A Crop Irrigation Demand Prediction Method Based on Intelligent Integrated Prefabricated Pumping Stations

Abstract

This invention relates to a method for predicting crop irrigation demand based on an intelligent integrated prefabricated pumping station, belonging to the field of artificial intelligence technology. It includes the following steps: acquiring crop images in an irrigation scenario using a prefabricated pumping station, obtaining auxiliary data, and calculating daily cumulative reference evapotranspiration values ​​to obtain a time-synchronized multimodal dataset; labeling and partitioning the multimodal dataset; constructing an irrigation demand prediction model, including a spectral reflectance enhancement module, a two-stream separation convolution module, a cross-gated fusion module, a region-sensitive pyramid module, and a coupled prediction head module; inputting the partitioned training set data into the model, training the model using a weighted total loss to obtain a trained model; inputting the image to be detected into the trained model to obtain soil moisture content and evapotranspiration prediction results; and making irrigation decisions based on the soil moisture content and evapotranspiration prediction results. This invention can improve the accuracy of irrigation prediction.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a method for predicting crop irrigation demand based on an intelligent integrated prefabricated pumping station. Background Technology

[0002] With the increasing demand for water-saving, efficient, and precise irrigation in agricultural production, traditional irrigation strategies that rely on manual experience or rough meteorological indicators are no longer sufficient to meet the requirements of modern agriculture for "timely and appropriate" water supply. Especially against the backdrop of increasing climate uncertainty and water scarcity, there is an urgent need for a system solution that integrates crop sensing, physiological modeling, and intelligent decision-making.

[0003] Chinese invention patent CN120236209A discloses a method for predicting irrigation water demand in irrigation districts using UAV remote sensing, comprising the following steps: Step S1: Data acquisition; Step S2: Cluster optimization calculation; Step S3: Irrigation water demand prediction based on clustering results. This invention constructs an optimization factor, assigning different weights to different features to reflect their different indicative roles in irrigation water demand, and considering the correlation between spectral features. Specifically, the weight of each band is determined by the correlation between spectral features and NDVI; bands with higher correlation to NDVI are assigned higher weights; NDVI is also assigned a higher weight due to its close relationship with crop growth and water conditions. The optimization factor constructed in this way can more accurately reflect the importance of different features in irrigation water demand prediction, improving the application accuracy of the K-means clustering algorithm in this scenario. Chinese invention patent CN120202905A discloses an irrigation prediction method, device, equipment, and medium. The method includes: acquiring an initial noise action sequence; when the acquired 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 predicted 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, acquiring the predicted 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, using the predicted action sequence corresponding to the current iteration step number as the current noise action sequence corresponding to the next iteration step number, and incrementing the current iteration step number by 1. This embodiment of the invention can improve the accuracy of irrigation prediction.

[0004] The following problems still exist in the existing technology: Conventional convolutional neural network models directly perform convolution processing on RGB images, resulting in the mixing of spectral and spatial information, which cannot effectively identify the weak spectral changes and local morphological changes of water stress; traditional methods predict water content and transpiration separately, lacking coupled modeling, and may output predicted values ​​that violate the crop transpiration mechanism; in multi-task training, conventional loss ignores the dimensional differences between soil moisture and transpiration, and does not consider the constraints of image structure on water distribution, causing the optimization objective to deviate from the actual needs; traditional resampling methods disrupt the time sequence or lead to information redundancy, making it difficult to adapt to the modeling needs of early or late-stage crop stress states, affecting the model's generalization ability and temporal continuity. Summary of the Invention

[0005] To achieve the above objectives, the present invention employs the following technical solution:

[0006] This invention provides a method for predicting crop irrigation demand based on an intelligent integrated prefabricated pumping station, comprising the following steps:

[0007] Crop images under the irrigation scenario of prefabricated pumping stations are collected, auxiliary data is obtained, and the daily cumulative reference evapotranspiration value is calculated. The auxiliary data is aligned with the crop image acquisition time to obtain a time-synchronized multimodal dataset. The multimodal dataset is then labeled and divided.

[0008] An irrigation demand prediction model is constructed, including a spectral reflectance enhancement module, a two-stream separation convolution module, a cross-gated fusion module, a region-sensitive pyramid module, and a coupled prediction head module. Data from the partitioned training set is input into the model, and the model is trained using a weighted total loss to obtain a trained model. The image to be detected is input into the trained model to obtain the predicted results of soil moisture content and transpiration. Irrigation decisions are made based on the predicted results of soil moisture content and transpiration.

[0009] Furthermore, in crop irrigation demand prediction tasks, crop RGB (Red-Green-Blue) images are easily affected by strong illumination changes and complex background interference. Conventional processing methods such as Z-score normalization are difficult to effectively eliminate shadow variations and soil background noise, resulting in the model being highly sensitive to illumination conditions and reducing prediction accuracy. This invention employs a spectral reflectance enhancement module, using hue-saturation-brightness space conversion and dynamic background mask generation methods to perform adaptive illumination correction and background separation on crop images. The specific steps are as follows:

[0010] Hue-Saturation-Lightness HSV Color Space Light Correction: Convert the original crop RGB image to HSV color space, use the pixel mean and standard deviation of the lightness channel to correct the light intensity distribution, combine the pixel mean of the hue channel for color shift compensation, and use the standard deviation of the saturation channel for adaptive gain compensation to output a uniformly lit corrected image and eliminate the influence of shadow variations.

[0011] Crop mask generation based on normalized differential vegetation index: The Euclidean distance between the RGB vector of each pixel in the original crop image and the pre-calculated mean RGB vector of a typical soil region is calculated. This distance is then compared with a normalized differential vegetation index threshold. If the Euclidean distance is greater than the threshold, the pixel is identified as a crop pixel and output as the foreground; otherwise, it is identified as a background soil pixel and output as the background. This process generates a binarized crop mask image, achieving separation between the crop foreground and the soil background.

