A crop planting fine management method and system

Abstract

This invention provides a method and system for refined management of crop planting, relating to the field of agricultural information technology. The method includes: conducting roving photography in a crop planting area to obtain sampled images of multiple crops in multiple zones, and inputting these images into a species identification model to obtain crop species information and growth stage information; summarizing the sampled images to obtain a set of sampled images of crops with the same species information; determining whether environmental parameters need adjustment based on the crop growth stage information and environmental parameters in the sampled images in the sampled image set; if adjustment is required, determining the adjusted environmental parameters based on a preset growth stage, growth stage information, and environmental parameters in multiple zones. According to this invention, it is possible to determine whether environmental parameters in each zone need adjustment, facilitating the management of crop growth stages and growth status, and promoting crops to reach the ideal growth stage.

Description

Technical Field

[0001] This invention relates to the field of agricultural information technology, and in particular to a method and system for refined management of crop planting. Background Technology

[0002] Among related technologies, CN118014758A relates to the field of smart agriculture technology and discloses an intelligent management system for crop planting, including: an information data acquisition module, a network transmission unit, an information data storage unit, a core processing module, a planting plan adjustment unit, an anomaly alarm unit, a plan feedback unit, and a terminal display module. The information data acquisition module is used to collect information data on factors affecting planting; the network transmission unit is used to transmit the information data collected by the information data acquisition module over the network. By collecting weather information, soil information, and crop information, it is possible to analyze and determine the weather conditions, soil texture, and crop status over a period of time. Furthermore, based on the analyzed weather conditions and soil texture information, it can match the most suitable crops for planting, adjust the planting plan, and then feed the adjusted planting plan back to the management personnel.

[0003] CN119048264A relates to the field of agricultural information technology, and in particular to a method and system for collecting and analyzing crop planting information. The method involves polling crop seeds to obtain historical germination data; obtaining growth monitoring data of crop seeds through a growth monitoring system and / or equipment; acquiring multiple stored germination experiment cycles from the crop seeds; wherein the germination experiment cycle is the difference between the storage time stamp and the experimental germination time stamp of each stored seed; and generating integrated vigor data of the crop seeds based on the vigor data, wherein the vigor data includes historical germination data, growth monitoring data, and germination experiment cycles. This solution can be used for superior seed selection of oilseed crops, improving the germination rate of oilseed crops, and thus promoting the improvement of oilseed crop breeding.

[0004] Therefore, while related technologies can monitor the planting and growth of crops through information technology, they are difficult to manage and optimize the planting and growth environment of crops, as well as the growth stages and states of crops, making it difficult to enable crops to reach the ideal growth stage.

[0005] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0006] This invention provides a method and system for refined management of crop planting, which can solve the technical problem that related technologies are difficult to use to manage the growth stages and growth status of crops.

[0007] According to a first aspect of the present invention, a method for refined management of crop cultivation is provided, comprising:

[0008] The photographer toured the crop-growing areas to capture sampled images of multiple crops in multiple zones.

[0009] Multiple sampled images are input into the species identification model to obtain information on the species and growth stage of crops in each partition;

[0010] By summarizing sampled images of the same species in different zones, a set of sampled images of crops with the same species information is obtained.

[0011] Based on the crop growth stage information in each sampled image in the sampled image set, and the environmental parameters in multiple partitions, determine whether the environmental parameters in each partition need to be adjusted.

[0012] If adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, the crop growth stage information in each sampled image, and the environmental parameters in multiple partitions.

[0013] According to a second aspect of the present invention, a precision management system for crop cultivation is provided, comprising:

[0014] The shooting module is used to take pictures in a circular manner in the crop planting area to obtain sampled images of multiple crops in multiple zones;

[0015] The identification module is used to input multiple sampled images into the species identification model to obtain information on the species and growth stage of crops in each partition;

[0016] The collection module is used to summarize sampled images of the same species in different partitions to obtain a collection of sampled images of crops with the same species information.

[0017] The judgment module is used to determine whether the environmental parameters in each partition need to be adjusted based on the crop growth stage information in each sampled image in the sampled image set and the environmental parameters in multiple partitions.

[0018] The adjustment module is used to determine the adjusted environmental parameters based on the preset growth stage, the crop growth stage information in each sampled image, and the environmental parameters in multiple partitions if adjustment is required.

[0019] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0020] According to the present invention, a species identification model can identify the types and growth stages of crops planted in multiple zones. This allows for statistical analysis of the growth stages of the same type of crops planted in different zones, and based on the statistical results, it can determine whether environmental parameters in each zone need adjustment to ensure that crops in each zone are in an ideal growth environment. This facilitates the management of crop growth stages and states, promoting crops to reach their ideal growth stages. When training the species identification model, a growth stage identification module can determine the crop growth stage feature vector. This feature vector, combined with the first sample feature map, is used to comprehensively determine the crop species and growth stage, thereby improving the accuracy of species identification for crops at different growth stages. Furthermore, during training, multiple training processes can be used to train the species identification model's capabilities in growth stage identification, species identification, and other aspects, improving training efficiency and the accuracy of the species identification model. When determining the realism loss function, adversarial training can be used to determine the realism loss function, thereby improving the realism of the images generated by the image generation model while simultaneously enhancing the discrimination ability of the discrimination model. This balances the performance of the two models and improves the realism of the first sample generated image. When determining the generation accuracy loss function, it can be composed of the cosine similarity between the generated species information and the labeled species information, and the cosine similarity between the generated growth stage information and the labeled growth stage information. This improves the consistency between the generated species information and the labeled information of the first sample image during training, while reducing the errors between the trained species information and the labeled information of the first sample image, the errors between the generated species information and the trained species information, and the errors between the generated growth stage information and the trained growth stage information. This simultaneously improves the accuracy of the image generation model and the accuracy of the species recognition model in identifying the species and growth stage of crops in the image. When determining the second comprehensive loss function, a second branch can be used to automatically supervise the training of the first branch. That is, the second branch participates in the accuracy competition of the first branch, thus urging the first branch to improve its accuracy during training. Furthermore, the second branch also continuously improves its accuracy during training, participating in the competition with a higher standard, thus supervising the training of the first branch with a higher standard, thereby improving the accuracy of the first branch, and ultimately improving the accuracy of the species recognition model and training efficiency.When determining the third comprehensive loss function, the appearance errors of crops in terms of morphology and color in the first sample image and the first sample generated image can be used to determine the third comprehensive loss function. In the solution process, the first training probability distribution is used as an amplification coefficient to reflect the influence of the realism of the image generated by the image generation model on the second morphological feature information and the second color feature information. This can be applied to the training of the image generation model and the species recognition model to amplify the appearance error. At the same time, it can also improve the first training probability distribution, further improve the realism of the image generated by the image generation model, and improve training efficiency.

