Crop yield estimation method and system based on hierarchical recognition and weight decision fusion

Abstract

The application discloses a crop yield estimation method and system based on hierarchical identification and weight decision fusion, and relates to the technical field of crop yield estimation. The method comprises the following steps: inputting a visible light image of a target mature period crop field block obtained to a yield grading model to obtain yield grade output values corresponding to yield grade values of the visible light image of the target mature period crop field block; the yield grade values are determined according to yield grade rules; confidence score values corresponding to each yield grade output value are calculated, the confidence score values are arranged in descending order, then the first m confidence score values are selected and normalized to obtain m yield grade weight values; the m yield grade weight values are multiplied by corresponding yield grade values respectively and accumulated to obtain a yield estimation value of the target mature period crop field block; and the application can accurately and quickly predict the yield value of the mature period crop field block.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of crop yield estimation, in particular to a crop yield estimation method and system based on hierarchical identification and weight decision fusion. BACKGROUND

[0002] Crop yield estimation is of great significance to food production and security. Crop yield estimation helps to grasp the production status of crops in time and guide agricultural production decisions. Traditional crop yield estimation methods are usually field sampling survey methods, which calculate yield estimates according to empirical formulas after sampling. This method has large errors and is time-consuming and labor-intensive.

[0003] With the development of unmanned aerial vehicle remote sensing technology, the existing crop yield estimation method has the following operation process: first, multi-spectral / hyperspectral images of crop field areas are collected by using unmanned aerial vehicles, then a plurality of vegetation indices are extracted based on the multi-spectral / hyperspectral images, then regression modeling is performed according to the plurality of extracted vegetation indices, and finally the established model is used to predict crop yield. This method is relatively complex in processing unmanned aerial vehicle images, and the average absolute percentage error of the yield estimation results of the current better model is still high, which needs to be further improved. SUMMARY

[0004] The purpose of the present application is to provide a crop yield estimation method and system based on hierarchical identification and weight decision fusion, which can accurately and quickly predict the yield value of a mature crop field.

[0005] To achieve the above purpose, the present application provides the following scheme:

[0006] In a first aspect, a crop yield estimation method based on hierarchical identification and weight decision fusion comprises:

[0007] Obtaining a visible light image of a target mature crop field;

[0008] Inputting the visible light image of the target mature crop field into a yield classification model to obtain yield level output values corresponding to yield level values of the visible light image of the target mature crop field; the yield level values are determined according to yield level rules; the yield level rules are that the interval of adjacent yield level values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield level values is not less than a set number;

[0009] Normalizing each yield level output value to obtain a confidence score value corresponding to each yield level output value, and arranging the confidence score values in descending order;

[0010] In the descendingly arranged confidence score values, the first m confidence score values are selected and normalized to obtain m yield level weight values;

[0011] multiplying the m yield level weight values with the corresponding yield level values respectively and accumulating, to obtain a yield estimation value of the target mature period crop field;

[0012] The yield grading model is obtained by training a deep convolutional neural network according to sample data; the sample data includes sample input data and corresponding label data; the sample input data is a sample mature period crop field image; and the label data is a yield level value of the sample mature period crop field image.

[0013] Optionally, the determination process of the sample data is as follows:

[0014] A visible light sequence photo of a sample mature period crop field with a known crop yield and a height not higher than a set value is collected by using a remote sensing device, and the visible light sequence photo of the sample mature period crop field is transmitted to a computer and saved to the computer in the form of a visible light sequence image;

[0015] The visible light sequence image is spliced and orthorectified according to image features to obtain a preprocessed visible light sequence image; the image features at least include photo heading features, front-back repeating features and side repeating features;

[0016] The preprocessed visible light sequence image is cropped according to different crop yields and boundaries of the sample mature period crop field, so as to obtain initial sample mature period crop field images corresponding to different known crop yields;

[0017] A yield grading value of the initial sample mature period crop field image is determined according to a crop yield corresponding to the initial sample mature period crop field image and a yield level rule;

[0018] The initial sample mature period crop field images with different known yield grading values are randomly sampled in equal size and without repetition to obtain sampled sample mature period crop field images and corresponding yield level values;

[0019] The sampled sample mature period crop field images and the corresponding yield level values are subjected to data format conversion, normalization and label assignment processing to obtain final sample mature period crop field images and corresponding label data.

