WO2021098472A1 - 一种深度时空特征联合学习的农作物产量估测方法 - Google Patents

一种深度时空特征联合学习的农作物产量估测方法 Download PDF

Info

Publication number
WO2021098472A1
WO2021098472A1 PCT/CN2020/124870 CN2020124870W WO2021098472A1 WO 2021098472 A1 WO2021098472 A1 WO 2021098472A1 CN 2020124870 W CN2020124870 W CN 2020124870W WO 2021098472 A1 WO2021098472 A1 WO 2021098472A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
yield
term memory
crop
short
Prior art date
Application number
PCT/CN2020/124870
Other languages
English (en)
French (fr)
Inventor
林涛
钟仁海
徐金凡
江昊
应义斌
丁冠中
Original Assignee
浙江大学
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 浙江大学 filed Critical 浙江大学
Priority to US17/777,053 priority Critical patent/US11741555B2/en
Publication of WO2021098472A1 publication Critical patent/WO2021098472A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION 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
    • G06Q50/00—Systems or methods specially adapted for specific business sectors, e.g. utilities or tourism
    • G06Q50/02—Agriculture; Fishing; Mining
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01W—METEOROLOGY
    • G01W1/00—Meteorology
    • G01W1/02—Instruments for indicating weather conditions by measuring two or more variables, e.g. humidity, pressure, temperature, cloud cover or wind speed
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/24—Querying
    • G06F16/245—Query processing
    • G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
    • G06F16/2474—Sequence data queries, e.g. querying versioned data
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/29—Geographical information databases
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06Q—INFORMATION 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
    • G06Q10/00—Administration; Management
    • G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"

