Method of training an encoder-decoder network and system for predicting crop growth status

Abstract

The present application relates to a computing model, and discloses a method for training an encoder-decoder network and a system for predicting a growth state of crops. The method for training an encoder-decoder network comprises: receiving an image pair of the same crop and a growth control condition associated with the growth of the crop; performing a forward diffusion process on the growth amount for a predetermined time step to diffuse the growth amount into an initial noise image; based on the initial noise image and the growth control condition, iteratively performing a reverse denoising process using the encoder-decoder network for a predetermined time step to determine a growth amount prediction value and a noise prediction value associated with each time step; determining a loss function associated with the encoder-decoder network; and adjusting parameters of the encoder-decoder network based on the loss function. The method according to the present application overcomes the limitation that the denoising diffusion model does not consider the time sequence logic of crop growth when learning input images and target images, and improves the accuracy of crop growth prediction.

Description

Technical Field

[0001] This invention generally relates to computer systems utilizing computational models, and more specifically to methods for training encoder-decoder networks and systems for predicting crop growth status. Background Technology

[0002] Crop growth simulation is a technique that predicts the future growth status of crops based on their current state. Existing techniques for crop growth simulation are simply transferring traditional diffusion models to agricultural scenarios.

[0003] One approach is to directly apply the Denoising Diffusion Probabilistic Model (DDPM) to natural image generation. However, since the current image of the crop and the target image at a future time point both belong to the same image data distribution covering all morphological forms of the same crop, the model cannot distinguish the temporal logic between the input and target images. Consequently, the predicted image generated by the model may not show a reasonable increase in leaf quantity and canopy size compared to the input image, and may even exhibit a reverse growth phenomenon of reduced leaves and shorter plant height, violating the natural growth pattern of crops.

[0004] Another approach is to use the Conditional Control Network (ControlNet) model, which uses textual information to constrain the properties of the generated image. However, existing textual information focuses on semantic descriptions and still cannot solve the problem caused by the input image and the target image belonging to the same data distribution.

[0005] There is a need in this field for crop growth status prediction technologies that can be improved at at least one of the aforementioned levels. Summary of the Invention

[0006] This invention is provided to offer a technique for further improving crop growth status prediction based on crop growth rate.

[0007] One aspect of the present invention provides a method for training an encoder-decoder network, comprising: S1: receiving an image pair of the same crop and growth control conditions associated with the growth of the crop, the image pair including a first image and a second image, wherein the second image was acquired later than the first image; S2: performing a forward diffusion process on the growth amount for a predetermined time step based on a growth amount term and a noise term to diffuse the growth amount into an initial noise image, the growth amount being determined based on the difference between the crop in the second image and the crop in the first image; S3: performing a reverse denoising process iteratively using the encoder-decoder network for the predetermined time step based on the initial noise image and the growth control conditions to determine a predicted growth amount and a predicted noise amount associated with each time step. S3 includes: S31: For the current time step in the predetermined time step, based on the noise image associated with the current time step and the growth control conditions, determining the predicted growth amount and the predicted noise value associated with the current time step; and S32: For the current time step in the predetermined time step, based on the noise image associated with the current time step, the predicted growth amount and the predicted noise value determined in the current time step, determining the noise image associated with the next time step, wherein the initial noise image is the noise image associated with the first time step; S4: Based on the growth amount, the predicted growth amount, the noise, and the predicted noise value, determining the loss function associated with the encoder-decoder network; and S5: Adjusting the parameters of the encoder-decoder network based on the loss function.

[0008] The method described above further includes: S6: determining the growth amount term based on the growth amount and retention ratio coefficient for the current time step in the predetermined time step.

[0009] As described above, the retention ratio is dynamically adjusted based on the characteristics of the first image.

[0010] As described above, step S32 includes: S321: for the current time step in the predetermined time step, determining a growth prediction term associated with the current time step based on the growth prediction value and the difference between the cumulative retention ratio coefficient of the current time step and the next time step corresponding to the forward diffusion process; S322: for the current time step in the predetermined time step, determining a noise prediction term associated with the current time step based on the noise prediction value and the difference between the cumulative noise coefficient of the current time step and the next time step corresponding to the forward diffusion process; and S323: for the current time step in the predetermined time step, determining a noise image associated with the next time step based on the noise image, growth prediction term, and noise prediction term associated with the current time step, wherein the initial noise image is a noise image associated with the first time step.

[0011] As described above, the loss function includes spatial structure loss for growth and noise, and Fourier statistical loss.

[0012] In the method described above, the first image and the second image include multispectral images.

[0013] As described above, the growth control conditions include environmental parameters, crop variety, and stress type.

[0014] The method described above further includes: S7: converting the growth control conditions into a growth control embedding vector; wherein, S31 includes: S311: for the current time step in the predetermined time step, based on the noise image associated with the current time step and the growth control embedding vector, determining the growth prediction value and noise prediction value associated with the current time step.

[0015] As described above, step S7 includes: S71: converting the growth control condition into a conditional code; S72: obtaining an intermediate embedding vector based on the conditional code using a common codebook and a dedicated codebook corresponding to the conditional code; and S73: obtaining the growth control embedding vector based on the intermediate embedding vector and the conditional code using a hybrid expert module.

[0016] Another aspect of the present invention provides a system for predicting crop growth status, comprising: computing resources; and an encoder-decoder network trained according to any of the methods described above, the encoder-decoder network invoking the computing resources to: acquire an input image, growth control conditions, and an initial noise image of the same size as the input image, the input image including a crop; and iteratively perform a reverse denoising process up to a predetermined time step to determine a predicted growth value and a predicted noise value associated with each time step, wherein the predicted growth value associated with the predetermined time step is determined as the crop growth, wherein the predicted crop image is determined based on the input image and the growth.

[0017] The system described above further includes a crop growth status evaluation model, which invokes the computing resources to: normalize the crop prediction image to obtain a normalized image; perform a Fourier transform on the normalized image to obtain a spectrogram; perform feature extraction on the normalized image and the spectrogram to obtain spatial and statistical features; and perform regression prediction based on the spatial and statistical features to determine a score for the crop prediction image, wherein the score indicates the growth status of the crop in the crop prediction image.

[0018] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described in any of the preceding claims.

[0019] Another aspect of the present invention provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described in any of the above embodiments.

[0020] The system and method of the present invention overcome the limitation that the denoising diffusion model cannot take into account the temporal logic of crop growth, and improve the accuracy of crop growth status prediction by performing the diffusion and denoising process based on the crop growth amount. Attached Figure Description

[0021] Various embodiments of the present invention are described in conjunction with the accompanying drawings.

[0022] Figure 1 This is a schematic diagram of the process of training an encoder-decoder network according to some embodiments of the present invention.