[0021] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0023] Figure 1 An exemplary flowchart of a refined management method for crop cultivation according to an embodiment of the present invention is shown.

[0024] Figure 2 An exemplary schematic diagram of the training of a species identification model according to an embodiment of the present invention is shown;

[0025] Figure 3 A block diagram of a precision crop planting management system according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0028] Figure 1 An exemplary flowchart illustrates a method for refined management of crop cultivation according to an embodiment of the present invention, the method comprising:

[0029] Step S101: Conduct roving photography in the crop planting area to obtain sampled images of multiple crops in multiple zones;

[0030] Step S102: Input multiple sampled images into the species recognition model to obtain species information and growth stage information of crops in each partition;

[0031] Step S103: Summarize the sampled images of the same species in different partitions to obtain a set of sampled images of crops with the same species information.

[0032] Step S104: Based on the crop growth stage information in each sampled image in the sampled image set and the environmental parameters in multiple partitions, determine whether the environmental parameters in each partition need to be adjusted.

[0033] Step S105: If adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, the crop growth stage information in each sampled image, and the environmental parameters in multiple partitions.

[0034] According to an embodiment of the present invention, the refined management method for crop planting can identify the types and growth stages of crops planted in multiple zones through a species identification model. This allows for the statistical analysis of the growth stage information of the same type of crops planted in different zones. Based on the statistical results, it can be determined whether the environmental parameters of each zone need to be adjusted so that the crops in each zone are in an ideal growth environment. This is beneficial for the management of the growth stages and growth status of crops and promotes the crops to reach the ideal growth stage.

[0035] According to one embodiment of the present invention, in step S101, the planting area can be a vegetable greenhouse or other planting area. The vegetable greenhouse may include multiple zones, each relatively independent, with independently adjustable environmental parameters such as air temperature, air humidity, soil moisture, light intensity, and daily light duration. Manual roving photography or the use of drones, robots, or other equipment can be used to obtain sampled images of crops within each zone. Each zone may only grow one type of crop, and when acquiring sampled images, multiple sampled images of crops can be obtained from each zone.

[0036] According to an embodiment of the present invention, in step S102, the species identification model can be used to obtain species information and growth stage information of crops in the sampled images. Both species information and growth stage information can be in vector form. The species identification model can obtain species information and growth stage information of multiple sampled images within a unified partition and solve for the average value to obtain the species information and growth stage information of crops within the partition.

[0037] Figure 2 An exemplary schematic diagram of the training of a species identification model according to an embodiment of the present invention is shown.

[0038] According to one embodiment of the present invention, the species identification model is a deep learning neural network model, such as a convolutional neural network model. The training steps of the species recognition model include: inputting a first sample image into the encoding module of the species recognition model to obtain a first sample feature map of the first sample image, wherein the first sample image is an image of a crop of a specific species at a specific growth stage; inputting the first sample feature map into the growth stage recognition module of the species recognition model to obtain a growth stage feature vector of the crop in the first sample image; inputting the growth stage feature vector and the first sample feature map into the decoding module of the species recognition model to obtain training species information and training growth stage information of the crop in the first sample image; obtaining a first sample generated image based on the training species information and the training growth stage information, and an image generation model; obtaining a first comprehensive loss function based on the first sample generated image, the discrimination model, and the species recognition model; determining a second comprehensive loss function based on the growth stage feature vector, the training species information, and the training growth stage information; obtaining a third comprehensive loss function based on the first sample generated image and the first sample image, and an image feature recognition model; and training the species recognition model based on the first comprehensive loss function, the second comprehensive loss function, and the third comprehensive loss function to obtain a trained species recognition model.

[0039] According to one embodiment of the present invention, during the training process, a first sample image can be used for training. The first sample image is an image of a specific species of crop at a specific growth stage, and the first sample image has annotation information, including annotation of species type information and annotation of growth stage information, which can be used as error-free species type information and growth stage information of the crop in the first sample image.

[0040] According to an embodiment of the present invention, the encoding module of the species identification model may include multiple layers such as convolutional layers, activation layers, and pooling layers, and can encode the first sample image to obtain a first sample feature map of the first sample image. The number of first sample feature maps is greater than that of the first sample image, but the resolution of each first sample feature map is lower than that of the first sample image. Each first sample feature map can be used to represent one aspect of the features of the first sample image.

[0041] According to one embodiment of the present invention, the growth stage identification module may also include multiple layers, such as convolutional layers, activation layers, pooling layers, etc., which can be used to process the first sample feature map. Multiple first sample feature maps can be input into the growth stage identification module respectively to obtain growth stage feature vectors. The growth stage feature vectors are vector-type data that can be used to describe which growth stage the crop is in, such as germination stage, seedling stage, tillering stage, jointing stage, spikelet growth stage, and grain filling stage.

