Crop irrigation demand prediction method based on intelligent integrated prefabricated pump station

Abstract

The invention relates to a crop irrigation demand prediction method based on an intelligent integrated prefabricated pump station, and belongs to the technical field of artificial intelligence. The method comprises the following steps: collecting a crop image in a prefabricated pump station irrigation scene, obtaining auxiliary data, calculating a daily accumulated reference evapotranspiration value, and obtaining a time-synchronized multi-modal data set; labeling and dividing the multi-modal data set; constructing an irrigation demand prediction model, wherein the irrigation demand prediction model comprises a spectral reflectivity enhancement module, a double-flow separation convolution module, a cross gating fusion module, a region sensitive pyramid module and a coupling prediction head module; inputting data in the divided training set into the model, and training the model through weighted total loss to obtain a trained model; inputting a to-be-detected image into the trained model to obtain a soil water content and transpiration prediction result; and performing irrigation decision based on the soil water content and transpiration prediction result. The accuracy of irrigation prediction can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a crop irrigation demand prediction method based on intelligent integrated prefabricated pump stations. BACKGROUND

[0002] With the increasing demand for water-saving, efficient and precise irrigation in agricultural production, the traditional irrigation strategy relying on manual experience or rough meteorological indicators has been difficult to meet the requirements of modern agriculture for "timely and appropriate" water supply. Especially under the background of increasing climate uncertainty and water resource shortage, an integrated system solution is urgently needed, which integrates crop perception, physiological modeling and intelligent decision-making.

[0003] The Chinese invention patent with publication number CN120236209A proposes a method for predicting irrigation water demand in irrigation areas based on unmanned aerial vehicle remote sensing, including the following steps: step S1: data acquisition; step S2: clustering optimization calculation; step S3: irrigation water demand prediction based on clustering results. The present application constructs an optimization factor, which gives different weights to different features to reflect their different indicating effects on irrigation water demand, and considers the correlation between spectral features. Specifically, the correlation between spectral features and NDVI is used to determine the weight of each band, and the band with higher correlation with NDVI will be given higher weight; NDVI will also be given higher weight due to its close relationship with crop growth conditions and water conditions. In this way, the optimization factor can more accurately reflect the importance of different features in irrigation water demand prediction, and improve the application accuracy of the K-means clustering algorithm in this scenario. The Chinese invention patent with publication number CN120202905A proposes an irrigation prediction method, device, equipment and medium. The method includes: obtaining an initial noise action sequence; when the current iteration step number is less than or equal to the total number of model iteration steps, determining the current noise action sequence corresponding to the current iteration step number; determining the current supplementary data corresponding to the current iteration step number; inputting the current supplementary data and the current noise action sequence into the irrigation prediction model to determine the prediction noise action sequence corresponding to the current iteration step number; when the current iteration step number is greater than the total number of model iteration steps, obtaining the prediction action sequence corresponding to the current iteration step number; in response to the current iteration step number being less than or equal to the total number of model iteration steps, taking the prediction action sequence corresponding to the current iteration step number as the current noise action sequence corresponding to the next iteration step number, and adding 1 to the current iteration step number. The embodiment of the present application can improve the accuracy of irrigation prediction.

[0004] There are still the following problems in the prior art: the conventional convolutional neural network model directly convolves the RGB image, resulting in spectral and spatial information aliasing, and being unable to effectively identify the weak spectral changes and local morphological changes of water stress; the traditional method separately predicts the water content and transpiration, lacks coupled modeling, and may output predicted values that violate the crop transpiration mechanism; the conventional loss in multi-task training ignores the dimensional differences of soil moisture and transpiration, and does not consider the restriction of image structure on water distribution, causing the optimization target to deviate from the actual demand; the traditional resampling method disrupts the time sequence or causes information redundancy, and is difficult to adapt to the modeling needs of the early or late stress state of crops, affecting the generalization ability and time continuity of the model. SUMMARY

[0005] To achieve the above-mentioned purpose, the following technical solutions are adopted: The application provides a crop irrigation demand prediction method based on an intelligent integrated prefabricated pump station, comprising the following steps: Crop images in the irrigation scene of the prefabricated pump station are collected, auxiliary data is obtained, and daily cumulative reference evapotranspiration values are calculated, the auxiliary data is aligned with the crop image collection time, a time-synchronized multi-modal data set is obtained, and the multi-modal data set is labeled and divided; An irrigation demand prediction model is constructed, including a spectral reflectance enhancement module, a double-flow separation convolution module, a cross-gate fusion module, a region-sensitive pyramid module, and a coupled prediction head module; the divided data in the training set is input into the model, the model is trained through a weighted total loss, and a trained model is obtained; the to-be-detected image is input into the trained model, and soil water content and transpiration prediction results are obtained; irrigation decisions are made based on the soil water content and transpiration prediction results.

[0006] Further, in the crop irrigation demand prediction task, the crop RGB (Red-Green-Blue) image is easily affected by strong light changes and complex background interference, and conventional processing methods such as the Z-score normalization method cannot effectively eliminate shadow changes and soil background noise, resulting in a high sensitivity of the model to light conditions and a reduction in prediction accuracy. The spectral reflectance enhancement module is adopted in the application, the crop image is adaptively corrected for light and separated from the background through a hue-saturation-lightness space conversion and a dynamic background mask generation method, and the specific steps are as follows: Illumination correction in the hue-saturation-lightness HSV space: convert the original crop RGB image to the HSV color space, correct the light intensity distribution using the pixel mean and standard deviation of the lightness channel, compensate for color shift using the pixel mean of the hue channel, and perform adaptive gain compensation using the standard deviation of the saturation channel, output an illumination-corrected image with uniform illumination, and eliminate the influence of shadow changes; Crop mask generation based on normalized difference vegetation index: calculate the Euclidean distance between the RGB vector of each pixel of the original crop image and the pre-computed RGB mean vector of the typical soil area, then compare this distance with the normalized difference vegetation index threshold; when the Euclidean distance is greater than the threshold, it is determined as a crop pixel, and the output is foreground; otherwise, it is determined as a background soil pixel, and the output is background; then generate a binary crop mask image to separate the crop foreground from the soil background; Background mask correction: multiply the illumination correction image and the crop mask image element by element, the pixel value of the background area is set to zero, and the pixel value of the crop foreground area remains unchanged, output the mask correction image with illumination robustness and pure crop foreground.