[0023] Figure 2 This is a schematic diagram of the forward denoising process and the reverse denoising process according to some embodiments of the present invention.

[0024] Figure 3 This is a schematic diagram of the training phase of an encoder-decoder network according to some embodiments of the present invention.

[0025] Figure 4 This is a schematic diagram of the inference phase of an encoder-decoder network according to some embodiments of the present invention.

[0026] Figure 5 This is a schematic diagram of a multispectral image according to some embodiments of the present invention.

[0027] Figure 6 This is a schematic diagram of the crop growth status evaluation process according to some embodiments of the present invention.

[0028] Figure 7 These are schematic diagrams of crop images and spectrograms according to some embodiments of the present invention.

[0029] Figure 8 These are schematic diagrams of different crop images and spectrograms of the same crop according to some embodiments of the present invention.

[0030] Figure 9 This is a flowchart of a first method for training an encoder-decoder network according to some embodiments of the present invention.

[0031] Figure 10 This is a flowchart of a first process associated with a first method for training an encoder-decoder network according to some embodiments of the present invention.

[0032] Figure 11 This is a flowchart of a second process associated with a first method for training an encoder-decoder network, according to some embodiments of the present invention.

[0033] Figure 12 This is a flowchart of a second method for predicting crop growth status according to some embodiments of the present invention.

[0034] Figure 13 This is a flowchart of a third method for evaluating crop growth status according to some embodiments of the present invention.

[0035] Figure 14 This is a block diagram of a system for predicting crop growth status according to some embodiments of the present invention.

[0036] Figure 15 This is a comparison image of a crop, showing a real image and a predicted image, according to some embodiments of the present invention.

[0037] Figure 16 These are the scoring results of crop images by a crop growth status evaluation model according to some embodiments of the present invention.

[0038] Figure 17 This is a block diagram of a computer-readable storage medium according to some embodiments of the present invention.

[0039] Figure 18 This is a block diagram of a computer program product according to some embodiments of the present invention. Detailed Implementation

[0040] In this application, ordinal numbers such as "first," "second," and "third" are used to distinguish different instances of objects with the same name. The ordinal numbers "first," "second," and "third" do not indicate a relative order of the indicated objects in time, space, sequence, or other aspects.

[0041] According to one aspect of the present invention, a process for training an encoder-decoder network is provided.

[0042] Figure 1 This is a schematic diagram of the process of training an encoder-decoder network according to some embodiments of the present invention.

[0043] At the first frame 102, the first image, the second image, and the growth control conditions are received.

[0044] At frame 104 in the second frame, the forward diffusion process is performed.

[0045] At position 106 in the third frame, the reverse denoising process is performed.

[0046] In box 4, at position 108, determine the loss function.

[0047] At box 5, position 110, adjust the parameters of the encoder-decoder network.

[0048] The following text combines Figure 2 Further details of the above process will be described.

[0049] Figure 2 This is a schematic diagram of the forward denoising process and the reverse denoising process according to some embodiments of the present invention.

[0050] The first image 202 and the second image 204 can be Figure 1 The first image 202 and the second image 204 are located in the first frame 102. In some embodiments, the first image 202 and the second image 204 may be images reflecting the growth status of the same crop, and they may have the same size. The second image 204 was acquired later than the first image 202. The second image 204 shows growth relative to the first image 202, for example, variations in leaf number, canopy size, plant height, etc. Growth control conditions 206 are information associated with crop growth, including crop characteristics and external conditions.

[0051] Forward diffusion process 210 can be Figure 1The forward diffusion process is performed at the second frame 104. In some embodiments, the forward diffusion process 210 is performed based on the growth amount to diffuse the growth amount into an initial noisy image 212. The growth amount can be determined based on the difference between the crop in the second image 204 and the crop in the first image 202. For example, the growth amount can be determined by subtracting the second image 204 from the first image 202 pixel by pixel.

[0052] The reverse denoising process 220 can be Figure 1 The reverse denoising process at position 106 in the third frame. Growth control condition 206 can be... Figure 1 The growth control conditions are defined at the first frame 102. In some embodiments, the reverse denoising process 220 can be performed using an encoder-decoder network. Based on the initial noisy image 212 and the growth control conditions 206, the reverse denoising process 220 is performed to determine the predicted growth value 222.

[0053] The loss function associated with the encoder-decoder network can be determined based on the growth rate and the predicted growth rate 222. The parameters of the encoder-decoder network can be adjusted based on the determined loss function, thereby enabling the training of the encoder-decoder network.

[0054] Some embodiments of the present invention perform forward diffusion and reverse denoising processes based on crop growth, so that the crop growth status prediction process fully considers the natural laws of crop growth and time-series correlation, and achieves accurate simulation of crop time-series growth processes under different conditions.

[0055] Figure 3 This is a schematic diagram of the training phase of an encoder-decoder network according to some embodiments of the present invention. The training phase of the encoder-decoder network includes a forward diffusion process and a backward denoising process.

[0056] During the forward diffusion process, based on the growth quantity and noise quantity, the growth quantity is subjected to a forward diffusion process for a predetermined time step to diffuse the growth quantity into an initial noise image.

[0057] The forward diffusion process starts with the growth amount and diffuses it from the growth amount to the initial noisy image. The growth amount is determined based on the difference between the crop in the second image 204 and the crop in the first image 202.

[0058] For time step t, the single-step diffusion formula can be expressed as follows:

[0059] (1);

[0060] in, This represents the image state obtained by diffusion at the current time step t. This represents the image state at the previous time step t-1 during the forward diffusion process. Indicates growth amount. This represents the noise at the current time step t. ~N(0,1), conforming to a standard normal distribution, is used for perturbations in the diffusion process.

[0061] In some embodiments, for the current time step in a predetermined time step, the growth amount term can be determined based on the growth amount and the retention ratio coefficient.

[0062] This represents the retention ratio coefficient, used to control the retention ratio of growth during diffusion at each time step, satisfying... In the table of retained scaling factors, for each time step 1 to T, T is the total number of time steps in the forward diffusion process, also known as the predetermined time steps.

[0063] Represents the noise figure, used to control the injection intensity of noise during the diffusion process at each time step, satisfying... This achieves noise intensity normalization. In the noise figure table, for each time step 1 to T, .

[0064] and The monotonic linear scheduling method can be used to determine this. It can be designed for linear decay, cosine decay, etc. For example, A grid search method can be used to verify the... The setting is denoted by k, where k is the step size coefficient when performing verification using the grid search method, and can be selected to have the highest peak signal-to-noise ratio. value. It can be designed to grow linearly, exponentially, etc.

[0065] Some embodiments of the present invention determine the growth quantity term based on growth quantity and retention ratio coefficients, thereby controlling the retention ratio of growth quantity. Compared with traditional diffusion models that control the mixing ratio of image and noise using only a single coefficient table, some embodiments of the present invention control the retention ratio of growth quantity and the injection rate of noise separately, enabling the learning of the variation pattern of crop growth quantity.