Definitions

  • the invention relates to the field of agrometeorology, in particular to a method for estimating crop yields based on joint learning of deep spatiotemporal features.
  • Constructing a crop yield estimation model is an important research method for quantitatively assessing the response of crop growth to changes in meteorological resources.
  • the impact of meteorological resources on crops has dynamic changes and cumulative effects in time series. Understanding these time series characteristics is beneficial to optimizing crop production decision-making; while meteorological resources have spatial heterogeneity in spatial distribution, which leads to the distribution of crop growth and yield. Differences in spatial distribution and affect the stability of the model. How to construct a deep learning model for joint learning of the temporal and spatial characteristics of crop growth and weather is the current technical difficulty and also a key breakthrough.
  • the current crop yield estimation model mainly relies on three approaches: (1) a physiological process-driven process mechanism model; (2) a statistical regression model; (3) a data-driven machine learning model.
  • Process mechanism models are difficult to apply in large spatial scales due to their over-parameterization and data requirements.
  • Statistical regression models are difficult to deal with nonlinear relationships and collinearity issues in the data.
  • the current machine learning methods only use machine learning models to analyze the data. Feature extraction and yield estimation tasks.
  • the prior art lacks a crop yield estimation method using machine learning, and a method that can combine spatio-temporal feature learning and spatial feature learning for accurate estimation.
  • the present invention provides a crop yield estimation method based on joint learning of deep spatiotemporal features.
  • the present invention adopts the following technical solutions, and the specific steps are as follows:
  • Step 1) Obtain and preprocess the historical crop yield data and meteorological data of the region, preprocess the meteorological data to obtain meteorological parameters, and preprocess the yield data to obtain de-trended yields, which are used as deep learning models for the spatiotemporal characteristics of subsequent crop yields.
  • Input and output data Obtain and preprocess the historical crop yield data and meteorological data of the region, preprocess the meteorological data to obtain meteorological parameters, and preprocess the yield data to obtain de-trended yields, which are used as deep learning models for the spatiotemporal characteristics of subsequent crop yields.
  • Step 2) Construct a deep learning model for the spatio-temporal characteristics of crop yields, and optimize the hyperparameters of the model;
  • Step 3) Take the meteorological parameters obtained in step 1) as input, and use the detrended yield obtained in step 1) as output to form training set samples to train the deep learning model of crop yield spatiotemporal characteristics, using Adam optimization method combined with back propagation method
  • the parameters of the model are obtained through training, and the optimal parameters are obtained after multiple rounds of training, and then the trained model is obtained;
  • Step 4) Input the meteorological parameters of the crop yield to be measured into the trained model, output the prediction result, and obtain the crop yield estimation result. In the specific implementation, it is also compared with the de-trending output, and processed to obtain the prediction effect of the model on the unknown sample.
  • the data preprocessing includes: constructing a yield-year unary linear regression equation based on historical crop yield data, detrending the crop yield data, and fitting the residuals obtained by fitting the yield-year unary linear regression equation As the de-trend yield, as the true value of the model output; extract the time series of crops sowed to mature meteorological parameters from the historical meteorological data and normalize them as the model input.
  • the meteorological parameters at each time during the corn growth period can be calculated according to historical meteorological data and normalized as the input of the model at each time, and time series iterative processing is performed during model training.
  • the historical meteorological data specifically includes daily maximum temperature, daily minimum temperature and daily average temperature in a fixed time period, and daily precipitation in a fixed time period.
  • the deep learning model of crop yield spatio-temporal characteristics is mainly composed of an input layer, a long and short-term memory neural network layer, an attention neural network layer, and a multi-task output layer connected in sequence;
  • the long and short-term memory neural network layer is composed of long and short-term memory neural units, and the meteorological parameter x t at each moment is input into the corresponding long and short-term memory group composed of three consecutively connected long and short-term memory neural units.
  • the first long-short-term memory neural unit in the long-short-term memory group corresponding to the meteorological parameter x t at all moments transmits and connects in turn along the time (the time corresponding to the long-short-term memory neural unit), sharing the parameters of the neural network; the weather at all moments
  • the second long and short-term memory neural unit in the long and short-term memory group corresponding to the parameter x t is transmitted and connected sequentially along time, sharing the parameters of the neural network; the meteorological parameter x t at all times corresponds to the third long and short-term memory group in the long and short-term memory group.
  • the short-term memory neural unit transmits and connects sequentially along time, sharing the parameters of the neural network; the meteorological parameters x t
  • the attention neural network layer adopts a fully connected neural network layer, inputs the hidden features h t extracted by the long and short-term memory neural network layer at each time t, and outputs the attention value ⁇ t at each time, and then according to the attention value ⁇ t hidden features hidden wherein h t of the time series according to the following formula to give attention hidden features weighting ⁇ t h t, give attention to the weighted sum of all combined into a sequence of feature vectors H:
  • W A and b A sub-tables represent the learnable weight matrix and deviation vector of the attention neural network layer
  • softmax() is the activation function, and the range of values is mapped to the (0,1) interval