[0007] Further, the present application adopts a dual-stream separation convolution module to extract frequency domain spectral features and spatial texture features using parallel processing streams to solve the spectral confusion problem. Through the dual-stream separation convolution module, multi-scale spectral-texture feature collaborative extraction is performed, and the specific steps are as follows: Spectral stream frequency domain feature extraction: learnable spectral weights are used to weight each RGB channel of the illumination correction image to realize spectral decoupling, sum the weighted three-channel results, and perform two-dimensional discrete cosine transform on the summed results to convert spatial information into frequency domain representation, obtain a spectral feature map that retains main low-frequency components, and capture frequency domain patterns related to water stress; Texture stream spatial feature extraction: depth separable convolution is used to extract spatial texture features from each RGB channel of the illumination correction image, batch normalization processing is performed on the extracted features, and a texture feature map is obtained through a rectified linear unit activation function to retain the crown layer spatial structure information.

[0008] Further, the present application adopts a cross-gated fusion module to realize adaptive selection of features, and optimizes feature expression through channel recalibration to reduce redundancy. Through the cross-gated fusion module, feature dynamic fusion and channel recalibration are performed. The specific steps are as follows: The spectral feature map and the texture feature map are spliced along the channel dimension, and the spliced feature map is processed through convolution operation and Sigmoid activation function to obtain a dynamic fusion weight map; the spectral feature map and the texture feature map are weighted using the dynamic fusion weight map to obtain a fusion feature map, realizing adaptive soft selection of spectral and texture features; the fusion feature map is globally averaged to obtain a global feature vector; the global feature vector is cross-channel interacted through a calibration matrix, and the interaction result is processed through a normalized exponential function, and then multiplied by the fusion feature map to obtain a recalibrated feature map, optimizing the importance distribution of each feature channel.

[0009] Further, the sensitivity of different regions of the crop canopy to water stress is significantly different, and conventional pooling operations can lose local key information, making it difficult for the model to accurately capture the spatial heterogeneity of water stress. The present application performs spatial context modeling and region importance weighting through a region-sensitive pyramid module, the specific steps being: The recalibrated feature map is subjected to average pooling operation with three step sizes to obtain three feature maps of different scales; the three feature maps of different scales are spliced along the channel dimension to obtain a multi-scale feature map; the multi-scale feature map is subjected to convolution operation and then applied with a normalized exponential function in the spatial dimension to obtain a spatial attention map, the pixel value of which quantifies the sensitivity of the spatial position to water stress; the recalibrated feature map is subjected to maximum pooling operation to obtain a global context feature, which is subjected to convolution operation and then up-sampled to the spatial resolution of the recalibrated feature map by bilinear interpolation, and then the up-sampled global context feature is multiplied element by element with the spatial attention map to obtain a context-enhanced feature map.

[0010] Further, in irrigation demand prediction, it is necessary to output soil water content and crop transpiration at the same time, and the conventional single output head method ignores the physical correlation of water transfer, resulting in prediction results that may violate the basic biophysical rule that transpiration increases with increasing soil water content. The present application performs multi-task prediction and physical constraint through a coupled prediction head module, the specific steps being: The context-enhanced feature map is flattened into a one-dimensional vector, which is mapped through a fully connected layer to obtain a soil water content prediction value; the context-enhanced feature map is subjected to global maximum pooling to retain the most prominent features of each channel, and the pooling result is subjected to a fully connected layer to generate a transpiration adjustment factor, and the soil water content prediction value is multiplied by the transpiration adjustment factor to obtain a crop transpiration prediction value.

[0011] Further, a single mean square error loss function cannot balance the dimensional difference between soil water content and crop transpiration, such as cubic meters per cubic meter and millimeters per day, and ignores the influence of meteorological factors on the transpiration process, resulting in unbalanced multi-task optimization and inaccurate physical laws. The present application adopts reference evapotranspiration as a meteorological prior to construct a hybrid loss function; through multi-objective weighting and spatial consistency constraint, the dual-task optimization process is coordinated, including: Soil water content regression loss: calculate the L2 norm between the soil water content prediction value and the soil water content true value to obtain the soil water content regression loss, which measures the deviation of the soil water content prediction; Normalized transpiration regression loss: the transpiration prediction value and the transpiration true value are normalized using the reference evapotranspiration value, and the L1 norm between the normalized prediction value and the normalized true value is calculated to obtain the normalized transpiration regression loss, which measures the deviation of the normalized transpiration prediction and eliminates the dimensional difference; Spatial gradient consistency loss: apply Sobel operator to the pixel-level soil moisture prediction value to obtain the spatial gradient distribution of the predicted soil moisture; apply Sobel operator to the grayscale image of the illumination correction image to obtain the spatial gradient distribution of the image; calculate the Kullback-Leibler divergence between the spatial gradient distribution of the predicted soil moisture and the spatial gradient distribution of the image to obtain the spatial gradient consistency loss, which constrains the consistency of the spatial distribution of the water gradient and the spatial distribution of the image visual gradient.

[0012] Further, the soil moisture regression loss, the normalized transpiration regression loss and the spatial gradient consistency loss are multiplied by preset weight coefficients respectively, and then the three weighted losses are added to obtain a multi-objective weighted loss.

[0013] Further, the uneven distribution of crop growth stages will lead to the model being biased towards the main growth stages, such as the vigorous growth period. Conventional resampling methods such as oversampling or undersampling will destroy the continuity of the time series, cannot preserve the time sequence correlation of crop growth dynamics, and also cannot enhance the model's generalization ability to rare growth stages such as the seedling stage or the mature stage. By dynamically adjusting the importance of samples through time decay weight: calculate the absolute time difference between the sample collection time point and the average time point of the current training batch, and combine the decay rate coefficient to calculate the time decay term; calculate the contrast factor of the illumination correction image, multiply the time decay term and the contrast factor to obtain the dynamic weight value of the sample.

[0014] Further, the multi-objective weighted loss of each sample in the training batch is calculated, the multi-objective weighted loss of each sample is multiplied by its corresponding dynamic sample weight, and then the average value of the weighted losses of all samples is calculated to obtain the final weighted total loss for model training.