[0066] In some embodiments, a predetermined time step T can be set empirically. As an example, the predetermined time step T can be set to 1000. In some embodiments, a sampler can be used to attempt to reduce the predetermined time step to obtain the minimum predetermined time step with acceptable prediction quality, thereby reducing the number of iterations and improving data processing speed.

[0067] When t=1, which is the first time step of the forward diffusion process, substituting t=1 into the above formula (1), we get ,in The image state at time step 1. = That is, the diffusion process begins with the growth rate. and This can be obtained from the retention ratio table and the noise figure table. Noise that conforms to a standard normal distribution.

[0068] Similarly, when t=T, which is the predetermined time step of the forward diffusion process, substituting t=T into the above formula (1), we obtain... ,in The image state at time step T. The image state at time step T-1, and This can be obtained from the retention ratio table and the noise figure table. Noise that conforms to a standard normal distribution.

[0069] After a T-step diffusion process, the growth in the image is removed, resulting in an initial noisy image conforming to Gaussian noise. .

[0070] like Figure 3 As shown in the forward diffusion process, from left to right, the growth rate... (i.e. This is the starting point of the forward diffusion process, and the initial noise image conforms to Gaussian noise. It is the endpoint of the forward diffusion process.

[0071] Some embodiments of the present invention perform a forward diffusion process based on growth amount. Compared with the traditional diffusion based on the crop image itself, it can learn the changing pattern of crop growth amount, rather than the overall shape of the target image of the crop. This solves the problem of temporal correlation failure caused by the consistent distribution of the input image and the target image of the crop.

[0072] In some embodiments, the retention ratio can be dynamically adjusted based on the characteristics of the first image.

[0073] In some embodiments, the retention ratio can be dynamically adjusted based on the brightness of the first image. For example, when the brightness of the first image is high, the retention ratio can be adjusted, thereby adjusting the noise figure to reduce the injection intensity of noise.

[0074] In some embodiments, the retention ratio can be dynamically adjusted based on the canopy density of the first image. For example, when the canopy density of the first image is high, leaves occlude each other, and new leaf growth mainly occurs within the canopy, resulting in less visible growth. Therefore, the retention ratio can be reduced to decrease the rate of growth stripping. When the canopy density of the first image is low, the canopy gaps are large, and new leaves are almost unobstructed, making visible growth significant. Therefore, the retention ratio can be maintained or slightly increased.

[0075] Some embodiments of the present invention dynamically adjust the coefficients of the forward diffusion process by taking into account the characteristics of crop images, which can improve the adaptability of the forward diffusion process to different crop images, make the diffusion process applicable to crop images in different growth states, and improve the prediction accuracy of the trained encoder-decoder network for various types of crop images.

[0076] In the reverse denoising process, based on the initial noisy image and growth control conditions, the encoder-decoder network iteratively performs the reverse denoising process up to a predetermined time step to determine the growth prediction and noise prediction associated with each time step.

[0077] During the training phase, the inverse denoising process uses the initial noisy image obtained from the forward diffusion process as a starting point to recover the growth amount and noise from the initial noisy image. The inverse denoising process can include a prediction process and a denoising process.

[0078] During the prediction process, for the current time step in the predetermined time step, the predicted growth amount and noise value associated with the current time step are determined based on the noise image and growth control conditions associated with the current time step.

[0079] For example, for the current time step t, firstly based on the image state of the current time step... Given growth control conditions, an encoder-decoder network is used to determine the predicted growth for the current time step t. and noise prediction values The denoising process removes noise from the initial noisy image, and the image state at each time step. It can also be called the noise image at each time step.

[0080] During the denoising process, for the current time step within a predetermined time step, a noise image associated with the next time step is determined based on the noise image associated with the current time step, the predicted growth value determined in the current time step, and the predicted noise value. The initial noise image can be the noise image associated with the first time step.

[0081] By removing the accumulated growth and noise from the image state at the current time step, the initial image state, i.e., the initial value of the growth, can be obtained. For time steps t and t-1, the following image reconstruction formula exists:

[0082] (2);

[0083] (3);

[0084] in, To predict the image state after removing accumulated growth and accumulated noise, The image state at time step t. This represents the image state at time step t-1, and also the image state at the next time step after the current time step t in the reverse denoising process. , is the cumulative retention ratio coefficient from time step 1 to time step t, which is the sum of the retention ratio coefficients corresponding to time steps 1 to t used in the forward diffusion process. , is the cumulative retention ratio coefficient from time step 1 to t-1, which is the sum of the retention ratio coefficients corresponding to time steps 1 to t-1 used in the forward diffusion process. , is the cumulative noise figure from time step 1 to time step t, which is the square root of the sum of squares of the noise figures corresponding to time steps 1 to t used in the forward diffusion process. , is the cumulative noise figure from time step 1 to t-1, which is the square root of the sum of squares of the noise figures corresponding to time steps 1 to t-1 used in the forward diffusion process. and These are the predicted growth value and the predicted noise value for time step t, respectively.

[0085] Since the left-hand sides of formulas (2) and (3) are the same, we can obtain the following by combining them:

[0086] (4).

[0087] Formula (4) above is also the single-step sampling formula. For the current time step t, it can be obtained from the image state at the current time step t. The image state at the next time step t-1 in the reverse denoising process is obtained by sampling. The next time step t-1 in the reverse denoising process corresponds to the previous time step t-1 in the forward diffusion process.

[0088] In some embodiments, for the current time step within a predetermined time step, a growth prediction term associated with the current time step can be determined based on the difference between the predicted growth value and the cumulative retention ratio coefficient between the current time step and the next time step corresponding to the forward diffusion process. For the current time step within a predetermined time step, a noise prediction term associated with the current time step can be determined based on the difference between the predicted noise value and the cumulative noise coefficient between the current time step and the next time step corresponding to the forward diffusion process. For the current time step within a predetermined time step, a noise image associated with the next time step can be determined based on the noise image, the growth prediction term, and the noise prediction term associated with the current time step. The initial noise image can be the noise image associated with the first time step.

[0089] When t=T, which is the first time step of the reverse denoising process, the encoder-decoder network is based on the image state at the first time step. Given growth control conditions, determine the predicted growth value for time step T. and noise prediction values .

[0090] Substituting t=T into formula (4) above, we get ,in The image state at time step T-1, This represents the image state at time step T, which is also the initial noise image obtained during the forward diffusion process. The cumulative retention ratio coefficient from time step 1 to time step T is the sum of the retention ratio coefficients used during the forward diffusion process corresponding to time steps 1 to T. The cumulative retention ratio coefficient from time step 1 to T-1 is the sum of the retention ratio coefficients corresponding to time steps 1 to T-1 used in the forward diffusion process. The cumulative noise figure from time step 1 to time step T is the square root of the sum of squares of the noise figures used during the forward diffusion process corresponding to time steps 1 to T. The cumulative noise figure for time step 1 to T-1 is the square root of the sum of squares of the noise figures corresponding to time step 1 to T-1 used in the forward diffusion process.

