage unit that attempts to store information for a long time, and according to the time-series order, the influence of all pictures and environmental attributes can be accumulated and stored in order.
  • the invention has the advantage that, as a method for predicting severe wheat disease based on the depth characteristics of multiple time-series attribute elements, compared with the prior art, starting from the time-series dimension of wheat disease occurrence, using image, environment, and other characteristic factors, time-series information storage is used.
  • the network and deep feature extraction network fuse multiple time series attribute elements of wheat severe disease, and automatically learn and know the degree of wheat disease in different time periods in the data sequence, so as to achieve prediction for wheat severe disease. Through the analysis and calculation of the existing factors, the development trend of wheat diseases was predicted.
  • the invention solves the problems of completing training and learning from a single feature or data source, and lacking a sequential model analysis of the occurrence time of wheat diseases, using a time series information storage network unit to respond to multiple input variables, and using it in disease severity
  • time series prediction the environment of wheat disease occurrence and the multi-time series attribute information of wheat images are fully utilized, so it can better predict the severity of wheat disease than the existing technology.
  • FIG. 1 is an overall flowchart of a wheat severe disease prediction method based on depth characteristics of multiple time-series attribute elements of the present invention.
  • the method for predicting severe wheat disease based on the depth characteristics of multiple time-series attribute elements includes the following steps:
  • the first step is to obtain basic data, to obtain multi-day image data sets and environmental information data taken by the drone.
  • the multi-day image data set can include image information of mild disease images, moderate disease images, and severe disease images, or approximate disease incidence, mild disease, and mild disease.
  • Related image information such as moderate disease images and severe disease images.
  • Environmental information data includes related environmental information such as temperature, humidity, soil moisture parameters, and historical disease forecast data.
  • the second step is to construct a wheat severe disease prediction model.
  • a deep convolutional neural network and a time series information storage network were used to fuse wheat disease occurrence environment, image semantics and location environment attributes to construct a wheat disease severity prediction model.
  • Convolutional neural networks are highly adaptable, which is very suitable for processing data with statistical stability and local correlation. It can implicitly learn the features of different shape regions from the training image data taken by the drone. It is suitable as a deep feature extraction network for Extraction of image feature information.
  • the temporal information storage network has great advantages in learning the long-term dependence and temporality in higher-level feature sequences, so combining the two can model severe diseases from time and space. The specific steps are as follows:
  • the wheat severe disease data characteristics are modeled, and the image feature information of the disease at different periods and the corresponding environmental information are established.
  • Several types of environmental information data that affect the occurrence of wheat diseases and image information data taken by drones are selected for time-domain joint learning.
  • the occurrence of wheat diseases is a time-series event, so the pixel environment is not stable.
  • the diseases are affected by factors such as different light; on the other hand, the diseases are limited by different appearance factors in different periods.
  • the context of the occurrence environment and the time of occurrence are generally relatively stable, which manifests as that certain types of crops will have corresponding diseases in a certain period of time and in a certain environment. Therefore, it is necessary to make full use of temporal context and environmental context, that is, various occurrence manifestations and perception information in different time periods to conduct research, specifically, the relationship between spatial information, time information, and climate information of disease occurrence.
  • K, L, and M respectively represent the species, soil type, and topographical characteristics. These are attribute elements that will not change over time.
  • time sequence information storage network update rules are set as follows:
  • a 1 is a constant environmental attribute element ⁇ K, L, M ⁇ in the time domain
  • a 2 is a variable attribute element ⁇ I t , C t , S t ⁇ in the time domain
  • T 1 and T 2 are The transformation matrix of the time-invariant attribute and the time-varying attribute.
  • the time-invariant environment attribute factor feature set of the time-series information storage network is (x 0 , x 1 , x 2 , ..., x T ).
  • A4. Set the number of layers of the time series information storage network model to be consistent with the time point, and each layer is provided with input and output.
  • each layer of the time series information storage network model is equivalent to the staged process of developing healthy wheat into severely diseased wheat. That is, in practical applications, if the staged process of developing healthy wheat into severely diseased wheat is divided into healthy (image), mild disease (image), moderate disease (image), and severe disease (image) stages, then the time series information
  • the number of layers of the storage network model is 4; if the healthy wheat develops into a severe disease, the staged process of wheat is divided into health (image), approximate disease incidence (image), mild disease (image), and mild to moderate disease (image) , Moderate disease (image) and severe disease (image) stages, the number of layers of the time series information storage network model is 6, and the number of model layers can also be divided according to the actual situation.
  • the first layer of the time series information storage network model inputs the time-invariant environment attribute factor feature x 0 , and its influence results are continuously transmitted to each moment of prediction. This can make the prediction of the second layer affected by the first layer and the prediction of the third layer affected by the second layer.
  • the time series information storage network model includes a storage unit that attempts to store information for a long time. According to the time series order, the effects of all pictures and environmental attributes can be accumulated and stored in order to facilitate the final prediction.
