Crop recognition model training method, crop recognition method and device
By integrating the data characteristics of the benchmark growth environment and reference growth environment in the crop recognition model, the growth curve characteristics are extracted to train the model, which solves the problem of instability of the crop recognition model's recognition ability under different environmental conditions, and improves the recognition accuracy and stability sex.
Patent Information
- Application Number
- CN202411239991.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-05
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-09-05
AI Technical Summary
The existing crop recognition model has unstable recognition capabilities and poor spatiotemporal migration capabilities under different environmental conditions, resulting in reduced accuracy and stability of classification results.
By obtaining sample data in the benchmark growth environment and the reference growth environment, the sample base growth curve and reference growth curve were constructed and extracted separately, and the feature extraction was integrated to train the crop recognition model to improve the adaptability of the model in different environments.
It effectively improves the recognition accuracy and stability of crop recognition models under different environmental conditions, and solves the problem of poor adaptability.
Smart Images

Figure CN118760935B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent recognition technology, and in particular to a crop recognition model training method, a crop recognition method and a crop recognition device. Background Art
[0002] Remote sensing technology can quickly and accurately monitor the types, distribution and growth conditions of crops in large areas of farmland, optimize planting layout, rationally allocate resources, and improve the efficiency and benefits of agricultural production. Different crops may have similar or overlapping spectral features in remote sensing images under the influence of different growth stages and growth environments, which can lead to classification confusion and misidentification. In recent years, crop classification based on remote sensing images based on artificial neural networks has become a development trend. However, existing crop recognition models are mostly based on multi-feature expressions of remote sensing data at different time phases. There are still problems such as poor spatiotemporal migration capabilities and unstable recognition capabilities under the influence of different stress environments, which reduces the accuracy and stability of crop classification results. Summary of the invention
[0003] In view of the above problems, the present invention provides a crop recognition model training method, a crop recognition method and a crop recognition device.
[0004] According to a first aspect of the present invention, a method for training a crop recognition model is provided, comprising:
[0005] Acquire training samples, the training samples include sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of sample target type crops in a reference growth environment, and classification labels, wherein the reference environmental conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment, the reference environmental conditions are determined based on stress environment parameters, and the classification labels characterize the types of sample target type crops; perform feature extraction on the sample benchmark growth curve and the sample reference growth curve respectively to obtain sample benchmark growth features and sample reference growth features, wherein the sample benchmark growth curve is determined based on the sample benchmark growth state data, and the sample reference growth curve is determined based on the sample reference growth state data; train a crop recognition model based on the sample benchmark growth features, the sample reference growth features, and the classification labels to obtain a trained crop recognition model.
[0006] Optionally, the sample reference growth state data includes sample reference radar data and sample reference spectrum data; wherein the sample reference growth curve is determined based on the following operations:
[0007] The sample reference radar data are processed according to the radar index function to obtain sample reference radar index data, wherein the sample reference radar index data characterizes the growth density state of the sample target type of crops; the sample reference spectral data are processed according to the spectral index function to obtain sample reference spectral index data, wherein the sample reference spectral index data characterizes the moisture content state of the sample target type of crops; according to the sample reference radar index data of the sample target type of crops in the reference growth environment, a first sub-reference growth curve corresponding to the stress environment parameter is determined; according to the sample reference spectral index data of the sample target type of crops in the reference growth environment, a second sub-reference growth curve corresponding to the stress environment parameter is determined; according to the first sub-reference growth curve and the second sub-reference growth curve, a sample reference growth curve of the sample target type of crops is determined.
[0008] Optionally, the sample benchmark growth state data includes sample benchmark radar data and sample benchmark spectral data; wherein the sample benchmark growth curve is determined based on the following operations:
[0009] Determine the number of pixels in the plot area, wherein the sample target type crops grow in the plot area; obtain a radar average value in the plot area based on the number of pixels in the plot area and the sample benchmark radar index data on each pixel, wherein the sample benchmark radar index data is obtained based on processing the sample benchmark radar data; use a spline interpolation function to process the radar average value to obtain radar index time series data; reconstruct the radar index time series data to generate a first sub-benchmark growth curve; obtain a spectral average value in the plot area based on the number of pixels in the plot area and the sample benchmark spectral index data on each pixel, wherein the sample benchmark spectral index data is obtained based on processing the sample benchmark spectral benchmark data; use a spline interpolation function to process the spectral average value to obtain spectral index time series data; reconstruct the spectral index time series data to generate a second sub-benchmark growth curve; obtain a sample benchmark growth curve of the sample target type crops based on the first sub-benchmark growth curve and the second sub-benchmark growth curve.
[0010] Optionally, the sample benchmark radar index data includes at least one of the following: radar ratio vegetation index data, normalized radar vegetation index data, normalized radar difference water index data and enhanced radar vegetation index data; the sample benchmark spectral index data includes at least one of the following: greenness vegetation index data, enhanced vegetation index data and vegetation moisture index data.
[0011] Optionally, the sample benchmark growth characteristics include at least one of the following: time series characteristics, curve statistical characteristics, curve waveform characteristics, and exponential combination characteristics; wherein the time series characteristics characterize the time variation characteristics of the index data corresponding to the sample target type of crops, the curve statistical characteristics characterize the growth morphological characteristics of the sample target type of crops, the curve waveform characteristics characterize the growth periodic variation characteristics of the sample target type of crops, and the exponential combination characteristics characterize the difference characteristics of the same index time series data generated by the sample target type of crops in the benchmark growth environment and the reference growth environment.
[0012] Optionally, the crop recognition model includes a convolutional neural network layer, a long short-term memory network layer, a fully connected layer, and an activation layer; wherein the crop recognition model is trained based on the sample baseline growth characteristics, the sample reference growth characteristics, and the classification label to obtain a trained crop recognition model, including:
[0013] The sample benchmark growth characteristics and the sample reference growth characteristics are input into the convolutional neural network layer, and the sample convolution characteristics with time series attributes are output; the sample convolution characteristics with time series attributes are input into the long short-term memory network layer, and the sample deep characteristics are output; the sample deep characteristics are input into the fully connected layer, and the sample fully connected characteristics are output; the sample fully connected characteristics are input into the activation layer, and the sample crop type recognition results are output; according to the sample crop type recognition results and classification labels, the crop recognition model is trained to obtain the trained crop recognition model.
[0014] Optionally, the convolutional neural network layer includes a convolution sublayer, a pooling sublayer, and a normalization sublayer; wherein the sample baseline growth feature and the sample reference growth feature are input into the convolutional neural network layer, and the sample convolution feature with time series attributes is output, including:
[0015] The sample baseline growth features and the sample reference growth features are input into the convolution sublayer, and the sample subconvolution features are output; the sample subconvolution features are input into the pooling sublayer, and the sample pooling features are output; the sample pooling features are input into the normalization sublayer, and the sample convolution features are output.
[0016] Optionally, a crop recognition model is trained according to the sample crop type recognition results and classification labels to obtain a trained crop recognition model, including:
[0017] The sample crop type recognition results and classification labels are processed according to the cross entropy loss function to obtain a loss value; according to the loss value, the model parameters of the crop recognition model are adjusted to obtain a trained crop recognition model.