[0012] Background mask correction: The illumination-corrected image is multiplied element-wise with the crop mask image. The pixel values ​​of the background area are set to zero, while the pixel values ​​of the crop foreground area remain unchanged, resulting in a mask-corrected image that is both illumination-robust and has a clean crop foreground.

[0013] Furthermore, 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 obfuscation problem. The specific steps for multi-scale spectral-texture feature co-extraction via the dual-stream separable convolution module are as follows:

[0014] Spectral flow frequency domain feature extraction: Learnable spectral weights are used to weight each RGB channel of the illumination-corrected image to achieve spectral decoupling. The weighted three-channel results are summed, and a two-dimensional discrete cosine transform is performed on the summation result to convert the spatial information into a frequency domain representation, thereby obtaining a spectral feature map that retains the main low-frequency components and capturing frequency domain patterns related to water stress.

[0015] Texture flow spatial feature extraction: Depth-separable convolution is used to extract spatial texture features from each RGB channel of the illumination-corrected image. The extracted features are batch normalized and then processed by a modified linear unit activation function to obtain a texture feature map, preserving the spatial structure information of the canopy.

[0016] Furthermore, this invention employs a cross-gated fusion module to achieve adaptive feature selection and optimizes feature representation through channel recalibration to reduce redundancy. Dynamic feature fusion and channel recalibration are performed through the cross-gated fusion module. The specific steps are as follows:

[0017] The spectral feature map and texture feature map are concatenated along the channel dimension. The concatenated feature map is then processed by convolution and the sigmoid activation function to obtain a dynamically fused weight map. The dynamically fused weight map is used to weight the spectral feature map and texture feature map to obtain a fused feature map, realizing adaptive soft selection of spectral and texture features. Global average pooling is performed on the fused feature map to obtain a global feature vector. The global feature vector is then interacted across channels through a calibration matrix. The interaction result is processed by a normalized exponential function and then multiplied with the fused feature map channel by channel to obtain a recalibrated feature map, optimizing the importance distribution of each feature channel.

[0018] Furthermore, different regions of the crop canopy exhibit significant differences in their sensitivity to water stress. Conventional pooling operations can lose crucial local information, making it difficult for models to accurately capture the spatial heterogeneity of water stress. This invention utilizes a region-sensitive pyramid module for spatial context modeling and region importance weighting. The specific steps are as follows:

[0019] Three different strides of average pooling are applied to the recalibrated feature map to obtain three different scales of feature maps. The three feature maps of different scales are concatenated along the channel dimension to obtain a multi-scale feature map. The multi-scale feature map is then convolved and a normalized exponential function is applied in the spatial dimension to obtain a spatial attention map, whose pixel values ​​quantify the sensitivity of the spatial location to water stress. Max pooling is applied to the recalibrated feature map to obtain global context features. The global context features are then convolved and upsampled to the spatial resolution of the recalibrated feature map through bilinear interpolation. The upsampled global context features are then multiplied element-wise with the spatial attention map to obtain a context enhancement feature map.

[0020] Furthermore, irrigation demand forecasting requires the simultaneous output of soil moisture content and crop transpiration. Conventional single-output head methods neglect the physical correlation of water transport, potentially leading to prediction results that violate the fundamental biophysical law that transpiration increases with increasing soil moisture content. This invention performs multi-task forecasting and physical constraints by coupling the forecast head module. The specific steps are as follows:

[0021] The context-enhanced feature map is flattened into a one-dimensional vector, and this one-dimensional vector is mapped through a fully connected layer to obtain the predicted soil moisture content. Global max pooling is performed on the context-enhanced feature map to retain the most salient features of each channel. The pooling result is passed through a fully connected layer to generate a transpiration regulation factor. The predicted soil moisture content is multiplied by the transpiration regulation factor to obtain the predicted crop transpiration.

[0022] Furthermore, a single mean squared error loss function cannot balance the dimensional differences between soil moisture content and crop transpiration, such as cubic meters per cubic meter versus millimeters per day, and ignores the influence of meteorological factors on the transpiration process, leading to imbalances in multi-task optimization and inaccuracies in physical laws. This invention uses reference evapotranspiration as a meteorological prior to construct a hybrid loss function; through multi-objective weighting and spatial consistency constraints, it coordinates the dual-task optimization process, including:

[0023] Soil moisture content regression loss: The L2 norm between the predicted soil moisture content and the actual soil moisture content is calculated to obtain the soil moisture content regression loss, which measures the deviation of the soil moisture content prediction.

[0024] Normalized evapotranspiration regression loss: The predicted and actual evapotranspiration values ​​are normalized using reference evapotranspiration values. The L1 norm between the normalized predicted and actual values ​​is calculated to obtain the normalized evapotranspiration regression loss, which measures the deviation of the normalized evapotranspiration prediction and eliminates dimensional differences.

[0025] Spatial gradient consistency loss: The Sobel operator is applied to the pixel-level soil moisture content prediction to obtain the spatial gradient distribution of the predicted soil moisture content; the Sobel operator is applied to the grayscale image of the illumination-corrected image to obtain the image spatial gradient distribution; 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, which constrains the consistency between the spatial distribution of the moisture gradient and the spatial distribution of the image visual gradient.

[0026] Furthermore, the soil moisture content regression loss, normalized evapotranspiration regression loss, and spatial gradient consistency loss are multiplied by preset weight coefficients, and then the three weighted losses are added together to obtain the multi-objective weighted loss.

[0027] Furthermore, the uneven distribution of crop growth stages can lead to training models biased towards major growth stages, such as the vigorous growth period. Conventional resampling methods, such as oversampling or undersampling, can disrupt the continuity of the 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. To address this, the importance of samples is dynamically adjusted using time decay weights: 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 using the decay rate coefficient; the contrast factor of the illumination-corrected image is calculated, and the time decay term is multiplied by the contrast factor to obtain the dynamic weight value of the sample.