  • the image is used as the input of the n + 1 layer of the time-series information storage network model, and the time-series information storage network model level is sequentially input in the order of different time-series shooting.
  • the deep feature extraction network is a deep convolutional neural network.
  • a neural network training model for wheat disease images is constructed, which includes several convolutional layers, several pooling layers, fully-linked layers, and output layers.
  • the input of the neural network is the captured image data, which can be normalized to the same pixel size in advance, and the output is the category probability to which the image belongs.
  • the categories are the various stages of the development of the healthy wheat described above into severely diseased wheat. The degree of disease development in different regions can be obtained by deep convolutional neural networks.
  • A6 Set the forward calculation of the timing information storage network element model.
  • the storage unit is used to store the previous state.
  • the key of the time series information storage network unit is the storage unit.
  • the storage unit runs through the entire process.
  • the storage unit carries information and adds or deletes information to the storage unit through the unit structure.
  • the forgetting unit determines how much information of the storage unit c t-1 will be forgotten, so the long-term storage unit c t-1 is set to forget the discarded information through the forgetting unit f t .
  • the forgetting unit f t is controlled by the external input x t at the current time, the short-term storage output h t-1 at the previous time, and the long-term storage c t-1 at the previous time.
  • the expressions are as follows:
  • W xf , W hf , W cf , and b f respectively represent weights and biases of external input, short-term storage and long-term storage, and are obtained through training and learning after the model is established.
  • the A64 the input unit decides how much information can flow into the storage unit, the new information at the current time from the input control unit i t Write a long-term storage unit to generate a new long-term storage c t with the following expression:
  • i t is controlled by x t , h t-1 , and c t-1 .
  • It t , c t , f t , and o t are input units, storage units, forgetting units, and output units on the timing information storage network unit, respectively.
  • ⁇ () represents an S-shaped activation function
  • represents a component component multiplication
  • W is a weight matrix connecting different units.
  • the output unit determines how much information in the storage unit is output. It is controlled by the output unit o t .
  • the relevant accumulated storage is generated from the currently accumulated storage c t. At this moment, the storage h t we are concerned about, and then this part Store for output,
  • the output unit o t is controlled by x t , h t-1 and the long-term storage c t at the current time.
  • the activation function introduces non-linear factors to the neuron, so that the neural network can arbitrarily approximate any non-linear function, so the neural network can be applied to many non-linear models, which is the role of tanh (c t ).
  • W and b are weight and offset terms, respectively.
  • the third step is the joint training of the temporal information storage network and the deep convolutional neural network.
  • Multi-day image data sets are used as training samples for deep convolutional neural networks, and environmental information data is used as training samples for time-series information storage networks to perform joint training of the two.
  • the deep convolutional neural network is trained end-to-end in the traditional way, and the model pre-trained using the ImageNet image set is initialized.
  • the method steps of the forward calculation training are the same as the forward calculation of the time series information storage network unit model in step A6.
  • the deep convolutional neural network is used to extract features from image data obtained at different time periods, and sequentially pass the time series information storage network
  • the output value of each neuron is calculated forward, and the time series information storage network calculates the values of the five vectors i t , c t , f t , o t , h t .
  • Back propagation is the traditional method.
  • Back propagation of the error term of the time series information storage network includes two directions: one is back propagation along time, starting from the current time t, calculating the error term at each time; one is the error
  • the term is propagated to the upper layer of the feature extraction network, and the gradient of each weight is calculated according to the corresponding error term.
  • the fourth step is to obtain the image to be predicted and the environmental information data to be predicted. In practical applications, it is used to obtain the image data captured by the drone during the routine inspection and the environmental information data provided with it.
  • the fifth step is the prediction of severe wheat disease.
  • the to-be-predicted image and the environmental information data to be predicted are input into the prediction model to obtain the prediction result of severe wheat disease.
  • the present invention extracts features from image data obtained at different time periods through a deep convolutional neural network, and models the data features based on the data features through a time series information storage network.
  • the two jointly study to establish a complete wheat severe disease prediction model. After that, multiple days of image data sets and environmental information data training samples of corresponding periods are used to perform the joint training of the two.
  • the error term ⁇ value of each neuron is calculated backward to complete the training.
  • the wheat severe disease prediction can be performed according to the trained prediction model: the image to be predicted and the environmental information data to be predicted are input into the model to obtain the prediction result of the wheat severe disease.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Business, Economics & Management (AREA)
  • General Physics & Mathematics (AREA)
  • Economics (AREA)
  • Human Resources & Organizations (AREA)
  • Strategic Management (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Marketing (AREA)
  • General Health & Medical Sciences (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Data Mining & Analysis (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Mining & Mineral Resources (AREA)
  • Game Theory and Decision Science (AREA)
  • Animal Husbandry (AREA)
  • Agronomy & Crop Science (AREA)
  • Quality & Reliability (AREA)
  • Primary Health Care (AREA)
  • Operations Research (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Marine Sciences & Fisheries (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Development Economics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Image Analysis (AREA)