[0020] Optionally, the yield level rule is that an interval between adjacent yield level values is not greater than 10% of an average value of all known sample crop yields, and the number of yield level values is not less than 10.

[0021] Optionally, the visible light image of the target mature period crop field block is input into a yield grading model to obtain yield grade output values corresponding to yield grade values of the visible light image of the target mature period crop field block, and the method specifically comprises the following steps:

[0022] The visible light image of the target mature period crop field block is spliced and ortho-rectified according to image features to obtain a preprocessed visible light image, and the image features at least include photo heading features, forward-backward repeating features and side repeating features.

[0023] The preprocessed visible light image is cropped according to the boundary of the target mature period crop field block to obtain a plurality of initial target mature period crop field images.

[0024] The plurality of initial target mature period crop field images are randomly sampled in equal size and without repetition to obtain sampled target mature period crop field images.

[0025] Data format conversion is performed on each of the sampled target mature period crop field images to obtain a plurality of data format converted target mature period crop field images.

[0026] The data format converted target mature period crop field images are input into a yield grading model to obtain yield grade output values corresponding to yield grade values of each of the data format converted target mature period crop field images.

[0027] Optionally, the yield grade output values are normalized to obtain confidence score values corresponding to each of the yield grade output values, and the confidence score values are arranged in descending order, and the method specifically comprises the following steps:

[0028] A first operation is performed on each of the data format converted target mature period crop field images.

[0029] The first operation is to normalize the yield grade output values corresponding to the data format converted target mature period crop field images to obtain confidence score values corresponding to each of the yield grade output values, and the confidence score values corresponding to the data format converted target mature period crop field images are arranged in descending order.

[0030] Optionally, in the descendingly arranged confidence score values, the first m confidence score values are selected and normalized to obtain m yield grade weight values, and the method specifically comprises the following steps:

[0031] A second operation is performed on each of the data format converted target mature period crop field images.

[0032] The second operation is: in the descending order arranged set corresponding to the data format converted target mature period crop field image, selecting the first m confidence score values and performing normalization processing to obtain m yield level weight values corresponding to the data format converted target mature period crop field image.

[0033] Optionally, the m yield level weight values are multiplied by the corresponding yield level values respectively and accumulated to obtain the yield estimation value of the target mature period crop field, specifically including:

[0034] A third operation is performed on each of the data format converted target mature period crop field images.

[0035] The third operation is: multiplying the m yield level weight values corresponding to the data format converted target mature period crop field image by the yield level weight values corresponding to the yield level values respectively and accumulating to obtain the yield estimation value of the field corresponding to the data format converted target mature period crop field image.

[0036] In a second aspect, the present application provides a crop yield estimation system based on hierarchical recognition and weight decision fusion, comprising:

[0037] A target field visible light image acquisition module is configured to acquire a visible light image of a target mature period crop field.

[0038] A target field yield level output value determination module is configured to input the visible light image of the target mature period crop field into a yield classification model to obtain yield level output values corresponding to yield level values of the visible light image of the target mature period crop field; the yield level values are determined according to a yield level rule; the yield level rule is that the interval of adjacent yield level values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield level values is not less than a set number.

[0039] A confidence score value arrangement module is configured to normalize each yield level output value to obtain a confidence score value corresponding to each yield level output value, and arrange the confidence score values in descending order.

[0040] A yield level weight value calculation module is configured to select the first m confidence score values from the descending order arranged confidence score values and perform normalization processing to obtain m yield level weight values.

[0041] A target field yield estimation module is configured to multiply the m yield level weight values by the corresponding yield level values respectively and accumulate to obtain a yield estimation value of the target mature period crop field.