[0091] Similarly, when t=1, which is the predetermined time step of the reverse denoising process, the encoder-decoder network calculates the image state based on the predetermined time step. Based on growth control conditions, determine the predicted growth value for time step 1. and noise prediction values .

[0092] After a T-step denoising process, the noise in the initial noisy image is removed.

[0093] like Figure 3 The reverse denoising process is shown, from right to left, the initial noisy image conforming to Gaussian noise. It is the starting point of the reverse denoising process, growth amount This is the endpoint of the reverse denoising process. It represents the image state based on the current time step t. And the predicted growth value determined for the current time step t The image state at the next time step t-1 can be obtained. .

[0094] Some embodiments of the present invention perform a reverse denoising process based on growth amount, jointly estimating growth amount and noise. Compared with the traditional method of directly denoising a noisy image into a crop image, this can strengthen the temporal correlation between the input image and the target image, and avoid the problem that the denoised image is unrelated to crop growth.

[0095] After completing the reverse denoising process, the loss function associated with the encoder-decoder network is determined based on the growth amount, the predicted growth amount, the noise, and the predicted noise.

[0096] For example, the loss function can be determined based on the predicted growth and noise values ​​determined at each time step of the reverse denoising process, along with the growth and noise used in the forward diffusion process, to adjust the parameters of the encoder-decoder network.

[0097] In some embodiments, the loss function may include spatial structure loss for growth and noise, and Fourier statistical loss.

[0098] Growth loss can be calculated as follows:

[0099] ;

[0100] Where E represents expectation. , which represents the weight of the spatial structure loss of growth. The amount of growth used in the forward diffusion process. This is the predicted growth value determined at a certain time step in the reverse denoising process. The growth rate is the Fourier loss weight. F() is the two-dimensional Fourier transform. The two-dimensional Fourier transform can extract statistical features of the image state to reduce spatial noise interference.

[0101] The Fourier statistical loss can be calculated as follows:

[0102] ;

[0103] Where E represents expectation. , which is the weight of the noise spatial structure loss. The noise used at each time step of the forward diffusion process is subject to variance constraints on the noise figure. The equivalent obtained by merging below. This is the predicted noise value determined at a certain time step in the reverse denoising process. The noise Fourier loss weights.

[0104] The loss function is the sum of growth loss and Fourier statistical loss: .

[0105] In some embodiments, a loss function may be determined for the predicted growth and noise values ​​determined at each time step of the reverse denoising process.

[0106] Some embodiments of the present invention integrate spatial structure loss and Fourier statistical loss in the loss function. By using Fourier transform to preserve statistical information such as the distribution pattern of pixel values, the constraints of spatial structure and statistical features can be balanced. This ensures that when there is a shooting error between the first and second images of the same crop, the growth prediction value determined by the network remains reasonable, thereby enhancing the robustness of the determined growth prediction value.

[0107] After determining the loss function associated with the encoder-decoder network, the parameters of the encoder-decoder network are adjusted based on the loss function to obtain the trained encoder-decoder network.

[0108] Figure 4 This is a schematic diagram of the inference stage of an encoder-decoder network according to some embodiments of the present invention. The inference stage of the encoder-decoder network includes a reverse denoising process.

[0109] The inverse denoising process starts with an initial noisy image and aims to recover the growth amount and noise from it. During the inference phase, the input image, growth control conditions, and an initial noisy image of the same size as the input image (including the crop) are acquired. The inverse denoising process is iteratively executed up to predetermined time steps to determine the predicted growth amount and noise value associated with each time step. The predicted growth amount associated with the predetermined time step is determined as the crop's growth amount. The predicted crop image is then determined based on the input image and the growth amount.

[0110] During the inference phase, input image 402 is acquired, which includes crops. Noise image 404 is an initial noise image with Gaussian noise of the same size as the received input image 402. Noise image 404 is used as the image state at time step T. The reverse denoising process is performed iteratively up to a predetermined time step to determine the growth prediction and noise prediction associated with each time step.

[0111] For time step t, first, based on the image state at the current time step... Given the received growth control conditions, an encoder-decoder network is used to determine the predicted growth value. and noise prediction values .

[0112] When t=T, which is the first time step of the reverse denoising process, the encoder-decoder network is based on the image state at the first time step. Given growth control conditions, determine the predicted growth value for time step T. and noise prediction values .

[0113] Substituting t=T into formula (4) above, we get The specific meaning and reference of this equation. Figure 3 The training phase of the encoder-decoder network described is similar.

[0114] Similarly, when t=1, which is the predetermined time step of the reverse denoising process, the encoder-decoder network calculates the image state based on the predetermined time step. Based on growth control conditions, determine the predicted growth value for time step 1. and noise prediction values .

[0115] After a T-step denoising process, the noise in the image is removed.

[0116] like Figure 4 As shown in the reverse denoising process, from right to left, the noisy image 404 is the starting point of the reverse denoising process. This is based on the image state at the current time step t. And the predicted growth value determined for the current time step t The image state at the next time step t-1 can be obtained. .

[0117] The crop prediction image 408 can be determined based on the input image 402 and the crop growth amount 406. For example, the input image 402 and the crop growth amount 406 can be added pixel by pixel to obtain the crop prediction image 408.

[0118] In some embodiments, the predicted growth value determined for a predetermined time step can be used. Used as crop growth amount 406. In some embodiments, the predicted growth amount determined at any suitable time step can be used as crop growth amount 406 as needed.

[0119] In some embodiments, the first and second images in the image pair used to train the encoder-decoder network can be images of the same crop spaced one day apart to reflect subtle differences in crop growth. Based on the trained encoder-decoder network, if the user expects the growth status of the input image one day later, one complete inference phase can be performed on the input image, i.e., one inverse denoising process from t=T to t=1. If the user expects the growth status of the input image N days later, N complete inference phases can be performed on the input image, i.e., N inverse denoising processes from t=T to t=1.

[0120] In some embodiments, the first and second images in the image pair used to train the encoder-decoder network can be images of the same crop spaced one month apart to highlight the crop's growth changes. Based on the trained encoder-decoder network, if the user expects the growth state of the input image one month later, one complete inference stage can be performed on the input image, i.e., one inverse denoising process from t=T to t=1. If the user expects the growth state of the input image N months later, N complete inference stages can be performed on the input image, i.e., N inverse denoising processes from t=T to t=1.