[0042] According to one embodiment of the present invention, growth stage feature vectors and a first sample feature map can be input into the decoding module of a species recognition model to obtain training species information and training growth stage information of crops in the first sample image. The decoding module may include multiple layers such as deconvolution layers and fully connected layers, which can improve the resolution of the first sample feature map, reduce the number of first sample feature maps, and finally output training species information and training growth stage information through a fully connected layer. The training species information can be used to describe the species of crops in the first sample image. Similar to the above-mentioned species information, the training species information can be data in vector form. Similarly, the training growth stage information can also be data in vector form, similar to the above-mentioned growth stage information, and can be used to determine the growth stage of the crop. For example, each component of the vector represents the probability of multiple stages, such as a 1% probability of being in the germination stage, a 1% probability of being in the seedling stage, a 1% probability of being in the tillering stage, a 1% probability of being in the jointing stage, a 1% probability of being in the heading stage, and a 95% probability of being in the fruiting stage. Therefore, the training growth stage information is... .

[0043] According to one embodiment of the present invention, the image generation model is also a deep learning neural network model, which can be used to generate images based on vector representation information. For example, it can generate images of corresponding crops based on training species information and training growth stage information, that is, generate a first sample generated image. Furthermore, it can determine the error of the training species information and training growth stage information output by the species recognition model based on the characteristics of the first sample generated image.

[0044] According to an embodiment of the present invention, obtaining a first comprehensive loss function based on the first sample generated image, a discriminant model, and a species identification model includes: inputting the first sample generated image into the discriminant model to obtain a first training probability distribution that the first sample generated image is a real image; inputting the first sample image into the discriminant model to obtain a second training probability distribution that the first sample image is a real image; inputting the first sample generated image into the species identification model to obtain the species type information and growth stage information of the first sample generated image; and determining the first comprehensive loss function based on the first training probability distribution, the second training probability distribution, the species type information, the growth stage information, and the annotation information of the first sample image.

[0045] According to one embodiment of the present invention, the discrimination model is a deep learning neural network model, such as a convolutional neural network model, which can be used to determine whether the input image is a real photographed image or a virtual image generated by the model, and can output the probability that the image is a real photographed image. A first sample generated image can be input into the discrimination model to obtain a first training probability distribution, and the first sample image can also be input into the discrimination model to obtain a second training probability distribution. Furthermore, the first sample generated image can be input into a species identification model to obtain information about the species type and growth stage of the first sample generated image. During training, the realism of images generated by the image generation model and the accuracy of the discrimination model can be continuously improved, resulting in a combined performance enhancement. This allows the image generation model to generate sufficiently realistic images, even when the discrimination model has high accuracy, making them difficult for the discrimination model to distinguish. Furthermore, during training, the species of crops in the images generated by the image generation model can be matched to the training descriptions of species and growth stages. This ensures that the species and growth stage information obtained by the species recognition model from the first sample image matches the labeled information of the first sample image. Since the first sample image is generated through the image generation model... The species type information and growth stage information are obtained based on the training of the first sample image, and the generated species type information and growth stage information are obtained by the species recognition model based on the first sample image. Therefore, by making the generated species type information and growth stage information consistent with the annotation information of the first sample image, the species recognition model and the image generation model can be mutually verified and trained simultaneously. That is, the errors between the training species type information and training growth stage information and the annotation information of the first sample image, the errors between the generated species type information and the training species type information, and the errors between the generated growth stage information and the training growth stage information can be reduced simultaneously. This can improve the accuracy of the image generated by the image generation model and the accuracy of the species recognition model in recognizing the species type and growth stage of crops in the image.

[0046] According to one embodiment of the present invention, determining a first comprehensive loss function based on the first training probability distribution, the second training probability distribution, the generated species information, the generated growth stage information, and the annotation information of the first sample image includes: obtaining the labeled species information and labeled growth stage information of crops in the first sample image based on the annotation information of the first sample image; determining a realism loss function based on the first training probability distribution and the second training probability distribution; determining a generation accuracy loss function based on the generated species information, the generated growth stage information, the labeled species information, and the labeled growth stage information; and determining the first comprehensive loss function based on the realism loss function and the generation accuracy loss function.

[0047] According to an embodiment of the present invention, determining the realism loss function based on the first training probability distribution and the second training probability distribution includes: determining the realism loss function according to formula (1). ,

[0048] (1)

[0049] in, The first training probability distribution is generated for the i-th first sample to be a real image. Let be the second training probability distribution where the i-th first sample image is a real image, and n be the number of first sample images in a training batch, i ≤ n, and both i and n are positive integers. Indicates according to Minimize the direction to adjust the parameters of the image generation model. Indicates according to The parameters of the discrimination model are adjusted according to the direction of maximization.

[0050] According to one embodiment of the present invention, theoretically, the probability of the discriminative model classifying the first sample image as a real image is 100%, and the probability of classifying the first sample generated image as a real image is 0. However, as the training process progresses, the realism of the generation model increases, leading to an increase in the probability that the discriminative model classifies the first sample generated image as a real image. Furthermore, as the discriminative model also improves its discrimination ability during training, the probability that the discriminative model classifies the first sample generated image as a real image decreases again. Ultimately, the performance of the discriminative model and the image generation model can be balanced, so that even when the discrimination accuracy of the discriminative model is very high, it is still difficult to determine whether the first sample generated image generated by the image generation model is a real image, that is, the realism of the image generated by the image generation model is high.

[0051] According to an embodiment of the present invention, following the training method described above that balances the performance of the image generation model and the discriminative model, the discriminative model can be adjusted in the direction that maximizes the realism loss function represented by formula (1), and the image generation model can be adjusted in the direction that maximizes the realism loss function. In formula (1), if the input image is the first sample image, then during the training of the discriminative model, the image generation model is adjusted in the direction that maximizes the realism loss function. Maximize, that is, get as close to 1 as possible, so that it can make To maximize the accuracy of the discriminative model in judging real images, if the input image is a generated image of the first sample, then during the training of the discriminative model, the accuracy of the model should be improved. Minimize, that is, get as close to 0 as possible, so that it can make Maximize this, thereby improving the accuracy of the discrimination model in judging the generated images.

[0052] According to one embodiment of the present invention, on the other hand, in formula (1), if the input image is a first sample generated image, then during the training of the image generation model, make Maximize, that is, get as close to 1 as possible, so that it can make Minimizing this increases the probability that the first sample image generated by the image generation model will be recognized as a real image by the discrimination model, thereby improving the realism of the images generated by the image generation model and improving the performance of the image generation model.