[0042] The yield grading model is obtained by training a deep convolutional neural network according to sample data; the sample data includes sample input data and corresponding label data; the sample input data is a sample mature crop field image; and the label data is a yield level value of the sample mature crop field image.

[0043] Optionally, the yield level rule is that the interval of adjacent yield level values is not greater than 10% of the average value of all known sample crop yields, and the number of yield level values is not less than 10.

[0044] According to the specific embodiments of the present application, the following technical effects are disclosed:

[0045] The present application provides a crop yield estimation method and system based on hierarchical recognition and weight decision fusion. The present application uses a deep convolutional neural network to grade the yield of a target mature crop field visible light image, to obtain yield level output values corresponding to each yield level value of the target mature crop field visible light image; then the yield level output values are normalized and screened to obtain m yield level weight values, and the m yield level weight values are multiplied by the corresponding yield level values and accumulated to obtain the yield estimation value of the target mature crop field. The present application not only has a simple estimation method, does not require manual image feature extraction, and can realize crop yield estimation based on simple visible light images. Through actual production verification, the average absolute error of all yield estimation values and yield true values is not greater than 550 kg / ha, and the average absolute percentage error is not greater than 5%, which has strong production application value. BRIEF DESCRIPTION OF DRAWINGS

[0046] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0047] Figure 1 The flowchart of the crop yield estimation method based on hierarchical recognition and weight decision fusion disclosed by the present application embodiment one;

[0048] Figure 2 The specific flowchart of the crop yield estimation method based on hierarchical recognition and weight decision fusion disclosed by the present application embodiment one;

[0049] Figure 3 The weight decision example diagram disclosed by the present application embodiment two;

[0050] Figure 4A structure schematic diagram of the crop yield estimation system based on hierarchical recognition and weight decision fusion disclosed in Embodiment Three of the present application. DETAILED DESCRIPTION

[0051] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0052] In order to make the above objectives, characteristics and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0053] Embodiment One

[0054] As shown in Figure 1 and Figure 2 , the present application provides a crop yield estimation method based on hierarchical recognition and weight decision fusion, specifically comprising:

[0055] Step 100: obtaining a visible light image of a target mature period crop field.

[0056] Step 200: inputting the visible light image of the target mature period crop field into a yield grading model to obtain yield grade output values corresponding to yield grade values of the visible light image of the target mature period crop field; the yield grade values are determined according to yield grade rules; the yield grade rules are that the interval of adjacent yield grade values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield grade values is not less than a set number.

[0057] Step 300: normalizing each yield grade output value to obtain a confidence score value corresponding to each yield grade output value, and arranging the confidence score values in descending order.

[0058] Step 400: selecting the first m confidence score values in the descendingly arranged confidence score values and normalizing them to obtain m yield grade weight values.

[0059] Step 500: multiplying the m yield grade weight values and the corresponding yield grade values respectively and accumulating them to obtain a yield estimation value of the target mature period crop field. The yield estimation value here is a unit yield value.

[0060] The yield grading model is obtained by training a deep convolutional neural network according to sample data; the sample data includes sample input data and corresponding label data; the sample input data is a sample mature period crop field image; and the label data is a yield level value of the sample mature period crop field image.

[0061] The determination process of the sample data is as follows:

[0062] Step 1: Use a remote sensing device such as a drone to collect a plurality of complete visible light sequence photos of sample mature period crop fields of known crop yields at a certain height and not higher than a set value (for example, 100 m), and transmit the complete visible light sequence photos to a computer. Save the complete visible light sequence photos to the computer in the form of visible light sequence images; wherein the visible light sequence photos are processed in the form of image data when processed by the computer.

[0063] Step 2: The computer splices and ortho-rectifies the visible light sequence images according to image features such as photo heading, front-back repetition and side repetition, to obtain pre-processed visible light sequence images, and crops the pre-processed visible light sequence images according to different crop yields and boundaries of the sample mature period crop fields, to obtain initial sample mature period crop field images corresponding to different known crop yields.