Abstract

基于多时序属性元素深度特征的小麦重度病害预测方法,包括第一步、基础数据的获取;第二步、小麦重度病害预测模型的构建;第三步、时序信息存储网络和深度卷积神经网络的联合训练;第四步、待预测图像和待预测环境信息数据的获取;第五步、小麦重度病害的预测。该方法将单条地址的匹配效率从1min左右降低到约2.2s;匹配结果在匹配度与精确度指标上更均衡,对推动智慧城市的构建具有较高的应用价值。该方法能自动学习和获知数据序列中不同时间段小麦病害的程度,从而实现针对于小麦重度病害的预测。通过对现有因素的分析计算,预测出小麦病害的发展趋势。

Description

基于多时序属性元素深度特征的小麦重度病害预测方法 技术领域
本发明涉及农业植保预测技术领域,具体涉及基于多时序属性元素深度特征的小麦重度病害预测方法。
背景技术
当下农业大数据正在驱动农业生产向精准化、智能化转变,数据逐渐成为现代农业生产中新兴的生产要素。围绕农田环境下小麦病害大数据表示、识别与预测的模型研究仍处于起步阶段,无论在理论上还是算法上,都还不够完善。特别是,传统的小麦病害识别技术只能识别或预测出病害和非病害小麦,而对于小麦的病害程度则无法判断,而在实际应用中,重度病害的预测对于小麦病害的前期防治有着重要的作用。
现有的小麦病害预测模型研究受限于以下两个方面:影响小麦病害发生的环境信息是复杂多因素,环境信息和获取的直观图像数据具有很高的相关性;其次,没有考虑到小麦病害数据时间观测之间的依赖关系,传统的基于线性回归或神经网络方法无法建模时序预测,以至于其无法预测出重度病害。
小麦病害发生的过程要经历多个状态,对应很多的时间阶段。病害在每个时间段也呈现出不同的特征状态,重度病害是基于从无到有,从轻度、中度演变而来的,不同时间点之间又具有很高的相关性。
因此,如何研发一种能够预测出小麦重度病害的方法已经成为急需解决的技术问题。
发明内容
本发明的目的在于提供基于多时序属性元素深度特征的小麦重度病害预测方法,解决现有技术无法分析小麦病害数据时间观测之间的依赖关系,没有实现建模时序预测,以至于无法通过现有信息预测出重度病害的缺陷。
所述的基于多时序属性元素深度特征的小麦重度病害预测方法,包括以下步骤:
第一步、基础数据的获取:所述基础数据包括拍摄的图像数据集和环境信息数据;
第二步、小麦重度病害预测模型的构建:利用深度卷积神经网络以及时序信息存储网络融合小麦病害发生的环境信息、图像的语义和位置环境属性后,构造出小麦病害重度预测模型;
第三步、时序信息存储网络和深度卷积神经网络的联合训练:将多日的图像数据集作为深度卷积神经网络的训练样本,将环境信息数据作为时序信息存储网络的训练样本,进行两者的联合训练;
第四步、待预测图像和待预测环境信息数据的获取;
第五步、小麦重度病害的预测:将待预测图像和待预测环境信息数据输入模型,得到小麦重度病害的预测结果。
优选的,所述第二步具体包括以下步骤:
S2.1、环境信息数据和图像信息数据的时域联合学习:对小麦重度病害数据特征建模,选取影响小麦病害发生的若干种环境信息数据以及拍摄的图像信息数据进行时域联合学习;
S2.2、将多次迭代的时序信息存储网络的网络单元最终隐藏层状态h(t)作为输入传递进输出层,利用sof tmax函数估计重度病害的概率分布y t
y t=soft max(W*h t+b),
其中,W、b分别为权值、偏置项。
优选的,所述步骤S2.1具体包括以下步骤:
S2.1.1、设定时域上不变的环境属性元素,K、L、M分别表示品种、土壤类型、地形特征;
设定时域上变化的属性元素,t=1至t=T时刻,时域上存在变换的气象C t、土壤墒情特征描述S t和图像I t
S2.1.2、时序信息存储网络更新规则设定如下:
x 0=T 1*A 1
x t=T 2*{I,A 2} t,t∈{1,...,T},
其中,A 1为时域上不变的环境属性元素{K,L,M},A 2为时域上变化的属性元素{I t,C t,S t},T 1和T 2分别为时序不变属性和时序变化属性的转换矩阵,时序信息存储网络的时序不变环境属性因子特征集为(x 0,x 1,x 2,...,x T);
S2.1.3、对特征数据进行归一化处理,把数据维度控制在0到1之间;
S2.1.4、设定时序信息存储网络模型的层数划分与时间点相一致,且每一层均设有输入和输出,时序信息存储网络模型首层输入时序不变环境属性因子特征x 0,其影响结果持续传递到每一刻的预测中;
S2.1.5、时序信息存储网络模型层第n+1层的输入:
图像经过深度特征提取网络后和时序变化环境属性因子融合共同作为时序信息存储网络模型第n+1层的输入,并按照不同时序拍摄的顺序依次输入时序信息存储网络模型的层次;
S2.1.6、设定时序信息存储网络单元模型的前向计算。
优选的,所述步骤S2.1.6的具体包括以下步骤:
S2.1.6-1、长时存储单元c t-1通过遗忘单元f t去遗忘丢弃信息;
S2.1.6-2、设定遗忘单元f t受当前时刻的外部输入x t、上一时刻的短时存储输出h t-1、上一时刻的长时存储c t-1的控制,其表达式如下:
f t=σ(W xfx t+W hfh t-1+W cfc t-1+b f),
W xf、W hf、W cf、b f分别表示外部输入,短时存储和长时存储的权重和偏置;
S2.1.6-3、由当前时刻外部输入x t和上一时刻的短时存储输出h t-1计算出当前时刻的新信息
Figure PCTCN2018103965-appb-000001
Figure PCTCN2018103965-appb-000002
S2.1.6-4、由输入单元i t控制将当前时刻的新信息
Figure PCTCN2018103965-appb-000003
写入长时存储单元,产生新的长时存储c t,其表达式如下:
i t=σ(W xix t+W hih t-1+W cic t-1+b i),
Figure PCTCN2018103965-appb-000004
其中,i t受x t、h t-1、c t-1的控制,i t、c t、f t、o t分别为时序信息存储网络单元上输入单元、存储单元、遗忘单元、输出单元;其中,σ()表示S形的激活函数,·表示组件分量乘法,W是连接不同单元的权值矩阵;
S2.1.6-5、激活长时存储单元c t,准备输出;