[0053] In this way, the realism loss function can be determined through adversarial training, thereby improving the realism of the images generated by the image generation model while enhancing the discrimination ability of the discrimination model. This balances the performance of the two models and improves the realism of the images generated by the first sample.

[0054] According to an embodiment of the present invention, determining a generation accuracy loss function based on the generated species information, the generated growth stage information, the labeled species information, and the labeled growth stage information includes: determining the generation accuracy loss function according to formula (2). ,

[0055] (2)

[0056] in, Generate species information for the image of the i-th first sample. Generate growth stage information for the i-th first sample image. The annotation species information for the i-th first sample image, This provides the labeled growth stage information for the i-th first sample image. for The transpose of , for The transpose of , where n is the number of the first sample images in a training batch, i ≤ n, and i and n are both positive integers. and Preset weights.

[0057] According to one embodiment of the present invention, The cosine similarity between the generated species type information and the labeled species type information can be used to describe the consistency between the generated and labeled species type information. The cosine similarity between the generated growth stage information and the labeled growth stage information can be used to describe the consistency between the generated growth stage information and the labeled growth stage information. When the generated species information and the generated growth stage information match the labeled information of the first sample image, the two cosine similarities mentioned above are theoretically both 1. Therefore, the error between 1 and the two cosine similarities mentioned above can form the generation accuracy loss function. That is, the error between 1 and the two cosine similarities mentioned above is weighted and summed, and the weighted sum value corresponding to each first sample image in a training batch is summed to obtain the generation accuracy loss function. During the training process, the generation accuracy loss function is reduced, thereby reducing the error between 1 and the two cosine similarities mentioned above, and improving the consistency between the generated species information and the generated growth stage information and the labeled information of the first sample image.

[0058] In this way, the generation accuracy loss function can be composed of the cosine similarity between the generated species information and the labeled species information, and the cosine similarity between the generated growth stage information and the labeled growth stage information. This improves the consistency between the generated species information and the labeled information of the first sample image during training, while reducing the errors between the trained species information and the labeled information of the first sample image, the errors between the generated species information and the trained species information, and the errors between the generated growth stage information and the trained growth stage information. This can simultaneously improve the accuracy of the image generation model in generating images and the accuracy of the species recognition model in identifying the species and growth stage of crops in images.

[0059] According to one embodiment of the present invention, the above-mentioned generation accuracy loss function and authenticity loss function can be weighted and summed to obtain a first comprehensive loss function.

[0060] According to one embodiment of the present invention, the accuracy of the growth stage feature vector can be significantly improved, thereby enhancing the accuracy of the training species information and training growth stage information obtained based on the growth stage feature vector.

[0061] According to an embodiment of the present invention, determining a second comprehensive loss function based on the growth stage feature vector, the training species information, and the training growth stage information includes: decoding the growth stage feature vector through a growth stage information decoding module to obtain first growth stage information; obtaining labeled species information and labeled growth stage information of crops in the first sample image based on the labeling information of the first sample image; and determining a second comprehensive loss function based on the first growth stage information, the training species information, the training growth stage information, the labeled species information, and the labeled growth stage information.

[0062] According to one embodiment of the present invention, the growth stage information decoding module may include layers such as deconvolution layers and fully connected layers, and can decode the growth stage feature vectors separately to obtain the first growth stage information.

[0063] According to an embodiment of the present invention, determining a second comprehensive loss function based on the first growth stage information, the training species information, the training growth stage information, the labeled species information, and the labeled growth stage information includes: determining the second comprehensive loss function according to formula (3). ,

[0064] (3)

[0065] in, This represents the training and growth stage information for the i-th first sample image. This provides the labeled growth stage information for the i-th first sample image. for The transpose of , This represents the information of the first growth stage corresponding to the i-th first sample image. for The transpose of , The training species information for the i-th first sample image. for The transpose of , Let represent the species information labeled for the i-th first sample image, and n be the number of first sample images in a training batch, where i ≤ n, and both i and n are positive integers. , and The preset weights are defined by 'if', which is a conditional function.

[0066] According to an embodiment of the present invention, the condition function Indicates that when the condition is met In this case, the value of the conditional function is Otherwise, the condition function value is This conditional function allows the two branches to be trained in a competitive manner. The first branch consists of an encoding module, a growth stage identification module, and a decoding module; that is, the branch that obtains the training species type information and the training growth stage information. The second branch consists of an encoding module, a growth stage identification module, and a growth stage information decoding module; that is, the branch that obtains the first growth stage information. When the above conditions of the conditional function are satisfied, the cosine similarity between the training growth stage information and the labeled growth stage information is greater than or equal to the cosine similarity between the first growth stage information and the labeled growth stage information. In other words, the first branch has higher accuracy. In this case, This represents the error between the training growth stage information and the labeled growth stage information. This represents the error between the training species information and the labeled species information. The error between the information from the first growth stage and the information from the labeled growth stage is represented by this function. These three errors are weighted and summed to obtain the conditional function value, which is then used as the second comprehensive loss function corresponding to the i-th first sample image. During training, this second comprehensive loss function is minimized, thereby reducing the three errors. Both branches can be trained simultaneously, and the training intensity of each branch is determined by the preset weights of the three errors. If the above conditions of the conditional function are not met, i.e., the accuracy of the second branch is higher, then... This amplifies the errors between training and labeled growth stage information, as well as the errors between training and labeled species information. The weighted sum of these amplified errors is then used as the conditional function value. As a magnification factor only, during training, only the errors between the training growth stage information and the labeled growth stage information, and the errors between the training species information and the labeled species information, participate in the training. Therefore, only the first branch is trained with a larger training intensity to improve its accuracy. Furthermore, using... As an amplification factor, it can be used to reflect the impact of inaccurate feature vectors during the growth stage on the error of the species identification model. Therefore, when training using the above method, if the accuracy of the first branch is lower than that of the second branch, the amplification factor can be used to increase the training intensity of the first branch, thereby improving its training efficiency. If the accuracy of the first branch is higher than or equal to that of the second branch, both the first and second branches are trained simultaneously to improve their accuracy at the same time. In other words, the second branch is used to automatically supervise the training of the first branch, or to participate in the accuracy competition of the first branch, thereby urging the first branch to improve its accuracy during training. Moreover, the second branch also continuously improves its accuracy during training, thus participating in the competition with a higher standard, or supervising the training of the first branch with a higher standard, thereby improving the accuracy of the species identification model.