[0064] Step 3: Determine the yield grading value of the initial sample mature period crop field image according to the crop yield corresponding to the initial sample mature period crop field image and the yield grading rule. The yield grading rule is that the interval between adjacent yield grading values is not greater than 10% of the average value of all known sample crop yields, and the number of yield grading values is not less than 10.

[0065] Step 4: Randomly sample the initial sample mature period crop field images of different known yield grading values to obtain sampled sample mature period crop field images and corresponding yield grading values; the sampled sample mature period crop field images are sample images with a size not less than 224x224, and the number of sample images in each yield grading is not less than 100.

[0066] Random sampling: This operation refers to randomly cropping or selecting images of the same size (such as 224x224) and without repetition from a sample mature period crop field image of a known yield as input data for a subsequent deep convolutional neural network.

[0067] The reason for random sampling is that the sizes of sample mature crop field image blocks are different, and the sample data input into the deep convolutional neural network needs to be uniform in size. For example, for a sample mature crop field image block with a known yield, equal-size and non-repetitive sampling is performed to obtain a number of sample images, and the label values of these sample images are the same, i.e., the yield classification value corresponding to the true yield value of the sample mature crop field block.

[0068] Step 5: Perform data format conversion, normalization, label assignment, etc. on the sampled sample mature crop field image and the corresponding yield level value to obtain the final sample mature crop field image and the corresponding label data. After the above processing, the image is a multi-dimensional data matrix.

[0069] For example, by using the ImageFolder function to perform tensor format conversion, normalization, and label data correspondence on the above data, the final sample mature crop field image and the corresponding label data are obtained, and the label data is the yield level value corresponding to the sample input data.

[0070] The training process of the yield classification model is as follows:

[0071] Step A: Construct a yield classification network.

[0072] Step B: Train the yield classification network using sample data to obtain a yield classification model.

[0073] Further, the step B specifically includes:

[0074] Step B1: Divide the sample data into a training set, a validation set, and a test set according to a certain proportion, for example, 6:2:2.

[0075] Step B2: Divide the sample data in the training set into not less than 8 batches, then input the batches into the deep convolutional neural network respectively to extract image features, and obtain the yield level value of the sample data in all training sets through classification.

[0076] In this process, the deep convolutional neural network outputs yield level output values corresponding to each yield level value, and the yield level value corresponding to the maximum yield level output value is selected as the predicted result value of the sample data, i.e., the yield level value corresponding to the sample data.

[0077] Step B3: Calculate the average loss value of the label data corresponding to the sample data in all training sets and the yield grade value of the sample data in all training sets by the nn.CrossEntropyLoss() cross-entropy loss function, and feed back the average loss value to the yield grading network to further optimize the model parameters, thereby completing a training process of the yield grading network.

[0078] The training process is not less than 50 times, and the yield grading network after each training is verified on the verification set until the yield grading accuracy of the yield grading network after training tends to be flat on the verification set, and the yield grading accuracy is not less than 80%, the parameters of the yield grading network after this training are saved, and the yield grading model of the embodiment of the application is constructed according to the parameters of the yield grading network after this training.

[0079] As a preferred embodiment of step 200, it specifically comprises:

[0080] Step 201: Splicing and orthographic image correction are performed on the visible light image of the target mature crop field according to image features, to obtain a pretreated visible light image; the image features at least include photo heading features, front-back repeating features and side repeating features.

[0081] Step 202: According to the boundary of the target mature crop field, the pretreated visible light image is cropped to obtain a plurality of initial target mature crop field images.

[0082] Step 203: Random sampling of equal size and non-repetition is performed on the plurality of initial target mature crop field images to obtain a sampled target mature crop field image.

[0083] Step 204: Data format conversion is performed on each of the sampled target mature crop field images to obtain a plurality of data format converted target mature crop field images.

[0084] Step 205: The data format converted target mature crop field image is input into the yield grading model to obtain a yield grade output value corresponding to each yield grade value of the data format converted target mature crop field image.