[0121] Although the above examples are based on images of the same crop at intervals of 1 day and 1 month, embodiments of the present invention can be based on images of the same crop at intervals of any length, such as several hours, several days, several months, etc., and are not limited to the above examples.

[0122] Some embodiments of the present invention use a trained encoder-decoder network to predict crop growth based on input images, thereby predicting the crop growth status. Based on the predicted crop images, the stress resistance of different varieties can be compared, providing a basis for breeding screening. Furthermore, the crop growth status can be monitored in real time based on the predicted crop images, allowing for timely artificial intervention in the crop growth environment as needed, achieving precision field management.

[0123] In some embodiments, the encoder-decoder network may employ a U-Net architecture. In some embodiments, the encoder-decoder network may employ a neural network based on a self-attention mechanism, such as the Transformer architecture.

[0124] In some embodiments, crops may include rice, wheat, soybeans, etc.

[0125] In some embodiments, the first image 202 and the second image 204 may include multispectral images.

[0126] Figure 5 This is a schematic diagram of a multispectral image according to some embodiments of the present invention.

[0127] A multispectral image is a spectral image of a crop that includes nine channels. For example... Figure 5 As shown, the nine channels—f0, nir, anthocyanin, red, green, blue, CHL, fm, and red edge—represent minimum fluorescence yield, near-infrared band, anthocyanin index, red band, green band, blue band, chlorophyll content index, maximum fluorescence yield / dry matter content, and red edge band, respectively. The nine channels are shown in color to facilitate understanding of the image characteristics reflected in each channel. Although the nine channels are shown in different colors, the colors are depicted solely for visual convenience, and the color within each channel is only used to visually reflect the relative magnitude of the pixel value, as indicated by the legend on the right side of each channel.

[0128] Some embodiments of the present invention use multispectral images to predict crop growth status, which can fully reflect the deep statistical laws contained in the multispectral information of crops that go beyond basic visual features, thereby improving the accuracy of crop growth status prediction.

[0129] In some embodiments, growth control conditions may include environmental parameters, crop variety, and stress type. Environmental parameters may include soil temperature, humidity, etc. Crop variety may include variety ID, genotype, etc. Stress type may include water deficiency, nitrogen deficiency, salt stress, etc. By accommodating multiple growth conditions, the influence of external environment and crop characteristics on growth can be more comprehensively considered in the growth prediction process, and the crop growth status can be simulated more accurately.

[0130] In some embodiments, growth control conditions can be converted into growth control embedding vectors. For the current time step in a predetermined time step, the predicted growth amount and noise value associated with the current time step can be determined based on the noise image associated with the current time step and the growth control embedding vector. By converting the growth control conditions into growth control embedding vectors, the control conditions can be transformed into inputs that the encoder-decoder network can understand.

[0131] In some embodiments, growth control conditions can first be converted into conditional codes.

[0132] In some embodiments, the environmental parameters in the growth control conditions can be normalized into real number vectors.

[0133] For example, a soil temperature of T=28 Numerical conditions such as humidity H=65% are normalized to the [0,1] interval, resulting in a real number vector c_num = [28 / 40, 65 / 100] = [0.7, 0.65]. Numerical conditions can be received based on triplet format (e.g., temperature "high / normal / low"), word text format (e.g., "twenty-eight"), or number format (e.g., "28").

[0134] In some embodiments, crop variety and stress type in growth control conditions can be converted into unique heat codes.

[0135] For example, the stress type nitrogen deficiency (N) and variety ID 024 can be converted into a unique heat code.

[0136] After conversion, the real-valued vector representing environmental parameters and the unique-heat code representing crop variety and stress type can be concatenated into a 256-dimensional primary vector as conditional encoding. The first two dimensions of this vector reflect the numerical values ​​of the environmental parameters, the third dimension reflects the stress type (nitrogen deficiency), the fourth dimension reflects the crop variety code, and the remaining bits are 0.

[0137] Some embodiments of the present invention transform multi-dimensional growth control conditions into conditional codes with a unified format, which can be adapted to the cross-attention mechanism of encoder-decoder networks, thereby enabling growth control conditions to guide growth simulation.

[0138] Then, based on conditional coding, an intermediate embedding vector can be obtained using a public codebook and a dedicated codebook corresponding to the conditional coding.

[0139] Codebooks can be used to map continuous feature spaces and high-dimensional data to discrete coding spaces, enabling data quantization, compression, and feature representation.

[0140] A public codebook can be trained on all crop data to learn general patterns of crop growth, such as "high temperatures accelerate growth." As an example, the public codebook was obtained by pre-training on 300,000 images of mixed varieties.

[0141] Dedicated codebooks can be trained on data from specific crop varieties or specific stress types to learn scenario-specific patterns, such as "Variety A experiences a faster decline in growth when water is scarce." As an example, the nitrogen deficiency-specific codebook was obtained through pre-training on 18,000 nitrogen deficiency images.

[0142] For example, the first 128 dimensions of the primary vector used as conditional encoding can be used as the query key to find the common codebook to obtain the common embedding vector, and the specific codebook corresponding to the stress type nitrogen deficiency can be found to obtain the specific embedding vector. The common embedding vector and the specific embedding vector can be combined to obtain the intermediate embedding vector.

[0143] Some embodiments of the present invention utilize a codebook mechanism that includes a public codebook and a dedicated codebook, fully considering the diversity of growth control conditions and providing a relatively broad interface for various growth control conditions. When a new control condition affecting crop growth is added, only the dedicated codebook needs to be trained for the new condition to achieve flexible expansion for different growth control conditions.

[0144] Finally, growth control embedding vectors can be obtained using a hybrid expert module based on intermediate embedding vectors and conditional encoding.

[0145] The Mixture of Experts (MoE) module comprises independent expert modules designed for each crop variety. For example, an expert module could be a two-layer multi-layer perceptron (MLP). These expert modules share parameters during early training to learn general conditional effects, and then optimize independently later to learn variety-specific effects. During inference, a gating network can invoke the expert module corresponding to the crop variety included in the conditional encoding. Based on the intermediate embedding vector, the growth control embedding vector is obtained using the expert module. The resulting growth control embedding vector can be input into the encoder-decoder network, for example, by performing cross-attention fusion with features at intermediate positions in the encoder-decoder network.

[0146] Some embodiments of the present invention decouple different crop varieties by using the MoE module, which can highlight the differences between different crop varieties based on the manual settings of different expert modules, thereby improving the ability to capture the growth patterns of different crop varieties.

[0147] Some embodiments of the present invention convert growth control conditions into growth control embedding vectors, enabling encoder-decoder networks to make predictions based on multidimensional factors (such as environment, variety, stress type, etc.), effectively simulating the real growth environment of crops, supporting the simulation of crop growth status under different scenarios, and improving the consistency between predicted values ​​and the real growth environment.