[0028] Furthermore, a multi-objective weighted loss is calculated for each sample in the training batch. The multi-objective weighted loss of each sample is multiplied by its corresponding dynamic sample weight, and then the average of the weighted losses of all samples is calculated to obtain the final weighted total loss, which is used for model training.

[0029] The advantages of this invention are:

[0030] This invention employs a parallel extraction mechanism of spectral and texture streams, combined with frequency domain analysis and spatial modeling, to accurately extract crop water stress signals and solve the feature redundancy problem caused by RGB channel confusion. It embeds the physical causal relationship of "increased water content - increased transpiration" into the soil moisture and transpiration prediction task, ensuring that the model output follows agricultural hydrology principles through multiplicative modeling, thus improving prediction reliability under extreme drought conditions. It uses reference evapotranspiration values ​​for transpiration normalization and leverages KL divergence to constrain the consistency between the spatial gradient of soil moisture and image edge structure, significantly improving the realism and coherence of the predicted spatial distribution. Finally, it employs a sample weighting mechanism based on time decay and image contrast to avoid disrupting the crop growth sequence using traditional resampling methods, enhancing the model's generalization ability to low-frequency stages such as seedling and maturity. Attached Figure Description

[0031] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0032] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0033] Figure 2 The images shown are the original crop images collected in a specific embodiment of the present invention.

[0034] Figure 3 This is an illumination correction image in a specific embodiment of the present invention;

[0035] Figure 4 This is a crop mask image in a specific embodiment of the present invention;

[0036] Figure 5 This is a mask correction image in a specific embodiment of the present invention;

[0037] Figure 6 This is a spectral feature map in a specific embodiment of the present invention;

[0038] Figure 7 This is a texture feature map in a specific embodiment of the present invention;

[0039] Figure 8 This is a fusion feature map in a specific embodiment of the present invention;

[0040] Figure 9 This is a graph showing the effect of soil moisture content time series prediction in a specific embodiment of the present invention;

[0041] Figure 10This is a comparison of the crop full growth cycle prediction performance in specific embodiments of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] Example 1

[0044] In this embodiment, as Figure 1 As shown, this invention provides a method for predicting crop irrigation demand based on an intelligent integrated prefabricated pumping station, the specific steps of which include:

[0045] S1. Acquire crop images in a prefabricated pumping station irrigation scenario.

[0046] In the scenario of prefabricated pumping station irrigation, crop image acquisition is completed in collaboration between fixed multispectral cameras deployed in the pumping station control area and mobile drones;

[0047] A fixed camera is installed on the support of the main irrigation pipeline of the pumping station at a top-down angle. It automatically triggers shooting at three time periods each day before, during and after irrigation (06:00, 12:00 and 18:00), covering the core crop area controlled by the pumping station, such as a radius of 50 meters.

[0048] The drones equipped with GPS positioning fly twice a week along a preset route, at an altitude of 3 meters above the canopy, and take pictures at a 45-degree tilt angle to supplement the perspective and eliminate blind spots from the fixed perspective.

[0049] During the crop image acquisition process, complete crop growth cycle data, i.e. from sowing to maturity, must be included. The original images are stored in uncompressed RGB format and cached locally through the pump station edge computing gateway.

[0050] In one embodiment, the acquired raw crop image is as follows: Figure 2 As shown.

[0051] S2. Obtain auxiliary data and calculate the daily cumulative reference evapotranspiration value.

[0052] In addition to crop images, real-time temperature, humidity, wind speed and sunshine duration data from weather stations were acquired simultaneously. Based on the weather station data, the daily cumulative reference evapotranspiration value was calculated using the FAO-Penman-Monteith method. Solar radiation data was obtained through calibration using a solar intensity meter on the top of the pump station. All auxiliary data were strictly aligned with the image acquisition time to form a time-synchronized multimodal dataset.

[0053] S3. Training Data Labeling and Training Dataset Construction

[0054] The collected crop images were labeled, and information such as soil moisture content and transpiration status during the data acquisition period was combined with the manual labeling by agronomic experts to label each crop image.

[0055] Specifically, soil moisture content levels (drought / normal / excessively wet) are classified based on data from soil moisture sensors measured simultaneously, and transpiration status (wilted / normal / saturated) is determined by combining canopy temperature and stomatal conductance measurement results.

[0056] The labeled data was divided into a dataset containing 70% training set (stratified sampling according to growth stage), 15% validation set, and 15% test set.

[0057] S4. Construction and Training of Irrigation Demand Forecasting Model

[0058] S401, the spectral reflectance enhancement module performs adaptive illumination correction and background separation for crop images.

[0059] In crop irrigation demand prediction tasks, 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 are difficult to effectively eliminate shadow changes and soil background noise, resulting in the model being highly sensitive to light conditions and reducing prediction accuracy.

[0060] This invention employs a spectral reflectance enhancement module, which achieves adaptive illumination correction and background separation through HSV (Hue-Saturation-Value) space conversion and dynamic background mask generation, eliminating illumination changes and background interference. The specific steps are as follows:

[0061] 1) Illumination correction in HSV space

[0062] The original crop RGB image is converted to the HSV color space. Then, the pixel mean and standard deviation of the light intensity channel are used to correct the unevenness of the light intensity distribution. At the same time, the pixel mean of the hue channel is used for color shift compensation, and the standard deviation of the saturation channel is used for adaptive gain compensation. The output is a uniformly lit, light-corrected image that eliminates the influence of shadow variations, represented as:

[0063] ,

[0064] In the formula, This is a crop image, specifically an original RGB image of the crop, with dimensions of [dimension number missing]. ; This is an illumination-corrected image that adaptively adjusts the illumination using luminance channel statistics in the HSV color space, and eliminates shadow variations by combining hue and saturation compensation terms, outputting a uniformly illuminated RGB image with dimensions of [missing information]. ; Image height; Image width; It is the pixel mean of the brightness channel in the HSV space, used to correct for unevenness in light intensity distribution; is the pixel standard deviation of the brightness channel in the HSV color space, representing the intensity of illumination fluctuations; It is a minimal constant and serves as a numerical stability term to prevent the denominator from being zero; its preferred value is [value missing]. ; The adaptive saturation gain coefficient is calculated as follows: ; It is a natural exponential function; The standard deviation of the HSV color space saturation channel is used to enhance color contrast in low light conditions. The hue compensation term is calculated as follows: ; This is the average value of the hue channels in the HSV color space, used to balance color shifts; This is a truncation function that restricts pixel values ​​to the range [0, 255].