[0085] As a preferred embodiment of step 300, it specifically comprises:

[0086] The first operation is performed on each of the data format converted target mature crop field images.

[0087] The first operation is to normalize each yield level output value corresponding to the data format converted target mature period crop field image, obtain a confidence score value corresponding to each yield level output value, and arrange the confidence score values corresponding to the data format converted target mature period crop field image in descending order.

[0088] As a preferred embodiment of step 400, it specifically comprises:

[0089] The second operation is performed on each data format converted target mature period crop field image.

[0090] The second operation is to select the first m confidence score values in the descending order arrangement set corresponding to the data format converted target mature period crop field image and perform normalization processing to obtain m yield level weight values corresponding to the data format converted target mature period crop field image; the descending order arrangement set is composed of confidence score values arranged in descending order.

[0091] As a preferred embodiment of step 500, it specifically comprises:

[0092] The third operation is performed on each data format converted target mature period crop field image.

[0093] The third operation is to multiply and accumulate the m yield level weight values corresponding to the data format converted target mature period crop field image and the yield level values corresponding to the yield level weight values, respectively, to obtain the yield estimation value of the field corresponding to the data format converted target mature period crop field image.

[0094] One example: using the yield grading model to grade the sample data on the test set to obtain the yield level output values corresponding to each sample data; after yield grading, the yield grading model performs normalization processing on each yield level output value O1, O2, O3, …, O n of each sample data, for example, using the Softmax normalization function, to convert it into the confidence score values S1, S2, S3, …, S i , …, S n , where the sum of the confidence score values of the yield grading model for each sample belonging to each yield level output value is 1. Arrange the confidence score values in descending order and take out the first m confidence score values, where m is not less than 3. Normalize the selected m confidence score values according to formula (1) to obtain the yield level weight values W1, W2, W3, …, W i , …, W m , and the sum of the yield level weight values is 1.

[0095]

[0096] For the obtained yield level weight value corresponding to the test set sample data, and the yield level value corresponding to the yield level weight value, the yield level weight value is multiplied by the corresponding yield level value according to formula (2) and accumulated, so as to obtain the yield estimation value Y of the mature crop field corresponding to the sample data p All sample data on the test set are obtained according to the method corresponding to the yield estimation value.

[0097]

[0098] Wherein, C i Indicates the yield level value corresponding to the i-th yield level weight value.

[0099] After comparison, the average absolute error of all obtained yield estimation values and yield true values is not more than 550 kg / ha, and the average absolute percentage error is not more than 5%.

[0100] Example two

[0101] Now taking rice as an example to illustrate the crop yield estimation method based on yield grading identification and weight decision fusion protected by the application.

[0102] The embodiment of the application provides a rice yield estimation method based on yield grading identification and weight decision fusion, which comprises the following steps:

[0103] First step: the heading repetition rate of the unmanned aerial vehicle is set to 60%, the lateral repetition rate of the unmanned aerial vehicle is set to 55%, the flight speed of the unmanned aerial vehicle is set to 2.5 m / s, and the flight height of the unmanned aerial vehicle is set to 25 m. A plurality of complete visible light images of mature rice fields are collected by using the unmanned aerial vehicle with the above configured parameters, and the yield true value of the mature rice field covers 3654.2-16978.5 kg / ha. Wherein, the model of the unmanned aerial vehicle is DJI Mavic2.

[0104] Second step: the visible light images are spliced and ortho-rectified according to the photo heading, front-back repetition and lateral repetition, and the mature rice field images of different crop yields are cut out, so as to obtain 131 mature rice field images of known crop yields.

