Field surface water depth measurement method and system based on image intelligent recognition

Abstract

The application relates to the technical field of image processing, and discloses a field water depth determination method and system based on image intelligent identification. The method comprises the following steps: performing multi-period image acquisition and processing on a field plot to obtain pretreatment image data; inputting the pretreatment image data into a two-stage network for matching to obtain water gauge three-dimensional coordinates; performing segmentation based on the water gauge coordinates to obtain scale reading and water surface intersection point data; performing interpolation processing on the reading data and field plot point cloud to obtain a water depth model and performing decision analysis to obtain a heterogeneity distribution map and a diagnosis matrix; and inputting the distribution map and the matrix into a learning framework for optimization processing to obtain an irrigation control parameter set. The application can accurately identify the water gauge scale by using the image intelligent identification technology under complex field environment conditions, realize high-precision automatic determination of the field water depth, organically combine the determination result with the irrigation control decision, form an intelligent closed-loop management system, and thus improve the utilization efficiency of irrigation water resources.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method and system for measuring field surface water depth based on intelligent image recognition. Background Art

[0002] In the agricultural irrigation sector, traditional methods for measuring field surface water depth rely primarily on manual observation of water gauge readings. Workers regularly visit fields, record water level readings on the water gauge, and derive water depth data through simple calculations. With technological advancements, some semi-automated methods have been implemented, such as installing water level sensors and buoy-type water level gauges, enabling automated collection of water level data. However, these methods are still limited to point measurements and cannot fully reflect the spatial distribution of water depth within a field. In recent years, remote sensing technology has achieved some application in large-scale water monitoring, but due to resolution limitations, it is difficult to meet the needs of refined irrigation management for small-scale farmland. Meanwhile, computer vision technology has made significant progress in industrial inspection, but its application in agricultural water conservancy is relatively lagging. In particular, specialized algorithms for water gauge recognition and water depth measurement remain imperfect.

[0003] Traditional technical methods have numerous shortcomings: First, manual observation methods are labor-intensive and inefficient, making frequent monitoring difficult to achieve, resulting in a lack of timely data support for irrigation management decisions. Second, point measurements cannot reflect the spatial distribution characteristics of water depth within a field, resulting in a lack of scientific basis for irrigation uniformity evaluation. Third, existing automated equipment is expensive and complex to maintain, facing economic barriers to large-scale application. Fourth, there is a lack of a closed-loop management mechanism that organically integrates water depth measurement results with irrigation decision optimization, limiting the scope for improving irrigation efficiency. Fifth, traditional methods are unable to cope with interfering factors in complex irrigation environments, such as light changes, water surface fluctuations, and weed obstruction, making it difficult to ensure the accuracy and stability of measurement results. These shortcomings have seriously restricted the scientific and precise level of farmland irrigation, affecting water resource utilization efficiency and increasing crop yields. Summary of the Invention

[0004] The present application provides a method and system for measuring field surface water depth based on image intelligent recognition, which is used to accurately identify the water gauge scale under complex field environmental conditions using image intelligent recognition technology, realize high-precision automatic measurement of field surface water depth, and organically combine the measurement results with irrigation control decisions to form an intelligent closed-loop management system, thereby improving the efficiency of irrigation water resource utilization.

[0005] In a first aspect, the present application provides a method for measuring field surface water depth based on image intelligent recognition, the method comprising: collecting multi-period images of a field provided with a water gauge and performing time series processing to obtain time series preprocessed image data; inputting the time series preprocessed image data into a two-stage feature pyramid network for multi-scale anchor frame matching to obtain three-dimensional coordinates and attitude angle data of the water gauge; performing instance segmentation based on the three-dimensional coordinates and attitude angle data of the water gauge to obtain water gauge scale readings and water surface boundary intersection data; combining the water gauge scale readings and water surface boundary intersection data with a high-density discrete point cloud of a field sampled by actual measurement, and obtaining a dynamic water depth distribution grid model through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing; performing Markov decision analysis based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix; inputting the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix into a deep reinforcement learning framework for multi-step decision optimization processing under constraints to obtain a set of regional irrigation control parameters.