  • the multi-task output layer constructs a geographic region-specific output layer based on spatial differences, and outputs the crop yield y in the corresponding geographic region r.
  • the processing is as follows:
  • r represents the index of the corresponding geographic area
  • w r and b r respectively represent the learnable weight matrix and deviation vector of the multi-task output layer corresponding to the geographic area r.
  • Long and short-term memory neural network layer used to process time series data and extract time features
  • Attention neural network layer used to quantify the importance of each time sequence, and assign weights to temporal features
  • the multi-task output layer is used to learn space-specific features and output the estimated crop yield value.
  • the multi-task output layer is divided into different regions according to the geographic area, and a fully connected output layer specific to the geographic region is constructed without an activation function.
  • the model hyperparameter optimization includes the number of layers of the long and short-term memory neural network, the number of layers of the attention neural network, the number of layers and the number of tasks of the multi-task output layer, the hidden feature dimensions of each layer, and the model training learning rate . After cross-validation, the optimal model hyperparameters are selected.
  • the number of layers of the long and short-term memory neural network is 3, and the hidden feature dimension is 32; the attention neural network layer structure is a fully connected neural network, and the activation function is softmax; the multi-task output layer is divided into different according to the research area Task, to construct a region-specific fully connected output layer without activation function.
  • the training set data is input into the model, and the Adam method combined with backpropagation algorithm training is used to obtain the weight of the model, specifically to obtain the parameters of W A , b A , W r , and b r until the training set The loss function value converges to the lowest value.
  • the output of the model is used to estimate the crop yield of the crop sample, for example, the yield per unit area of corn, wheat, and rice.
  • the time series of meteorological parameters of each region can be input, and the crop yield of each region can be estimated through the model output.
  • the invention is based on a deep learning method, learns temporal and spatial features from historical meteorological data and crop yield data, improves crop yield estimation accuracy at large spatial scales; extracts data through a long and short-term memory neural network embedded with an attention mechanism Then use the multi-task learning method to construct a region-specific output layer to learn the spatial-specific features of each region and output the estimated crop yield.
  • the present invention is based on the long and short-term memory neural network, the attention mechanism and the multi-task learning method, constructs a deep learning model framework that simultaneously learns the temporal characteristics and spatial specific characteristics of the crop growth process, and realizes the joint learning of temporal and spatial characteristics , Improve the accuracy and stability of the model's estimation of crop yield.
  • the embedded attention mechanism can visualize the process of the model's processing of time series features, which is conducive to judging the key growth period to assist decision-making.
  • the invention extracts the specific spatial characteristics of each region through a multi-task learning method. In the case of large spatial differences in the study area, the crop yield estimation accuracy of the invention is higher and the stability is better.
  • the present invention combines temporal feature learning and spatial feature learning to realize the estimation of crop yield.
  • the crop yield estimation accuracy of the present invention is higher and more stable. better.
  • Figure 1 is a modeling flow chart of the method of the present invention
  • Figure 2 is a schematic diagram of the model structure of the method of the present invention.
  • Fig. 3 is a box diagram drawn by visualizing the attention value of the embodiment.
  • This embodiment is applied to the quantitative estimation of corn yield at the county level.
  • the selected research area is the Corn Belt region of the United States, including county-level data of 11 states: Minnesota (MN), Wisconsin (WI), Michigan (MI), Wyoming (NE), Iowa (IA), Illinois (IL) , Indiana (IN), Ohio (OH), Kansas (KS), Missouri (MO) and Kentucky (KY).
  • the data used are county-level corn production and meteorological data from 1981 to 2016, all from public data sets.
  • the meteorological indicators selected in this embodiment include: effective accumulated temperature indicator Growing Degree Days (GDD), high temperature stress indicator Killing Degree Days (KDD), and accumulated rainfall PRCP.
  • GDD Geographical temperature indicator
  • KDD high temperature stress indicator Killing Degree Days
  • accumulated rainfall PRCP The corn production and meteorological indicators of a county each year are the same, with a total of 34,403 samples.
  • the samples from 1981-2014 are used as the training set, a total of 32778; the samples from 2015-2016 are used as the test set, a total of 1625.
  • the minimum detrended yield was -7.63 tons/ha
  • the maximum detrended yield was 4.13 tons/ha
  • the average detrended yield was -0.01 tons/ha
  • the standard deviation was 1.32 tons/ha.
  • the structure of the model is shown in Figure 2, which specifically includes: three layers of long and short-term memory neural network layer, attention neural network layer and multi-task output layer.
  • the time sequence of the model is 20 weeks from the start of sowing to the maturity of the corn, and the hidden feature dimension of the three-layer long and short-term memory neural network is set to 32.
  • the structure of the attention neural network layer is a single-layer fully connected neural network, which takes the hidden features extracted by the long and short-term memory neural network at each time sequence as input, and outputs the attention value of each time sequence, which is used to quantify the attention of the model to each time sequence. degree.
  • the output layer is three region-specific output layers, which output the estimated crop yield values of samples in their respective regions.
  • the optimized hyperparameters contained in it include: the number of layers of the long and short-term memory network (3 layers) and hidden feature dimensions (32), the number of layers of the attention neural network (1 layer), and the number of layers of the multitasking output layer (1) And the number of tasks (3), and the learning rate is set to 0.00001.