[0067] In this way, the second branch can be used to automatically supervise the training of the first branch. That is, the second branch participates in the accuracy competition of the first branch, thereby urging the first branch to improve its accuracy during training. Furthermore, the second branch also continuously improves its accuracy during training, thereby participating in the competition with a higher standard. That is, it supervises the training of the first branch with a higher standard, thereby improving the accuracy of the first branch. In other words, it improves the accuracy of the species recognition model and improves training efficiency.

[0068] According to one embodiment of the present invention, a species identification model and an image generation model can be trained based on the similarity between a first sample generated image and a first sample image. Since the first sample generated image is generated by the image generation model based on training species information and training growth stage information, and the training species information and training growth stage information are obtained by the species identification model based on the first sample image, when the accuracy of the species identification model and the image generation model is high, the consistency of the species and growth stage of the crops contained in the first sample image and the first sample generated image is high. If there is an error in the species and growth stage of the two images, then there is an error in the appearance of the crops contained in the two images. Therefore, a third comprehensive loss function can be obtained based on this appearance error, and the species identification model and the image generation model can be further trained using the third comprehensive loss function to improve the accuracy of the species identification model and the image generation model.

[0069] According to an embodiment of the present invention, a third comprehensive loss function is obtained based on the first sample generated image, the first sample image, and an image feature recognition model, including: inputting the first sample image into the image feature recognition model to obtain first morphological feature information and first color feature information of crops in the first sample image; inputting the first sample generated image into the image feature recognition model to obtain second morphological feature information and second color feature information of crops in the first sample generated image; and determining the third comprehensive loss function based on the first morphological feature information, the first color feature information, the second morphological feature information, the second color feature information, and the first training probability distribution.

[0070] According to one embodiment of the present invention, an image feature recognition model can be used to identify the morphological and color feature information (both in vector form) of a target contained in an input image. For example, the morphology of multiple parts (e.g., leaves, flowers, stems, etc.) of a crop in an image can correspond to one or more components of the morphological feature information, and the color of multiple parts of a crop in an image can correspond to one or more components of the color feature information. Therefore, the image feature recognition model can obtain the first morphological feature information and the first color feature information of a first sample image, and can also obtain the second morphological feature information and the second color feature information of the image generated from the first sample. The image feature recognition model is a deep learning neural network model, such as a convolutional neural network model. The present invention does not limit the specific type of image feature recognition model.

[0071] According to an embodiment of the present invention, determining a third comprehensive loss function based on the first morphological feature information, the first color feature information, the second morphological feature information, the second color feature information, and the first training probability distribution includes: determining the third comprehensive loss function according to formula (4). ,

[0072] (4)

[0073] in, The first morphological feature information of the i-th first sample image, Generate the second morphological feature information of the image for the i-th first sample. The first color feature information of the i-th first sample image, Generate the second color feature information of the image for the i-th first sample. The first training probability distribution is generated for the i-th first sample to be a real image. and The preset weights are denoted as n, where n is the number of the first sample images in a training batch, i ≤ n, and both i and n are positive integers.

[0074] According to one embodiment of the present invention, in formula (4), The error between the first morphological feature information and the second morphological feature information. The error between the first and second color feature information can be weighted and summed to reduce the result of the weighted sum during training, thereby reducing these two errors and improving the accuracy of the species recognition model and the image generation model. Furthermore, the first training probability distribution can be used as an amplification factor to improve training efficiency. Using the first training probability distribution as an amplification factor can reflect the impact of the realism of the images generated by the image generation model on the second morphological feature information and the second color feature information. This can be applied to the training of the image generation model and the species recognition model to amplify the aforementioned errors, improve training efficiency, and simultaneously enhance the first training probability distribution, further improving the realism of the images generated by the image generation model.

[0075] In this way, the appearance errors of crops in terms of morphology and color in the first sample image and the first sample generated image can be used to determine the third comprehensive loss function. In the solution process, the first training probability distribution is used as an amplification coefficient to reflect the influence of the realism of the image generated by the image generation model on the second morphological feature information and the second color feature information. This can be applied to the training of the image generation model and the species recognition model to amplify the appearance error. At the same time, it can also improve the first training probability distribution, further improve the realism of the image generated by the image generation model, and improve training efficiency.

[0076] According to one embodiment of the present invention, the species identification model can be trained based on the first comprehensive loss function, the second comprehensive loss function, and the third comprehensive loss function. That is, the species identification model is trained through the above training methods. Furthermore, during the training process through each loss function, the parameters of the species identification model can be adjusted during backpropagation of the loss function. Thus, the training of the species identification model can be assisted by various other models (image generation model, discrimination model, image feature recognition model, growth stage information decoding module), thereby improving the training efficiency and training effect of the species identification model. After multiple training sessions, a trained species identification model can be obtained, which can be used for refined identification of crop species and growth stages.

[0077] According to an embodiment of the present invention, in step S103, sampled images of the same species in different partitions can be summarized to obtain a set of sampled images of crops with the same species information, and the partitions corresponding to these sampled images can also be labeled.

[0078] According to an embodiment of the present invention, in step S104, it can be determined whether the environmental parameters in each partition need to be adjusted. Based on the crop growth stage information in each sampled image of the sampled image set and the environmental parameters in multiple partitions, determining whether the environmental parameters in each partition need to be adjusted includes: obtaining preset growth stage vector information corresponding to a preset growth stage; determining the growth stage similarity between the crop growth stage information in each sampled image and the preset growth stage vector information; setting adjustment judgment conditions for the growth stage similarity; and determining whether the environmental parameters in each partition need to be adjusted if the growth stage similarity corresponding to each sampled image meets the adjustment judgment conditions.