[0015] The advantages of the present application are: The present application adopts a spectral flow and texture flow parallel extraction mechanism, combines frequency domain analysis and spatial modeling, accurately extracts crop water stress signals, solves the feature redundancy problem caused by RGB channel confusion; embeds the physical causal relationship of "increased soil moisture-increased transpiration" in the soil moisture and transpiration prediction task, ensures that the model output follows the principles of agricultural hydrology through multiplication modeling, and improves the prediction reliability in extreme drought situations; uses reference evapotranspiration values for transpiration normalization, uses KL divergence to constrain the consistency of soil water spatial gradient and image edge structure, and significantly improves the authenticity and continuity of the prediction spatial distribution; uses a sample weight mechanism based on time decay and image contrast to avoid disrupting the crop growth time sequence with traditional resampling methods and enhance the model's generalization ability to low-frequency stages such as the seedling stage and the mature stage. BRIEF DESCRIPTION OF DRAWINGS

[0016] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of the specification, illustrate embodiments of the application, and are used to explain the present application, but are not intended to limit the present application.

[0017] Figure 1 Flow chart of steps of the method of the present application; Figure 2 Raw crop image collected in a specific embodiment of the present application; Figure 3 Illumination correction image in a specific embodiment of the present application; Figure 4 Crop mask image in a specific embodiment of the present application; Figure 5 Mask correction image in a specific embodiment of the present application; Figure 6 Spectral feature map in a specific embodiment of the present application; Figure 7 Texture feature map in a specific embodiment of the present application; Figure 8 Fusion feature map in a specific embodiment of the present application; Figure 9 Soil moisture time series prediction effect diagram in a specific embodiment of the present application; Figure 10 Crop full growth cycle prediction performance comparison in a specific embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0019] Embodiment 1 In this embodiment, as shown in the accompanying drawings, Figure 1 The present application provides a crop irrigation demand prediction method based on intelligent integrated prefabricated pump station, and the specific steps include: S1, collecting crop images under prefabricated pump station irrigation scene Under the prefabricated pump station irrigation scene, the crop image collection is completed by the cooperation of the fixed multispectral camera deployed in the pump station control area and the mobile unmanned aerial vehicle; The fixed camera is installed on the pump station irrigation main pipeline support at a top-down angle, and automatically triggers shooting at three time periods (06:00, 12:00, 18:00) before, during and after irrigation every day, covering the core crop area controlled by the pump station, such as a radius of 50 meters; The unmanned aerial vehicle equipped with GPS positioning flies along the preset route twice a week, at a height of 3 meters from the canopy layer, and takes supplementary views at a 45-degree tilt angle to eliminate the blind area of the fixed view.

[0020] During the process of crop image collection, the data of the entire crop growth cycle, from the seeding period to the mature period, needs to be included. The original images are stored in uncompressed RGB format and cached locally through the pump station edge computing gateway.

[0021] In one embodiment, the collected original crop images are as shown in Figure 2

[0022] S2, Obtain auxiliary data and calculate daily cumulative reference evapotranspiration value In addition to crop images, real-time temperature, humidity, wind speed, and sunshine hours are collected by weather stations. Based on weather station data, the FAO-Penman-Monteith method is used to calculate the daily cumulative reference evapotranspiration value, where solar radiation data is obtained through a pyranometer on the top of the pump station. All auxiliary data are strictly aligned with the image collection time to form a time-synchronized multi-modal data set.

[0023] S3, Training data labeling and training data set construction The collected crop images are labeled, and the soil moisture, transpiration state, and other information during the data acquisition period are combined to label each crop image through manual labeling by agricultural experts. Specifically, the soil moisture level (drought / normal / excessive wetness) is divided according to the synchronous measurement of soil humidity sensor data, and the transpiration state (wilting / normal / saturation) is determined in combination with the crown temperature and stomatal conductance instrument measurement results.

[0024] The labeled data is divided into a data set, which includes 70% training set (sampled by growth stage), 15% validation set, and 15% test set.

[0025] S4, Irrigation demand prediction model construction and training S401, Spectral reflectance enhancement module for adaptive illumination correction and background separation of crop images In the task of crop irrigation demand prediction, crop RGB (Red-Green-Blue) images are easily affected by strong light changes and complex background interference. Conventional processing methods such as Z-score normalization method cannot effectively eliminate shadow changes and soil background noise, resulting in high sensitivity of the model to light conditions and reduced prediction accuracy.

[0026] ​The application adopts a spectral reflectance enhancement module, and realizes adaptive illumination correction and background separation through HSV (Hue-Saturation-Value) space conversion and dynamic background mask generation, eliminates illumination change and background interference, and the specific steps are as follows: 1) Illumination correction in HSV space The original crop RGB image is converted to the HSV color space, then the pixel mean and standard deviation of the brightness channel are used to correct the non-uniformity of illumination intensity distribution, the pixel mean of the hue channel is combined to compensate for color deviation, and the standard deviation of the saturation channel is used for adaptive gain compensation, an illumination-corrected image with uniform illumination is output, the influence of shadow change is eliminated, and is expressed as: , In the formula, is the crop image, is the original crop RGB image, and the dimension is ; is the illumination-corrected image, which is adjusted adaptively by the brightness channel statistics of the HSV space, and the compensation terms of hue and saturation are combined to eliminate shadow change, an RGB image with uniform illumination is output, and the dimension is ; is the image height; is the image width; is the pixel mean of the brightness channel in the HSV space, which is used to correct the non-uniformity of illumination intensity distribution; is the pixel standard deviation of the brightness channel in the HSV space, which represents the illumination fluctuation intensity; is a minimum constant, which is a numerical stability term to prevent the denominator from being zero, and the preferred value is ; is an adaptive saturation gain coefficient, and the calculation method is expressed as: ; is a natural exponential function; is the standard deviation of the saturation channel in the HSV space, which is used to enhance the color contrast under weak illumination; is a hue compensation term, and the calculation method is expressed as: ; is the mean of the hue channel in the HSV space, which is used to balance the color deviation; is a truncation function, which limits the pixel value to the range of [0, 255].

[0027] In one embodiment, the illumination-corrected image is as shown in Figure 3 .