[0018] A second aspect of the present invention provides a crop identification method, comprising:
[0019] Acquire growth status data of the target crop; extract features from the growth curve to obtain growth features, wherein the growth curve is generated based on the growth status data; input the growth features into a trained crop recognition model, and output a crop type recognition result of the target crop, wherein the trained crop recognition model is obtained according to the above training method.
[0020] A third aspect of the present invention provides a training device for a crop recognition model, comprising:
[0021] A first acquisition module is used to acquire training samples, the training samples including sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of sample target type crops in a reference growth environment, and classification labels, wherein the reference environmental conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment, the reference environmental conditions are determined based on stress environment parameters, and the classification labels represent the types of sample target type crops;
[0022] A first feature extraction module is used to extract features from the sample benchmark growth curve and the sample reference growth curve respectively to obtain the sample benchmark growth feature and the sample reference growth feature, wherein the sample benchmark growth curve is determined based on the sample benchmark growth state data, and the sample reference growth curve is determined based on the sample reference growth state data;
[0023] The training module is used to train the crop recognition model based on the sample baseline growth characteristics, the sample reference growth characteristics and the classification label to obtain the trained crop recognition model.
[0024] A fourth aspect of the present invention provides a crop identification device, comprising:
[0025] The second acquisition module is used to acquire the growth status data of the target crop;
[0026] A second feature extraction module is used to extract features from the growth curve to obtain growth features, wherein the growth curve is generated based on the growth state data;
[0027] The recognition module is used to input the growth characteristics into a trained crop recognition model and output a crop type recognition result of the target crop, wherein the trained crop recognition model is obtained according to the above training method.
[0028] According to the training method, crop identification method and device of the crop identification model provided by the present invention, a sample benchmark growth curve is constructed according to the sample benchmark growth state data under the benchmark growth environment, and a sample reference growth curve is constructed according to the respective sample reference growth state data under multiple reference growth environment conditions, and features are extracted from the sample benchmark growth curve and the sample reference growth curve respectively to obtain the sample benchmark growth feature and the sample reference growth feature. Since the sample benchmark growth curves of different sample target types of crops under the benchmark growth environment and the sample reference growth curves under the reference growth environment constitute a diverse curve library information, it is possible to extract features from a multi-dimensional level to comprehensively characterize the growth feature difference information caused by different sample target types of crops under different growth environments, and the crop identification model is trained based on the sample benchmark growth feature and the sample reference growth feature, which solves the problem of poor adaptability of the crop identification model under different environmental conditions, and effectively improves the accuracy and stability of large-scale sample target type crop identification. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The above and other objects, features and advantages of the present invention will become more apparent through the following description of the embodiments of the present invention with reference to the accompanying drawings, in which:
[0030] Figure 1 A flow chart showing a method for training a crop recognition model according to an embodiment of the present invention is shown;
[0031] Figure 2 An example schematic diagram of determining a sample crop type identification result according to an embodiment of the present invention is shown;
[0032] Figure 3 A flow chart of a crop identification method according to an embodiment of the present invention is shown;
[0033] Figure 4 A structural block diagram of a training device for a crop recognition model according to an embodiment of the present invention is shown;
[0034] Figure 5 A structural block diagram of a crop identification device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION
[0035] Below, embodiments of the present invention will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present invention. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of embodiments of the present invention. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of concepts of the present invention.
[0036] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.
[0037] All terms (including technical and scientific terms) used herein have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.
[0038] When using expressions such as "at least one of A, B, and C, etc.", they should generally be interpreted according to the meaning of the expression commonly understood by those skilled in the art (for example, "a system having at least one of A, B, and C" should include but is not limited to a system having A alone, B alone, C alone, A and B, A and C, B and C, and / or A, B, C, etc.).
[0039] In the process of implementing the present invention, it is found that the spectral features of different crops in remote sensing images may be similar or overlapped under the influence of different growth stages and growth environments, which may lead to classification confusion and misidentification. In recent years, remote sensing image crop classification based on artificial neural networks has become a development trend. However, existing crop recognition models are mostly based on multi-feature expressions of remote sensing data in different phases, and still have problems such as poor spatiotemporal migration capabilities and unstable recognition capabilities under the influence of different stress environments, which reduces the accuracy and stability of crop classification results.
[0040] In view of this, an embodiment of the present invention provides a training method for a crop recognition model, a crop recognition method and a device. The method comprises: obtaining training samples, the training samples comprising sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of sample target type crops in a reference growth environment and classification labels; performing feature extraction on the sample benchmark growth curve and the sample reference growth curve respectively to obtain sample benchmark growth features and sample reference growth features, wherein the sample benchmark growth curve is determined based on the sample benchmark growth state data, and the sample reference growth curve is determined based on the sample reference growth state data; training a crop recognition model based on the sample benchmark growth features, the sample reference growth features and the classification labels to obtain a trained crop recognition model.
[0041] In the technical solution of the present invention, the user information (including but not limited to user personal information, user image information, user device information, such as location information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved are all information and data authorized by the user or fully authorized by all parties, and the collection, storage, use, processing, transmission, provision, disclosure and application of the relevant data comply with relevant laws, regulations and standards, take necessary confidentiality measures, do not violate public order and good morals, and provide corresponding operation entrances for users to choose to authorize or refuse.
[0042] It should be noted that the sequence numbers of the operations in the following method are only used as representations of the operations for the purpose of description, and should not be regarded as representing the execution order of the operations. Unless explicitly stated, the method does not need to be executed completely in the order shown.
[0043] Figure 1 A flow chart of a method for training a crop recognition model according to an embodiment of the present invention is shown.
[0044] like Figure 1 As shown, the method 100 includes operations S110 to S130.
[0045] In operation S110, a training sample is acquired.
[0046] Optionally, the training samples include sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of sample target type crops in a reference growth environment, and classification labels.
[0047] Optionally, the classification label represents the type of the sample target category crop. For example, the classification label of the sample target category crop is 1, representing the corn type.
[0048] Optionally, the sample target type crops may be four major crops, for example, rice, corn, wheat and soybean. It should be noted that the crops in the embodiment of the present invention may include crops, and may also include other types of plant type crops such as rapeseed and potato, and the embodiment of the present invention does not limit the specific types of crops.
[0049] Optionally, the reference environment conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment. The benchmark growth conditions are set to simulate the benchmark growth environment, which is a normal environment suitable for the sample target type of crop to grow.
[0050] Optionally, the reference environmental conditions are determined based on stress environmental parameters, and the reference growth environment may be drought, salinity, low temperature, high temperature, disease, nutrient deficiency, etc. The stress environmental parameters may be soil parameters, meteorological parameters, and pest and disease parameters. A variety of reference environmental conditions are determined by adjusting soil parameters, meteorological parameters, and pest and disease degree parameters individually or in combination, thereby simulating a single or combined reference growth environment and the corresponding reference growth environment intensity, and recording changes in stress environmental parameters according to environmental monitoring time intervals. For example, a single reference growth environment is a drought environment, and a combined reference growth environment may be a drought and high temperature environment.
[0051] For example, soil parameters may be soil water content parameters, potential of hydrogen (pH) value parameters, conductivity parameters, nutrient content parameters, etc. Meteorological parameters may be temperature parameters, humidity parameters, light intensity parameters, precipitation parameters, etc. Pest and disease parameters may be pest and disease type parameters and severity parameters. A drought reference growth environment is simulated by adjusting the precipitation parameter alone to be lower than the precipitation threshold; a saline reference growth environment is simulated by adjusting the soil pH value parameter alone.