S2.1.6-6、由输出单元o t控制,将至目前积累下来的存储c t选出部分相关的存储生成这一时刻关注的存储h t,再把这部分存储进行输出y t
o t=σ(W xox t+W hoh t-1+W coc t1+b o),
h t=o t·tanh(c t),
其中,输出单元o t受x t、h t-1和当前时刻的长时存储c t的控制。
优选的,所述步骤三的具体包括以下步骤:
S3.1、对深度卷积神经网络进行端到端训练,使用ImageNet图像集预训练的模型进行初始化;
S3.2、时序信息存储网络的前向计算训练:
使用预训练的深度卷积神经网络对不同时间段获得的图像数据提取特征,并按顺序通过时序信息存储网络前向计算每个神经元的输出值,时序信息存储网络计算i t、c t、f t、o t、h t五个向量的值;
S3.3、同时微调深度卷积神经网络和时序信息存储网络的所有参数,反向计算每个神经元的误差项δ值;
时序信息存储网络误差项的反向传播包括两个方向:一个是沿时间的反向传播,从当前t时刻开始,计算每个时刻的误差项;一个是 将误差项向特征提取网络上层传播,根据相应的误差项,计算每个权重的梯度。
优选的,所述图像数据集包括轻度病害图像、中度病害图像和重度病害图像的图像信息,所述环境信息数据包括温度、湿度、土壤墒情参数、历史病害测报数据。
优选的,所述图像数据集包括近似病害发病、轻度病害、轻中度病害、中度病害图像和重度病害的图像信息,所述环境信息数据包括温度、湿度、土壤墒情参数、历史病害测报数据。
优选的,所述步骤S2.1.4中,时序信息存储网络模型包含一个尝试将信息储存较久的存储单元,根据时序顺序可以将所有图片和环境属性的影响按顺序依次累积存储。
本发明的优点在于:作为基于多时序属性元素深度特征的小麦重度病害预测方法,与现有技术相比从小麦病害发生的时序维度上以图像、环境等多种特征因素出发,利用时序信息存储网络以及深度特征提取网络融合小麦重度病害多时序属性元素,自动学习和获知数据序列中不同时间段小麦病害的程度,从而实现针对于小麦重度病害的预测。通过对现有因素的分析计算,预测出小麦病害的发展趋势。
本发明解决了从单一特征或数据源出发完成训练学习、缺少小麦病害发生时间上先后顺序模型分析的问题,利用时序信息存储网络单元能够应多个输入变量的问题,将其利用在病害严重度的时序预测上,充分利用了小麦病害发生的环境以及小麦图像的多时序属性信息,因此能比现有技术更好的实现对小麦病害严重度的预测。
附图说明
图1为本发明基于多时序属性元素深度特征的小麦重度病害预测方法的整体流程图。
具体实施方式
下面对照附图,通过对实施例的描述,对本发明具体实施方式作进一步详细的说明,以帮助本领域的技术人员对本发明的发明构思、技术方案有更完整、准确和深入的理解。
如图1所示,本发明所述的基于多时序属性元素深度特征的小麦重度病害预测方法,包括以下步骤:
第一步,基础数据的获取,获取无人机拍摄的多日图像数据集以及环境信息数据。
其中,根据实际应用现场环境的需要,通常多日图像数据集可以包括轻度病害图像、中度病害图像和重度病害图像的图像信息,又或者包括近似病害发病、轻度病害、轻中度病害、中度病害图像和重度病害图像等相关图像信息。环境信息数据包括温度、湿度、土壤墒情参数、历史病害测报数据等相关环境信息。
第二步,小麦重度病害预测模型的构建。
利用深度卷积神经网络以及时序信息存储网络融合小麦病害发生的环境、图像的语义和位置环境属性后,构造出小麦病害重度预测模型。
卷积神经网络适应性强,非常适合处理具有统计平稳性和局部关联性的数据,能够隐式地从无人机拍摄的训练图像数据中学习不同形状区域的特征,作为深度特征提取网络适用于图像特征信息的提取。时序信息存储网络在学习更高级别特征序列中的长期依赖性和时序性上有着很大的优势,所以将两者相结合能从时间和空间对重度病害建模。其具体步骤如下:
A、环境信息数据和图像信息数据的时域联合学习。
对小麦重度病害数据特征建模,对不同时期病害的图像特征信息与相应的环境信息建立关联模型。选取影响小麦病害发生的若干种环境信息数据以及无人机拍摄的图像信息数据进行时域联合学习。
小麦病害发生是时序事件,所以像素环境是不太稳定的,一方面,病害受不同的光照等因素影响;另一方面,病害受不同时期不同外观因素的限制。但是,对同一病害来说,其发生环境和发生时间上下文关系则一般较为稳定,表现为某种类型的作物在一定的时间段和一定的环境下会发生相应的病害。因此需要充分利用时间上下文和环境上下文即不同时间段的各种发生表现和感知信息来进行研究,具体来说是利用病害发生的空间信息、时间信息、气候信息等之间的关系。
其包括以下步骤:
A1、设定时域上不变的环境属性元素,K、L、M分别表示品种、土壤类型、地形特征,其均为随着时间推移上不会发生变换的属性元素。
设定时域上变化的属性元素,t=1至t=T时刻,时域上存在变换 的气象C t、土壤墒情特征描述S t和图像I t,其均为随着时间推移上会发生变换的属性元素。
A2、时序信息存储网络更新规则设定如下:
x 0=T 1*A 1
x t=T 2*{I,A 2} t,t∈{1,...,T},
其中,A 1为时域上不变的环境属性元素{K,L,M},A 2为时域上变化的属性元素{I t,C t,S t},T 1和T 2分别为时序不变属性和时序变化属性的转换矩阵,时序信息存储网络的时序不变环境属性因子特征集为(x 0,x 1,x 2,...,x T)。
A3、对特征数据进行归一化处理,把数据维度控制在0到1之间。由于特征数据中几个参数的维度不同,所以需要先对数据进行归一化处理,把数据维度控制在到1之间,转化为无量纲表达式,有利于消除各维度之间的量纲影响。
A4、设定时序信息存储网络模型的层数划分与时间点相一致,且每一层均设有输入和输出。
在此,时序信息存储网络模型的每一层,相当于健康小麦发展成重度病害小麦的阶段性过程。即,在实际应用中,若健康小麦发展成重度病害小麦的阶段性过程分为健康(图像)、轻度病害(图像)、中度病害(图像)和重度病害(图像)阶段,则时序信息存储网络模型的层数为4层;若健康小麦发展成重度病害小麦的阶段性过程分为健康(图像)、近似病害发病(图像)、轻度病害(图像)、轻中度病害(图像)、中度病害(图像)和重度病害(图像)阶段,则时序 信息存储网络模型的层数为6层,模型层数还可根据实际情况的等级划分。