[0079] According to one embodiment of the present invention, the preset growth stage is an expected growth stage estimated based on the duration of crop planting, and the preset growth stage vector information is a growth stage vector corresponding to the expected growth stage. The format of the vector is similar to that of the training growth stage information described above, and will not be repeated here.

[0080] According to one embodiment of the present invention, if the environmental parameters are suitable, the crop of this type can theoretically grow to the preset growth stage within the above-mentioned planting period. However, environmental parameters can affect the growth of crops. Therefore, the growth stages of crops in different zones may differ due to different environments. The growth stage information determined by the above-mentioned species identification model can be compared with the preset growth stage vector information. If the similarity of the growth stages between the two (e.g., cosine similarity) is low, then the environmental parameters in that zone are not suitable for crop growth, and the environmental parameters should be adjusted.

[0081] According to one embodiment of the present invention, setting adjustment judgment conditions for growth stage similarity includes: determining the average value and standard deviation of growth stage similarity, and the maximum value and minimum value of growth stage similarity; and determining the adjustment judgment conditions as at least one of the following: the average value of growth stage similarity is less than or equal to the average value threshold; the standard deviation of growth stage similarity is greater than or equal to the annotation difference threshold; the maximum value of growth stage similarity is less than or equal to the maximum value threshold; and the minimum value of growth stage similarity is less than or equal to the minimum value threshold.

[0082] According to one embodiment of the present invention, if the average value of the growth stage similarity is less than or equal to the average value threshold, it indicates that the average level of environmental parameter suitability in each partition is not up to standard, resulting in unsatisfactory crop growth in each partition. If the standard deviation of the growth stage similarity is greater than or equal to the standard deviation threshold, it indicates that the environments of each partition differ significantly, resulting in large differences in crop growth in each partition. If the maximum value of the growth stage similarity is less than or equal to the maximum value threshold, it indicates that the environmental parameters of each partition are unsuitable. If the minimum value of the growth stage similarity is less than or equal to the minimum value threshold, it indicates that there are partitions with unsuitable environmental parameters. Therefore, if any of the above adjustment judgment conditions are met, it can be determined whether the environmental parameters in each partition need to be adjusted. During adjustment, all partitions may need to be adjusted, or only some partitions may need to be adjusted.

[0083] According to an embodiment of the present invention, in step S105, if adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, the growth stage information and species information of crops in each sampled image, and the environmental parameters in multiple zones. This includes: inputting the growth stage information and species information of crops in the sampled images, as well as the planting duration of crops in each zone during this round of shooting, into an environmental parameter estimation model to obtain predicted environmental parameters for each zone; determining the loss function of the environmental parameter estimation model based on the environmental parameters of each zone and the predicted environmental parameters; training the environmental parameter estimation model based on the loss function of the environmental parameter estimation model to obtain a trained environmental parameter estimation model; obtaining the preset growth stage vector information corresponding to the preset growth stage; and inputting the planting duration of crops in each zone during the next round of shooting, as well as the preset growth stage vector information, into the trained environmental parameter estimation model to determine the adjusted environmental parameters.

[0084] According to one embodiment of the present invention, the environmental parameter estimation model is a deep learning neural network model, such as a backpropagation (BP) neural network model. Suitable environmental parameters can be determined using the environmental parameter estimation model. The environmental parameter estimation model can be trained using the crop growth stage information and planting duration of each of the aforementioned zones. The environmental parameter estimation model can derive predicted environmental parameters for each zone; that is, it estimates under what environmental parameters a particular type of crop can grow to that condition within the aforementioned planting duration.

[0085] According to one embodiment of the present invention, the environmental parameters output by the environmental parameter estimation model may contain errors. These errors can be compared with the environmental parameters actually detected in each partition to obtain the loss function of the environmental parameter estimation model. The environmental parameter estimation model is then trained using the loss function of the environmental parameter estimation model to reduce the error between the predicted environmental parameters output by the environmental parameter estimation model and the environmental parameters actually detected in each partition, thereby improving the accuracy of the predicted environmental parameters. The weighted summation of the relative errors between multiple predicted environmental parameters and multiple environmental parameters can be used as the loss function. After backpropagation of the loss function determined by the relative errors between multiple predicted environmental parameters and multiple environmental parameters in multiple partitions, the trained environmental parameter estimation model is obtained.

[0086] According to one embodiment of the present invention, preset growth stage vector information and the planting duration of crops in each partition during the next round of shooting can be input into the trained environmental parameter estimation model to determine the adjusted environmental parameters. The environmental parameters of partitions with growth stage similarity less than or equal to the similarity threshold are adjusted, that is, set to the adjusted environmental parameters so that the growth stage of crops in these partitions reaches or approaches the preset growth stage during the next round of shooting.

[0087] Figure 3 An exemplary block diagram of a precision crop cultivation management system according to an embodiment of the present invention is shown, the system comprising:

[0088] The shooting module is used to take pictures in a circular manner in the crop planting area to obtain sampled images of multiple crops in multiple zones;

[0089] The identification module is used to input multiple sampled images into the species identification model to obtain information on the species and growth stage of crops in each partition;

[0090] The collection module is used to summarize sampled images of the same species in different partitions to obtain a collection of sampled images of crops with the same species information.

[0091] The judgment module is used to determine whether the environmental parameters in each partition need to be adjusted based on the crop growth stage information in each sampled image in the sampled image set and the environmental parameters in multiple partitions.

[0092] The adjustment module is used to determine the adjusted environmental parameters based on the preset growth stage, crop growth stage information, species information, and environmental parameters in multiple zones in each sampled image if adjustment is required.