[0052] Optionally, based on the set satellite shooting time interval and the growth cycle of the sample target type of crops, sample baseline growth state data of the sample target type of crops in the baseline growth environment and sample reference growth state data of the sample target type of crops in different reference growth environments are obtained from the remote sensing data shot by the satellite. For example, the shooting time interval can be five days.
[0053] In operation S120, feature extraction is performed on the sample benchmark growth curve and the sample reference growth curve respectively to obtain a sample benchmark growth feature and a sample reference growth feature.
[0054] Optionally, the sample baseline growth curve is determined based on the sample baseline growth status data, and the sample reference growth curve is determined based on the sample reference growth status data.
[0055] Optionally, a sample benchmark growth curve is reconstructed based on sample benchmark growth state data corresponding to different growth periods of the sample target type of crops; a sample reference growth curve is reconstructed based on sample reference growth state data corresponding to different growth periods of the sample target type of crops. The sample reference growth curve corresponds to the reference growth environment one by one, and each sample target type of crop obtains different sample reference growth state data under different reference growth environments, and thus corresponds to different sample reference growth curves.
[0056] Optionally, stress environment parameters in different reference growth environments are normalized to ensure data comparability, thereby reconstructing sample reference growth curves that correspond one-to-one to different reference growth environment intensities during the growth cycle of different sample target types of crops.
[0057] Optionally, feature extraction is performed on the sample baseline growth curve to obtain the sample baseline growth feature; feature extraction is performed on the sample reference growth curve to obtain the sample reference growth feature. The sample baseline growth feature and the sample reference growth feature characterize the growth characteristics of the sample target type of crops in different growth cycles and different growth environments. For example, the growth rate characteristics corresponding to the sample target type of crops in different growth cycles and different growth environments.
[0058] In operation S130 , a crop recognition model is trained based on the sample baseline growth characteristics, the sample reference growth characteristics, and the classification labels to obtain a trained crop recognition model.
[0059] Optionally, the crop identification model can be constructed based on machine learning or deep learning algorithms. The sample baseline growth characteristics and the sample reference growth characteristics are fused and input into the crop identification model to obtain the type identification result of the sample target type crop.
[0060] Optionally, the performance of the model can be evaluated based on the type recognition results and classification labels of the sample target crop types, and then the parameters of the crop recognition model can be adjusted, and the crop recognition model can be iteratively trained. When the performance meets a preset performance threshold, the training is stopped to obtain a trained crop recognition model.
[0061] Optionally, since the sample benchmark growth curves of different sample target types of crops under the benchmark growth environment and the sample reference growth curves under the reference growth environment constitute a diverse curve library information, features can be extracted from multiple dimensions to comprehensively characterize the growth characteristic difference information caused by different sample target types of crops under different growth environments. The crop recognition model is trained based on the sample benchmark growth characteristics and the sample reference growth characteristics, which solves the problem of poor adaptability of the crop recognition model under different environmental conditions and effectively improves the accuracy and stability of large-scale recognition of target types of crops in different years.
[0062] Optionally, the sample benchmark growth state data includes sample benchmark radar data and sample benchmark spectral data; wherein the sample benchmark growth curve is determined based on the following operations:
[0063] Determine the number of pixels in the plot area, wherein the sample target type crops grow in the plot area; obtain a radar average value in the plot area based on the number of pixels in the plot area and the sample benchmark radar index data on each pixel, wherein the sample benchmark radar index data is obtained based on processing the sample benchmark radar data; use a spline interpolation function to process the radar average value to obtain radar index time series data; reconstruct the radar index time series data to generate a first sub-benchmark growth curve; obtain a spectral average value in the plot area based on the number of pixels in the plot area and the sample benchmark spectral index data on each pixel, wherein the sample benchmark spectral index data is obtained based on processing the sample benchmark spectral benchmark data; use a spline interpolation function to process the spectral average value to obtain spectral index time series data; reconstruct the spectral index time series data to generate a second sub-benchmark growth curve; obtain a sample benchmark growth curve of the sample target type crops based on the first sub-benchmark growth curve and the second sub-benchmark growth curve.
[0064] Optionally, the sample benchmark radar data comes from the synthetic aperture radar (SAR) data of the microwave remote sensing satellite (Senitnel-1), and the sample benchmark spectral data comes from the multispectral data of the high-resolution multispectral imaging satellite (Senitnel-2).
[0065] Optionally, preprocessing the sample benchmark radar data may include: converting the sample benchmark radar data into radiation intensity data representing multiple backscatter coefficients, reducing the speckle noise data by applying speckle filtering, removing invalid data at the edge, ensuring data quality, and finally performing geocoding and regularization. The preprocessed sample benchmark radar data is processed according to the radar index function to obtain the sample benchmark radar index data.
[0066] Optionally, download the land parcel boundary area division data, and determine multiple land parcel areas and the number of pixels in each land parcel area according to the land parcel boundary area division data. A pixel represents the smallest unit for satellites to take pictures and sample ground scenes. For example, each land parcel may include multiple 10m*10m pixels. The smaller the pixel, the higher the data resolution.
[0067] Optionally, for one of the plot areas collected at a certain data shooting time, the plot area contains multiple pixels, and the total sum of the sample benchmark radar index data corresponding to the multiple pixels is determined; then the radar average value on the plot area is determined based on the ratio of the total sum of the sample benchmark radar index data to the number of pixels in the plot area.
[0068] Optionally, in one acquisition cycle, based on the satellite shooting time interval, sample baseline radar index data corresponding to each of a plurality of interval shooting moments are obtained, thereby obtaining radar average values corresponding to each of different land areas at the shooting moment.
[0069] For example, if the collection period is 365 days and the satellite shooting time interval is 5 days, the sample benchmark radar index data corresponding to multiple shooting moments with an interval of 5 days are obtained, and the radar average values corresponding to different land areas at the shooting time are obtained.
[0070] Optionally, each plot area has a radar average value corresponding to the shooting time in a collection cycle, and the radar average values corresponding to multiple shooting times in a collection cycle are processed using a spline interpolation function to obtain the radar average values corresponding to multiple non-shooting times in a collection cycle, thereby obtaining continuous radar index time series data in a collection cycle. For example, using a spline interpolation function for interpolation can use a low-order polynomial (such as a cubic polynomial) to interpolate between known points (radar average values corresponding to multiple shooting times).
[0071] Optionally, the radar index time series data is reconstructed to generate a first sub-reference growth curve. The horizontal axis of the first sub-reference growth curve represents the acquisition period, and the vertical axis represents the radar index time series data obtained by growing the sample target type of crops in this plot area based on the reference growth environment.
[0072] Optionally, preprocessing the sample benchmark spectral data may include: converting the sample benchmark spectral data into surface reflectance data characterizing reflectances of multiple bands, performing atmospheric correction, removing clouds and cloud shadow areas, and resampling the resolution of the first ten bands to 10m using the nearest neighbor method to ensure image quality and consistent resolution. The preprocessed sample benchmark spectral data is processed according to the spectral index function to obtain sample benchmark spectral index data.