[0065] In one embodiment, the illumination-corrected image is as follows: Figure 3 As shown.

[0066] 2) Crop mask generation based on normalized difference vegetation index

[0067] The Euclidean distance between the RGB vector of each pixel in the original crop image and the pre-calculated mean RGB vector of a typical soil region is calculated. This distance is then compared to a normalized differential vegetation index threshold. If the Euclidean distance is greater than the threshold, the pixel is identified as a crop pixel and output as the foreground; otherwise, it is identified as a background soil pixel and output as the background. This process generates a binary crop mask image, achieving separation between the crop foreground and the soil background, as shown below:

[0068] ,

[0069] In the formula, This is a crop mask image, with elements being binary numerical values ​​of 0 / 1, and a dimension of . ; The pre-calculated RGB mean vector of a typical soil region is obtained by pre-collecting RGB images of typical soil regions without vegetation cover and calculating their pixel RGB mean vector. The calculation method is expressed as follows: ; This represents the number of pixels in a typical soil area without vegetation cover. For the first Pixel values ​​of a typical soil area without vegetation cover; For pixel value index; It uses the L2 norm, and is calculated in the same way as the Euclidean distance. The term calculates the RGB vector of each pixel of the crop and The Euclidean distance represents vegetation cover; The normalized differential vegetation index threshold is used to determine crop pixels, with a preferred value of 0.3. This is an indicator function that outputs 1 when the condition is true and 0 otherwise.

[0070] In one embodiment, the crop mask image is as follows: Figure 4 As shown.

[0071] It should be noted that, The term represents pixel-level binary judgment. When the condition is met, it outputs 1, representing the crop area; otherwise, it outputs 0, representing the background area. This enables accurate vegetation segmentation in complex farmland backgrounds.

[0072] It should also be noted that the preferred threshold value for the normalized difference vegetation index is 0.3. Experiments have verified that this threshold can effectively separate crops from soil, such as the canopy reflectance characteristics of corn and wheat.

[0073] 3) Background mask correction

[0074] The illumination-corrected RGB image is multiplied element-wise with the generated binarized crop mask. Pixel values ​​in the background area are set to zero, while pixel values ​​in the crop foreground area remain unchanged. The output is a mask-corrected image that combines illumination robustness with a clean crop foreground, represented as:

[0075] ,

[0076] 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... .

[0077] In one embodiment, the masked image is as follows: Figure 5 As shown.

[0078] S402, dual-stream separation and convolution module for multi-scale spectral-texture feature co-extraction

[0079] 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:

[0080] 1) Spectral flow frequency domain feature extraction

[0081] 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:

[0082] ,

[0083] 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.

[0084] In one embodiment, the spectral feature map is as follows: Figure 6 As shown.

[0085] It should be noted that water stress causes changes in the cell structure of crop leaves, such as stomatal closure, which manifests as abnormal reflectance in specific mid-to-high frequency domains. The term uses two-dimensional discrete cosine transform to convert spatial texture into frequency domain energy distribution. The microscopic changes of water stress exhibit significant characteristics in the frequency domain, such as the attenuation of high-frequency components, while conventional spatial convolution is difficult to capture such patterns.

[0086] 2) Texture flow spatial feature extraction

[0087] For each RGB channel of the illumination-corrected image, a depthwise separable convolution is independently applied to extract spatial texture features. Then, batch normalization is performed on the extracted features, and a modified linear unit activation function is applied to enhance nonlinearity. The resulting texture feature map preserves the canopy spatial structure information and is represented as follows:

[0088] ,

[0089] In the formula, For depthwise separable convolutions oriented towards the first The spatial convolution kernels with channels are learnable parameters; the dimension is . ; For depthwise convolution operations, spatial convolution is performed independently for each channel; This is a batch normalization layer used to standardize the feature distribution; To modify the activation function of the linear unit; The texture feature map has the following dimensions: ; The number of spatial feature channels is 32, with the preferred value being 32.

[0090] In one embodiment, the texture feature map is as follows: Figure 7 As shown.

[0091] It should be noted that the spectral characteristic map The characterization uses frequency domain features directly extracted via discrete cosine transform, focusing on reflectance frequency shifts caused by moisture stress, such as near-infrared band attenuation, and texture feature maps. By employing depthwise separable convolution to preserve the canopy spatial structure, such as leaf curling morphology, the separable design solves the spectral confusion problem caused by RGB channel coupling. It can decouple frequency domain analysis from spatial modeling, making frequency domain flow more sensitive to early weak signals of water stress, such as mild wilting, while spatial flow is more robust to macroscopic deformation under severe stress. The two complement each other and significantly improve the model's discriminative power for stress stages.

[0092] It should also be noted that spectral weighting As a learnable scalar, the three RGB channels are independently weighted to eliminate inter-channel coupling. For example, the strong correlation between the red and green channels in the green leaf region is eliminated. Channel decoupling is performed before frequency domain transformation, so that the two-dimensional discrete cosine transform is more focused on the stress-sensitive channel. This enables the model to automatically enhance the contribution of the moisture response channel. For example, if the weight of the first channel is increased, the second channel containing noise is suppressed, which effectively improves the frequency domain feature discrimination.

[0093] It should also be noted that maintaining spatial resolution constant, and Independent settings, preferred settings , This allows the model to allocate feature capacity on demand, resulting in a higher number of texture stream channels to capture complex canopy structures, and a lower number of spectral stream channels due to the greater sparsity of frequency domain features. Forced spectral flow to learn key frequency domain modes of water stress, avoiding redundant features, large This enhances the ability of texture flow to model local details, such as the ability to model local details like wilting leaf edges.