[0093] According to one embodiment of the present invention, based on the crop growth stage information in each sampled image of the sampled image set and environmental parameters in multiple partitions, determining whether the environmental parameters in each partition need to be adjusted includes:

[0094] Obtain the preset growth stage vector information corresponding to the preset growth stage;

[0095] Determine the similarity of growth stages between the crop growth stage information in each sampled image and the preset growth stage vector information;

[0096] Set adjustment criteria for the similarity of growth stages;

[0097] If the similarity of the growth stages corresponding to each sampled image meets the adjustment judgment conditions, determine whether the environmental parameters in each partition need to be adjusted.

[0098] According to an embodiment of the present invention, setting adjustment judgment conditions for growth stage similarity includes:

[0099] Determine the mean and standard deviation of the growth stage similarity, as well as the maximum and minimum values ​​of the growth stage similarity;

[0100] The adjustment judgment condition is determined to be at least one of the following:

[0101] The average similarity of growth stages is less than or equal to the average threshold;

[0102] The standard deviation of similarity at different growth stages is greater than or equal to the label difference threshold.

[0103] The maximum similarity during the growth stages is less than or equal to the maximum threshold.

[0104] The minimum similarity of growth stages is less than or equal to the minimum threshold.

[0105] According to one embodiment of the present invention, if adjustment is required, the adjusted environmental parameters are determined based on a preset growth stage, crop growth stage information and species information in each sampled image, and environmental parameters within multiple zones, including:

[0106] The growth stage information and species information of crops in the sampled images, as well as the planting time of crops in each zone during this round of shooting, are input into the environmental parameter estimation model to obtain the predicted environmental parameters for each zone.

[0107] Based on the environmental parameters of each partition and the predicted environmental parameters, the loss function of the environmental parameter estimation model is determined;

[0108] Based on the loss function of the environmental parameter estimation model, the environmental parameter estimation model is trained to obtain the trained environmental parameter estimation model.

[0109] Obtain the preset growth stage vector information corresponding to the preset growth stage;

[0110] The planting duration of crops in each zone during the next round of filming, along with the preset growth stage vector information, are input into the trained environmental parameter estimation model to determine the adjusted environmental parameters.

[0111] According to embodiments of the present invention, the refined management method for crop planting can identify the types and growth stages of crops planted in multiple zones through a species identification model. This allows for statistical analysis of the growth stage information of the same type of crop planted in different zones. Based on the statistical results, it can determine whether environmental parameters in each zone need adjustment to ensure that crops in each zone are in an ideal growth environment. This facilitates the management of crop growth stages and states, promoting the achievement of ideal growth stages. When training the species identification model, a growth stage identification module can determine the crop's growth stage feature vector. Then, the growth stage feature vector and a first sample feature map are used to comprehensively determine the crop species and growth stage, thereby improving the accuracy of species identification for crops at different growth stages. Furthermore, during training, multiple training processes can be used to train the species identification model's capabilities in growth stage identification, species identification, and other aspects, improving training efficiency and the accuracy of the species identification model. When determining the realism loss function, adversarial training can be used to determine the realism loss function, thereby improving the realism of the images generated by the image generation model while enhancing the discrimination ability of the discrimination model. This balances the performance of the two models and improves the realism of the first sample image. When determining the generation accuracy loss function, it can be composed of the cosine similarity between the generated species information and the labeled species information, and the cosine similarity between the generated growth stage information and the labeled growth stage information. During training, this improves the consistency between the generated species information and the labeled information of the first sample image, while reducing the errors between the trained species information and the labeled information of the first sample image, the errors between the generated species information and the trained species information, and the errors between the generated growth stage information and the trained growth stage information. This simultaneously improves the accuracy of the images generated by the image generation model and the accuracy of the species recognition model in identifying the species and growth stage of crops in the images. When determining the second comprehensive loss function, the second branch can be used to automatically supervise the training of the first branch. That is, the second branch participates in the accuracy competition of the first branch, thereby urging the first branch to improve its accuracy during training. Furthermore, the second branch also continuously improves its accuracy during training, thereby participating in the competition with a higher standard. That is, it supervises the training of the first branch with a higher standard, thereby improving the accuracy of the first branch. In other words, it improves the accuracy of the species recognition model and improves training efficiency.When determining the third comprehensive loss function, the appearance errors of crops in terms of morphology and color in the first sample image and the first sample generated image can be used to determine the third comprehensive loss function. In the solution process, the first training probability distribution is used as an amplification coefficient to reflect the influence of the realism of the image generated by the image generation model on the second morphological feature information and the second color feature information. This can be applied to the training of the image generation model and the species recognition model to amplify the appearance error. At the same time, it can also improve the first training probability distribution, further improve the realism of the image generated by the image generation model, and improve training efficiency.

[0112] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and any variations or modifications may be made to the implementation of the present invention without departing from the stated principles.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for refined management of crop cultivation, characterized in that, include: The photographer toured the crop-growing areas to capture sampled images of multiple crops in multiple zones. Multiple sampled images are input into the species identification model to obtain information on the species and growth stage of crops in each partition; By summarizing sampled images of the same species in different zones, a set of sampled images of crops with the same species information is obtained. Based on the crop growth stage information in each sampled image in the sampled image set, and the environmental parameters in multiple partitions, determine whether the environmental parameters in each partition need to be adjusted. If adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, the crop growth stage information, species information, and environmental parameters in multiple zones in each sampled image. The training steps of the species recognition model include: inputting a first sample image into the encoding module of the species recognition model to obtain a first sample feature map of the first sample image, wherein the first sample image is an image of a crop of a specific species at a specific growth stage; inputting the first sample feature map into the growth stage recognition module of the species recognition model to obtain a growth stage feature vector of the crop in the first sample image; inputting the growth stage feature vector and the first sample feature map into the decoding module of the species recognition model to obtain training species information and training growth stage information of the crop in the first sample image; obtaining a first sample generated image based on the training species information and the training growth stage information, and an image generation model; obtaining a first comprehensive loss function based on the first sample generated image, the discrimination model, and the species recognition model; determining a second comprehensive loss function based on the growth stage feature vector, the training species information, and the training growth stage information; obtaining a third comprehensive loss function based on the first sample generated image and the first sample image, and an image feature recognition model; and training the species recognition model based on the first comprehensive loss function, the second comprehensive loss function, and the third comprehensive loss function to obtain a trained species recognition model. Determining a second comprehensive loss function based on the growth stage feature vector, the training species information, and the training growth stage information includes: decoding the growth stage feature vector using a growth stage information decoding module to obtain first growth stage information; obtaining labeled species information and labeled growth stage information of crops in the first sample image based on the annotation information of the first sample image; and determining a second comprehensive loss function based on the first growth stage information, the training species information, the training growth stage information, the labeled species information, and the labeled growth stage information. Based on the first growth stage information, the training species information, the training growth stage information, the labeled species information, and the labeled growth stage information, a second comprehensive loss function is determined, including: determining the second comprehensive loss function according to the following formula. , in, This represents the training and growth stage information for the i-th first sample image. This provides the labeled growth stage information for the i-th first sample image. for The transpose of , This represents the information of the first growth stage corresponding to the i-th first sample image. for The transpose of , The training species information for the i-th first sample image. for The transpose of , Let represent the species information labeled for the i-th first sample image, and n be the number of first sample images in a training batch, where i ≤ n, and both i and n are positive integers. , and The preset weights are defined by 'if', which is a conditional function.