[0073] Optionally, for one of the plot areas collected at a certain data shooting time, the plot area contains multiple pixels, and the sum of the sample benchmark spectral index data corresponding to the plot area is determined based on the cumulative sum of the sample benchmark spectral index data corresponding to the multiple pixels; and then the spectral average value of the plot area is determined based on the ratio of the sum of the sample benchmark spectral index data to the number of pixels in the plot area.
[0074] Optionally, in one acquisition cycle, according to the satellite shooting time interval, sample benchmark spectral index data corresponding to each of the multiple interval shooting moments are obtained, so as to obtain the spectral average values corresponding to each of the different plot areas at the shooting moment; each plot area has a spectral average value corresponding to the shooting moment in one acquisition cycle, and the spectral average values corresponding to each of the multiple shooting moments in one acquisition cycle are processed using a spline interpolation function to obtain the spectral average values corresponding to each of the multiple non-shooting moments in one acquisition cycle, so as to obtain continuous spectral index time series data in one acquisition cycle.
[0075] Optionally, the spectral index time series data is reconstructed to generate a second sub-reference growth curve. The horizontal axis of the second sub-reference growth curve represents the acquisition period, and the vertical axis represents the spectral index time series data of the sample target type of crops grown in this plot area based on the reference growth environment.
[0076] Optionally, the sample benchmark growth curve includes a first sub-benchmark growth curve and a plurality of second sub-benchmark growth curves.
[0077] Optionally, a first sub-benchmark growth curve with multiple features reconstructed based on radar index time series data and a second sub-benchmark growth curve with multiple features reconstructed based on spectral index time series data are normalized to the same growth stage through time shift and shape matching, thereby overcoming the limitations of using samples from the same period for modeling and identification.
[0078] Optionally, the sample benchmark radar index data includes at least one of the following: radar ratio vegetation index data, normalized radar vegetation index data, normalized radar difference water index data and enhanced radar vegetation index data; the sample benchmark spectral index data includes at least one of the following: greenness vegetation index data, enhanced vegetation index data and vegetation moisture index data.
[0079] Optionally, radar ratio vegetation index data characterizes the growth density status of sample target type crops, normalized radar vegetation index data characterizes the growth structure characteristics and crop type of sample target type crops, normalized radar difference water index data characterizes the water content status of sample target type crops, and enhanced radar vegetation index data characterizes the quantity difference status of sample target type crops.
[0080] In one embodiment, the radar ratio vegetation index data RRI is as shown in formula (1):
[0081] (1);
[0082] in, Characterizes the backscattering coefficient after VH polarization, It represents the backscattering coefficient after VV polarization, VH represents the vertical-horizontal polarization mode, and VV represents the vertical-vertical polarization mode.
[0083] In one embodiment, the normalized radar vegetation index data NRVI is as shown in formula (2):
[0084] (2);
[0085] in, Characterizes the backscattering coefficient after VH polarization, It represents the backscattering coefficient after VV polarization, VH represents the vertical-horizontal polarization mode, and VV represents the vertical-vertical polarization mode.
[0086] In one embodiment, the normalized radar difference water index data NDWIR is as shown in formula (3):
[0087] (3);
[0088] in, Characterizes the backscattering coefficient after HH polarization, It represents the backscattering coefficient after VV polarization, HH represents the horizontal-horizontal polarization mode, and VV represents the vertical-vertical polarization mode.
[0089] In one embodiment, the enhanced radar vegetation index data RVIE is as shown in formula (4):
[0090] (4);
[0091] in, Characterizes the backscattering coefficient after VH polarization, It represents the backscattering coefficient after VV polarization, VH represents the vertical-horizontal polarization mode, and VV represents the vertical-vertical polarization mode.
[0092] Optionally, green vegetation index data characterizes the greenness information of sample target type crops, enhanced vegetation index data characterizes the contrast of growth reflectance of sample target type crops, further corrects the atmospheric influence and surface reflectance change, and vegetation moisture index data characterizes the moisture content state of vegetation canopy of sample target type crops. Vegetation heterogeneity index data enhances the detection of sample target type crop types and reduces the influence of atmosphere. Red edge normalized difference vegetation index enhances the sensitive identification of vegetation state.
[0093] In one embodiment, the green vegetation index data NDVI is as shown in formula (5):
[0094] (5);
[0095] in, Characterize the reflectivity of blue light band, Characterize the reflectivity in the near infrared band, Characterizes the reflectivity in the red light band.
[0096] In one embodiment, the enhanced vegetation index data EVI is shown in formula (6):
[0097] (6);
[0098] in, Characterize the reflectivity in the near infrared band, Characterizes the reflectance of the red light band. In one embodiment, the vegetation moisture index data NDWI is shown in formula (7):
[0099] (7);
[0100] in, Characterizes the reflectivity in the shortwave infrared band.
[0101] In one embodiment, the vegetation heterogeneity index data VARI is as shown in formula (8):
[0102] (8);
[0103] in, Characterizes the green band reflectivity.
[0104] In one embodiment, the red edge normalized difference vegetation index data RENDVI is as shown in formula (9):
[0105] (9);
[0106] in, Characterizes the red edge band reflectivity.
[0107] Optionally, the sample benchmark growth feature includes at least one of the following: time series feature, curve statistical feature, curve waveform feature, exponential combination feature;
[0108] Among them, the time series feature characterizes the time variation characteristics of the index data corresponding to the sample target type of crops, the curve statistical feature characterizes the growth morphological characteristics of the index data corresponding to the sample target type of crops, the curve waveform feature characterizes the growth periodic variation characteristics of the index data corresponding to the sample target type of crops, and the index combination feature characterizes the combination of different spectra and radar indices corresponding to the sample target type of crops to generate new composite features to better reflect the characteristics of the crops.
[0109] Optionally, the time series features may be time series features extracted from radar index time series data and spectral index time series data according to an acquisition period.
[0110] Optionally, the curve statistical features may be key points in the sample baseline growth curve, such as peak values, inflection points, etc., and growth slopes at different time periods, to identify growth acceleration or deceleration stages.
[0111] Optionally, the curve waveform characteristics may be morphological characteristics of the sample benchmark growth curve, such as duration and area of the morphological characteristics, and may also include periodic characteristics of the growth cycle of the sample target type of crop, such as whether there are multiple growth peaks.
[0112] Optionally, the index combination feature can be a combination of different radar index time series data and spectral index time series data (such as the ratio of normalized radar vegetation index data NDVI to normalized radar vegetation index data NRVI) to generate a new composite feature, which can better reflect the characteristics of crops. For example, the difference between the same index time series data generated by the same sample target type of crops in the benchmark growth environment and the reference growth environment is calculated to generate a stress response feature.
[0113] Optionally, the sample reference growth characteristics also include at least one of the following: time series characteristics, curve statistical characteristics, curve waveform characteristics, and exponential combination characteristics.
[0114] Optionally, based on the sample benchmark growth curves of different sample target types of crops under a benchmark growth environment and the sample reference growth curves under a reference growth environment, features are extracted from the time domain, statistics, waveform and combination dimensions to comprehensively characterize the characteristic difference information caused by different sample target types of crops under different growth environments, thereby improving the recognition accuracy of the crop recognition model and solving the problem of poor adaptability of the crop recognition model under different environmental conditions.