[0094] S403, cross-gating fusion module performs feature dynamic fusion and channel recalibration

[0095] Directly concatenating spectral and texture features leads to information redundancy, while conventional attention mechanisms ignore cross-modal correlations, which is detrimental to feature fusion. This invention employs a cross-gated fusion module to achieve adaptive feature selection and optimizes feature representation through channel recalibration to reduce redundancy. The specific steps are as follows:

[0096] 1) Cross-gating feature fusion

[0097] The spectral and texture feature maps are concatenated along the channel dimension. A 1x1 convolution operation is then applied to the concatenated feature map, and the convolution output is passed through a sigmoid activation function to generate a dynamic fusion weight map ranging from 0 to 1. This weight map is then used to weight the spectral and texture feature maps, resulting in a fusion feature map. This achieves adaptive soft selection of spectral and texture features, as shown below:

[0098] ,

[0099] ,

[0100] In the formula, This is for channel splicing operations. The output dimension of the item is ; For dynamically fused weight graphs, dimensions The Sigmoid function compresses the value to... Control the mixing ratio of spectral / texture features; It is a 1×1 convolutional layer. The item compresses the number of channels to ; Use the Sigmoid activation function; To fuse feature maps, the dimension is .

[0101] In one embodiment, the fused feature map is as follows: Figure 8 As shown.

[0102] 2) Channel-level recalibration

[0103] Global average pooling is performed on the fused feature map to obtain a global feature vector. This global feature vector is then interacted across channels using a calibration matrix. The interaction result is processed using a normalized exponential function. The resulting channel weight vector is broadcast to the spatial dimension and then multiplied channel-by-channel with the fused feature map to output a recalibrated feature map. This optimizes the importance distribution of each feature channel, as shown below:

[0104] ,

[0105] In the formula, For global average pooling, Xiang Jiang Compress to vector; The calibration matrix is ​​a learnable parameter that enables cross-channel interaction; its dimension is [missing information]. ; It is a normalized exponential function; This is a channel multiplication operation used to broadcast the weight vector to the spatial dimension and then multiply it with the feature map channel by channel; To recalibrate the feature map, the dimension is .

[0106] It should be noted that the Sigmoid gate enables soft selection between spectral and texture features, rather than hard switching, allowing the model to dynamically adjust weights based on local region characteristics. For example, the canopy center is dominated by spectral features. Approaching 1, the edges are mainly characterized by texture features. Approaching 0, while adopting a spatial adaptive fusion method to solve the information redundancy problem of splicing operation, in the transition region of water stress, such as semi-wilted leaves, the gating mechanism automatically mixes dual-flow features to generate a more continuous stress distribution map and avoid spatial abrupt changes in the prediction results.

[0107] It should also be noted that the calibration matrix Modeling cross-modal channel correlations, such as the association between the low-frequency component of the spectral flow and the edge detection channel of the texture flow, and realizing global channel interaction through a fully connected matrix, solves the problem that conventional attention mechanisms ignore cross-modal dependencies. The stress response channel and the spatial structure channel are synergistically enhanced, such as the synergistic enhancement of the mid-frequency component of the two-dimensional discrete cosine transform and the leaf texture features, so that the fused features still maintain high discriminability under low light conditions.

[0108] It should also be noted that global average pooling and Convolution can significantly reduce computational overhead, enabling feature recalibration through lightweight operations. The low computational load allows the model to process high-definition images in real time, meeting the needs of online field monitoring.

[0109] S404, the region-sensitive pyramid module performs spatial context modeling and region importance weighting.

[0110] Different regions of the crop canopy exhibit significant differences in their sensitivity to water stress. Conventional pooling operations can lose crucial local information, making it difficult for models to accurately capture the spatial heterogeneity of water stress. This invention employs a region-sensitive pyramid module, which enhances the feature representation of water stress-sensitive regions through multi-scale contextual feature fusion and adaptive region weighting. The specific steps are as follows:

[0111] 1) Constructing multi-scale features

[0112] Three different step sizes of average pooling are applied to the recalibrated feature maps to obtain feature maps of three different scales. These three feature maps are then concatenated along the channel dimension to output a multi-scale feature map containing multi-scale contextual information such as local leaf details, plant structure, and canopy group patterns, represented as:

[0113] ,

[0114] In the formula, The step size is Average pooling operation; This is the step size for the pooling operation. This represents the pooling step size set, which controls the feature map downsampling rate. This is a splicing operation along the channel dimension. Indicates the splicing step size along the channel dimension. The average pooling result; This is a multi-scale feature map with the following dimensions: ; This is the sum of the number of feature channels at each scale.

[0115] 2) Generate regional importance weights

[0116] A 3x3 convolutional layer is applied to the multi-scale feature pyramid. Then, a normalized exponential function is applied to the convolutional result in the spatial dimension to output a spatial attention map. The pixel values ​​of this map quantify the sensitivity of the spatial location to water stress, as shown below:

[0117] ,

[0118] In the formula, This is a spatial attention graph with dimensions [missing information]. The range is Quantify pixel-level moisture stress sensitivity; It is a 3×3 convolutional layer; For the Softmax function along the spatial dimension;

[0119] 3) Contextual Feature Enhancement

[0120] Max pooling with a stride of 4 is performed on the recalibrated feature map to obtain global context features. Then, a 1x1 convolutional layer is applied to the global context features, and the result is upsampled to the original image spatial resolution through bilinear interpolation. Finally, the upsampled global context features are multiplied element-wise with the spatial attention map to output a context enhancement feature map, which enhances the feature representation of the stress-sensitive region, as shown below:

[0121] ,

[0122] In the formula, Context-enhanced feature maps, with dimensions of ; This is max pooling with a step size of 4; The spatial resolution of the bilinear interpolation upsampling to the recalibrated feature map is, i.e., the resolution is... .