Abstract

一种深度时空特征联合学习的农作物产量估测方法。获取地区的历史作物产量数据和气象数据并进行预处理，对气象数据进行预处理获得气象参数，对产量数据进行预处理获得去趋势产量，分别作为后续作物单产时空特征深度学习模型的输入和输出数据；构建作物单产时空特征深度学习模型，并对模型的超参数进行优化；将气象参数作为输入，将去趋势产量作为输出，形成训练集样本进而训练作物单产时空特征深度学习模型获得模型的参数，将待测作物产量的气象参数输入训练后的模型，输出估测结果，获得农作物产量估测结果。该方法联合了时间特征学习和空间特征学习，在空间差异较大且复杂的研究区域内，该方法的作物产量估测精度更高，稳定性更好。

Description

一种深度时空特征联合学习的农作物产量估测方法 技术领域
本发明涉及农业气象领域，具体涉及一种深度时空特征联合学习的农作物产量估测方法。
背景技术
构建作物产量估测模型，是当前量化评估作物生长对于气象资源变化的响应的重要研究方法。气象资源对作物的影响存在时间序列上的动态变化和累积效应，理解这些时序特征有利于优化作物生产决策；而气象资源在空间分布上存在空间异质性，进而导致作物生长情况和产量分布在空间上的分布差异，并影响模型的稳定性。如何构建一个深度学习模型对作物生长与气象关联的时序特征与空间特征进行联合学习是目前的技术难点，也是关键的突破口。
当前作物产量估测模型主要依赖于三种途径：(1)生理过程驱动的过程机理模型；(2)统计回归模型；(3)数据驱动的机器学习模型。过程机理模型因其过于参数化和数据要求难以在大空间尺度范围下应用，统计回归模型难以处理数据中的非线性关系和共线性问题，而当前的机器学习方法仅是使用机器学习模型对数据特征进行提取，进行产量估测任务，现有技术中缺少了利用机器学习的作物产量估测方法，更缺少了能够联合利用时空特征学习与空间特征学习进行准确估测的方法。
发明内容
为了解决现有技术中存在的不足，本发明提供了一种深度时空特征联合学习的农作物产量估测方法。
本发明采用如下技术方案，具体步骤如下：
步骤1)：获取地区的历史作物产量数据和气象数据并进行预处理，对气象数据进行预处理获得气象参数，对产量数据进行预处理获得去趋势产量，分别作为后续作物单产时空特征深度学习模型的输入和输出数据；
步骤2)：构建作物单产时空特征深度学习模型，并对模型的超参数进行优化；
步骤3)：将步骤1)获得的气象参数作为输入，将步骤1)获得的去趋势产量作为输出，形成训练集样本进而训练作物单产时空特征深度学习模型，采用Adam优化方法结合反向传播方法训练获得模型的参数，经过多轮训练后得到最优参数，进而获得训练后的模型；
步骤4)：将待测作物产量的气象参数输入训练后的模型，输出预测结果，获得农作物产量估测结果。具体实施中，还与去趋势产量对比，处理获得模型对未知样本的预测效果。
所述的步骤1)中，数据预处理包括：根据历史作物产量数据构建产量-年份一元线性回归方程，对作物产量数据进行去趋势处理，将产量-年份一元线性回归方程拟合得到的残差作为去趋势产量，作为模型输出的真实值；从历史的气象数据中提取时间序列的作物播种至成熟的气象参数并进行归一化处理作为模型输入。
具体实施中可根据历史气象数据，计算玉米生育期内各时刻的气象参数并进行归一化处理作为模型各时刻的输入，在模型训练时进行时序迭代处理。
所述历史气象数据具体包括固定时间段内的日最高温度、日最低温度和日平均温度以及固定时间段内的日降水量。
所述步骤2)中，如图2所示，作物单产时空特征深度学习模型主要由输入层、长短期记忆神经网络层、注意力神经网络层和多任务输出层依次连接构成；
所述的输入层中输入作物生长期间内的气象参数时间序列，并采用min-max归一化处理将各气象参数值域变换为[0,1]；
所述的长短期记忆神经网络层由长短期记忆神经单元组成，每个时刻的气象参数x t输入到各自对应的由三个连续依次连接的长短期记忆神经单元组成的长短期记忆组中，并且所有时刻的气象参数x t所对应的长短期记忆组中第一个长短期记忆神经单元沿时间(长短期记忆神经单元对应的时间)依次传递连接，共享神经网络的参数；所有时刻的气象参数x t所对应的长短期记忆组中第二个长短期记忆神经单元沿时间依次传递连接，共享神经网络的参数；所有时刻的气象参数x t所对应的长短期记忆组中第三个长短期记忆神经单元沿时间依次传递连接，共享神经网络的参数；最后各个时刻的气象参数x t经长短期记忆组处理输出各自的隐层特征h t；
所述的注意力神经网络层采用一层全连接神经网络层，输入长短期记忆神经网络层在各时刻t提取的隐藏特征h t，输出各时刻的注意力值α t，之后根据注意力值α t对整个时间序列的隐藏特征h t按照以下公式进行处理，得到注意力加权的隐藏特征α th t，得到所有时序的注意力加权的隐藏特征合并成一特征向量H：
α t＝softmax(W A*h t+b A)
H＝α 1h 1+α 2h 2+…+α th t
其中，W A和b A分表表示注意力神经网络层的可学习的权重矩阵和偏差向量，softmax()是激活函数，将值的范围映射到(0,1)区间内；
所述的多任务输出层根据空间差异构建地理区域特异的输出层，输出对应地理区域r内的作物产量y，处理如以下公式：
y＝W r*H+b r
其中，r表示对应地理区域的索引，w r和b r分别表示对应于地理区域r的多任务输出层的可学习的权重矩阵和偏差向量。
长短期记忆神经网络层，用于处理时间序列数据，提取时间特征；
注意力神经网络层，用于量化各时序的重要性，对时间特征进行权值分配；