[0115] Optionally, the sample reference growth state data includes sample reference radar data and sample reference spectrum data; wherein the sample reference growth curve is determined based on the following operations:
[0116] The sample reference radar data are processed according to the radar index function to obtain sample reference radar index data, wherein the sample reference radar index data characterizes the growth density state and structural characteristics of the sample target type of crops; the sample reference spectral data are processed according to the spectral index function to obtain sample reference spectral index data, wherein the sample reference spectral index data characterizes the greenness, health and moisture content state of the sample target type of crops; according to the sample reference radar index data of the sample target type of crops in the reference growth environment, a first sub-reference growth curve corresponding to the stress environment parameter is determined; according to the sample reference spectral index data of the sample target type of crops in the reference growth environment, a second sub-reference growth curve corresponding to the stress environment parameter is determined; according to the first sub-reference growth curve and the second sub-reference growth curve, a sample reference growth curve of the sample target type of crops is determined.
[0117] Optionally, the sample reference radar data are from synthetic aperture radar data of a microwave remote sensing satellite (Senitnel-1), and the sample reference spectral data are from multispectral data of a high-resolution multispectral imaging satellite (Senitnel-2).
[0118] For example, the sample reference radar data after preprocessing is processed according to the radar index function such as formula (1) to formula (4) to obtain the sample reference radar index data; the sample reference spectral data after preprocessing is processed according to the spectral index function such as formula (5) to formula (9) to obtain the sample reference spectral index data.
[0119] Optionally, the sample reference radar index data includes at least the following: radar ratio vegetation index data, normalized radar vegetation index data, normalized radar difference water index data and enhanced radar vegetation index data; the sample reference spectral index data includes at least the following: greenness vegetation index data, enhanced vegetation index data, vegetation moisture index data, vegetation heterogeneity index data and red edge normalized difference vegetation index data.
[0120] Optionally, determining the first sub-reference growth curve corresponding to the stress environment parameter may include: determining the number of pixels in the plot area; obtaining the reference radar average value in the plot area according to the number of pixels in the plot area and the sample reference radar index data on each pixel; processing the reference radar average value using a spline interpolation function to obtain reference radar index time series data; and reconstructing the reference radar index time series data to generate the first sub-reference growth curve. The sample target crop type corresponds to a first sub-reference growth curve under multiple stress environment parameters.
[0121] Optionally, determining the second sub-reference growth curve corresponding to the stress environment parameter may include: determining the number of pixels in the plot area; obtaining the reference spectrum average value in the plot area according to the number of pixels in the plot area and the sample reference spectrum index data on each pixel; processing the reference spectrum average value using a spline interpolation function to obtain reference spectrum index time series data; and reconstructing the reference spectrum index time series data to generate the second sub-reference growth curve. The sample target crop type corresponds to a second sub-reference growth curve under multiple stress environment parameters.
[0122] Optionally, the sample reference growth curve includes a first sub-reference growth curve and a plurality of second sub-reference growth curves.
[0123] Optionally, the crop recognition model includes a convolutional neural network layer, a long short-term memory network layer, a fully connected layer, and an activation layer; wherein the crop recognition model is trained based on the sample baseline growth characteristics, the sample reference growth characteristics, and the classification label to obtain a trained crop recognition model, including:
[0124] The sample benchmark growth characteristics and the sample reference growth characteristics are input into the convolutional neural network layer, and the sample convolution characteristics with time series attributes are output; the sample convolution characteristics with time series attributes are input into the long short-term memory network layer, and the sample deep characteristics are output; the sample deep characteristics are input into the fully connected layer, and the sample fully connected characteristics are output; the sample fully connected characteristics are input into the activation layer, and the sample crop type recognition results are output; according to the sample crop type recognition results and classification labels, the crop recognition model is trained to obtain the trained crop recognition model.
[0125] Optionally, the convolutional neural network layer may be a Convolutional Neural Networks (CNN), and the long short-term memory network layer may be a long short-term memory network (Long Short-Term Memory, LSTM).
[0126] Optionally, the sample baseline growth characteristics and the sample reference growth characteristics are normalized so that the characteristic values are distributed between 0 and 1 or -1 and 1 to accelerate convergence and improve model stability.
[0127] Optionally, the normalized sample baseline growth features and sample reference growth features are fused and then input into the convolutional neural network layer to output sample convolution features with time series attributes.
[0128] Optionally, the long short-term memory network layer includes multiple long short-term memory network sub-layers, each long short-term memory network sub-layer includes an input gate, a forget gate and an output gate, and the time step of the long short-term memory network sub-layer is determined according to the growth cycle of the sample target type of crops and the data resolution.
[0129] Optionally, the convolutional features of samples with time series properties are processed according to the long short-term memory network layer to effectively capture the long-term dependencies in the time series and output deep features of samples with strong time series properties.
[0130] Optionally, add one or more fully connected layers after the LSTM layer to integrate feature information on the time series. The fully connected layers process the deep features of the samples and output the fully connected features of the samples to prepare for the classification task.
[0131] Optionally, the activation layer can be constructed according to the activation function softmax, the fully connected features of the sample are processed according to the activation layer, the output dimension is determined according to the number of sample target crop types, and the sample crop type recognition result is output. For example, the sample target crop types include rice, corn, wheat and soybeans. The sample crop type recognition result can be that the probability of belonging to the rice type is 0.8, the probability of belonging to the corn type is 0.1, the probability of belonging to the wheat type is 0.05, and the probability of belonging to the soybean type is 0.05.
[0132] Optionally, the sample crop species identification results and classification labels are processed according to the loss function to obtain the model recognition accuracy. The model performance can also be evaluated using indicators such as precision and recall to ensure that the model has good classification performance. Hyperparameters are tuned using methods such as grid search, random search, or Bayesian optimization to find the optimal model configuration. For example, model parameters may include but are not limited to the number of convolutional layers, convolution kernel size, number of LSTM layers, number of LSTM units, number of neurons in the fully connected layer, etc.
[0133] Optionally, during the training process, the adaptive moment estimation (ADAM) optimizer is used for optimization, and the mini-batch gradient descent method is used for model training. The size of each batch is determined according to the available computing resources. The initial learning rate is 0.0001, and the learning rate decay strategy is used to gradually reduce the learning rate as the number of training iterations increases. The model performance is monitored on the validation set, and training is stopped when the performance no longer improves to avoid overfitting.
[0134] Optionally, the local temporal characteristics and global temporal dependencies of different sample target types of crops under different phenological and growing environments are considered, and the spatial feature extraction capability of the convolutional neural network layer and the temporal dependency processing capability of the long short-term memory network layer are integrated to more comprehensively capture complex features and improve the stability of identifying sample target types of crops in different plots.
[0135] Optionally, the convolutional neural network layer includes a convolution sublayer, a pooling sublayer, and a normalization sublayer; wherein the sample baseline growth feature and the sample reference growth feature are input into the convolutional neural network layer, and the sample convolution feature with time series attributes is output, including:
[0136] The sample baseline growth features and the sample reference growth features are input into the convolution sublayer, and the sample subconvolution features are output; the sample subconvolution features are input into the pooling sublayer, and the sample pooling features are output; the sample pooling features are input into the normalization sublayer, and the sample convolution features are output.