[0123] It should be noted that, through step size Multi-scale pooling, when Extracting local leaf details in real time Extracting plant structure at the time, Extract canopy group patterns in real time and combine them with spatial attention maps. Weighted stress-sensitive regions, such as wilted leaf areas, are enhanced with spatial attention by combining multi-scale context with spatial importance. This addresses the information homogenization problem of conventional pooling, enabling the model to still function effectively even in canopy-occluded scenarios. Large-scale context inference of the stress state of hidden regions and attention focus on local salient areas such as leaf tip scorch can improve the robustness of complex farmland environments.

[0124] S405, Coupled prediction head module performs multi-task prediction and physical constraints.

[0125] In irrigation demand forecasting, it is necessary to simultaneously output soil moisture content and crop transpiration. Conventional single-output head methods ignore the physical correlation of water transport, which may lead to prediction results that violate the fundamental biophysical law that transpiration increases with increasing soil moisture content. This invention employs a coupled prediction head module, embedding physical constraints on soil moisture content and transpiration to simultaneously output dual-objective prediction values. The specific steps are as follows:

[0126] 1) Soil moisture content prediction

[0127] The context enhancement feature map is flattened into a one-dimensional vector, and then this one-dimensional vector is mapped through a fully connected layer to output the predicted soil moisture content, represented as:

[0128] ,

[0129] In the formula, This is a spatial flattening operation, where the flattened vector is a one-dimensional vector. This is a fully connected layer used for mapping to a soil moisture content scalar; This is a predicted value for soil moisture content; it is a scalar quantity, and the unit is 1000 ppm. .

[0130] 2) Evapotranspiration coupled prediction

[0131] Global max pooling is performed on the context-enhanced feature map to retain the most salient features of each channel. The pooling result is then passed through a fully connected layer to generate a transpiration modulator. The predicted soil moisture content is multiplied by this transpiration modulator to output the predicted crop transpiration. The physical constraint that soil moisture content is positively correlated with transpiration is embedded, which is represented as:

[0132] ,

[0133] In the formula, To achieve global max pooling, the most salient features of each channel are preserved; As a fully connected layer, it generates transpiration regulation factors; This is a predicted value for crop transpiration.

[0134] It should be noted that by embedding physical constraints into the model in the way crop transpiration is predicted, the positive correlation between soil moisture content and crop transpiration is forced. The output of the two tasks is coupled by multiplication, which solves the problem of independent prediction heads violating biophysical laws. When data is scarce, physical constraints serve as strong priors to prevent the model from outputting contradictory values, such as the contradiction of negative correlation between soil moisture content and crop transpiration, thereby improving the reliability of prediction under extreme drought conditions.

[0135] S406, Multi-objective weighted loss calculation

[0136] A single mean squared error loss function cannot balance the dimensional differences between soil moisture content and crop transpiration, such as cubic meters per cubic meter versus millimeters per day, and it ignores the influence of meteorological factors on the transpiration process, leading to imbalances in multi-task optimization and inaccuracies in physical laws. This invention uses reference evapotranspiration as a meteorological prior, constructs a hybrid loss function, and coordinates the dual-task optimization process through multi-objective weighting and spatial consistency constraints. The specific steps are as follows:

[0137] 1) Soil moisture content regression loss

[0138] The L2 norm between the predicted and actual soil moisture content is calculated to obtain the soil moisture content regression loss, which measures the bias of the soil moisture content prediction and is expressed as:

[0139] ,

[0140] In the formula, This represents the regression loss of soil moisture content; This represents the actual soil moisture content.

[0141] 2) Normalized evapotranspiration regression loss

[0142] The predicted and actual evapotranspiration values ​​are normalized using reference evapotranspiration values. Then, the L1 norm between the normalized predicted and actual values ​​is calculated to obtain the regression loss of normalized evapotranspiration, which measures the bias in the prediction of normalized evapotranspiration and eliminates dimensional differences. This is expressed as:

[0143] ,

[0144] 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.

[0145] 3) Spatial gradient consistency loss

[0146] 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:

[0147] ,

[0148] 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.

[0149] 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.

[0150] 4) Multi-objective weighted fusion

[0151] 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:

[0152] ,

[0153] 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.

[0154] 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.

[0155] S407, Sample Importance Weighting

[0156] 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:

[0157] 1) Calculate dynamic sample weights

[0158] 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:

[0159] ,

[0160] 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, obtained from the illumination-corrected image. The calculated value reflects the overall brightness of the image.

[0161] It should be noted that, The term representing time decay strengthens the weight of recent samples, maintains the continuity of the growth sequence, and at the same time... As a contrast factor, the term favors high-contrast images, which show more pronounced stress symptoms. The two can integrate temporal locality and visual information to solve the problem of resampling disrupting temporal correlation. In the seedling stage of crops, the increased weight of high-contrast samples forces the model to learn subtle stress features, such as cotyledon chlorosis, thus alleviating growth stage bias.

[0162] 2) Total Loss Calculation

[0163] 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, and then calculate the average of the weighted losses of all samples. The final weighted total loss is output for model training, expressed as:

[0164] ,

[0165] In the formula, Weighted total loss; For the first Multi-objective weighted loss for each sample. The total number of samples.

[0166] S408. Parameter Update and Iterative Training of Irrigation Demand Forecasting Model

[0167] The model training employs a phased optimization strategy, as detailed below:

[0168] 1) Initialization phase

[0169] Set the convolutional layer weights using the Xavier initialization method and start the Adam optimizer with a learning rate of 0.001.

[0170] 2) Iterative update

[0171] Each batch of input consists of 32 images and associated auxiliary data, and the weighted total loss is calculated using forward propagation.

[0172] Meanwhile, parameters are updated through backpropagation algorithm, and the learning rate is dynamically adjusted every 10 batches, with a learning rate decay factor of 0.95.