[0028] 2) Generation of crop mask based on normalized difference vegetation index The Euclidean distance of the RGB vector of each pixel of the original crop image and the pre-computed typical soil region RGB mean vector is calculated, and then the distance is compared with the normalized difference vegetation index threshold value. When the Euclidean distance is greater than the threshold value, it is determined that the pixel is a crop pixel, and the output is foreground. Otherwise, it is determined that the pixel is a background soil pixel, and the output is background. Then, a binary crop mask image is generated, and the separation of crop foreground and soil background is realized, which is represented as: , In the formula, is a crop mask image, which is a binary value with elements of 0 / 1, and the dimension is ; is a pre-computed typical soil region RGB mean vector, which is obtained by pre-collecting the RGB image of a typical soil region without vegetation coverage, and the calculation method is represented as: ; is the number of pixels in the typical soil region without vegetation coverage; is the pixel value of the typical soil region without vegetation coverage; is the pixel value index; is the L2 norm, which is the same as the Euclidean distance calculation method, calculates the Euclidean distance of the RGB vector of each crop pixel and , which represents the vegetation coverage; is the normalized difference vegetation index threshold value, which is used to determine the crop pixel, and the preferred value is 0.3; is an indicator function, which outputs 1 when the condition is met, and otherwise outputs 0.

[0029] In one embodiment, the crop mask image is as shown in Figure 4 .

[0030] It should be noted that represents a pixel-level binary judgment, which outputs 1 when the condition is met, indicating a crop region, and otherwise outputs 0, indicating a background region, which can achieve accurate vegetation segmentation in complex farmland backgrounds.

[0031] It should also be noted that the preferred value of the normalized difference vegetation index threshold value is 0.3, which has been verified by experiments to effectively separate crops and soil, such as corn, wheat, and other canopy reflectance characteristics.

[0032] 3) Background mask correction The light-corrected RGB image and the generated binary crop mask are multiplied element by element. The pixel values of the background region are set to zero, and the pixel values of the crop foreground region remain unchanged. The output is a mask-corrected image that has both light robustness and pure crop foreground, which is represented as: , in, The Hadamard product represents element-wise multiplication. To mask-correct the image, achieving both illumination robustness and a clean crop foreground, the dimension is... .

[0033] In one embodiment, the masked image is as follows: Figure 5 As shown.

[0034] S402, dual-stream separation and convolution module for multi-scale spectral-texture feature co-extraction When extracting crop features, conventional convolutional neural networks using direct convolution operations lose local frequency domain features related to water stress. Furthermore, the coupling of the RGB three channels leads to spectral information confusion, impacting model performance. This invention employs a dual-stream separable convolution module, using parallel processing streams to extract frequency domain spectral features and spatial texture features separately, thus resolving the spectral confusion problem. The specific steps are as follows: 1) Spectral flow frequency domain feature extraction For each channel of the illumination-corrected image, a corresponding learnable spectral weight is used for weighting to achieve spectral decoupling. Then, the weighted three-channel results are summed, and a two-dimensional discrete cosine transform is applied to the summation result to convert the spatial information into a frequency domain representation. The output is a spectral feature map that retains the main low-frequency components, capturing the frequency domain patterns related to water stress, represented as: , In the formula, The first image for illumination correction One channel, with dimensions of ; For image channel indexing, These correspond to the three channels of RGB respectively; For the first The spectral weight of the i-th channel is a scalar and a learnable parameter that acts on the i-th channel. One channel to achieve spectral decoupling; This is a standard two-dimensional convolution operation, here it is Convolution kernels are used to weight single-channel images; The spatial features are converted into a frequency domain representation using a two-dimensional discrete cosine transform, and frequency domain patterns related to water stress are extracted. This is a spectral feature map with dimensions of [dimensional value missing]. ; The number of frequency domain characteristic channels is determined by the number of discrete cosine transform coefficients, and is used to retain the previous... The low-frequency component ensures that the energy proportion is >95%, which is the preferred value. It can cover the main frequency band of moisture response.

[0035] In one embodiment, the spectral feature map is as follows:Figure 6 as shown.

[0036] It should be noted that water stress causes changes in the structure of crop leaf cells, such as stomatal closure, which is manifested as abnormal reflectivity in the medium and high frequency specific frequency domain, Item 3) Texture flow spatial feature extraction

[0037] 2) Texture flow spatial feature extraction For each RGB channel of the light correction image, independently apply a depth separable convolution to extract spatial texture features, then perform batch normalization on the extracted features, and apply a rectified linear unit activation function to enhance nonlinearity, output a texture feature map that preserves the spatial structure information of the canopy, denoted as: , In the formula, is the spatial convolution kernel facing the i-th channel in the depth separable convolution, and is a learnable parameter; the dimension is ; is a depth convolution operation, and each channel is independently spatially convolved; is a batch normalization layer used to standardize the feature distribution; is a rectified linear unit activation function; is a texture feature map, with a dimension of: ; is the number of spatial feature channels, and the preferred value is 32.

[0038] In one embodiment, the texture feature map is as shown in Figure 7 .

[0039] It should be noted that the spectral feature map characterizes the frequency domain features directly extracted by discrete cosine transform, focusing on the reflectivity frequency shift caused by water stress, such as attenuation in the near-infrared band, and the texture feature map By using a depth separable convolution to preserve the spatial structure of the canopy, such as leaf curling morphology, the separation design solves the spectral confusion problem caused by RGB channel coupling, enabling the decoupling of frequency domain analysis and spatial modeling, making the frequency domain flow more sensitive to early and weak signals of water stress, such as mild wilting, while the spatial flow is more robust to macroscopic deformation under severe stress, and the two complement each other significantly improve the model's ability to distinguish stress stages.

[0040] It should also be noted that the spectral weight ​For the learnable scalar, the RGB three channels are independently weighted to eliminate the coupling between channels, such as the strong correlation of red / green channels in green leaf area. Channel decoupling is performed before frequency domain transformation, so that the two-dimensional discrete cosine transformation is more focused on stress-sensitive channels, and the contribution of the model to the water response channel is automatically enhanced. For example, if the weight of the first channel is increased, the second channel containing noise is suppressed, so that the frequency domain feature distinction is effectively improved.

[0041] It should be further pointed out that the spatial resolution is kept unchanged, and is independently set, preferably set to , , allowing the model to allocate feature capacity as needed, with a higher number of texture flow channels to capture complex canopy structure, and a smaller number of spectral flow channels due to the sparsity of frequency domain features. Force the spectral flow to learn the key frequency domain patterns of water stress and avoid redundant features, and the large enhances the modeling ability of the texture flow to local details, such as the local detail modeling ability of leaf edge wilting.