[0137] Optionally, the input dimension of the convolution sublayer is determined according to the number of features and the length of the time series of the sample baseline growth feature and the sample reference growth feature. The convolution sublayer uses multiple convolution kernels, each of which is responsible for extracting different local features.
[0138] Optionally, the convolution sublayer is connected to an activation function (Rectified Linear Unit, ReLU), which outputs the sample subconvolution features, introduces a nonlinear activation function, and enhances the model's expressiveness. A regularization function can also be connected after the convolution sublayer to randomly discard some network connections to prevent the model from overfitting.
[0139] Optionally, a maximum pooling layer is used after the convolution layer to process the sample sub-convolution features according to the pooling sub-layer and output the sample pooling features, thereby reducing the number of parameters of the maximum pooling layer and reducing the risk of overfitting.
[0140] Optionally, the pooling layer can be a batch normalization network layer, which inputs the sample pooling features into the normalization sublayer and outputs the sample convolution features to accelerate model training and improve generalization ability.
[0141] Optionally, a crop recognition model is trained according to the sample crop type recognition results and classification labels to obtain a trained crop recognition model, including:
[0142] The sample crop type recognition results and classification labels are processed according to the cross entropy loss function to obtain a loss value; according to the loss value, the model parameters of the crop recognition model are adjusted to obtain a trained crop recognition model.
[0143] Optionally, the sample crop type identification results and classification labels are processed according to a cross entropy loss function to obtain a loss value of the model, and the loss value is used to measure the difference between the sample crop type identification results and the classification labels, and then evaluate the classification performance of the model.
[0144] Optionally, adjust the model parameters of the crop recognition model based on the loss value, use grid search, random search, or Bayesian optimization to perform hyperparameter tuning, and stop training when the performance no longer improves.
[0145] Figure 2 An example schematic diagram of determining a sample crop type identification result according to an embodiment of the present invention is shown.
[0146] like Figure 2As shown, sample benchmark growth state data of sample target type crops in a benchmark growth environment and sample reference growth state data of sample target type crops in a reference growth environment are obtained. The sample benchmark growth state data includes sample benchmark radar data 201 and sample benchmark spectrum data 202, and the sample reference growth state data includes sample reference radar data 203 and sample reference spectrum data 204. Sample benchmark radar index data 205 is obtained based on processing the sample benchmark radar data 201, and sample benchmark spectrum index data 206 is obtained based on processing the sample benchmark spectrum data 202; sample reference radar index data 207 is obtained based on processing the sample reference radar data 203, and sample reference spectrum index data 208 is obtained based on processing the sample reference spectrum data 204.
[0147] Optionally, the sample benchmark radar index data 205 is reconstructed to generate a first sub-benchmark growth curve 209, and the sample benchmark spectral index data 206 is reconstructed to generate a second sub-benchmark growth curve 210; the sample reference radar index data 207 is reconstructed to generate a first sub-reference growth curve 211, and the sample reference spectral index data 208 is reconstructed to generate a second sub-reference growth curve 212; feature extraction is performed on the first sub-benchmark growth curve 209 and the second sub-benchmark growth curve 210 to obtain a sample benchmark growth feature 213, and feature extraction is performed on the first sub-reference growth curve 211 and the second sub-reference growth curve 212 to obtain a sample reference growth feature 214.
[0148] Optionally, the crop recognition model includes a convolutional neural network layer, a long short-term memory network layer, a fully connected layer, and an activation layer. The sample baseline growth characteristics 213 and the sample reference growth characteristics 214 are processed according to the crop recognition model to obtain the sample crop type recognition result 215.
[0149] Figure 3 A flow chart of a crop identification method according to an embodiment of the present invention is shown.
[0150] like Figure 3 As shown, the method 300 includes operations S310 to S330.
[0151] In operation S310, growth status data of a target crop is acquired.
[0152] In operation S320, feature extraction is performed on the growth curve to obtain growth features.
[0153] In operation S330, the growth characteristics are input into the trained crop recognition model, and a crop type recognition result of the target crop is output.
[0154] Optionally, the growth status data of the target crop in each plot area is obtained from the remote sensing data taken by the satellite. The target crop represents the crop grown in the plot area. The growth status data represents the state data such as the morphology and water content of the target crop in the growth environment in the plot area. The growth status data is processed to obtain a growth curve.
[0155] Optionally, feature extraction is performed on the growth curve to obtain growth features, which characterize the characteristics of the growth curve in the time domain, statistics, waveform and other dimensions. The growth features are input into the trained crop recognition model to output the crop type recognition results of the target crops in different plot areas. For example, the crop type recognition result of the target crop in plot area 1 is corn, and the crop type recognition result of the target crop in plot area 2 is soybean.
[0156] Optionally, multiple crop type recognition results are classified to obtain the crop types in each plot area, which can be divided into: rice, corn, wheat, soybeans, other crops and non-crops. Finally, the classification results are used to draw a detailed spatial distribution map of regional crops.
[0157] Based on the above-mentioned crop recognition model training method, the present invention also provides a crop recognition model training device. Figure 4 The device is described in detail.
[0158] Figure 4 A structural block diagram of a crop recognition model training device according to an embodiment of the present invention is shown.
[0159] like Figure 4 As shown, the crop recognition model training device 400 of this embodiment includes a first acquisition module 410 , a first feature extraction module 420 and a training module 430 .
[0160] The first acquisition module 410 is used to acquire training samples, the training samples including sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of sample target type crops in a reference growth environment, and classification labels, wherein the reference environmental conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment, the reference environmental conditions are determined based on the stress environment parameters, and the classification labels characterize the types of sample target type crops. In one embodiment, the first acquisition module 410 can be used to perform the operation S110 described above, which will not be repeated here.
[0161] The first feature extraction module 420 is used to extract features from the sample baseline growth curve and the sample reference growth curve respectively to obtain the sample baseline growth feature and the sample reference growth feature, wherein the sample baseline growth curve is determined based on the sample baseline growth state data, and the sample reference growth curve is determined based on the sample reference growth state data. In one embodiment, the first feature extraction module 420 can be used to perform the operation S120 described above, which will not be repeated here.
[0162] The training module 430 is used to train the crop recognition model based on the sample baseline growth characteristics, the sample reference growth characteristics and the classification label to obtain a trained crop recognition model. In one embodiment, the training module 430 can be used to perform the operation S130 described above, which will not be repeated here.
[0163] Optionally, the first feature extraction module 420 includes a first determination submodule, a second determination submodule, a third determination submodule, a fourth determination submodule and a fifth determination submodule.
[0164] The first determination submodule is used to process the sample reference radar data according to the radar index function to obtain the sample reference radar index data, wherein the sample reference radar index data represents the growth density state of the sample target type of crops.
[0165] The second determination submodule is used to process the sample reference spectral data according to the spectral index function to obtain sample reference spectral index data, wherein the sample reference spectral index data represents the moisture content state of the sample target type of crop.
[0166] The third determination submodule is used to determine a first sub-reference growth curve corresponding to the stress environment parameter according to the sample reference radar index data of the sample target type crop in the reference growth environment.
[0167] The fourth determination submodule is used to determine a second sub-reference growth curve corresponding to the stress environment parameter according to the sample reference spectral index data of the sample target type crop in the reference growth environment.