多任务输出层，用于学习空间特异的特征，输出估测的作物产量值。多任务输出层根据地理区域划分成不同，构地理建区域特异的全连接输出层，无激活函数。
所述步骤2)中，模型超参数优化包括长短期记忆神经网络层的层数、注意力神经网络层数、多任务输出层的层数及任务数目、各层隐藏特征维度和模型训练学习率。在经过交叉验证后，选取最优的模型超参数。
具体实施中，长短期记忆神经网络层层数为3，隐藏特征维度为32；注意力神经网络层结构为一层全连接神经网络，激活函数为softmax；多任务输出层根据研究区域划分成不同任务，构建区域特异的全连接输出层，无激活函数。
所述步骤3)中，将训练集数据输入到模型中，采用Adam方法结合反向传播算法训练获得模型的权重，具体是获得W A、b A、W r、b r的参数，直至训练集损失函数值收敛至最低值。
所述步骤4)中，通过模型输出估测农作物样本的作物产量，例如可以是玉米、小麦、水稻的单位面积产量。
所述步骤4)中，可以输入各地区的气象参数时间序列，通过模型输出估测各地区的作物产量。
本发明基于深度学习方法，从历史气象数据和作物产量数据中学习时间特征和空间特征，提高了大空间尺度下的作物产量估测精度；通过嵌入了注意力机制的长短期记忆神经网络提取数据中的时序特征，进而采用多任务学习方法构建区域特异的输出层来学习各区域的空间特异性特征，并输出估测的作物产量。
本发明的有益效果是：
本发明基于长短期记忆神经网络、注意力机制和多任务学习方法，构建了 同时学习作物生长过程中的时序特征和空间特异性特征的深度学习模型框架，实现了时间特征和空间特征的联合学习，提高了模型对作物产量估测的精度和稳定性。嵌入的注意力机制可以可视化模型对时序特征的处理过程，有利于判断关键生长时期以辅助决策。本发明通过多任务学习方法提取各区域特异的空间特征，在研究区域空间差异大的情况下，本发明的作物单产估测精度更高，稳定性更好。
综合来说，本发明联合了时间特征学习和空间特征学习用于实现了农作物产量的估测，在空间差异较大且复杂的研究区域内，本发明的作物产量估测精度更高，稳定性更好。
附图说明
图1为本发明方法的建模流程图；
图2为本发明方法的模型结构示意图；
图3为将实施例的注意力值可视化绘制出的箱体图。
具体实施方式
为更好理解本发明，下面结合实施例对本发明做进一步详细说明，但本发明要求保护的范围并不局限于实施例表示的范围。以下进行的实施例，在Python软件上运行。下面结合附图和实施例对本发明做进一步说明。
本实施例应用于县级玉米产量的定量估测。所选取的研究区域是美国玉米带地区，一共包括11个州的县级数据：Minnesota(MN),Wisconsin(WI),Michigan(MI),Nebraska(NE),Iowa(IA),Illinois(IL),Indiana(IN),Ohio(OH),Kansas(KS),Missouri(MO)和Kentucky(KY)。采用数据为1981-2016年县级玉米产量和气象数据，均来自公开数据集。本实施例所选择的气象指标包括：有效积温指标Growing Degree Days(GDD),高温胁迫指标Killing Degree Days(KDD)和累积降雨量PRCP。每一年一个县的玉米产量及气象指标为一样本，共有34403个样本。根据年份，将1981-2014年的样本作为训练集，共32778个；将2015-2016年的样本作为测试集，共1625个。所有样本中，最小去趋势产量为-7.63吨/公顷，最大去趋势产量为4.13吨/公顷，平均去趋势产量为-0.01吨/公顷，标准差为1.32吨/公顷。
如图1所示步骤，实施例过程如下：
1)将训练集数据输入到模型中。
2)优化模型的超参数，模型的结构如图2所示，具体包括：三层长短期记忆神经网络层、注意力神经网络层和多任务输出层。模型的时序为从播种开始到玉米成熟的20周，三层长短期记忆神经网络的隐藏特征维度设置为32。注意 力神经网络层的结构为单层全连接神经网络，以长短期记忆神经网络在各时序所提取的隐藏特征作为输入，输出各时序的注意力值，用于量化表征模型对各时序的关注程度。输出层为三个区域特异的输出层，输出各自区域内样本的估测作物产量值。其中含有的优化过后的超参数包括：长短期记忆网络的层数(3层)和隐藏特征维度(32)、注意力神经网络层数(1层)、多任务输出层的层数(1)和任务数(3)，以及学习率设置为0.00001。
3)模型的超参数优化完毕后，再次使用所有的训练集数据来训练模型的参数。通过反向传播算法在训练集上训练模型的参数，模型训练迭代直至训练集损失函数收敛，从初始化至训练完成共15830轮。
4)模型训练好后，终止训练，保存最优的模型，并在测试集上测试模型的效果。根据测试集各样本的预测值和真实值的RMSE评估模型精度。
5)通过训练集数据训练传统的LASSO模型和Random Forest模型，并在测试集上测试，得到的结果是LASSO模型的RMSE是1.16吨/公顷，Random Forest模型的RMSE是1.07吨/公顷，而通过本发明提出的方法的RMSE是0.87吨/公顷。通过比较可以看出，该方法在作物产量估测的精度上高于传统方法。
6)并且，通过将注意力网络计算得到的注意力值α t可视化，绘制出每个州的注意力值在各时刻的箱体图，如图3所示。可以发现模型提取到了一种时间维度上的特征，且这种特征在各州之间表现一致性。从第1周到第20周，注意力值不断增大，代表模型对后面时刻更加关注。此结果表明嵌入的注意力机制可以可视化模型对时序特征的处理过程，提高模型的可解释性。