[0105] Third step: Set the interval of adjacent yield level values as 900 kg / ha, set the yield level values as 36500 kg / ha, 45500 kg / ha, 54500 kg / ha, 63500 kg / ha, 72500 kg / ha, 81500 kg / ha, 90500 kg / ha, 99500 kg / ha, 108500 kg / ha, 117500 kg / ha, 126500 kg / ha, 135500 kg / ha, 144500 kg / ha, 153500 kg / ha, 162500 kg / ha, 17250 kg / ha, a total of 16 yield level values. On the basis of each yield level value, set the upper and lower floating values (for example, the floating value is 450 kg / ha), and then use each yield level value after setting the floating value to classify the yield of 131 known crop yield images of mature rice fields, such as 3200-4100 kg / ha of rice belonging to the yield level of 3650 kg / ha, and randomly and repeatedly sampling the mature rice field image of each yield level with a size of 224x224, and obtaining 600 sample data for each yield level.

[0106] Fourth step: All sample data are divided into training data, validation data and test data according to 6:2:2. Based on the open source functions ImageFolder, DataLoader, etc. in the computer programming language Python, the tensor format conversion, normalization, label assignment and other processing steps are completed for all sample data, so as to form the training set, validation set and test set of the yield classification model. The yield classification model is built by using the Pytorch framework. In this embodiment, the publicly disclosed ShuffleNetV2 pre-trained model is selected, and the output dimension of the full connection layer is set as the number of yield levels, i.e. 16. The nn.CrossEntropyLoss cross-entropy loss function is selected for loss calculation in the training process of the yield classification model, the torch.optim.Adam function is used as the optimizer, the epoch running number is set as 50, the batch size parameter is set as 64, and the learning rate is set as 0.005. On this basis, the test set and the validation set are input into the model, and the yield classification model starts to be iteratively trained. After 40 epochs, the loss value of the yield classification model on the validation set tends to be stable, the best accuracy of the training set in the 50-epoch training process is 98.5%, the best accuracy of the validation set is 87.5%, the accuracy of the test set is 87.3%, and the model parameters with the best accuracy are saved.

[0107] Fifth step: calling the yield grading model trained in the fourth step, using the yield grading model to grade the samples on the test set, and the yield level output value of the sample data on the test set can be obtained. The yield level output value O of each sample data on the test set is normalized by the Softmax normalization function, and is converted into the confidence score value corresponding to each yield level output value of the sample by the trained yield grading model, wherein the sum of the confidence score values corresponding to each yield level output value of a sample by the trained yield grading model is 1. The top three largest confidence score values are selected, and the three confidence score values are normalized to obtain yield level weight values, wherein the sum of the three yield level weight values is 1. The three yield level weight values are multiplied by the corresponding yield level values and then added to obtain the yield estimation value of the rice field corresponding to the sample.

[0108] Taking a sample with a yield of 10850kg / ha identified by preliminary yield grading as an example:

[0109] As shown in Figure 3 , the sample is input into the ShuffleNetV2 model trained in the fourth step to obtain the yield level output value O corresponding to the yield level value C; the yield level output value O is converted into the confidence score value S by the Softmax normalization function, and the top three largest confidence score values are taken out. The three confidence score values are multiplied by the reciprocal of the sum of the three confidence score values (i.e. 1.209), and then normalized so that the sum of the three confidence score values is 1, thereby obtaining the three yield level weight values W of the sample; the three yield level weight values are multiplied by the corresponding yield level values and then added to obtain the yield estimation value of the rice field corresponding to the sample.

[0110] Sixth step: according to the above steps, the yield estimation of all test set samples is performed, and the average absolute error of the obtained yield estimation value and the true value is 402.13kg / ha, and the average absolute percentage error is 3.87%.

[0111] Example three

[0112] In order to perform the method corresponding to the above-mentioned example one, to realize the corresponding functions and technical effects, a crop yield estimation system based on yield grading identification and weight decision fusion is provided below.

[0113] As shown in Figure 4 , the crop yield estimation system based on yield grading identification and weight decision fusion provided by the embodiment of the present application comprises:

[0114] The target field visible light image acquisition module 1 is used to acquire the visible light image of the target mature crop field.

[0115] The target field plot yield level output value determination module 2 is configured to input the visible light image of the target mature crop field plot into a yield grading model to obtain yield level output values corresponding to yield level values of the visible light image of the target mature crop field plot; the yield level values are determined according to a yield level rule; the yield level rule is that the interval between adjacent yield level values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield level values is not less than a set number.