在此,时序信息存储网络模型首层输入时序不变环境属性因子特征x 0,其影响结果持续传递到每一刻的预测中。这样可以使得第二层的预测受到第一层的影响、第三层的预测受到第二层的影响。同时,时序信息存储网络模型包含一个尝试将信息储存较久的存储单元,根据时序顺序可以将所有图片和环境属性的影响按顺序依次累积存储以利于最后的预测。
A5、时序信息存储网络模型层第n+1层(第2层、第3层)的输入。
图像经过深度特征提取网络后和时序变化环境属性因子融合共同作为时序信息存储网络模型第n+1层的输入,并按照不同时序拍摄的顺序依次输入时序信息存储网络模型的层次。
深度特征提取网络为深度卷积神经网络,以深度卷积神经网络模型为基础构建小麦病害图像的神经网络训练模型,其包含若干卷积层、若干池化层、全链接层与输出层。该神经网络的输入是拍摄的图像数据,可预先归一为相同像素大小,而输出是该图像所属的类别概率,类别为上述的健康小麦发展成重度病害小麦的各个阶段。由深度卷积神经网络可得到各时间段病害在不同区域的发展程度。
A6、设定时序信息存储网络单元模型的前向计算。
前向计算的具体步骤如下:
A61、存储单元用来存储之前的状态,时序信息存储网络单元的 关键是存储单元,存储单元贯穿整个过程,上面承载着信息,通过单元结构对存储单元添加或者删除信息。遗忘单元决定存储单元c t-1有多少信息会被遗忘,因此设定长时存储单元c t-1通过遗忘单元f t去遗忘丢弃信息。
A62、设定遗忘单元f t受当前时刻的外部输入x t、上一时刻的短时存储输出h t-1、上一时刻的长时存储c t-1的控制,其表达式如下:
f t=σ(W xfx t+W hfh t-1+W cfc t-1+b f),
其中,W xf、W hf、W cf、b f分别表示外部输入,短时存储和长时存储的权重和偏置,建立模型后通过训练学习得到。
A63、由当前时刻外部输入x t和上一时刻的短时存储输出h t-1计算出当前时刻的新信息
Figure PCTCN2018103965-appb-000005
Figure PCTCN2018103965-appb-000006
A64、输入单元决定有多少信息可以流入存储单元,由输入单元i t控制将当前时刻的新信息
Figure PCTCN2018103965-appb-000007
写入长时存储单元,产生新的长时存储c t,其表达式如下:
i t=σ(W xix t+W hih t-1+W cic t-1+b i),
Figure PCTCN2018103965-appb-000008
其中,i t受x t、h t-1、c t-1的控制,i t、c t、f t、o t分别为时序信息存储网络单元上输入单元、存储单元、遗忘单元、输出单元;其中,σ()表示S形的激活函数,·表示组件分量乘法,W是连接不同单元的权值矩阵。
A65、激活长时存储单元c t,准备输出。
A66、输出单元决定存储单元内多少信息被输出,由输出单元o t控制,将至目前积累下来的存储c t选出部分相关的存储生成这一时刻我们关注的存储h t,再把这部分存储进行输出,
o t=σ(W xox t+W hoh t-1+W coc t1+b o),
h t=o t·tanh(c t),
其中,输出单元o t受x t、h t-1和当前时刻的长时存储c t的控制。激活函数给神经元引入了非线性因素,使得神经网络可以任意逼近任何非线性函数,这样神经网络就可以应用到众多的非线性模型中,就是tanh(c t)的作用。
B、将多次迭代的时序信息存储网络网络单元最终隐藏层状态h(t)作为输入传递进输出层,利用sof tmax函数估计重度病害的概率分布y t
y t=soft max(W*h t+b),
其中,W、b分别为权值、偏置项。
第三步,时序信息存储网络和深度卷积神经网络的联合训练。
将多日的图像数据集作为深度卷积神经网络的训练样本,将环境信息数据作为时序信息存储网络的训练样本,进行两者的联合训练。
首先,利用传统方式对深度卷积神经网络进行端到端训练,使用ImageNet图像集预训练的模型进行初始化。
其次,时序信息存储网络的前向计算训练。
前向计算训练的方法步骤与步骤A6的设定时序信息存储网络单元模型的前向计算相同,使用深度卷积神经网络对不同时间段获得的 图像数据提取特征,并按顺序通过时序信息存储网络前向计算每个神经元的输出值,时序信息存储网络计算i t、c t、f t、o t、h t五个向量的值。
最后,同时微调深度卷积神经网络和时序信息存储网络的所有参数,反向计算每个神经元的误差项δ值。
反向传播均为传统方式,时序信息存储网络误差项的反向传播包括两个方向:一个是沿时间的反向传播,从当前t时刻开始,计算每个时刻的误差项;一个是将误差项向特征提取网络上层传播,根据相应的误差项,计算每个权重的梯度。
第四步,待预测图像和待预测环境信息数据的获取。在实际应用中为获得无人机在日常巡检过程中拍到的图像数据和相配套提供的环境信息数据。
第五步,小麦重度病害的预测,将待预测图像和待预测环境信息数据输入预测模型,得到小麦重度病害的预测结果。
本发明通过深度卷积神经网络对不同时间段获得的图像数据提取特征,并通过时序信息存储网络依据数据特征建模。二者联合学习建立完整的小麦重度病害预测模型。之后再利用多日的图像数据集和相应时段的环境信息数据训练样本,进行两者的联合训练。通过微调深度卷积神经网络和时序信息存储网络的所有参数,反向计算每个神经元的误差项δ值,完成训练。之后就能依据训练后的预测模型进行小麦重度病害的预测:将待预测图像和待预测环境信息数据输入模型,得到小麦重度病害的预测结果。
在本说明书的描述中,参考术语“一个实施例”、“一些实施例”、“示例”、“具体示例”、或“一些示例”等的描述意指结合该实施例或示例描述的具体特征、结构、材料或者特点包含于本发明的至少一个实施例或示例中。在本说明书中,对上述术语的示意性表述不必须针对的是相同的实施例或示例。而且,描述的具体特征、结构、材料或者特点可以在任一个或多个实施例或示例中以合适的方式结合。此外,在不相互矛盾的情况下,本领域的技术人员可以将本说明书中描述的不同实施例或示例以及不同实施例或示例的特征进行结合和组合。
上面结合附图对本发明进行了示例性描述,显然本发明具体实现并不受上述方式的限制,只要采用了本发明方法构思和技术方案进行的各种非实质性的改进,或未经改进将本发明构思和技术方案直接应用于其它场合的,均在本发明保护范围之内。