[0042] S403, cross-gated fusion module for dynamic feature fusion and channel recalibration Directly splicing spectral and texture features will lead to information redundancy, and the conventional attention mechanism ignores the cross-modal association, which is not conducive to feature fusion. The present application uses a cross-gated fusion module to realize adaptive selection of features, and optimizes feature expression through channel recalibration to reduce redundancy. The specific steps are as follows: 1) Cross-gated feature fusion Splice the spectral feature map and the texture feature map along the channel dimension, then apply a 1x1 convolution operation to the spliced feature map, and pass the convolution output through a Sigmoid activation function to generate a dynamic fusion weight map ranging from 0 to 1. Then use the weight map to weight the spectral feature map and the texture feature map, and then output the fused feature map, realizing adaptive soft selection of spectral and texture features, represented as: , , In the formula, is a channel splicing operation, the output dimension of the term is ; is a dynamic fusion weight map, with a dimension , the value is compressed to by the Sigmoid function, controlling the mixing ratio of spectral / texture features; is a 1x1 convolution layer, the term compresses the channel number to ; is a Sigmoid activation function; is a fused feature map, with dimension .

[0043] In one embodiment, the fused feature map is as shown in Figure 8 .

[0044] 2) Channel-level re-calibration The global feature vector is obtained by performing global average pooling on the fused feature map, then the global feature vector is interacted across channels by a calibration matrix, and the interaction result is processed by a normalized exponential function, the obtained channel weight vector is broadcast to the spatial dimension, then multiplied with the fused feature map channel by channel, and a re-calibration feature map is output, which optimizes the importance distribution of each feature channel, and is represented as: , In the formula, is global average pooling, the term compresses into a vector; is a calibration matrix, which is a learnable parameter, and realizes cross-channel interaction, with a dimension of ; is a normalized exponential function; is a channel multiplication operation, which is used to multiply the weight vector with the feature map channel by channel after broadcasting to the spatial dimension; is a re-calibration feature map, with a dimension of .

[0045] It should be noted that the Sigmoid gate realizes soft selection of spectral and texture features, rather than hard switching, allowing the model to dynamically adjust the weight according to the characteristics of the local region, for example, the crown center mainly uses spectral features, tending to 1, and the edge mainly uses texture features tending to 0, while using a spatial adaptive fusion method to solve the information redundancy problem of the splicing operation, in the transition area of water stress, such as semi-wilting leaves, the gating mechanism automatically mixes the double-flow features to generate a stress distribution map with stronger continuity, avoiding spatial mutations in the prediction results.

[0046] It should also be noted that the calibration matrix models the cross-modal channel correlation, such as the correlation between the low-frequency component of the spectral flow and the edge detection channel of the texture flow, which realizes global channel interaction through a fully connected matrix, solving the problem of ignoring cross-modal dependence in the conventional attention mechanism, and the synergistic enhancement of the stress response channel and the spatial structure channel, such as the synergistic enhancement of the mid-frequency component in the two-dimensional discrete cosine transform and the leaf texture feature, so that the fused feature still maintains high discriminability under weak light.

[0047] It should also be noted that the global average pooling and Convolution can significantly reduce the computational overhead to realize feature recalibration with lightweight operation, and low computational load enables the model to process high-definition images in real time, meeting the demand for online monitoring in the field.

[0048] S404, the region sensitive pyramid module performs spatial context modeling and region importance weighting The sensitivity of different regions of the crop canopy to water stress is significantly different, and conventional pooling operations can lose local key information, making it difficult for the model to accurately capture the spatial heterogeneity of water stress. The region sensitive pyramid module is used in the present application to strengthen the feature expression of water stress sensitive regions through multi-scale context feature fusion and adaptive region weighting. The specific steps are as follows: 1) Constructing multi-scale features The recalibration feature map is subjected to average pooling operation with three step sizes, respectively, to obtain three feature maps of different scales, and then the three feature maps of different scales are spliced along the channel dimension to output a multi-scale feature map containing multi-scale context information such as local leaf details, plant structure and canopy group patterns, denoted as: , In the formula, is the average pooling operation with a step size of ; is the step size of the pooling operation, denotes the pooling step set, which controls the feature map downsampling rate; is the splicing operation along the channel dimension, denotes the average pooling result along the channel dimension with a splicing step size ; is the multi-scale feature map, with a dimension of: ; is the sum of the channel numbers of each scale feature.

[0049] 2) Generating region importance weight A 3x3 convolution layer is applied to the multi-scale feature pyramid, and then the convolution result is applied to the spatial dimension to output a spatial attention map, whose pixel value quantifies the sensitivity of the spatial position to water stress, denoted as: , In the formula, is the spatial attention map, with a dimension of , a value range of , and quantifying the pixel-level water stress sensitivity; is a 3x3 convolution layer; is a Softmax function along the spatial dimension; 3) Context feature enhancement The global context feature is obtained by maximum pooling with a step size of 4 on the re-calibrated feature map, then a 1x1 convolution layer is applied to the global context feature, and the result is up-sampled to the original image spatial resolution by bilinear interpolation, then the up-sampled global context feature is multiplied element by element with the spatial attention map, and the context-enhanced feature map is output, which enhances the feature expression of stress-sensitive areas, and is represented as: , In the formula, is the context-enhanced feature map, and the dimension is ; is the maximum pooling with a step size of 4; is the bilinear interpolation up-sampling to the spatial resolution of the re-calibrated feature map, that is, the resolution is .

[0050] It should be noted that through multi-scale pooling with a step size , local leaf details are extracted when , plant structures are extracted when , and canopy group patterns are extracted when , and the spatial attention map weights the stress-sensitive areas, such as wilting leaf area spatial attention enhancement. The combination of multi-scale context and spatial importance solves the information homogenization problem of conventional pooling, and enables the model to infer the stress state of hidden areas through large-scale context under the canopy shielding scene, and focus on local significant areas such as leaf tip scorching through attention, thereby improving the robustness of complex farmland environments.

[0051] S405, coupling prediction head module for multi-task prediction and physical constraint In irrigation demand prediction, it is necessary to output soil moisture content and crop transpiration at the same time. The conventional single-output head method ignores the physical correlation of water transmission, which may lead to prediction results that violate the basic biophysical law that transpiration increases with increasing soil moisture content. The present application adopts a coupled prediction head module to simultaneously output double-target prediction values by embedding the physical constraints of soil moisture content and transpiration. The specific steps are as follows: 1) Soil moisture content prediction The context-enhanced feature map is flattened into a one-dimensional vector, and then the one-dimensional vector is mapped through a fully connected layer to output the soil moisture content prediction value, represented as: , In the formula, is a spatial flattening operation, flattened into a one-dimensional vector; is a fully connected layer for mapping to a soil moisture content scalar; is the soil moisture content prediction value, which is a scalar, and the unit is .