[0006] In a second aspect, the present application provides a field surface water depth measurement system based on image intelligent recognition, the field surface water depth measurement system based on image intelligent recognition comprising:

[0007] A processing module is used to collect images of the field with the water gauge in multiple time periods and perform time series processing to obtain time series pre-processed image data;

[0008] A matching module is used to input the temporal preprocessed image data into a two-stage feature pyramid network to perform multi-scale anchor frame matching to obtain the three-dimensional coordinates and posture angle data of the water ruler;

[0009] A segmentation module is used to perform instance segmentation based on the three-dimensional coordinates and posture angle data of the water gauge to obtain water gauge scale readings and water surface boundary intersection data;

[0010] A correction module is used to combine the water gauge scale readings and water surface boundary intersection data with the high-density discrete point cloud of the field sampled by actual measurement, and obtain a dynamic water depth distribution grid model through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing;

[0011] an analysis module for performing Markov decision analysis based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix;

[0012] The input module is used to input the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix into the deep reinforcement learning framework to perform multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set.

[0013] In a third aspect, a field surface water depth measurement device based on image intelligent recognition is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory so that the field surface water depth measurement device based on image intelligent recognition executes the above-mentioned field surface water depth measurement method based on image intelligent recognition.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned method for measuring field surface water depth based on intelligent image recognition.

[0015] In the technical solution provided in this application, by applying image intelligent recognition technology to the field of field water depth measurement, the entire process from image acquisition to irrigation control is automated, effectively solving the technical problems of low efficiency and poor accuracy of traditional water depth measurement methods. At the same time, the combination of deep reinforcement learning framework and Markov decision analysis provides intelligent decision-making support for irrigation optimization. Specifically, multi-period image acquisition and time series processing technology ensure the continuity and reliability of the original data, providing a high-quality data foundation for subsequent analysis; the introduction of the two-stage feature pyramid network greatly improves the accuracy and robustness of water gauge positioning, especially under complex lighting conditions and partial occlusion. The multi-scale anchor frame matching mechanism effectively extracts water gauge features and accurately obtains its three-dimensional coordinates and attitude angle data; the instance segmentation method based on water gauge coordinates solves the problem of water gauge scale recognition, and can accurately read the scale value even under the interference of water surface fluctuations; the combination of variational adaptive thin plate spline interpolation algorithm and hydraulic simulation correction realizes the accurate conversion from discrete points to continuous water depth distribution, overcoming the limitation of traditional point measurement that cannot express spatial distribution; Markov decision analysis introduces time series data processing capabilities, and accurately diagnoses irrigation anomalies through comprehensive analysis of historical status and current observations; the deep reinforcement learning framework adaptively generates the optimal irrigation control strategy through multi-step decision optimization under constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0017] Figure 1 This is a schematic diagram of an embodiment of a method for measuring field surface water depth based on image intelligent recognition in an embodiment of the present application;

[0018] Figure 2An embodiment of a field surface water depth measurement system based on image intelligent recognition in the embodiments of the present application is shown in the figure;

[0019] Figure 3 An embodiment of a field surface water depth measurement system based on image intelligent recognition in the embodiments of the present application is shown in the figure; DETAILED DESCRIPTION

[0020] The embodiments of the present application provide a field surface water depth measurement method and system based on image intelligent recognition. The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0021] For the sake of understanding, the specific process of the embodiments of the present application is described below. Please refer to Figure 1 An embodiment of a field surface water depth measurement method based on image intelligent recognition in the embodiments of the present application includes:

[0022] Step S101, multi-period image acquisition is performed on the field plot provided with a water gauge and time sequence processing is performed to obtain time sequence pre-processing image data;

[0023] Step S102, multi-scale anchor frame matching is performed on the time sequence pre-processing image data input into a two-stage feature pyramid network to obtain water gauge three-dimensional coordinate and attitude angle data;

[0024] Step S103, instance segmentation is performed based on the water gauge three-dimensional coordinate and attitude angle data to obtain water gauge scale reading and water surface boundary intersection point data;

[0025] Step S104, the water gauge scale reading and water surface boundary intersection point data are combined with the high-density discrete point cloud of the field plot actually measured, and through variational self-adaptive thin-plate spline interpolation and hydraulic simulation correction processing, a dynamic water depth distribution grid model is obtained;

[0026] Step S105, Markov decision analysis is performed based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix;

[0027] Step S106: Input the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix into the deep reinforcement learning framework to perform multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set.

[0028] It is understandable that the execution subject of this application can be a field surface water depth measurement system based on image intelligent recognition, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0029] Specifically, multi-time image acquisition is performed on the fields where the water gauge is installed. This process uses a camera fixed to the edge of the field, controlling the shooting frequency to once an hour to obtain a sequence of original images. These images contain water gauge and water surface information, but there are often interference factors such as water surface ripples and weed obstruction. These interferences are eliminated through noise filtering to obtain clear images of the field water surface. These images are then histogram equalized to balance the brightness differences under different lighting conditions and standardize the image data. Perspective correction is then performed based on the reference points at the four corners of the field to make the image conform to the actual proportional relationship. The corrected image is subjected to detail enhancement processing to sharpen the edge features of the water gauge scale for subsequent identification. Finally, the enhanced image is combined with the acquisition time, field number and geographic location information to form time-series preprocessed image data.