Claims (6)

  1. 一种深度时空特征联合学习的农作物产量估测方法，其特征在于该方法包含如下步骤：
    步骤1)：获取地区的历史作物产量数据和气象数据并进行预处理，对气象数据进行预处理获得气象参数，对产量数据进行预处理获得去趋势产量，分别作为后续作物单产时空特征深度学习模型的输入和输出数据；
    步骤2)：构建作物单产时空特征深度学习模型，并对模型的超参数进行优化；
    步骤3)：将步骤1)获得的气象参数作为输入，将步骤1)获得的去趋势产量作为输出，形成训练集样本进而训练作物单产时空特征深度学习模型，采用Adam优化方法结合反向传播方法训练获得模型的参数，经过多轮训练后得到最优参数，进而获得训练后的模型；
    步骤4)：将待测作物产量的气象参数输入训练后的模型，输出预测结果，获得农作物产量估测结果。
  2. 根据权利要求1所述的一种深度时空特征联合学习的农作物产量估测方法，其特征在于：所述的步骤1)中，数据预处理包括：根据历史作物产量数据构建产量-年份一元线性回归方程，对作物产量数据进行去趋势处理，将产量-年份一元线性回归方程拟合得到的残差作为去趋势产量，作为模型输出的真实值；从历史的气象数据中提取时间序列的作物播种至成熟的气象参数并进行归一化处理作为模型输入。
  3. 根据权利要求1所述的一种深度时空特征联合学习的农作物产量估测方法，其特征在于：所述历史气象数据具体包括固定时间段内的日最高温度、日最低温度和日平均温度以及固定时间段内的日降水量。
  4. 根据权利要求1所述的一种深度时空特征联合学习的农作物产量估测方法，其特征在于：所述步骤2)中，作物单产时空特征深度学习模型主要由输入层、长短期记忆神经网络层、注意力神经网络层和多任务输出层依次连接构成；
    所述的输入层中输入作物生长期间内的气象参数时间序列，并采用min-max归一化处理将各气象参数值域变换为[0,1]；
    所述的长短期记忆神经网络层由长短期记忆神经单元组成，每个时刻的气象参数x t输入到各自对应的由三个连续依次连接的长短期记忆神经单元组成的长短期记忆组中，并且所有时刻的气象参数x t所对应的长短期记忆组中第一个长短期记忆神经单元沿时间依次传递连接，共享神经网络的参数；所有时刻的 气象参数x t所对应的长短期记忆组中第二个长短期记忆神经单元沿时间依次传递连接，共享神经网络的参数；所有时刻的气象参数x t所对应的长短期记忆组中第三个长短期记忆神经单元沿时间依次传递连接，共享神经网络的参数；最后各个时刻的气象参数x t经长短期记忆组处理输出各自的隐层特征h t；
    所述的注意力神经网络层采用一层全连接神经网络层，输入长短期记忆神经网络层在各时刻t提取的隐藏特征h t，输出各时刻的注意力值α t，之后根据注意力值α t对整个时间序列的隐藏特征h t按照以下公式进行处理，得到注意力加权的隐藏特征α th t，得到所有时序的注意力加权的隐藏特征合并成一特征向量H：
    α t＝softmax(W A*h t+b A)
    H＝α 1h 1+α 2h 2+…+α th t
    其中，W A和b A分表表示注意力神经网络层的可学习的权重矩阵和偏差向量，softmax()是激活函数；
    所述的多任务输出层输出对应地理区域r内的作物产量y，处理如以下公式：
    y＝W r*H+b r
    其中，r表示对应地理区域的索引，w r和b r分别表示对应于地理区域r的多任务输出层的可学习的权重矩阵和偏差向量。
  5. 根据权利要求1所述的一种深度时空特征联合学习的农作物产量估测方法，其特征在于：所述步骤2)中，模型超参数优化包括长短期记忆神经网络层的层数、注意力神经网络层数、多任务输出层的层数及任务数目、各层隐藏特征维度和模型训练学习率。
  6. 根据权利要求1所述的一种深度时空特征联合学习的农作物产量估测方法，其特征在于：所述步骤4)中，通过模型输出估测农作物样本的作物产量。
PCT/CN2020/124870 2019-11-19 2020-10-29 一种深度时空特征联合学习的农作物产量估测方法 WO2021098472A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/777,053 US11741555B2 (en) 2019-11-19 2020-10-29 Crop yield estimation method based on deep temporal and spatial feature combined learning

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201911134971.4 2019-11-19
CN201911134971.4A CN111027752B (zh) 2019-11-19 2019-11-19 一种深度时空特征联合学习的农作物产量估测方法

Publications (1)