[0173] 3) Convergence condition judgment

[0174] An early stopping mechanism is used on the validation set. If the change in the weighted total loss value is less than a preset threshold, such as 0.01, the iterative training is terminated, which means that the model training is complete.

[0175] S5. Crop irrigation demand prediction in prefabricated pump station irrigation scenarios.

[0176] The prediction process consists of four stages, as follows:

[0177] 1) Real-time data acquisition

[0178] The pump station camera captures RGB images of the current irrigation area, and simultaneously reads temperature, humidity, wind speed, and soil sensor data from the weather station, generating a timestamp-aligned data packet.

[0179] 2) Model Inference

[0180] The image is input into the trained prediction model, and the crop image is then subjected to adaptive illumination correction and background separation through the spectral reflectance enhancement module, spectral-texture feature extraction through dual-stream separation convolution, dynamic weighted feature addition through the cross-gated fusion module, stress-sensitive area enhancement through the region-sensitive pyramid, and soil moisture content and transpiration output through the coupled prediction head.

[0181] 3) Irrigation decision

[0182] The central controller receives the prediction results and makes irrigation decisions. For example, when the predicted soil moisture content is lower than the crop water requirement threshold and the predicted crop transpiration is higher than the environmental evapotranspiration capacity, the pump station frequency converter is triggered to adjust the pump speed, speed up irrigation, and meet the real-time irrigation control requirements.

[0183] Example 2

[0184] In this embodiment, as Figure 9 As shown, a time-series prediction effect analysis was conducted to verify the long-term accuracy of soil moisture content prediction and validate the physical constraint effect of the coupled prediction head module. The experiment demonstrates the change in soil moisture content over 30 days (unit: cubic meters / cubic meter, representing the volume percentage of water in each unit volume of soil), including four irrigation events (marked by blue dashed lines) and one drought period (red shaded area). The black solid line represents the actual measured value. The traditional method (using a conventional convolutional neural network, red dashed line) significantly overestimated the value during the drought period and showed a delayed response after irrigation. The dual-stream network (green dotted line) still showed bias during the drought period. This technology (blue solid line) closely tracked the actual value throughout the entire process, with the smallest bias during the drought period and a rapid and accurate response after irrigation. The figure shows that the prediction of this technology is closest to the actual value at around 22 days, while the traditional method has the largest bias during the 8-10 day drought period. This indicates that the coupled prediction head predicts soil moisture content and transpiration by multiplicatively associating them, embedding the biophysical law that "increased moisture content leads to increased transpiration." Global max pooling preserves significant stress characteristics, and fully connected layers generate transpiration regulation factors, ensuring that the prediction results conform to agricultural hydrological patterns and maintain reliability under extreme conditions such as drought.

[0185] Example 3

[0186] In this embodiment, as Figure 10 As shown, a comparison of crop growth cycle prediction performance is presented, demonstrating the comprehensive performance over a complete 120-day growth cycle. The horizontal axis represents the number of growth days (unit: days), and the vertical axis represents the comprehensive prediction error index (a dimensionless index that integrates errors in soil moisture content and transpiration).

[0187] In this embodiment, five growth stages are marked with a gray background, and the gray background blocks clearly divide the five growth stages: 1) Seedling stage (1-25 days): The crop canopy is sparse, and traditional methods are easily affected by the soil background; 2) Jointing stage (26-50 days): The stem grows rapidly, and the water demand increases dramatically; 3) Heading stage (51-75 days): The critical period of reproductive growth, with high water sensitivity; 4) Grain filling stage (76-100 days): The grains are full, and stress symptoms are hidden; 5) Maturity stage (101-120 days): The leaves are senescent, and feature extraction is difficult.

[0188] In this embodiment, the curve is smoothed using a Gaussian filter, and the scatter plots are samples of the original data points. Traditional convolutional neural networks (gray curve) show a significant peak at the heading stage (around day 55), reflecting their insufficient adaptability to the rapid growth phase. The method of this invention (yellow curve) remains close to the bottom of the coordinate axis throughout, demonstrating its stability.

[0189] In this embodiment, the scatter plot data in the figure represents the original data points extracted every 10 days, for a total of 12 sampling points. This is used to reveal the fluctuation characteristics of the original data. For example, the ResNet method has high scatter plot dispersion in the seedling stage, indicating that the prediction results are unstable and verifying the reliability of the smoothing curve. All scatter plots are distributed near the corresponding curve, eliminating over-smoothing distortion. The scatter plots can also verify the advantages of the present invention. The yellow scatter plots are densely distributed in the ordinate range of 0.15-0.20, and the fluctuation range is significantly smaller than that of other methods.

[0190] In this embodiment, three main phenomena were observed through experimental results: 1) The method of this invention (yellow curve) remained consistently within the lowest error range throughout the entire process; 2) The error fluctuation was minimal during the seedling stage (1-25 days) and the maturity stage (101-120 days); 3) The green-filled region quantifies the sustained performance advantage of this invention over ResNet. Based on these three main phenomena, three main conclusions can also be drawn: 1) The coupled prediction head module maintains prediction consistency at the end of the growth stage through physical constraints; 2) The time decay weight enhances the temporal continuity of recent samples; 3) The multi-objective weighted loss coordinates dual-task optimization. The significant width of the green region in the figure during the seedling stage verifies the generalization enhancement effect of dynamic sample weights on the early growth stage.