[0030] The temporal preprocessed image data is fed into a two-stage feature pyramid network. The network first extracts feature maps of different scales through the first-stage network to generate a set of coordinates for candidate water gauge regions. Non-maximum suppression is performed on these candidate regions, and a confidence threshold of 0.85 is set to select high-confidence bounding boxes for the water gauge regions. The image within these bounding boxes is then cropped and fed into the second-stage network for fine-grained feature extraction, generating detailed contour data for the water gauge. A Hough transform is then performed on this contour data to detect the vertical edges of the water gauge and obtain its orientation and angle values ​​in the image plane. These angle values ​​are then mapped to pre-calibrated reference points to establish a correspondence between image space and physical space, resulting in a spatial transformation matrix. Finally, combined with the image resolution scale factor calculation, triangulation is used to determine the three-dimensional coordinates and attitude angle data of the water gauge.

[0031] Instance segmentation is performed based on the three-dimensional coordinates and attitude angle data of the water gauge. First, the image region of the water gauge is subjected to image enhancement processing, and the contrast parameter is adjusted to obtain a high-contrast water gauge image. Then, adaptive binary processing is performed to separate the water gauge scale and background by setting a dynamic threshold to generate a scale feature binary image. The image is input into a convolutional neural network for feature extraction to identify the main scale lines and numbers on the water gauge and establish water gauge scale reference system data. Based on this parameter system data, edge detection operations are performed to locate the intersection line between the water surface and the water gauge and obtain the water surface boundary line coordinate point set. Robustness analysis and processing are performed on these coordinate points to eliminate fluctuations and the effects of reflections, and stable water surface line position data is obtained. Finally, linear interpolation calculation is performed on the water surface line position data and the water gauge scale reference system data to determine the water gauge scale reading and the water surface boundary intersection point data. The water gauge scale reading and the water surface boundary intersection point data are combined with the high-density discrete point cloud of the field to process. The field point cloud measured by RTK-GPS technology is preprocessed, and abnormal points are detected and corrected by a sliding window to obtain a set of terrain feature points. These point sets are input into a variational adaptive thin-plate spline interpolation algorithm, and a 0.5m x 0.5m resolution parameter is set to construct a continuous field digital elevation model. Regional analysis is performed on the model to identify special terrain elements such as dikes and drainage ditches, and complete terrain structure data is formed. The water gauge scale reading and the water surface boundary intersection point data are spatially registered with the terrain structure data to establish a unified elevation reference, and the water surface-terrain correspondence relationship is obtained. Based on this relationship, raster calculation is performed to compare the water level elevation and the field elevation to generate initial water depth distribution data. Finally, the hydrodynamic model is input for flow field analysis, and the water surface imbalance factor is corrected by considering the water flow direction and flow rate parameters to obtain a dynamic water depth distribution grid model.

[0032] Markov decision analysis is performed based on the dynamic water depth distribution grid model. First, the dynamic water depth distribution grid model is compared with the crop irrigation water demand database to extract the current crop ideal irrigation water depth parameters and obtain water depth deviation distribution data. The uniformity index of these data is calculated to quantitatively evaluate the consistency of water depth distribution and obtain an irrigation uniformity evaluation index. This index is combined with the irrigation deficiency rate and the irrigation excess rate to obtain a comprehensive irrigation efficiency index through a comprehensive evaluation method. Based on this index, spatial clustering analysis is performed to identify the distribution pattern of water depth abnormal regions and form a water depth abnormal region feature map. This map is input into the Markov decision model to obtain a classification result of irrigation abnormal reasons by pattern matching with the expert knowledge base. Finally, the classification result is combined with the abnormal region feature map to establish a correspondence between irrigation problems and regional distribution, and an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix are obtained.

[0033] The irrigation heterogeneity distribution map and irrigation system fault diagnosis matrix are input into the deep reinforcement learning framework for optimization processing. First, the irrigation heterogeneity distribution map and fault diagnosis matrix are integrated, the water dynamic relationship of the field is established, and the irrigation control demand data is obtained. These data are input into the multi-objective optimization algorithm, considering irrigation uniformity, water resource utilization rate and operation cost, to obtain irrigation control candidate schemes. Based on these candidate schemes, classification processing is performed to distinguish different irrigation problem types and form differentiated irrigation optimization schemes. The schemes are prioritized, considering the implementation difficulty and expected effect, to obtain an irrigation execution task list. The list is matched with the field characteristic parameters to calculate the best irrigation time and flow, and the regional irrigation parameters are obtained. Finally, the parameters are arranged in time sequence to generate operation instruction sequence and execution schedule, and the regional irrigation control parameter set is obtained.