Publication Number Publication Date
WO2021098472A1 true WO2021098472A1 (zh) 2021-05-27

Family

ID=70200622

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2020/124870 WO2021098472A1 (zh) 2019-11-19 2020-10-29 一种深度时空特征联合学习的农作物产量估测方法

Country Status (3)

Country Link
US (1) US11741555B2 (zh)
CN (1) CN111027752B (zh)
WO (1) WO2021098472A1 (zh)

Cited By (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420513A (zh) * 2021-07-01 2021-09-21 西北工业大学 一种基于深度学习的水下圆柱体扰流分区流场预测方法
CN113836783A (zh) * 2021-07-16 2021-12-24 东南大学 斜拉桥主梁温致挠度监测基准值的数字回归模型建模方法
CN114994294A (zh) * 2022-05-20 2022-09-02 燕山大学 基于注意力与窗口选通机制的水泥熟料游离钙软测量方法
CN115237896A (zh) * 2022-07-12 2022-10-25 四川大学 一种基于深度学习预报空气质量的数据前处理方法及系统
CN117556956A (zh) * 2023-11-24 2024-02-13 中国科学院空天信息创新研究院 一种水稻单位面积产量估计方法及装置

Families Citing this family (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111027752B (zh) * 2019-11-19 2022-06-21 浙江大学 一种深度时空特征联合学习的农作物产量估测方法
CN112183575A (zh) * 2020-08-24 2021-01-05 盐城工学院 一种基于区域气象参数预测土豆产量的方法
CN113052433B (zh) * 2021-02-22 2024-02-06 中国科学院空天信息创新研究院 基于关键时相和农田景观特征参量的作物单产估算方法
CN113255999B (zh) * 2021-05-25 2024-04-12 中国农业科学院农业信息研究所 基于多模型的农作物产量评估方法、装置、设备及介质
CN114925947B (zh) * 2022-03-04 2023-04-07 武汉大学 一种物候自适应的作物生理指标深度学习估测方法及系统
CN114894619B (zh) * 2022-05-10 2024-04-05 中南大学 基于长短期记忆网络的钢管混凝土柱轴压荷载-应变曲线预测方法
CN115564145B (zh) * 2022-11-17 2023-06-30 北京市农林科学院信息技术研究中心 基于分布式数据的农作物品种产量预测方法及装置
CN115829162B (zh) * 2023-01-29 2023-05-26 北京市农林科学院信息技术研究中心 作物产量预测方法、装置、电子设备及介质
CN115860269B (zh) * 2023-02-20 2023-05-16 南京信息工程大学 一种基于三重注意力机制的农作物产量预测方法
CN117114190A (zh) * 2023-08-29 2023-11-24 华中科技大学 一种基于混合深度学习的河道径流预测方法及设备
CN116842351B (zh) * 2023-09-01 2023-11-10 山东省海洋资源与环境研究院(山东省海洋环境监测中心、山东省水产品质量检验中心) 滨海湿地碳汇评估模型构建方法、评估方法和电子设备
CN117133129B (zh) * 2023-10-26 2024-01-30 山东科技大学 一种基于多分量注意力图神经网络的交通速度预测方法
CN117471575B (zh) * 2023-12-28 2024-03-08 河海大学 基于bo-lstm神经网络模型的台风浪波高预报方法
CN117575111B (zh) * 2024-01-16 2024-04-12 安徽农业大学 一种基于迁移学习的农业遥感图像时空序列预测方法

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886496A (zh) * 2019-02-27 2019-06-14 中南大学 一种基于气象信息的农产量预测方法
CN110443420A (zh) * 2019-08-05 2019-11-12 山东农业大学 一种基于机器学习的作物产量预测方法
CN111027752A (zh) * 2019-11-19 2020-04-17 浙江大学 一种深度时空特征联合学习的农作物产量估测方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104089911B (zh) * 2014-06-27 2016-05-25 桂林电子科技大学 基于一元线性回归的光谱模型传递方法
US20180284741A1 (en) * 2016-05-09 2018-10-04 StrongForce IoT Portfolio 2016, LLC Methods and systems for industrial internet of things data collection for a chemical production process
US11327475B2 (en) * 2016-05-09 2022-05-10 Strong Force Iot Portfolio 2016, Llc Methods and systems for intelligent collection and analysis of vehicle data
EP3273387A1 (en) * 2016-07-19 2018-01-24 Siemens Healthcare GmbH Medical image segmentation with a multi-task neural network system
WO2018187632A1 (en) * 2017-04-05 2018-10-11 Carnegie Mellon University Deep learning methods for estimating density and/or flow of objects, and related methods and software
US11442445B2 (en) * 2017-08-02 2022-09-13 Strong Force Iot Portfolio 2016, Llc Data collection systems and methods with alternate routing of input channels
CN209085657U (zh) * 2017-08-02 2019-07-09 强力物联网投资组合2016有限公司 用于与化工生产工艺有关的或工业环境的数据收集系统
US11592594B2 (en) * 2020-04-13 2023-02-28 X Development Llc Subsurface lithological model with machine learning
US11625627B2 (en) * 2020-04-21 2023-04-11 Microsoft Technology Licensing, Llc Predicted forecast offset from remote location sensor