Claims (8)

  1. 基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:包括以下步骤:
    第一步、基础数据的获取:所述基础数据包括拍摄的图像数据集和环境信息数据;
    第二步、小麦重度病害预测模型的构建:利用深度卷积神经网络以及时序信息存储网络融合小麦病害发生的环境信息、图像的语义和位置环境属性后,构造出小麦病害重度预测模型;
    第三步、时序信息存储网络和深度卷积神经网络的联合训练:将多日的图像数据集作为深度卷积神经网络的训练样本,将环境信息数据作为时序信息存储网络的训练样本,进行两者的联合训练;
    第四步、待预测图像和待预测环境信息数据的获取;
    第五步、小麦重度病害的预测:将待预测图像和待预测环境信息数据输入模型,得到小麦重度病害的预测结果。
  2. 根据权利要求1所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述第二步具体包括以下步骤:
    S2.1、环境信息数据和图像信息数据的时域联合学习:对小麦重度病害数据特征建模,选取影响小麦病害发生的若干种环境信息数据以及拍摄的图像信息数据进行时域联合学习;
    S2.2、将多次迭代的时序信息存储网络的网络单元最终隐藏层状态h(t)作为输入传递进输出层,利用softmax函数估计重度病害的概率分布y t
    y t=softmax(W*h t+b),
    其中,W、b分别为权值、偏置项。
  3. 根据权利要求2所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述步骤S2.1具体包括以下步骤:
    S2.1.1、设定时域上不变的环境属性元素,K、L、M分别表示品种、土壤类型、地形特征;
    设定时域上变化的属性元素,t=1至t=T时刻,时域上存在变换的气象C t、土壤墒情特征描述S t和图像I t
    S2.1.2、时序信息存储网络更新规则设定如下:
    x 0=T 1*A 1
    x t=T 2*{I,A 2} t,t∈{1,...,T},
    其中,A 1为时域上不变的环境属性元素{K,L,M},A 2为时域上变化的属性元素{I t,C t,S t},T 1和T 2分别为时序不变属性和时序变化属性的转换矩阵,时序信息存储网络的时序不变环境属性因子特征集为(x 0,x 1,x 2,...,x T);
    S2.1.3、对特征数据进行归一化处理,把数据维度控制在0到1之间;
    S2.1.4、设定时序信息存储网络模型的层数划分与时间点相一致,且每一层均设有输入和输出,时序信息存储网络模型首层输入时序不变环境属性因子特征x 0,其影响结果持续传递到每一刻的预测中;
    S2.1.5、时序信息存储网络模型层第n+1层的输入:
    图像经过深度特征提取网络后和时序变化环境属性因子融合共同作为时序信息存储网络模型第n+1层的输入,并按照不同时序拍摄 的顺序依次输入时序信息存储网络模型的层次;
    S2.1.6、设定时序信息存储网络单元模型的前向计算。
  4. 根据权利要求3所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述步骤S2.1.6的具体包括以下步骤:
    S2.1.6-1、长时存储单元c t-1通过遗忘单元f t去遗忘丢弃信息;
    S2.1.6-2、设定遗忘单元f t受当前时刻的外部输入x t、上一时刻的短时存储输出h t-1、上一时刻的长时存储c t-1的控制,其表达式如下:
    f t=σ(W xfx t+W hfh t-1+W cfc t-1+b f),
    W xf、W hf、W cf、b f分别表示外部输入,短时存储和长时存储的权重和偏置;
    S2.1.6-3、由当前时刻外部输入x t和上一时刻的短时存储输出h t-1计算出当前时刻的新信息
    Figure PCTCN2018103965-appb-100001
    Figure PCTCN2018103965-appb-100002
    S2.1.6-4、由输入单元i t控制将当前时刻的新信息
    Figure PCTCN2018103965-appb-100003
    写入长时存储单元,产生新的长时存储c t,其表达式如下:
    i t=σ(W xix t+W hih t-1+W cic t-1+b i),
    Figure PCTCN2018103965-appb-100004
    其中,i t受x t、h t-1、c t-1的控制,i t、c t、f t、o t分别为时序信息存储网络单元上输入单元、存储单元、遗忘单元、输出单元;其中,σ( )表示S形的激活函数,·表示组件分量乘法,W是连接不同单元的权值矩阵;
    S2.1.6-5、激活长时存储单元c t,准备输出;
    S2.1.6-6、由输出单元o t控制,将至目前积累下来的存储c t选出部分相关的存储生成这一时刻关注的存储h t,再把这部分存储进行输出y t
    o t=σ(W xox t+W hoh t-1+W coc t1+b o),
    h t=o t·tanh(c t),
    其中,输出单元o t受x t、h t-1和当前时刻的长时存储c t的控制。
  5. 根据权利要求4所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述步骤三的具体包括以下步骤:
    S3.1、对深度卷积神经网络进行端到端训练,使用ImageNet图像集预训练的模型进行初始化;
    S3.2、时序信息存储网络的前向计算训练:
    使用预训练的深度卷积神经网络对不同时间段获得的图像数据提取特征,并按顺序通过时序信息存储网络前向计算每个神经元的输出值,时序信息存储网络计算i t、c t、f t、o t、h t五个向量的值;
    S3.3、同时微调深度卷积神经网络和时序信息存储网络的所有参数,反向计算每个神经元的误差项δ值;
    时序信息存储网络误差项的反向传播包括两个方向:一个是沿时间的反向传播,从当前t时刻开始,计算每个时刻的误差项;一个是将误差项向特征提取网络上层传播,根据相应的误差项,计算每个权重的梯度。
  6. 根据权利要求1所述的基于多时序属性元素深度特征的小麦 重度病害预测方法,其特征在于:所述图像数据集包括轻度病害图像、中度病害图像和重度病害图像的图像信息,所述环境信息数据包括温度、湿度、土壤墒情参数、历史病害测报数据。
  7. 根据权利要求1所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述图像数据集包括近似病害发病、轻度病害、轻中度病害、中度病害图像和重度病害的图像信息,所述环境信息数据包括温度、湿度、土壤墒情参数、历史病害测报数据。
  8. 根据权利要求3所述的基于多时序属性元素深度特征的小麦重度病害预测方法,其特征在于:所述步骤S2.1.4中,时序信息存储网络模型包含一个尝试将信息储存较久的存储单元,根据时序顺序可以将所有图片和环境属性的影响按顺序依次累积存储。
PCT/CN2018/103965 2018-09-04 2018-09-04 基于多时序属性元素深度特征的小麦重度病害预测方法 WO2020047739A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201880024483.5A CN110622182A (zh) 2018-09-04 2018-09-04 基于多时序属性元素深度特征的小麦重度病害预测方法
PCT/CN2018/103965 WO2020047739A1 (zh) 2018-09-04 2018-09-04 基于多时序属性元素深度特征的小麦重度病害预测方法