[0116] The confidence score value arrangement module 3 is configured to normalize each yield level output value to obtain a confidence score value corresponding to each yield level output value, and arrange the confidence score values in descending order.

[0117] The yield level weight value calculation module 4 is configured to select the first m confidence score values in the descendingly arranged confidence score values and perform normalization to obtain m yield level weight values.

[0118] The target field plot yield estimation module 5 is configured to multiply the m yield level weight values and the corresponding yield level values respectively and accumulate to obtain a yield estimation value of the target mature crop field plot.

[0119] The yield grading model is obtained by training a deep convolutional neural network according to sample data; the sample data includes sample input data and corresponding label data; the sample input data is a sample mature crop field plot image; and the label data is a yield level value of the sample mature crop field plot image.

[0120] Further, the yield level rule is that the interval between adjacent yield level values is not greater than 10% of the average value of all known sample crop yields, and the number of yield level values is not less than 10.

[0121] In the specification, each embodiment is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same or similar parts between the embodiments can be referred to each other.

[0122] The principles and implementation manners of the present application are described by using specific examples in this paper, and the above embodiment description is only used to help understand the method of the present application and its core idea; meanwhile, for those skilled in the art, according to the idea of the present application, the specific implementation manner and application range can be changed. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A crop yield estimation method based on hierarchical recognition and weight decision fusion, characterized in that, The method comprises the following steps: obtaining a visible light image of a target mature crop field; inputting the visible light image of the target mature crop field into a yield grading model to obtain yield grade output values corresponding to yield grade values of the visible light image of the target mature crop field; the yield grade values are determined according to yield grade rules; the yield grade rules are that the interval between adjacent yield grade values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield grade values is not less than a set number; normalizing each yield grade output value to obtain a confidence score value corresponding to each yield grade output value, and arranging the confidence score values in descending order; selecting the first m confidence score values in the descending order of the confidence score values and normalizing them to obtain m yield grade weight values; multiplying the m yield grade weight values by the corresponding yield grade values and accumulating them to obtain a yield estimation value of the target mature crop field; wherein the yield grading model is obtained by training a deep convolutional neural network based on sample data; the sample data includes sample input data and corresponding label data; the sample input data is a sample mature crop field image; the label data is a yield grade value of the sample mature crop field image; on the basis of each yield grade value, a floating value is set, and each yield grade value after the floating value is set is used to grade the yield of the sample mature crop field image. 2.The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 1, characterized in that, The determination process of the sample data is as follows: using a remote sensing device to collect visible light sequence photos of sample mature crop fields with known crop yields and a certain height not higher than a set value, and transmitting the visible light sequence photos of the sample mature crop fields to a computer and saving them to the computer in the form of visible light sequence images; splicing and orthographic image correction are performed on the visible light sequence images according to image features to obtain preprocessed visible light sequence images; the image features at least include photo heading features, forward-backward repetition features and lateral repetition features; according to different crop yields and the boundaries of the sample mature crop fields, the preprocessed visible light sequence images are cropped to obtain initial sample mature crop field images corresponding to different known crop yields; determining yield grading values of the initial sample mature crop field images according to the crop yields corresponding to the initial sample mature crop field images and yield grading rules; randomly sampling the initial sample mature crop field images with different known yield grading values to obtain sampled sample mature crop field images and corresponding yield grading values; performing data format conversion, normalization and label assignment processing on the sampled sample mature crop field images and corresponding yield grading values to obtain final sample mature crop field images and corresponding label data. 3.The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 1, characterized in that, The yield grading rules are that the interval between adjacent yield grade values is not greater than 10% of the average value of all known sample crop yields, and the number of yield grade values is not less than 10.