[0052] 2) Transpiration coupling prediction Global max pooling is performed on the context-enhanced feature map to retain the most significant features in each channel. Then, the pooled results are passed through a fully connected layer to generate a transpiration adjustment factor. The soil moisture prediction value is multiplied by the transpiration adjustment factor to output the crop transpiration prediction value, embedding the physical constraint of the positive correlation between soil moisture and transpiration, which is represented as: , In the formula, is global max pooling that retains the most significant features in each channel; is a fully connected layer that generates a transpiration adjustment factor; is the crop transpiration prediction value.

[0053] It should be noted that by embedding the physical constraint in the form of crop transpiration prediction, the model is forced to have a positive correlation between soil moisture and crop transpiration. The multiplication of the two tasks solves the problem of independent prediction heads violating biological and physical laws. In the case of data scarcity, the physical constraint serves as a strong prior to prevent the model from outputting contradictory values, such as a negative correlation between soil moisture and crop transpiration, thereby improving the reliability of predictions under extreme drought conditions.

[0054] S406, multi-objective weighted loss calculation The single mean square error loss function cannot balance the dimensional difference between soil moisture and crop transpiration, such as cubic meters per cubic meter and millimeters per day, and ignores the influence of meteorological factors on the transpiration process, leading to unbalanced multi-task optimization and inaccurate physical laws. The present application uses reference evapotranspiration as a meteorological prior to construct a hybrid loss function. Through multi-objective weighting and spatial consistency constraints, the dual-task optimization process is coordinated. The specific steps are as follows: 1) Soil moisture regression loss The L2 norm between the soil moisture prediction value and the true value of soil moisture is calculated to obtain the soil moisture regression loss, which measures the bias of soil moisture prediction, represented as: , In the formula, is the soil moisture regression loss; is the true value of soil moisture.

[0055] 2) Normalized transpiration regression loss The reference evapotranspiration value is used to normalize the transpiration prediction value and the true value of transpiration, respectively. Then, the L1 norm between the normalized prediction value and the normalized true value is calculated to obtain the normalized transpiration regression loss, which measures the bias of normalized transpiration prediction and eliminates the dimensional difference, represented as: , In the formula, The normalized evapotranspiration regression loss; This represents the true value of crop transpiration. To reference evapotranspiration values ​​and characterize the impact of meteorological conditions on evapotranspiration, the values ​​were calculated using the FAO-Penman-Monteith method based on the acquired temperature / humidity / wind speed / sunshine data. Given the L1 norm, calculate the absolute error.

[0056] 3) Spatial gradient consistency loss By applying the Sobel operator to the pixel-level soil moisture content prediction map, its spatial gradient map is calculated. Then, by applying the Sobel operator to the grayscale image of the illumination-corrected image, its spatial gradient map is calculated. Finally, the Kullback-Leibler divergence between the spatial gradient distribution of the predicted soil moisture content and the spatial gradient distribution of the image is calculated to obtain the spatial gradient consistency loss. This constraint ensures the consistency between the spatial distribution of the moisture gradient and the spatial distribution of the image visual gradient, and is expressed as: , In the formula, Spatial gradient consistency loss; To predict the spatial gradient map of soil moisture content, the spatial gradient is obtained by calculating the spatial gradient of the pixel-level soil moisture content prediction map using the Sobel operator, with dimensions of [dimensional value missing]. ; This is the spatial gradient map of the illumination-corrected image, obtained by analyzing the illumination-corrected image. The gradient of the grayscale image is obtained by calculating the gradient using the Sobel operator, with dimension . ; This is a function for calculating the Kullback-Leibler divergence, which measures the spatial consistency of two gradient distributions.

[0057] It should be noted that the dimension of the pixel-level soil moisture content prediction map is [dimension missing]. Define pixel-level soil moisture content prediction map as The calculation method is expressed as That is, by using a 1×1 convolutional layer to compress the number of channels to 1, the output dimension is reduced. The feature map is then used to constrain the predicted value to the [0,1] interval using the Sigmoid activation function, which corresponds to the normalized range of soil moisture content.

[0058] 4) Multi-objective weighted fusion The soil moisture content regression loss, normalized transpiration regression loss, and spatial gradient consistency loss are each multiplied by a preset weighting coefficient. Then, the three weighted losses are summed to output the final multi-objective weighted loss used for model optimization, expressed as: , In the formula, Multi-objective weighted loss; The weighting coefficient for soil moisture loss represents the dominant optimization direction, with an optimal value of 0.7. The weighting coefficient for transpiration loss balances the dimensional differences, with an optimal value of 0.3. is the weighting coefficient for spatial consistency loss, which strengthens physical constraints, and the preferred value is 0.1.

[0059] It should be noted that this is in conjunction with the reference evapotranspiration value. Normalize transpiration to eliminate dimensional differences, and at the same time... This project utilizes KL divergence loss to constrain the spatial consistency between moisture gradient and image gradient, such as abrupt changes in moisture at crop row boundaries. It combines meteorological priors with visual spatial structure to solve the imbalance problem in multi-task optimization.

[0060] S407, Sample Importance Weighting Uneven distribution of crop growth stages can lead to training models biased towards dominant growth stages, such as the vigorous growth period. Conventional resampling methods, such as oversampling or undersampling, can disrupt the continuity of time series, failing to preserve the temporal correlation of crop growth dynamics and hindering the model's ability to generalize to rare growth stages, such as the seedling or maturity stages. This invention addresses the problem of uneven distribution of growth stages by dynamically adjusting sample importance using time decay weights. The specific steps are as follows: 1) Calculate dynamic sample weights Calculate the absolute time difference between the sample acquisition time point and the average time point of the current training batch, and combine this with the attenuation rate coefficient to calculate the time decay term. Then, calculate the contrast factor of the illumination-corrected image, multiply the time decay term by the contrast factor, and output the dynamic weight value of the sample, expressed as: , In the formula, For the first The dynamic weight values ​​of each sample are used in the weighted loss function to adjust the importance of the samples. The decay rate coefficient controls the intensity of time decay, with a preferred value of 0.01. For the first The collection date sequence of each sample, that is, the time point when the sample was collected, expressed in days; For sample index; This represents the average daily sequence of the current training batch, which is the arithmetic mean of the collection times of all samples in the batch. The standard deviation of the image, from the illumination-corrected image. The calculated result reflects the image's contrast or noise level. The average value of the image is obtained from the illumination correction image The calculation result reflects the overall brightness of the image.