[0168] The fifth determination submodule is used to determine the sample reference growth curve of the sample target type crop according to the first sub-reference growth curve and the second sub-reference growth curve.
[0169] Optionally, the first feature extraction module 420 further includes a first extraction submodule, a second extraction submodule, a third extraction submodule, a fourth extraction submodule, a fifth extraction submodule, a sixth extraction submodule, a seventh extraction submodule and an eighth extraction submodule.
[0170] The first extraction submodule is used to determine the number of pixels in the plot area, wherein the sample target type of crops grows in the plot area.
[0171] The second extraction submodule is used to obtain the radar average value on the land plot area according to the number of pixels in the land plot area and the sample benchmark radar index data on each pixel, wherein the sample benchmark radar index data is obtained based on the processing of the sample benchmark radar data.
[0172] The third extraction submodule is used to process the radar average value using a spline interpolation function to obtain radar index time series data.
[0173] The fourth extraction submodule is used to reconstruct the radar index time series data to generate a first sub-reference growth curve.
[0174] The fifth extraction submodule is used to obtain the spectral average value of the plot area according to the number of pixels in the plot area and the sample benchmark spectral index data of each pixel, wherein the sample benchmark spectral index data is obtained based on the processing of the sample benchmark spectral benchmark data.
[0175] The sixth extraction submodule is used to process the spectrum average value using a spline interpolation function to obtain spectrum index time series data.
[0176] The seventh extraction submodule is used to reconstruct the spectral index time series data to generate a second sub-reference growth curve.
[0177] The eighth extraction submodule is used to obtain a sample benchmark growth curve of the sample target type of crop according to the first sub-benchmark growth curve and the second sub-benchmark growth curve.
[0178] Optionally, the training module 430 includes a first training submodule, a second training submodule, a third training submodule, a fourth training submodule and a fifth training submodule.
[0179] The first training submodule is used to input the sample baseline growth features and the sample reference growth features into the convolutional neural network layer, and output the sample convolution features with time series attributes.
[0180] The second training submodule is used to input the sample convolution features with time series attributes into the long short-term memory network layer and output the deep features of the samples.
[0181] The third training submodule is used to input the deep features of the sample into the fully connected layer and output the fully connected features of the sample.
[0182] The fourth training submodule is used to input the fully connected features of the samples into the activation layer and output the sample crop type recognition results.
[0183] The fifth training submodule is used to train the crop recognition model according to the sample crop type recognition results and classification labels to obtain a trained crop recognition model.
[0184] Optionally, the first training submodule includes a first training unit, a second training unit and a third training unit.
[0185] The first training unit is used to input the sample baseline growth feature and the sample reference growth feature into the convolution sublayer, and output the sample subconvolution feature.
[0186] The second training unit is used to input the sample sub-convolution features into the pooling sub-layer and output the sample pooling features.
[0187] The third training unit is used to input the sample pooling features into the normalization sublayer and output the sample convolution features.
[0188] Optionally, the fifth training submodule includes a fourth training unit and a fifth training unit.
[0189] The fourth training unit is used to process the sample crop type recognition results and classification labels according to the cross entropy loss function to obtain a loss value.
[0190] The fifth training unit is used to adjust the model parameters of the crop recognition model according to the loss value to obtain a trained crop recognition model.
[0191] Based on the above crop identification method, the present invention also provides a crop identification device. Figure 5 The device is described in detail.
[0192] Figure 5 A structural block diagram of a crop identification device according to an embodiment of the present invention is shown.
[0193] like Figure 5 As shown, the crop identification device 500 of this embodiment includes a second acquisition module 510 , a second feature extraction module 520 and an identification module 530 .
[0194] The second acquisition module 510 is used to acquire the growth status data of the target crop. In one embodiment, the second acquisition module 510 can be used to perform the operation S310 described above, which will not be described in detail here.
[0195] The second feature extraction module 520 is used to extract features from the growth curve to obtain growth features, wherein the growth curve is generated based on the growth state data. In one embodiment, the second feature extraction module 520 can be used to perform the operation S320 described above, which will not be described in detail here.
[0196] The identification module 530 is used to input the growth characteristics into the trained crop identification model and output the crop type identification result of the target crop. In one embodiment, the identification module 530 can be used to perform the operation S330 described above, which will not be described in detail here.
[0197] Optionally, any multiple modules among the modules, submodules, units, and subunits can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. Optionally, at least one of the modules, submodules, units, and subunits can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or implemented in any one of the three implementation methods of software, hardware, and firmware, or in any appropriate combination of any of them. Alternatively, at least one of the modules, submodules, units, and subunits can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be executed.
[0198] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present invention. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box may also occur in an order different from that marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flowchart, and the combination of boxes in the block diagram or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions. It can be understood by those skilled in the art that the features recorded in the various embodiments and / or claims of the present invention can be combined and / or combined in various ways, even if such a combination or combination is not explicitly recorded in the present invention. In particular, without departing from the spirit and teachings of the present invention, the features described in the various embodiments and / or claims of the present invention may be combined and / or combined in a variety of ways. All of these combinations and / or combinations fall within the scope of the present invention.
[0199] The embodiments of the present invention are described above. However, these embodiments are only for the purpose of illustration, and are not intended to limit the scope of the present invention. Although each embodiment is described above, it does not mean that the measures in each embodiment cannot be used in combination. The scope of the present invention is defined by the attached claims and their equivalents. Without departing from the scope of the present invention, those skilled in the art may make various substitutions and modifications, which should all fall within the scope of the present invention.
Claims
1. A method for training a crop recognition model, characterized in that: The method comprises: Acquire training samples, the training samples including sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of the sample target type crops in a reference growth environment, and classification labels, wherein reference environmental conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment, the reference environmental conditions are determined based on stress environment parameters, and the classification labels characterize the types of the sample target type crops; Feature extraction is performed on the sample benchmark growth curve and the sample reference growth curve respectively to obtain the sample benchmark growth feature and the sample reference growth feature, wherein the sample benchmark growth curve is determined based on the sample benchmark growth state data, the sample reference growth curve is determined based on the sample reference growth state data, the sample benchmark growth state data includes sample benchmark radar data and sample benchmark spectral data, and the sample reference growth state data includes sample reference radar data and sample reference spectral data, Wherein, the sample reference growth curve is determined based on the following operations: Processing the sample reference radar data according to a radar index function to obtain sample reference radar index data, wherein the sample reference radar index data characterizes the growth density state of the sample target type of crops; processing the sample reference spectral data according to a spectral index function to obtain sample reference spectral index data, wherein the sample reference spectral index data characterizes the moisture content state of the sample target type of crops; determining a first sub-reference growth curve corresponding to the stress environment parameter according to the sample reference radar index data of the sample target type of crops in a reference growth environment; determining a second sub-reference growth curve corresponding to the stress environment parameter according to the sample reference spectral index data of the sample target type of crops in a reference growth environment; determining the sample reference growth curve of the sample target type of crops according to the first sub-reference growth curve and the second sub-reference growth curve; The sample baseline growth curve is determined based on the following operations: Determine the number of pixels in a plot area, wherein the sample target type crop grows in the plot area; obtain a radar average value in the plot area according to the number of pixels in the plot area and the sample benchmark radar index data on each pixel, wherein the sample benchmark radar index data is obtained based on processing the sample benchmark radar data; process the radar average value using a spline interpolation function to obtain radar index time series data; reconstruct the radar index time series data to generate a first sub-benchmark growth curve; obtain a spectral average value in the plot area according to the number of pixels in the plot area and the sample benchmark spectral index data on each pixel, wherein the sample benchmark spectral index data is obtained based on processing the sample benchmark spectral data; process the spectral average value using the spline interpolation function to obtain spectral index time series data; reconstruct the spectral index time series data to generate a second sub-benchmark growth curve; obtain the sample benchmark growth curve of the sample target type crop according to the first sub-benchmark growth curve and the second sub-benchmark growth curve; Based on the sample baseline growth characteristics, the sample reference growth characteristics and the classification labels, a crop recognition model is trained to obtain a trained crop recognition model.