[0148] To assess the growth status of crops in crop images, existing methods focus on single-dimensional features / indicators. One approach is single-indicator visualization evaluation, which assesses crop status by manually measuring plant height and leaf number, or by extracting plant height features based on a simple convolutional neural network model. Another approach is a single-branch feature evaluation model, which extracts spatial features from the image based on a convolutional neural network model and outputs an evaluation score through a fully connected layer. However, these methods target only a single evaluation indicator and are sensitive to image noise, failing to comprehensively reflect the health status of the crop.

[0149] According to another aspect of the present invention, a crop growth status evaluation model is provided.

[0150] Figure 6 This is a schematic diagram of the crop growth status evaluation process according to some embodiments of the present invention.

[0151] First, normalization 610 is performed on crop image 602 to obtain normalized image 612.

[0152] For example, the pixel values ​​of crop images can be scaled to the [0,1] range, and the image size can be adjusted to a uniform size. Normalized images can preserve the original spatial structure, such as leaf edges and canopy distribution.

[0153] In some embodiments, crop image 602 may be a real image of the crop.

[0154] In some embodiments, crop image 602 may be a crop prediction image. For example, according to Figure 4 The process shown yields crop prediction image 408.

[0155] In some embodiments, crop image 602 may be a multispectral image.

[0156] Then, a Fourier transform 620 is performed on the normalized image 612 to obtain a spectrogram 622.

[0157] For example, a two-dimensional Fourier transform can be performed on the normalized image, and then the absolute value can be taken to obtain the spectrogram, and the size of the spectrogram can be adjusted to a uniform size.

[0158] Spectrum maps can downplay spatial structural details and highlight statistical information, such as the energy proportion of different frequency components, to reflect the distribution patterns of pixel values. For example, latent growth problems such as abnormal pixel brightness caused by localized canopy water shortage or abnormal leaf pixel values ​​caused by early nitrogen deficiency are difficult to identify through spatial structure because they do not cause obvious changes in leaf morphology or plant height. However, in a spectrum map, these problems can manifest as subtle differences in the statistical patterns of pixel values, such as an increased standard deviation of pixel brightness in water-deficient areas. Therefore, spectrum maps can help effectively identify latent growth problems in crops, allowing for early detection of abnormalities before growth deteriorates.

[0159] The following will combine Figure 7 , Figure 8 Further description of crop images and spectrograms.

[0160] Next, a first feature extraction 630 is performed on the normalized image 612 to obtain spatial features 632. A second feature extraction 640 is performed on the spectrogram 622 to obtain statistical features 642.

[0161] For example, depthwise separable convolution operations can be used to extract spatial features from crop images. For example, MobileNetV3 can be used to extract spatial features from crop images. For example, self-attention-based neural networks can be used to extract statistical features from spectrograms. For example, the Transformer architecture can be used to extract statistical features from spectrograms.

[0162] Finally, based on spatial features 632 and statistical features 642, regression prediction 650 is performed to determine the growth status score 652.

[0163] For example, spatial features can be concatenated with statistical features, or they can interact through an attention mechanism to obtain fused features. Then, a regression predictor is used to output a score on the fused features. The regression predictor could be, for example, an MLP or a multi-output regression head.

[0164] The growth status score of 652 indicates the growth status of crops in crop images. For example, the growth status score can be a value between 0 and 100. The higher the score, the healthier the crop is, and it can comprehensively reflect the crop's chlorophyll content, canopy uniformity, stress level, and other conditions.

[0165] In some embodiments, the growth status score may include multiple sub-scores, which respectively indicate the growth status of chlorophyll, canopy, etc., to support more refined growth analysis.

[0166] Therefore, some embodiments of the present invention, by extracting spatial and statistical features from crop images and performing dual-branch parallel processing on these features, can comprehensively consider the spatial structure and statistical information of the crop. For example, spatial features such as plant height and spatial structure, as well as statistical features such as the uniformity of pixel value distribution, can be analyzed, enabling the growth status score to cover the comprehensive structural and physiological state of the crop. Furthermore, by performing regression prediction on the features, an accurate mapping from the crop's multimodal features to the growth score is achieved, thereby realizing a comprehensive and robust evaluation of the crop's growth status.

[0167] During the training of the regression predictor, a training dataset can be used. This dataset can include crop images labeled with measured growth states (e.g., chlorophyll content, biomass) and their corresponding scores. Training the regression predictor using labeled data allows the scores to be closely correlated with the actual growth state of the crop, enabling rapid and accurate growth assessment of crop images.

[0168] Figure 7 These are schematic diagrams of crop images and spectrograms according to some embodiments of the present invention. For example, for such... Figure 7The crop image in the green channel of the multispectral image shown on the left can be obtained by Fourier transforming the crop image as follows: Figure 7 The amplitude spectrum diagram shown in the middle and as shown in the figure Figure 7 The phase spectrum is shown on the right. The amplitude spectrum describes the strength of each frequency component in a signal. The phase spectrum describes the phase angle distribution of each frequency component of the signal.

[0169] In addition, converting crop images into spectrograms can reduce the interference of image acquisition conditions on crop images, such as shooting angle, light changes, and soil background noise, thereby improving the stability of the scoring.

[0170] Figure 8 These are schematic diagrams of different crop images and spectrograms of the same crop according to some embodiments of the present invention. For example, when images of the same crop are acquired at different times, due to certain errors in the acquisition process, the two crop images may have spatial offsets, such as... Figure 8 The first and second rows are shown on the left. However, after converting the crop image to a spectrogram, the differences in the spectrograms are small, as shown... Figure 8 The first and second rows are shown on the right. Therefore, the spectrogram of crop images can help eliminate spatial errors during image acquisition, allowing subsequent analysis to focus more on the crop morphology itself.

[0171] According to another aspect of the present invention, a method for training an encoder-decoder network is provided.

[0172] Figure 9 This is a flowchart of a first method for training an encoder-decoder network according to some embodiments of the present invention.

[0173] The first method may include step S1: receiving a pair of images of the same crop and growth control conditions associated with the crop's growth. The image pair includes a first image and a second image, wherein the second image was acquired later than the first image.

[0174] The first method may include step S2: performing a forward diffusion process on the growth amount for a predetermined time step based on the growth amount and noise amount to diffuse the growth amount into an initial noise image. The growth amount is determined based on the difference between the crop in the second image and the crop in the first image.

[0175] The first method may include step S3: based on the initial noisy image and growth control conditions, using an encoder-decoder network to iteratively perform a reverse denoising process up to a predetermined time step to determine the growth prediction and noise prediction associated with each time step.

[0176] In some embodiments, step S3 may include step S31: for the current time step in a predetermined time step, based on the noise image and growth control conditions associated with the current time step, determine the predicted growth amount and noise value associated with the current time step.