[0191] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention 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 of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting crop irrigation demand based on an intelligent integrated prefabricated pumping station, characterized in that, Includes the following steps: Crop images are collected under the irrigation scenario of a prefabricated pumping station; auxiliary data is acquired and the daily cumulative reference evapotranspiration value is calculated. The auxiliary data is aligned with the crop image acquisition time to obtain a time-synchronized multimodal dataset; the multimodal dataset is labeled and divided. An irrigation demand prediction model is constructed, including a spectral reflectance enhancement module, a two-stream separation convolution module, a cross-gated fusion module, a region-sensitive pyramid module, and a coupled prediction head module. Data from a segmented training set is input into the model, and the model is trained using a weighted total loss to obtain a trained model. The image to be detected is input into the trained model to obtain the predicted soil moisture content and evapotranspiration. Irrigation decisions are made based on the predicted soil moisture content and evapotranspiration. The specific process is as follows: Multi-scale spectral-texture feature co-extraction is performed using a dual-stream separable convolution module: spectral stream frequency domain feature extraction: learnable spectral weights are used to weight each RGB channel of the illumination-corrected image, the weighted three-channel results are summed, and a two-dimensional discrete cosine transform is performed on the summation result to obtain the spectral feature map; texture stream spatial feature extraction: depthwise separable convolution is used to extract spatial texture features from each RGB channel of the illumination-corrected image, the extracted features are batch normalized, and the texture feature map is obtained after passing through a modified linear unit activation function. Dynamic feature fusion and channel recalibration are performed through a cross-gated fusion module: the spectral feature map and the texture feature map are concatenated along the channel dimension, and the concatenated feature map is processed by convolution and the Sigmoid activation function to obtain a dynamic fusion weight map; the dynamic fusion weight map is used to weight the spectral feature map and the texture feature map to obtain a fused feature map; Global average pooling is performed on the fused feature map to obtain a global feature vector. The global feature vector is then interacted across channels through a calibration matrix. The interaction result is processed by a normalized exponential function and then multiplied channel by channel with the fused feature map to obtain a recalibrated feature map. Spatial context modeling and region importance weighting are performed using a region-sensitive pyramid module: Average pooling with three different strides is applied to the recalibrated feature map to obtain feature maps of three different scales; these three feature maps are concatenated along the channel dimension to obtain a multi-scale feature map; the multi-scale feature map is then convolved and a normalized exponential function is applied in the spatial dimension to obtain a spatial attention map; max pooling is performed on the recalibrated feature map to obtain global context features; these global context features are then convolved and upsampled to the spatial resolution of the recalibrated feature map using bilinear interpolation; finally, the upsampled global context features are element-wise multiplied with the spatial attention map to obtain a context-enhanced feature map. Multi-task prediction and physical constraints are performed by coupling the prediction head module: the context enhancement feature map is flattened into a one-dimensional vector, and the one-dimensional vector is mapped through a fully connected layer to obtain the predicted soil moisture content. Global max pooling is performed on the context enhancement feature map, and the pooling result is passed through a fully connected layer to generate an evaporation regulation factor; The predicted soil moisture content is multiplied by the transpiration regulator to obtain the predicted crop transpiration.

2. The crop irrigation demand prediction method based on intelligent integrated prefabricated pumping stations according to claim 1, characterized in that, In the spectral reflectance enhancement module, adaptive illumination correction and background separation of crop images are performed using hue-saturation-brightness HSV space conversion and dynamic background mask generation methods. The original crop RGB image is converted to HSV color space. The light intensity distribution is corrected by using the pixel mean and standard deviation of the lightness channel. Color shift compensation is performed by combining the pixel mean of the hue channel and adaptive gain compensation is performed by using the standard deviation of the saturation channel. The output is a light-corrected image. The Euclidean distance between the RGB vector of each pixel in the original crop image and the pre-calculated mean RGB vector of a typical soil region is calculated. This distance is then compared with a normalized differential vegetation index threshold. If the Euclidean distance is greater than the threshold, the pixel is identified as a crop pixel and output as the foreground; otherwise, it is identified as a background soil pixel and output as the background. A binarized crop mask image is then generated. The illumination-corrected image is multiplied element-wise with the crop mask image, and the pixel values ​​of the background region are set to zero while the pixel values ​​of the crop foreground region remain unchanged. The mask-corrected image is then output.

3. The crop irrigation demand prediction method based on intelligent integrated prefabricated pumping stations according to claim 2, characterized in that, Coordinate the dual-task optimization process by using multi-objective weighting and spatial consistency constraints: Soil moisture content regression loss: Calculate the L2 norm between the predicted soil moisture content and the actual soil moisture content to obtain the soil moisture content regression loss; Normalized evapotranspiration regression loss: Normalize the predicted evapotranspiration value and the actual evapotranspiration value using the reference evapotranspiration value, calculate the L1 norm between the normalized predicted value and the normalized actual value, and obtain the normalized evapotranspiration regression loss. Spatial gradient consistency loss: The Sobel operator is applied to the pixel-level soil moisture content prediction to obtain the spatial gradient distribution of the predicted soil moisture content; the Sobel operator is applied to the grayscale image of the illumination-corrected image to obtain the image spatial gradient distribution; the Kullback-Leibler divergence between the spatial gradient distribution of the predicted soil moisture content and the image spatial gradient distribution is calculated to obtain the spatial gradient consistency loss.

4. The crop irrigation demand prediction method based on intelligent integrated prefabricated pumping stations according to claim 3, characterized in that, The soil moisture content regression loss, normalized transpiration regression loss, and spatial gradient consistency loss are multiplied by preset weight coefficients, and then the three weighted losses are added together to obtain the multi-objective weighted loss.

5. The crop irrigation demand prediction method based on intelligent integrated prefabricated pumping stations according to claim 4, characterized in that, During training, the importance of samples is dynamically adjusted using time decay weights. Calculate the absolute time difference between the sample collection time point and the average time point of the current training batch, and combine it with the decay rate coefficient to calculate the time decay term; Calculate the contrast factor of the illumination-corrected image, and multiply the time decay term by the contrast factor to obtain the dynamic weight value of the sample.

6. The crop irrigation demand prediction method based on intelligent integrated prefabricated pumping stations according to claim 5, characterized in that, For each sample in the training batch, calculate the multi-objective weighted loss, multiply the multi-objective weighted loss of each sample by the corresponding dynamic sample weight, and then calculate the average of the weighted losses of all samples to obtain the final weighted total loss, which is used for model training.

Images