2. The method according to claim 1, characterized in that The sample benchmark radar index data includes at least one of the following: radar ratio vegetation index data, normalized radar vegetation index data, normalized radar difference water index data and enhanced radar vegetation index data; the sample benchmark spectral index data includes at least one of the following: greenness vegetation index data, enhanced vegetation index data and vegetation moisture index data.
3. The method according to claim 1, characterized in that The sample benchmark growth characteristics include at least one of the following: time series characteristics, curve statistical characteristics, curve waveform characteristics, and exponential combination characteristics; Among them, the time series characteristics represent the time change characteristics of the index data corresponding to the sample target type of crops, the curve statistical characteristics represent the growth morphological characteristics of the sample target type of crops, the curve waveform characteristics represent the growth periodic change characteristics of the sample target type of crops, and the index combination characteristics represent the difference characteristics of the same index time series data generated by the sample target type of crops under the benchmark growth environment and the reference growth environment.
4. The method according to claim 1, characterized in that: The crop recognition model includes a convolutional neural network layer, a long short-term memory network layer, a fully connected layer, and an activation layer; The step of training a crop recognition model based on the sample baseline growth characteristics, the sample reference growth characteristics and the classification label to obtain a trained crop recognition model includes: Inputting the sample baseline growth feature and the sample reference growth feature into the convolutional neural network layer, and outputting a sample convolution feature with a time series attribute; Input the sample convolution features with time series attributes into the long short-term memory network layer, and output the sample deep features; Input the sample deep features into the fully connected layer, and output the sample fully connected features; Inputting the sample fully connected features into the activation layer, and outputting the sample crop type recognition result; A crop recognition model is trained according to the sample crop type recognition results and the classification labels to obtain a trained crop recognition model.
5. The method according to claim 4, characterized in that The convolutional neural network layer includes a convolution sublayer, a pooling sublayer, and a normalization sublayer; The sample baseline growth feature and the sample reference growth feature are input into the convolutional neural network layer, and the sample convolution feature with time series attributes is output, including: Inputting the sample baseline growth feature and the sample reference growth feature into the convolution sublayer, and outputting the sample subconvolution feature; Input the sample subconvolution features into the pooling sublayer, and output the sample pooling features; The sample pooling features are input to the normalization sublayer, and the sample convolution features are output.
6. The method according to claim 4, characterized in that Training a crop recognition model according to the sample crop type recognition result and the classification label to obtain a trained crop recognition model includes: Processing the sample crop type identification result and the classification label according to a cross entropy loss function to obtain a loss value; According to the loss value, the model parameters of the crop recognition model are adjusted to obtain a trained crop recognition model.
7. A crop identification method, characterized in that: The method comprises: Obtain growth status data of target crops; Extracting features from a growth curve to obtain growth features, wherein the growth curve is generated based on the growth status data; The growth characteristics are input into a trained crop identification model, and a crop type identification result of the target crop is output, wherein the trained crop identification model is obtained according to any one of claims 1 to 6.
8. A training device for a crop recognition model, characterized in that: include: A first acquisition module is used to acquire training samples, wherein the training samples include sample benchmark growth state data of sample target type crops in a benchmark growth environment, sample reference growth state data of the sample target type crops in a reference growth environment, and classification labels, wherein the reference environmental conditions of the reference growth environment are different from the benchmark growth conditions of the benchmark growth environment, the reference environmental conditions are determined based on stress environment parameters, and the classification labels characterize the types of the sample target type crops; a first feature extraction module, for extracting features from a sample benchmark growth curve and a sample reference growth curve respectively, to obtain a sample benchmark growth feature and a sample reference growth feature, wherein the sample benchmark growth curve is determined based on the sample benchmark growth state data, the sample reference growth curve is determined based on the sample reference growth state data, the sample benchmark growth state data includes sample benchmark radar data and sample benchmark spectral data, the sample reference growth state data includes sample reference radar data and sample reference spectral data, Among them, the first feature extraction module includes a first determination submodule, a second determination submodule, a third determination submodule, a fourth determination submodule and a fifth determination submodule: A first determination submodule is used to process the sample reference radar data according to a radar index function to obtain sample reference radar index data, wherein the sample reference radar index data represents the growth density state of the sample target type crop; A second determination submodule is used to process the sample reference spectral data according to a spectral index function to obtain sample reference spectral index data, wherein the sample reference spectral index data represents the moisture content state of the sample target type crop; A third determination submodule is used to determine a first sub-reference growth curve corresponding to the stress environment parameter according to the sample reference radar index data of the sample target type crop in the reference growth environment; A fourth determination submodule is used to determine a second sub-reference growth curve corresponding to the stress environment parameter according to the sample reference spectral index data of the sample target type crop in the reference growth environment; a fifth determination submodule, configured to determine the sample reference growth curve of the sample target type crop according to the first sub-reference growth curve and the second sub-reference growth curve; The first feature extraction module further includes a first extraction submodule, a second extraction submodule, a third extraction submodule, a fourth extraction submodule, a fifth extraction submodule, a sixth extraction submodule, a seventh extraction submodule and an eighth extraction submodule: A first extraction submodule is used to determine the number of pixels in a plot area, wherein the sample target type of crop grows in the plot area; A second extraction submodule is used to obtain a radar average value on the land area according to the number of pixels in the land area and the sample baseline radar index data on each pixel, wherein the sample baseline radar index data is obtained based on processing the sample baseline radar data; A third extraction submodule is used to process the radar average value using a spline interpolation function to obtain radar index time series data; A fourth extraction submodule, used for reconstructing the radar index time series data to generate a first sub-reference growth curve; a fifth extraction submodule, configured to obtain a spectrum average value on the land area according to the number of pixels in the land area and sample benchmark spectrum index data on each pixel, wherein the sample benchmark spectrum index data is obtained based on processing the sample benchmark spectrum data; a sixth extraction submodule, configured to process the spectrum average value by using the spline interpolation function to obtain spectrum index time series data; a seventh extraction submodule, configured to reconstruct the spectral index time series data to generate a second sub-reference growth curve; an eighth extraction submodule, configured to obtain the sample benchmark growth curve of the sample target crop type according to the first sub-benchmark growth curve and the second sub-benchmark growth curve; The training module is used to train a crop recognition model based on the sample baseline growth characteristics, the sample reference growth characteristics and the classification label to obtain a trained crop recognition model.