[0034] For example, for a certain field, first, the image is collected by the camera installed at the edge of the field, and after image processing, a clear water gauge image is obtained. The two-stage feature pyramid network identifies the water gauge position and determines its three-dimensional coordinates as (X=23.56 meters, Y=45.78 meters, Z=102.35 meters) and an attitude angle of 15.7 degrees. The instance segmentation network identifies the water gauge scale reading as 54.3 cm, and the water surface boundary intersection point is located between the 5th and 6th scales. Combined with the RTK-GPS collected field terrain point cloud data, a 0.5m x 0.5m resolution field digital elevation model is constructed, and after spatial registration and hydraulic simulation, the water depth values of each point in the field are obtained to form a dynamic water depth distribution grid model. Markov decision analysis finds that the water depth in the northwest corner of the field is significantly lower than the average value, which is diagnosed as a dike leakage problem, generating an irrigation heterogeneity distribution map and a fault diagnosis matrix. Based on these data, the deep reinforcement learning framework calculates that the northwest corner needs to increase 25% of the irrigation time, and generates a regional irrigation control parameter set, including the irrigation flow, duration and sequence of each region, forming an irrigation execution scheme.

[0035] In the embodiments of the present application, by applying image intelligent recognition technology to the field of water depth measurement, the whole process automation from image acquisition to irrigation control is realized, effectively solving the technical problems of low efficiency and poor precision of traditional water depth measurement methods. At the same time, the combination of deep reinforcement learning framework and Markov decision analysis provides intelligent decision support for irrigation optimization. Specifically, the multi-period image acquisition and timing processing technology ensures the continuity and reliability of the original data, providing a high-quality data foundation for subsequent analysis; the introduction of the two-stage feature pyramid network greatly improves the accuracy and robustness of water gauge positioning, especially under complex lighting conditions and partial occlusion, the multi-scale anchor box matching mechanism effectively extracts the water gauge feature, and accurately obtains its three-dimensional coordinate and attitude angle data; the instance segmentation method based on water gauge coordinates solves the problem of water gauge scale recognition, and even under water surface fluctuation interference, the scale value can be accurately read; the method combining the variational adaptive thin plate spline interpolation algorithm with the hydraulic simulation correction realizes the accurate conversion from discrete points to continuous water depth distribution, overcoming the limitation that traditional point measurement cannot express spatial distribution; Markov decision analysis introduces timing data processing capability, and through comprehensive analysis of historical state and current observation, irrigation abnormal phenomena are accurately diagnosed; the deep reinforcement learning framework optimizes multi-step decision-making under constraints, and adaptively generates the optimal irrigation control strategy.

[0036] In a specific embodiment, the process of performing step S101 can specifically include the following steps:

[0037] Adjust the illumination parameters of the camera equipment fixedly installed on the edge of the field, control the shooting frequency to be once per hour, and obtain a raw image sequence containing a water gauge and a water surface;

[0038] Perform noise filtering processing on the raw image sequence, eliminate water surface ripples and weed interference, and obtain a field water surface image;

[0039] Input the field water surface image into a histogram equalization processing unit, balance the image brightness under different lighting conditions, and obtain standardized image data;

[0040] Perform perspective correction based on the standardized image data, perform coordinate transformation by using the four corner reference points of the field, and obtain an orthographic image;

[0041] Perform detail enhancement processing on the orthographic image, sharpen the edge features of the water gauge scale, and obtain an enhanced image;

[0042] Combine the enhanced image with the acquisition time, field number and geographic location information for processing, and obtain timing preprocessed image data.

[0043] Specifically, the light parameter adjustment is performed on the camera equipment fixedly installed at the edge of the field block, the original image sequence containing the water gauge and the water surface is obtained by controlling the shooting frequency to be once per hour, the field water surface image is obtained by eliminating the water surface ripple and the weed interference through noise filtering processing on the original image sequence, the standardized image data is obtained by balancing the image brightness under different light conditions through inputting the field water surface image into the histogram equalization processing unit, the orthographic image is obtained through coordinate transformation by using the four corner reference points of the field block based on the standardized image data, the enhanced image is obtained through sharpening the water gauge scale edge features by performing detail enhancement processing on the orthographic image, and the time sequence pretreatment image data is obtained through combination processing of the enhanced image and the collection time, the field block number and the geographic location information.