[0177] In some embodiments, step S3 may include step S32: for the current time step in a predetermined time step, determining a noise image associated with the next time step based on the noise image associated with the current time step, the predicted growth value determined in the current time step, and the predicted noise value. The initial noise image is the noise image associated with the first time step.

[0178] The first method may include step S4: determining a loss function associated with the encoder-decoder network based on the growth amount, the predicted growth amount, the noise, and the predicted noise.

[0179] The first method may include step S5: adjusting the parameters of the encoder-decoder network based on the loss function.

[0180] Some embodiments of the present invention solve the problem of the failure of association between the input image and the target image caused by the uniformity of crop temporal morphological distribution by performing the forward diffusion process and the reverse denoising process in the denoising diffusion model based on the crop growth amount, and realize the accurate simulation of crop temporal growth images under different conditions.

[0181] In some embodiments, the first method may further include step S6: determining a growth amount term based on the growth amount and retention ratio coefficient for the current time step in a predetermined time step.

[0182] In some embodiments, the retention ratio can be dynamically adjusted based on the characteristics of the first image.

[0183] Some embodiments of the present invention determine the growth quantity based on the growth quantity and the retention ratio coefficient, thereby achieving independent control over the growth quantity and enabling the learning of the growth quantity variation pattern.

[0184] In some embodiments, the first method may further include step S7: converting growth control conditions into growth control embedding vectors.

[0185] Step S31 may include step S311: for the current time step in the predetermined time step, based on the noise image and growth control embedding vector associated with the current time step, determine the growth amount prediction value and noise prediction value associated with the current time step.

[0186] Figure 10 This is a flowchart of a first process associated with a first method for training an encoder-decoder network according to some embodiments of the present invention. The first process may be... Figure 9The specific implementation of step S7 in the first method is shown, but the scope of the present invention is not limited thereto.

[0187] The first process may include step S71: converting growth control conditions into condition codes.

[0188] The first process may include step S72: based on conditional coding, using a public codebook and a dedicated codebook corresponding to the conditional coding, to obtain an intermediate embedding vector.

[0189] The first process may include step S73: obtaining the growth control embedding vector based on the intermediate embedding vector and conditional coding using a hybrid expert module.

[0190] In some embodiments, step S71 may include normalizing the environmental parameters in the growth control conditions into a real number vector.

[0191] In some embodiments, step S71 may include converting crop variety and stress type in growth control conditions into unique heat codes.

[0192] Some embodiments of the present invention convert growth control conditions into growth control embedding vectors to achieve multi-dimensional condition control of factors such as environment, variety, and stress type of the predicted image, effectively simulating the real growth environment of crops and improving the consistency between predicted values ​​and the real growth environment.

[0193] In some embodiments, growth control conditions may include environmental parameters, crop variety, and stress type.

[0194] Figure 11 This is a flowchart of a second process associated with a first method for training an encoder-decoder network, according to some embodiments of the present invention. The second process may be... Figure 9 The specific implementation of step S32 in the first method is shown, but the scope of the present invention is not limited thereto.

[0195] The second process may include step S321: for the current time step in the predetermined time step, based on the growth prediction value and the difference between the cumulative retention ratio coefficient of the current time step and the next time step corresponding to the forward diffusion process, determine the growth prediction item associated with the current time step.

[0196] The second process may include step S322: for the current time step in the predetermined time step, based on the noise prediction value and the difference between the cumulative noise figure of the current time step and the next time step corresponding to the forward diffusion process, determine the noise prediction term associated with the current time step.

[0197] The second process may include step S323: for the current time step in the predetermined time steps, based on the noise image, growth prediction term, and noise prediction term associated with the current time step, determine the noise image associated with the next time step. The initial noise image is the noise image associated with the first time step.

[0198] In some embodiments, the loss function may include spatial structure loss for growth and noise, and Fourier statistical loss.

[0199] Some embodiments of the present invention can balance the constraints of spatial structure and statistical features by mixing spatial structure loss and Fourier statistical loss, thereby enhancing the robustness of the determined growth prediction values.

[0200] In some embodiments, the first image and the second image may include multispectral images.

[0201] Figure 12 This is a flowchart of a second method for predicting crop growth status according to some embodiments of the present invention.

[0202] The second method may include step S81: acquiring an input image, growth control conditions, and an initial noisy image of the same size as the input image. The input image includes the crop.

[0203] The second method may include step S82: iteratively performing a reverse denoising process up to a predetermined time step to determine the predicted growth value and noise value associated with each time step. The predicted growth value associated with the predetermined time step is determined as the crop growth. The crop prediction image is determined based on the input image and the growth value.

[0204] In some embodiments, the second method may be independent of the foregoing. Figures 9-11 The first method described is executed. In some embodiments, the second method may be combined with the above. Figures 9-11 The first method described is to be executed.

[0205] Figure 13 This is a flowchart of a third method for evaluating crop growth status according to some embodiments of the present invention.

[0206] The third method may include step S91: normalizing the crop prediction image to obtain a normalized image.

[0207] The third method may include step S92: performing a Fourier transform on the normalized image to obtain a spectrogram.

[0208] The third method may include step S93: performing feature extraction on the normalized image and spectrogram to obtain spatial features and statistical features.

[0209] The third method may include step S94: performing regression prediction based on spatial and statistical features to determine a score for the crop prediction image. The score indicates the growth status of the crop in the crop prediction image.

[0210] In some embodiments, the third method may be independent of the foregoing. Figure 12 The second method described is used to perform this. In some embodiments, the third method may be combined with the above. Figure 12 The second method described is used for execution.

[0211] Some embodiments of the present invention evaluate crop images by using a dual-branch approach targeting both spatial and statistical features, while also considering spatial structural features and pixel distribution patterns. This solves the problem of one-sided evaluation caused by single-index evaluation methods and can identify latent growth problems in crops, enabling growth status scoring to cover the comprehensive structural and physiological state of crops.

[0212] Figure 14 This is a block diagram of a system 1400 for predicting crop growth status according to some embodiments of the present invention.

[0213] System 1400 may be a local or remote computer, server, etc. System 1400 may include computing resources 1410 and encoder-decoder network 1420.

[0214] In some embodiments, computing resources 1410 may include a central processing unit (CPU), a graphics processing unit (GPU), and various other processing units or cores (e.g., arithmetic logic units, integer units, floating-point units, tensor units, ray tracing cores, etc.).

[0215] The encoder-decoder network 1420 can be based on the above combination Figures 9-11 The first method described is used for training.

[0216] The encoder-decoder network 1420 can be configured to invoke computing resource 1410 to perform corresponding operations. The encoder-decoder network 1420 can be configured to invoke computing resource 1410 to perform the operations described above. Figure 12 The second method described.

[0217] In some embodiments, system 1400 may further include crop growth status evaluation model 1430.