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/103965 WO2020047739A1 (zh) 2018-09-04 2018-09-04 基于多时序属性元素深度特征的小麦重度病害预测方法

Publications (1)

Publication Number Publication Date
WO2020047739A1 true WO2020047739A1 (zh) 2020-03-12

Family

ID=68921045

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/103965 WO2020047739A1 (zh) 2018-09-04 2018-09-04 基于多时序属性元素深度特征的小麦重度病害预测方法

Country Status (2)

Country Link
CN (1) CN110622182A (zh)
WO (1) WO2020047739A1 (zh)

Cited By (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738138A (zh) * 2020-06-19 2020-10-02 安徽大学 一种耦合气象特征区域尺度的小麦条绣病严重度遥感监测方法
CN111968003A (zh) * 2020-09-04 2020-11-20 郑州轻工业大学 一种基于农作物本体概念响应的农作物病害预测方法
CN112101265A (zh) * 2020-09-22 2020-12-18 四川大学 一种鲁棒的作物病害诊断系统
CN112287294A (zh) * 2020-09-10 2021-01-29 河海大学 一种基于深度学习的时空双向土壤含水量插值方法
CN112784906A (zh) * 2021-01-26 2021-05-11 中国科学院半导体研究所 基于多条件时间序列的农机监测数据清洗方法与装置
CN113159125A (zh) * 2021-03-16 2021-07-23 北京农业智能装备技术研究中心 一种施药方法及系统
CN113158754A (zh) * 2021-02-04 2021-07-23 安徽农业大学 一种番茄病害图像识别方法
CN113222211A (zh) * 2021-03-31 2021-08-06 中国科学技术大学先进技术研究院 一种多区域柴油车污染物排放因子预测方法及系统
CN113610327A (zh) * 2021-09-13 2021-11-05 贵州春芯科技有限公司 辣椒病害预测方法、装置、电子设备及存储介质
CN113887799A (zh) * 2021-09-29 2022-01-04 北京许继电气有限公司 一种水电厂安全生产人工智能报警方法
CN116310391A (zh) * 2023-05-18 2023-06-23 安徽大学 一种用于茶叶病害的识别方法
CN116541767A (zh) * 2023-07-05 2023-08-04 中国农业大学 一种基于图神经网络的多元温室环境参数预测方法及系统
CN117351293A (zh) * 2023-12-04 2024-01-05 天津医科大学口腔医院 一种联合学习的牙周病图像分类方法及设备
CN117852726A (zh) * 2024-03-07 2024-04-09 中国农业科学院烟草研究所(中国烟草总公司青州烟草研究所) 一种蔬菜病虫害暴发流行预测方法、介质及系统
CN117912645A (zh) * 2023-03-29 2024-04-19 安徽医科大学第一附属医院 一种基于物联网的血液保存全流程监管方法及系统

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111814866A (zh) * 2020-07-02 2020-10-23 深圳市万物云科技有限公司 一种病虫害预警方法、装置、计算机设备及存储介质
CN113436733B (zh) * 2021-08-26 2021-11-30 肾泰网健康科技(南京)有限公司 一种基于融合经验的血透方案生成模型的特征构造方法
CN114638445B (zh) * 2022-05-19 2022-07-29 北京佳格天地科技有限公司 作物病害预防的方法、装置、介质,及电子设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104598908A (zh) * 2014-09-26 2015-05-06 浙江理工大学 一种农作物叶部病害识别方法
CN106022467A (zh) * 2016-05-17 2016-10-12 浪潮集团有限公司 一种基于神经网络的农作物疾病检测系统
CN106682704A (zh) * 2017-01-20 2017-05-17 中国科学院合肥物质科学研究院 一种基于融合上下文信息的混合卷积神经网络的病害图像识别方法
CN106991428A (zh) * 2017-02-24 2017-07-28 中国科学院合肥物质科学研究院 基于自适应池化模型的害虫图像识别方法
CN106997475A (zh) * 2017-02-24 2017-08-01 中国科学院合肥物质科学研究院 一种基于并行卷积神经网络的害虫图像识别方法
CN109064460A (zh) * 2018-08-01 2018-12-21 中国科学院合肥物质科学研究院 基于多时序属性元素深度特征的小麦重度病害预测方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104598908A (zh) * 2014-09-26 2015-05-06 浙江理工大学 一种农作物叶部病害识别方法
CN106022467A (zh) * 2016-05-17 2016-10-12 浪潮集团有限公司 一种基于神经网络的农作物疾病检测系统
CN106682704A (zh) * 2017-01-20 2017-05-17 中国科学院合肥物质科学研究院 一种基于融合上下文信息的混合卷积神经网络的病害图像识别方法
CN106991428A (zh) * 2017-02-24 2017-07-28 中国科学院合肥物质科学研究院 基于自适应池化模型的害虫图像识别方法
CN106997475A (zh) * 2017-02-24 2017-08-01 中国科学院合肥物质科学研究院 一种基于并行卷积神经网络的害虫图像识别方法
CN109064460A (zh) * 2018-08-01 2018-12-21 中国科学院合肥物质科学研究院 基于多时序属性元素深度特征的小麦重度病害预测方法

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
XIE, CHENJUN ET AL.: "Multi-level Learning Features for Automatic Classification of Field Crop Pests", COMPUTERS AND ELECTRONICS IN AGRICULTURE, 20 July 2018 (2018-07-20), pages 233 - 241, XP055689655 *