[0061] It should be noted that, The term represents the time decay term, which enhances the weight of recent samples and maintains the continuity of the growth time sequence, while, The term serves as a contrast factor, which prefers high-contrast images and forces more significant symptoms, both of which can integrate temporal locality and visual information content, solving the problem of destroying temporal correlation caused by resampling. In the crop seedling stage, the weight of high-contrast samples is increased, forcing the model to learn subtle stress features, such as cotyledon discoloration, and alleviating growth stage bias.

[0062] 2) Total loss calculation For each sample in the training batch, calculate its multi-objective weighted loss, multiply the multi-objective weighted loss of each sample by its corresponding dynamic sample weight, then calculate the average of all sample weighted losses, and output the final weighted total loss for model training, denoted as: , In the formula, is the weighted total loss; is the multi-objective weighted loss of the i-th sample. is the total number of samples.

[0063] S408, irrigation demand prediction model parameter update and iterative training The model training adopts a phased optimization strategy, as follows: 1) Initialization phase The Xavier initialization method is used to set the convolution layer weights, and the Adam optimizer is started with a learning rate of 0.001.

[0064] 2) Iterative update Each batch input 32 images and associated auxiliary data, forward propagation to calculate the weighted total loss; At the same time, the parameters are updated through the back propagation algorithm, and the learning rate is dynamically adjusted every 10 batches, with a decay factor of 0.95.

[0065] 3) Convergence condition judgment The early stopping mechanism is adopted on the validation set, and the loss value change of the weighted total loss is less than the preset threshold, such as 0.01, for 5 consecutive rounds, then the iterative training is terminated, indicating that the model training is completed.

[0066] S5, crop irrigation demand prediction in prefabricated pump station irrigation scenario The prediction process consists of four parts, as follows: 1) Real-time data acquisition ​The pump station camera captures the current irrigation area RGB image, synchronously reads the weather station temperature and humidity, wind speed and soil sensor data, and generates a timestamp-aligned data packet.

[0067] 2) Model inference The image is input into the trained prediction model, which sequentially performs adaptive illumination correction and background separation on the crop image through the spectral reflectance enhancement module, extracts spectral-texture features through double-flow separation convolution, dynamically weights the features through the cross-gate fusion module, strengthens the stress-sensitive area through the region-sensitive pyramid, and outputs the soil water content and transpiration through the coupled prediction head.

[0068] 3) Irrigation decision The central controller receives the prediction results to make irrigation decisions. For example, when the predicted soil water content is lower than the crop water threshold and the predicted crop transpiration is higher than the environmental evapotranspiration capacity, the frequency converter of the pump station is triggered to adjust the water pump speed to increase the irrigation speed, meeting the real-time irrigation control requirements.

[0069] Example 2 In this example, as shown in Figure 9 , the time series prediction effect analysis is performed to test the long-term accuracy of soil water content prediction and verify the physical constraint effect of the coupled prediction head module. The experiment shows the soil water content change over 30 days (unit: cubic meter / cubic meter, representing the volume ratio of water in each unit volume of soil), including four irrigation events (blue dashed line mark) and a drought period (red shaded area). The black solid line represents the true measured value. The traditional method (using a conventional convolutional neural network, red dashed line) significantly overestimates during the drought period and lags in response after irrigation. The dual-flow network (green dot-dash line) still has a deviation during the drought period. The present technology (blue solid line) closely tracks the true value throughout, with the smallest deviation during the drought period and rapid and accurate response after irrigation. It can be seen from the figure that the present technology predicts the closest to the true value at about 22 days, while the traditional method has the largest deviation during the drought period at 8-10 days, indicating that the coupled prediction head predicts soil water content and transpiration by multiplication, embedding the biophysical law that "increased water content leads to increased transpiration." Global max pooling retains the stress significant features, and the fully connected layer generates the transpiration adjustment factor, ensuring that the prediction results conform to the agricultural hydrological law and remain reliable under extreme conditions such as drought.

[0070] Example 3 In this example, as shown in Figure 10 , the crop full growth cycle prediction performance comparison is performed to show the comprehensive performance over a 120-day complete growth cycle. The horizontal axis represents the growth days (unit: days), and the vertical axis represents the comprehensive prediction error index (a dimensionless index integrating the soil water content and transpiration error).

[0071] In this embodiment, the gray background marks five growth stages, and the gray blocks clearly divide the five growth stages: 1) seedling stage (1-25 days): the crop canopy is sparse, and the traditional method is easily disturbed by the soil background; 2) jointing stage (26-50 days): stem growth is rapid, and water demand increases dramatically; 3) heading stage (51-75 days): the key period of reproductive growth, with high water sensitivity; 4) filling stage (76-100 days): grain filling, with hidden stress symptoms; and 5) mature stage (101-120 days): leaf senescence, with difficulty in feature extraction.

[0072] In this embodiment, the curve is smoothed by Gaussian filtering, and the scattered points are original data points sampled. The traditional convolutional neural network (gray curve) has a clear peak at the heading stage (around 55 days), reflecting its inadequate adaptability to the rapid growth stage. The method of the present application (yellow curve) is close to the bottom of the coordinate axis throughout, proving its stability.

[0073] In this embodiment, the scattered data points in the figure represent every 10 days of original data points, a total of 12 sampling points, used to reveal the original data fluctuation characteristics. For example, the ResNet method has high dispersion in the seedling stage, indicating that the prediction result is unstable, and verifying the reliability of the smoothed curve. All scattered points are distributed near the corresponding curve, excluding excessive smoothing distortion. The yellow scattered points are densely distributed in the vertical coordinate interval of 0.15-0.20, with a fluctuation range significantly smaller than other methods.

[0074] In this embodiment, three main phenomena are obtained from the experimental results: 1) the method of the present application (yellow curve) is stable throughout the lowest error interval; 2) the error fluctuation amplitude is the smallest in the seedling stage (1-25 days) and the mature stage (101-120 days); and 3) the green filled area quantitatively shows the sustained performance advantage of the present application relative to ResNet. Based on the three main phenomena, three main conclusions can be drawn: 1) the coupling prediction head module maintains prediction consistency at the end of growth through physical constraints; 2) the time decay weight enhances the time continuity of recent samples; and 3) the multi-objective weighted loss coordinates dual-task optimization. The significant width of the green area in the seedling stage verifies the generalization enhancement effect of the dynamic sample weight on the early growth stage.