[0218] The crop growth status evaluation model 1430 can be configured to invoke computing resource 1410 to perform corresponding operations. The crop growth status evaluation model 1430 can be configured to invoke computing resource 1410 to perform the operations described above. Figure 13 The third method described.

[0219] Figure 15 This is a comparison chart of real and predicted images of a crop according to some embodiments of the present invention. The first row shows real images of the same crop on different dates, and the second row shows predicted images of the corresponding dates obtained by system 1400 based on the input image of the crop according to some embodiments of the present invention.

[0220] Visually, the predicted images match the real images in terms of plant height, tillering, and uniformity of growth, while also exhibiting high morphological plausibility.

[0221] Furthermore, by analyzing statistical indicators such as the overall brightness of the real and predicted images, the sum and variance of the saturation of all pixels, the sum of pixel values ​​of each channel, percentiles at each stage, and four-digit intervals, the KS pass rate reached 84% at a 10% significance level. These indicators comprehensively reflect the high similarity between the distribution of the predicted and real images.

[0222] Figure 16 These are the scoring results of crop images by a crop growth status evaluation model according to some embodiments of the present invention.

[0223] In the two crop prediction images in the first row, the plants are healthy and intact, receiving higher scores. In the actual prediction process, the growth control conditions were indeed set to no additional stress and suitable temperature, thus the scores matched the crop's growth status. Furthermore, the plant on the left received a higher score due to its better chlorophyll content and canopy uniformity.

[0224] In the two crop prediction images in the second row, the plant leaves show signs of lodging, indicating that their growth status is clearly affected by environmental stress, resulting in a lower score compared to the images in the first row. In reality, the growth control condition was indeed set to water shortage during the prediction process, thus the score matches the crop's growth status.

[0225] According to another aspect of the present invention, a computer-readable storage medium is provided.

[0226] Figure 17 This is a block diagram of a computer-readable storage medium 1700 according to some embodiments of the present invention.

[0227] A computer-readable storage medium 1700 stores a computer program 1750. When executed by a processor, the computer program 1750 implements the above-mentioned... Figures 9-13 The steps of each method or process described.

[0228] According to another aspect of the present invention, a computer program product is provided.

[0229] Figure 18 This is a block diagram of a computer program product 1800 according to some embodiments of the present invention.

[0230] Computer program product 1800 may include computer program 1750. Computer program 1750, when executed by a processor, implements the above-mentioned... Figures 9-13 The steps of each method or process described.

[0231] Embodiments of the invention have been described with reference to the accompanying drawings. These embodiments are illustrative and not restrictive.

Claims

1. A method of training an encoder-decoder network, the method comprising: comprising: S1: receiving an image pair of a same crop and a growth control condition associated with growth of the crop, the image pair comprising a first image and a second image, wherein a collection time of the second image is later than that of the first image; S2: performing a forward diffusion process on a growth amount based on a growth amount term and a noise term for a predetermined time step to diffuse the growth amount into an initial noise image, the growth amount being determined based on a difference between the crop in the second image and the crop in the first image; S3: iteratively performing a backward denoising process using the encoder-decoder network based on the initial noise image and the growth control condition for the predetermined time step to determine a growth amount prediction value and a noise prediction value associated with each time step, the S3 comprising: S31: determining, for a current time step in the predetermined time step, the growth amount prediction value and the noise prediction value associated with the current time step based on a noise image associated with the current time step and the growth control condition; and S32: determining, for the current time step in the predetermined time step, a noise image associated with a next time step based on the noise image associated with the current time step, the growth amount prediction value and the noise prediction value determined in the current time step, wherein the initial noise image is the noise image associated with a first time step; S4: determining a loss function associated with the encoder-decoder network based on the growth amount, the growth amount prediction value, the noise and the noise prediction value; and S5: adjusting parameters of the encoder-decoder network based on the loss function.

2. The method of claim 1, wherein, further comprising: S6: determining, for a current time step in the predetermined time step, the growth amount term based on the growth amount and a retention ratio coefficient.

3. The method of claim 2, wherein, the retention ratio coefficient is dynamically adjusted based on a characteristic of the first image.

4. The method of claim 1, wherein the S32 comprising: S321: determining, for a current time step in the predetermined time step, a growth amount prediction term associated with the current time step based on the growth amount prediction value and a difference between a cumulative retention ratio coefficient corresponding to the forward diffusion process for the current time step and a next time step; S322: determining, for a current time step in the predetermined time step, a noise prediction term associated with the current time step based on the noise prediction value and a difference between a cumulative noise coefficient corresponding to the forward diffusion process for the current time step and a next time step; and S323: determining, for a current time step in the predetermined time step, a noise image associated with a next time step based on the noise image associated with the current time step, the growth amount prediction term and the noise prediction term, wherein the initial noise image is the noise image associated with a first time step.

5. The method of claim 1, wherein, the loss function comprises a spatial structure loss of the growth amount and the noise and a Fourier statistic loss.

6. The method of claim 1, wherein, the first image and the second image comprise multispectral images.

7. The method of claim 1, wherein, the growth control condition comprises an environmental parameter, a crop variety and a stress type.

8. The method of claim 1, wherein, the method further comprising: S7: converting the growth control condition into a growth control embedding vector; wherein the S31 comprises: S311: determining, for a current time step of the predetermined time steps, a growth amount prediction value and a noise prediction value associated with the current time step based on a noise image associated with the current time step and the growth control embedding vector.

9. The method of claim 8, wherein, the S7 comprises: S71: converting the growth control condition into a condition code; S72: obtaining, based on the condition code, an intermediate embedding vector using a public codebook and a private codebook corresponding to the condition code; and S73: obtaining, based on the intermediate embedding vector and the condition code, the growth control embedding vector using a hybrid expert module.

10. A system for predicting a growth state of a crop, characterized in that, comprises: a computing resource; and an encoder-decoder network trained according to any one of claims 1-9, the encoder-decoder network invoking the computing resource to: obtain an input image, a growth control condition, and an initial noise image having a same size as the input image, the input image comprising a crop; and iteratively perform a reverse denoising process for a predetermined time steps to determine a growth amount prediction value and a noise prediction value associated with each time step, wherein the growth amount prediction value associated with the predetermined time steps is determined as a growth amount of the crop, wherein a crop prediction image is determined based on the input image and the growth amount.

11. The system of claim 10, wherein, The system further comprises a crop growth status evaluation model, the crop growth status evaluation model invoking the computing resource to: perform normalization on the crop prediction image to obtain a normalized image; perform Fourier transform on the normalized image to obtain a frequency spectrum image; perform feature extraction on the normalized image and the frequency spectrum image to obtain spatial features and statistical features; and perform regression prediction based on the spatial features and the statistical features to determine a score for the crop prediction image, wherein the score indicates a growth status of the crop in the crop prediction image.

Images