Cited By (25)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111738138A (zh) * 2020-06-19 2020-10-02 安徽大学 一种耦合气象特征区域尺度的小麦条绣病严重度遥感监测方法
CN111738138B (zh) * 2020-06-19 2024-02-02 安徽大学 一种耦合气象特征区域尺度的小麦条绣病严重度遥感监测方法
CN111968003A (zh) * 2020-09-04 2020-11-20 郑州轻工业大学 一种基于农作物本体概念响应的农作物病害预测方法
CN111968003B (zh) * 2020-09-04 2023-11-24 郑州轻工业大学 一种基于农作物本体概念响应的农作物病害预测方法
CN112287294A (zh) * 2020-09-10 2021-01-29 河海大学 一种基于深度学习的时空双向土壤含水量插值方法
CN112287294B (zh) * 2020-09-10 2024-02-27 河海大学 一种基于深度学习的时空双向土壤含水量插值方法
CN112101265B (zh) * 2020-09-22 2023-04-25 四川大学 一种鲁棒的作物病害诊断系统
CN112101265A (zh) * 2020-09-22 2020-12-18 四川大学 一种鲁棒的作物病害诊断系统
CN112784906A (zh) * 2021-01-26 2021-05-11 中国科学院半导体研究所 基于多条件时间序列的农机监测数据清洗方法与装置
CN112784906B (zh) * 2021-01-26 2024-02-02 中国科学院半导体研究所 基于多条件时间序列的农机监测数据清洗方法与装置
CN113158754A (zh) * 2021-02-04 2021-07-23 安徽农业大学 一种番茄病害图像识别方法
CN113159125B (zh) * 2021-03-16 2024-04-05 北京农业智能装备技术研究中心 一种施药方法及系统
CN113159125A (zh) * 2021-03-16 2021-07-23 北京农业智能装备技术研究中心 一种施药方法及系统
CN113222211B (zh) * 2021-03-31 2023-12-12 中国科学技术大学先进技术研究院 一种多区域柴油车污染物排放因子预测方法及系统
CN113222211A (zh) * 2021-03-31 2021-08-06 中国科学技术大学先进技术研究院 一种多区域柴油车污染物排放因子预测方法及系统
CN113610327A (zh) * 2021-09-13 2021-11-05 贵州春芯科技有限公司 辣椒病害预测方法、装置、电子设备及存储介质
CN113887799A (zh) * 2021-09-29 2022-01-04 北京许继电气有限公司 一种水电厂安全生产人工智能报警方法
CN117912645A (zh) * 2023-03-29 2024-04-19 安徽医科大学第一附属医院 一种基于物联网的血液保存全流程监管方法及系统
CN116310391A (zh) * 2023-05-18 2023-06-23 安徽大学 一种用于茶叶病害的识别方法
CN116310391B (zh) * 2023-05-18 2023-08-15 安徽大学 一种用于茶叶病害的识别方法
CN116541767A (zh) * 2023-07-05 2023-08-04 中国农业大学 一种基于图神经网络的多元温室环境参数预测方法及系统
CN117351293A (zh) * 2023-12-04 2024-01-05 天津医科大学口腔医院 一种联合学习的牙周病图像分类方法及设备
CN117351293B (zh) * 2023-12-04 2024-02-06 天津医科大学口腔医院 一种联合学习的牙周病图像分类方法及设备
CN117852726A (zh) * 2024-03-07 2024-04-09 中国农业科学院烟草研究所(中国烟草总公司青州烟草研究所) 一种蔬菜病虫害暴发流行预测方法、介质及系统
CN117852726B (zh) * 2024-03-07 2024-06-04 中国农业科学院烟草研究所(中国烟草总公司青州烟草研究所) 一种蔬菜病虫害暴发流行预测方法、介质及系统

Also Published As

Publication number Publication date
CN110622182A (zh) 2019-12-27

Similar Documents

Publication Publication Date Title
WO2020047739A1 (zh) 基于多时序属性元素深度特征的小麦重度病害预测方法
Elmaz et al. CNN-LSTM architecture for predictive indoor temperature modeling
CN109816095B (zh) 基于改进门控循环神经网络的网络流量预测方法
Paras et al. A feature based neural network model for weather forecasting
CN107273782B (zh) 使用递归神经网络的在线动作检测
CN109064460B (zh) 基于多时序属性元素深度特征的小麦重度病害预测方法
CN111639787A (zh) 一种基于图卷积网络的时空数据预测方法
Tijskens et al. Neural networks for metamodelling the hygrothermal behaviour of building components
Jia et al. Physics guided recurrent neural networks for modeling dynamical systems: Application to monitoring water temperature and quality in lakes
CN110322009B (zh) 基于多层卷积长短时记忆神经网络的图像预测方法
CN107544904B (zh) 一种基于深度cg-lstm神经网络的软件可靠性预测方法
Dubois et al. Data-driven predictions of the Lorenz system
US11182676B2 (en) Cooperative neural network deep reinforcement learning with partial input assistance
CN108182260B (zh) 一种基于语义选择的多变量时间序列分类方法
Song et al. Air quality prediction based on LSTM-Kalman model
Haque et al. Crop yield prediction using deep neural network
Kişi et al. Modeling monthly pan evaporations using fuzzy genetic approach
Shaby The open-faced sandwich adjustment for MCMC using estimating functions
CN111198966A (zh) 基于多智能体边界感知网络的自然语言视频片段检索方法
Gharghory Deep network based on long short-term memory for time series prediction of microclimate data inside the greenhouse
Huang et al. A data-driven method for hybrid data assimilation with multilayer perceptron
Zhang et al. Autoregressive matrix factorization for imputation and forecasting of spatiotemporal structural monitoring time series
Anupam et al. Data driven modelling and simulation of wetland dynamics
CN116484904A (zh) 一种基于人工智能深度学习的监管数据处理实现方法
Wang et al. Modeling of a three-dimensional dynamic thermal field under grid-based sensor networks in grain storage

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 18932388

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 16.06.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18932388

Country of ref document: EP

Kind code of ref document: A1