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109886496A (zh) * 2019-02-27 2019-06-14 中南大学 一种基于气象信息的农产量预测方法
CN110443420A (zh) * 2019-08-05 2019-11-12 山东农业大学 一种基于机器学习的作物产量预测方法
CN111027752A (zh) * 2019-11-19 2020-04-17 浙江大学 一种深度时空特征联合学习的农作物产量估测方法

Cited By (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113420513A (zh) * 2021-07-01 2021-09-21 西北工业大学 一种基于深度学习的水下圆柱体扰流分区流场预测方法
CN113420513B (zh) * 2021-07-01 2023-03-07 西北工业大学 一种基于深度学习的水下圆柱体扰流分区流场预测方法
CN113836783A (zh) * 2021-07-16 2021-12-24 东南大学 斜拉桥主梁温致挠度监测基准值的数字回归模型建模方法
CN113836783B (zh) * 2021-07-16 2022-11-18 东南大学 斜拉桥主梁温致挠度监测基准值的数字回归模型建模方法
CN114994294A (zh) * 2022-05-20 2022-09-02 燕山大学 基于注意力与窗口选通机制的水泥熟料游离钙软测量方法
CN115237896A (zh) * 2022-07-12 2022-10-25 四川大学 一种基于深度学习预报空气质量的数据前处理方法及系统
CN115237896B (zh) * 2022-07-12 2023-07-11 四川大学 一种基于深度学习预报空气质量的数据前处理方法及系统
CN117556956A (zh) * 2023-11-24 2024-02-13 中国科学院空天信息创新研究院 一种水稻单位面积产量估计方法及装置

Also Published As

Publication number Publication date
CN111027752A (zh) 2020-04-17
US11741555B2 (en) 2023-08-29
US20220405864A1 (en) 2022-12-22
CN111027752B (zh) 2022-06-21

Similar Documents

Publication Publication Date Title
WO2021098472A1 (zh) 一种深度时空特征联合学习的农作物产量估测方法
CN106815971B (zh) 一种基于rbf神经网络的滑坡灾害预报方法
Tian et al. An IPSO-BP neural network for estimating wheat yield using two remotely sensed variables in the Guanzhong Plain, PR China
CN102183621B (zh) 水产养殖溶解氧浓度在线预测方法及系统
CN110109193B (zh) 一种基于drnn神经网络的茄子温室温度智能化检测装置
CN109002917A (zh) 基于lstm神经网络的粮食总产量多维时间序列预测方法
CN107168402A (zh) 基于can现场总线的鸡舍环境温度智能监测系统
CN104523264B (zh) 一种心电信号处理方法
WO2018014658A1 (zh) 一种柔鱼类的中心渔场预测方法
CN107085732A (zh) 基于无线传感器网络的牛舍环境氨气智能监测系统
CN107945182A (zh) 基于卷积神经网络模型GoogleNet的玉米叶片病害识别方法
CN110119767B (zh) 一种基于lvq神经网络的黄瓜温室温度智能化检测装置
CN110069032A (zh) 一种基于小波神经网络的茄子温室环境智能检测系统
Tagarakis et al. A fuzzy inference system to model grape quality in vineyards
CN110119086A (zh) 一种基于anfis神经网络的番茄温室环境参数智能监测装置
CN112116145A (zh) 基于遗传算法优化bp神经网络的奶牛采食量评估方法
CN113159439A (zh) 一种作物产量的预测方法、系统、存储介质和电子设备
CN114529097A (zh) 多尺度农作物物候期遥感降维预测方法
CN115205695A (zh) 一种根据种植数据确定种植策略的方法和系统
Aubakirova et al. APPLICATION OF ARTIFICIAL NEURAL NETWORK FOR WHEAT YIELD FORECASTING.
Yang et al. Prediction of corn variety yield with attribute-missing data via graph neural network
Cheng et al. Research on prediction method based on ARIMA-BP combination model
CN113537645B (zh) 基于机器学习融合卫星与天气数据的大豆产量预测方法
Geetha et al. A time-series based yield forecasting model using stacked lstm to predict the yield of paddy in cauvery delta zone in tamilnadu
CN115064276A (zh) 一种使用自注意力机制的疾病预测方法

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: 20890494

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 20890494

Country of ref document: EP

Kind code of ref document: A1