2. The method for refined management of crop planting according to claim 1, characterized in that, Based on the crop growth stage information in each sampled image of the sampled image set, and the environmental parameters within multiple partitions, determine whether the environmental parameters within each partition need adjustment, including: Obtain the preset growth stage vector information corresponding to the preset growth stage; Determine the similarity of growth stages between the crop growth stage information in each sampled image and the preset growth stage vector information; Set adjustment criteria for the similarity of growth stages; If the similarity of the growth stages corresponding to each sampled image meets the adjustment judgment conditions, determine whether the environmental parameters in each partition need to be adjusted.

3. The method for refined management of crop planting according to claim 2, characterized in that, Define the adjustment criteria for growth stage similarity, including: Determine the mean and standard deviation of the growth stage similarity, as well as the maximum and minimum values ​​of the growth stage similarity; The adjustment judgment condition is determined to be at least one of the following: The average similarity of growth stages is less than or equal to the average threshold; The standard deviation of similarity at different growth stages is greater than or equal to the label difference threshold. The maximum similarity during the growth stages is less than or equal to the maximum threshold. The minimum similarity of the growth stages is less than or equal to the minimum threshold.

4. The method for refined management of crop planting according to claim 1, characterized in that, If adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, crop growth stage information, species information, and environmental parameters within multiple zones in each sampled image. These parameters include: The growth stage information and species information of crops in the sampled images, as well as the planting time of crops in each zone during this round of shooting, are input into the environmental parameter estimation model to obtain the predicted environmental parameters for each zone. Based on the environmental parameters of each partition and the predicted environmental parameters, the loss function of the environmental parameter estimation model is determined; Based on the loss function of the environmental parameter estimation model, the environmental parameter estimation model is trained to obtain the trained environmental parameter estimation model. Obtain the preset growth stage vector information corresponding to the preset growth stage; The planting duration of crops in each zone during the next round of filming, along with the preset growth stage vector information, are input into the trained environmental parameter estimation model to determine the adjusted environmental parameters.

5. A precision management system for crop cultivation, used to execute the method as described in any one of claims 1-4, characterized in that, include: The shooting module is used to take pictures in a circular manner in the crop planting area to obtain sampled images of multiple crops in multiple zones; The identification module is used to input multiple sampled images into the species identification model to obtain information on the species and growth stage of crops in each partition; The collection module is used to summarize sampled images of the same species in different partitions to obtain a collection of sampled images of crops with the same species information. The judgment module is used to determine whether the environmental parameters in each partition need to be adjusted based on the crop growth stage information in each sampled image in the sampled image set and the environmental parameters in multiple partitions. The adjustment module is used to determine the adjusted environmental parameters based on the preset growth stage, crop growth stage information, species information, and environmental parameters in multiple zones in each sampled image if adjustment is required.

6. The precision management system for crop planting according to claim 5, characterized in that, Based on the crop growth stage information in each sampled image of the sampled image set, and the environmental parameters within multiple partitions, determine whether the environmental parameters within each partition need adjustment, including: Obtain the preset growth stage vector information corresponding to the preset growth stage; Determine the similarity of growth stages between the crop growth stage information in each sampled image and the preset growth stage vector information; Set adjustment criteria for the similarity of growth stages; If the similarity of the growth stages corresponding to each sampled image meets the adjustment judgment conditions, determine whether the environmental parameters in each partition need to be adjusted.

7. The precision management system for crop planting according to claim 6, characterized in that, Define the adjustment criteria for growth stage similarity, including: Determine the mean and standard deviation of the growth stage similarity, as well as the maximum and minimum values ​​of the growth stage similarity; The adjustment judgment condition is determined to be at least one of the following: The average similarity of growth stages is less than or equal to the average threshold; The standard deviation of similarity at different growth stages is greater than or equal to the label difference threshold. The maximum similarity during the growth stages is less than or equal to the maximum threshold. The minimum similarity of the growth stages is less than or equal to the minimum threshold.

8. The precision management system for crop planting according to claim 5, characterized in that, If adjustment is required, the adjusted environmental parameters are determined based on the preset growth stage, crop growth stage information, species information, and environmental parameters within multiple zones in each sampled image. These parameters include: The growth stage information and species information of crops in the sampled images, as well as the planting time of crops in each zone during this round of shooting, are input into the environmental parameter estimation model to obtain the predicted environmental parameters for each zone. Based on the environmental parameters of each partition and the predicted environmental parameters, the loss function of the environmental parameter estimation model is determined; Based on the loss function of the environmental parameter estimation model, the environmental parameter estimation model is trained to obtain the trained environmental parameter estimation model. Obtain the preset growth stage vector information corresponding to the preset growth stage; The planting duration of crops in each zone during the next round of filming, along with the preset growth stage vector information, are input into the trained environmental parameter estimation model to determine the adjusted environmental parameters.

Images