4. The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 1, characterized in that, The visible light image of the target mature period crop field block is input into a yield grading model to obtain yield level output values corresponding to yield level values of the visible light image of the target mature period crop field block, and the yield level output values are normalized to obtain confidence score values corresponding to each yield level output value, and the confidence score values are arranged in descending order. According to the image features, the visible light image of the target mature period crop field block is spliced and orthographic image corrected to obtain a preprocessed visible light image; the image features at least include photo heading features, front-back repeating features and side repeating features; According to the boundary of the target mature period crop field block, the preprocessed visible light image is cropped to obtain a plurality of initial target mature period crop field block images; The plurality of initial target mature period crop field block images are randomly sampled in equal size and without repetition to obtain sampled target mature period crop field block images; The data format of each of the sampled target mature period crop field block images is converted to obtain a plurality of data format converted target mature period crop field block images; The data format converted target mature period crop field block images are input into a yield grading model to obtain yield level output values corresponding to yield level values of each of the data format converted target mature period crop field block images.

5. The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 4, characterized in that, The yield level output values are normalized to obtain confidence score values corresponding to each yield level output value, and the confidence score values are arranged in descending order, and the confidence score values are arranged in descending order. A first operation is performed on each of the data format converted target mature period crop field block images; The first operation is to normalize the yield level output values corresponding to the data format converted target mature period crop field block images to obtain confidence score values corresponding to each yield level output value, and the confidence score values corresponding to the data format converted target mature period crop field block images are arranged in descending order.

6. The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 5, characterized in that, In the descendingly arranged confidence score values, the first m confidence score values are selected and normalized to obtain m yield level weight values, and the confidence score values are arranged in descending order. A second operation is performed on each of the data format converted target mature period crop field block images; The second operation is to select the first m confidence score values in the descendingly arranged set corresponding to the data format converted target mature period crop field block images and normalize them to obtain m yield level weight values corresponding to the data format converted target mature period crop field block images; the descendingly arranged set is composed of descendingly arranged confidence score values.

7. The crop yield estimation method based on hierarchical identification and weight decision fusion according to claim 6, characterized in that, The m yield level weight values are multiplied by the corresponding yield level values and accumulated to obtain the yield estimation value of the target mature period crop field block, and the confidence score values are arranged in descending order. A third operation is performed on each of the data format converted target mature period crop field block images; The third operation is to multiply the m yield level weight values corresponding to the data format converted target mature period crop field block images by the yield level values corresponding to the yield level weight values and accumulate them to obtain the yield estimation value of the field block corresponding to the data format converted target mature period crop field block image. 8.A crop yield estimation system based on hierarchical recognition and weight decision fusion, characterized in that, It includes: The target field block visible light image acquisition module is configured to acquire a visible light image of a target mature crop field block. The target field block yield level output value determination module is configured to input the visible light image of the target mature crop field block into a yield grading model to obtain yield level output values corresponding to yield level values of the visible light image of the target mature crop field block. The yield level values are determined according to a yield level rule. The yield level rule is that the interval between adjacent yield level values is not greater than a set percentage of the average value of all known sample crop yields, and the number of yield level values is not less than a set number. The confidence score value arrangement module is configured to normalize each yield level output value to obtain a confidence score value corresponding to each yield level output value, and arrange the confidence score values in descending order. The yield level weight value calculation module is configured to select the first m confidence score values in the descendingly arranged confidence score values and perform normalization processing to obtain m yield level weight values. The target field block yield estimation module is configured to multiply the m yield level weight values and the corresponding yield level values, respectively, and accumulate them to obtain a yield estimation value of the target mature crop field block. The yield grading model is obtained by training a deep convolutional neural network according to sample data. The sample data includes sample input data and corresponding label data. The sample input data is a sample mature crop field image, and the label data is a yield level value of the sample mature crop field image. On the basis of each yield level value, an upper and lower floating value is set, and each yield level value after floating value setting is used to grade the yield of the sample mature crop field image. 9.The crop yield estimation system based on hierarchical identification and weight decision fusion of claim 8, wherein, The yield level rule is that the interval between adjacent yield level values is not greater than 10% of the average value of all known sample crop yields, and the number of yield level values is not less than 10.

Images