[0075] Finally, it should be noted that the above description is only a preferred embodiment of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some technical features. Any modification, equivalent substitution, improvement, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

Claims

1. A method for predicting crop irrigation needs based on intelligent integrated prefabricated pump station, characterized by, The method comprises the following steps: Collecting crop images in a pre-prepared pumping station irrigation scene; obtaining auxiliary data and calculating daily cumulative reference evapotranspiration values, the auxiliary data being aligned with the crop image collection time to obtain a time-synchronized multi-modal data set; labeling and dividing the multi-modal data set; Constructing an irrigation demand prediction model, including a spectral reflectance enhancement module, a double-flow separation convolution module, a cross-gate fusion module, a region-sensitive pyramid module, and a coupled prediction head module; inputting the divided data in the training set into the model, training the model through a weighted total loss, and obtaining a trained model; inputting the image to be detected into the trained model to obtain soil moisture content and transpiration prediction results; Making irrigation decisions based on the soil moisture content and transpiration prediction results.

2. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 1, characterized in that, In the spectral reflectance enhancement module, adaptive illumination correction and background separation are performed on the crop image through hue-saturation-value (HSV) space conversion and a dynamic background mask generation method: Convert the original crop RGB image to the HSV color space, correct the illumination intensity distribution using the mean and standard deviation of the pixels in the value channel, compensate for color deviation using the mean of the pixels in the hue channel, and perform adaptive gain compensation using the standard deviation of the pixels in the saturation channel to output an illumination-corrected image; Calculate the Euclidean distance between the RGB vector of each pixel of the original crop image and the pre-calculated RGB mean vector of the typical soil region, then compare this distance with the normalized difference vegetation index threshold; when the Euclidean distance is greater than the threshold, it is determined to be a crop pixel, and the output is foreground; otherwise, it is determined to be a background soil pixel, and the output is background; then a binary crop mask image is generated; multiply the illumination-corrected image and the crop mask image element by element, set the pixel values of the background region to zero, and keep the pixel values of the crop foreground region unchanged to output a mask-corrected image.

3. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 2, characterized in that, Multi-scale spectral-texture feature collaborative extraction is performed through the double-flow separation convolution module: Spectral flow frequency domain feature extraction: use learnable spectral weights to weight each RGB channel of the illumination-corrected image, sum the weighted three-channel results, and perform two-dimensional discrete cosine transformation on the summed results to obtain a spectral feature map; Texture flow spatial feature extraction: use a depth separable convolution to extract spatial texture features from each RGB channel of the illumination-corrected image, perform batch normalization on the extracted features, and pass them through a rectified linear unit activation function to obtain a texture feature map.

4. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 3, characterized in that, Dynamic feature fusion and channel recalibration are performed through the cross-gate fusion module: Concatenate the spectral feature map and the texture feature map along the channel dimension, process the concatenated feature map through convolution and a Sigmoid activation function to obtain a dynamic fusion weight map; use the dynamic fusion weight map to weight the spectral feature map and the texture feature map to obtain a fused feature map; Perform global average pooling on the fused feature map to obtain a global feature vector, cross-channel interact the global feature vector through a calibration matrix, process the interaction result through a normalized exponential function, and then multiply it with the fused feature map channel by channel to obtain a recalibrated feature map.

5. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 4, characterized in that, Spatial context modeling and region importance weighting are performed by a region-sensitive pyramid module: The recalibrated feature map is subjected to average pooling operations with three different step sizes to obtain three feature maps with different scales; the three feature maps with different scales are spliced along the channel dimension to obtain a multi-scale feature map; the multi-scale feature map is subjected to a convolution operation and then a normalized exponential function is applied to the spatial dimension to obtain a spatial attention map; the recalibrated feature map is subjected to a max-pooling operation to obtain a global context feature, which is subjected to a convolution operation and then up-sampled to the spatial resolution of the recalibrated feature map by bilinear interpolation, and then the up-sampled global context feature is multiplied element by element with the spatial attention map to obtain a context-enhanced feature map.

6. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 5, characterized in that, Multi-task prediction and physical constraints are performed by a coupled prediction head module: The context-enhanced feature map is flattened into a one-dimensional vector, which is mapped by a fully connected layer to obtain a soil moisture prediction value; the context-enhanced feature map is subjected to global max-pooling, and the pooling result is subjected to a fully connected layer to generate a transpiration adjustment factor; The soil moisture prediction value and the transpiration adjustment factor are multiplied to obtain a crop transpiration prediction value.

7. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 6, characterized in that, Through multi-objective weighted sum and spatial consistency constraints, the dual-task optimization process is coordinated: Soil moisture regression loss: the L2 norm between the soil moisture prediction value and the true value of the soil moisture is calculated to obtain the soil moisture regression loss; Normalized transpiration regression loss: the transpiration prediction value and the true value of the transpiration are normalized using the reference evapotranspiration value, and the L1 norm between the normalized prediction value and the normalized true value is calculated to obtain the normalized transpiration regression loss; Spatial gradient consistency loss: the Sobel operator is applied to the pixel-level soil moisture prediction value to obtain the spatial gradient distribution of the predicted soil moisture; the Sobel operator is applied to the grayscale image of the light correction image to obtain the image spatial gradient distribution; the Kullback-Leibler divergence between the spatial gradient distribution of the predicted soil moisture and the image spatial gradient distribution is calculated to obtain the spatial gradient consistency loss.

8. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 7, characterized in that, The soil moisture regression loss, normalized transpiration regression loss, and spatial gradient consistency loss are multiplied by the preset weight coefficients, respectively, and then the three weighted losses are added to obtain the multi-objective weighted loss.

9. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 8, characterized in that, In the training process, the importance of the samples is dynamically adjusted by a time decay weight: The absolute time difference between the sample collection time point and the average time point of the current training batch is calculated, and the time decay term is calculated by combining the decay rate coefficient; The contrast factor of the light correction image is calculated, and the time decay term is multiplied by the contrast factor to obtain the dynamic weight value of the sample.

10. The crop irrigation requirement prediction method based on the intelligent integrated prefabricated pump station according to claim 9, characterized in that, The multi-objective weighted loss of each sample in the training batch is calculated, each sample's multi-objective weighted loss is multiplied by the corresponding dynamic sample weight, and then the average value of all sample weighted losses is calculated to obtain the final weighted total loss, which is used for model training.

Images