[0044] The camera equipment is fixed at the best observation position of the field block edge, and the aperture, shutter speed and ISO sensitivity and other parameters of the camera equipment are adjusted to adapt to different weather and light conditions. The camera equipment adopts a high-resolution camera module, and the shooting frequency is set to once per hour, and the system automatically records the shooting time. When the light condition changes, the camera equipment will adjust according to the preset automatic exposure parameters to ensure the image clarity. The obtained original image sequence is saved as an RGB three-channel color image with a resolution of 3840x2160 pixels.

[0045] For noise filtering processing of the original image sequence, an improved Gaussian filtering algorithm is adopted to specially process the ripple reflection and water grass shielding and other interference factors specific to the farmland water surface. In the filtering process, the kernel size is set to 5x5 pixels, and the filtering strength parameter is dynamically adjusted according to the image noise degree. This processing can effectively remove the high-frequency interference and random noise points generated by the water surface ripple, while retaining the clear outline of the water gauge edge, and the water gauge scale in the obtained field water surface image is more clear and distinguishable.

[0046] After inputting the field water surface image into the histogram equalization processing unit, the system analyzes the image brightness distribution, calculates the pixel brightness histogram, and then applies the adaptive histogram equalization algorithm for processing. This algorithm divides the image into multiple small regions for local equalization, and then combines the processing results through bilinear interpolation, effectively solving the problem of uneven image brightness caused by different light conditions at different times. The processed standardized image data has a more balanced brightness distribution, and whether the image is taken on a cloudy day or a sunny day, it can present similar brightness characteristics.

[0047] When performing perspective correction based on standardized image data, the system first identifies reference marker points pre-set at the four corners of the field, which have accurate geographic coordinate information. The perspective transformation matrix is calculated through these reference points to perform geometric correction on the image, eliminating the perspective distortion caused by the installation angle of the camera. In the corrected orthographic image, the actual physical size of the water gauge scale is proportional to the pixel distance in the image, providing accurate spatial reference for subsequent water depth measurement.

[0048] When performing detail enhancement processing on the orthographic image, a nonlinear sharpening algorithm is used to enhance the edge features of the water gauge scale lines. During processing, selective enhancement is performed on the specific frequency range of the water gauge scale lines, while suppressing background textures and noise. In the enhanced image, the edges of the water gauge scale lines are sharper and the digital markings are clearer, facilitating subsequent scale recognition and water level reading. The enhanced image is combined with related metadata to add accurate collection timestamps, field numbers, geographic position coordinates (latitude and longitude), and camera device numbers to each image, forming time-series preprocessed image data with complete marker information. These metadata are stored in the image file header information, facilitating subsequent data analysis and tracing. The time-series preprocessed image data serves as the basic data input of the system, providing high-quality image sources for subsequent water gauge recognition and positioning.

[0049] In a specific embodiment, the process of performing step S102 can specifically include the following steps:

[0050] Inputting the time-series preprocessed image data into a first-stage network containing a backbone network and a feature pyramid structure to obtain a water gauge candidate region coordinate set through feature map extraction of different scales;

[0051] Performing non-maximum suppression operation on the water gauge candidate region coordinate set and screening through setting a 0.85 confidence threshold to obtain a high-confidence water gauge region bounding box;

[0052] Inputting the image within the high-confidence water gauge region bounding box into a second-stage network to obtain water gauge fine contour data through fine-grained feature extraction;

[0053] Performing Hough transform calculation based on the water gauge fine contour data to obtain the direction angle value of the water gauge in the image plane by detecting the vertical edge line of the water gauge;

[0054] Mapping and transforming the direction angle value of the water gauge in the image plane and the pre-calibrated reference scale point to obtain the spatial transformation matrix of the water gauge by establishing the corresponding relationship between the image space and the physical space;

[0055] Combining the spatial transformation matrix with the image resolution scale factor to obtain the three-dimensional coordinates and attitude angle data of the water gauge through the principle of triangulation.

[0056] Specifically, the time-series preprocessed image data is input into the first-stage network comprising a backbone network and a feature pyramid structure, and the coordinate set of the candidate water ruler area is obtained by extracting feature maps of different scales; a non-maximum suppression operation is performed on the coordinate set of the candidate water ruler area, and a high-confidence water ruler area bounding box is obtained by setting a confidence threshold of 0.85; the image cropping within the high-confidence water ruler area bounding box is input into the second-stage network, and the fine contour data of the water ruler is obtained by extracting fine-grained features; Hough transform calculation is performed based on the fine contour data of the water ruler, and the directional angle value of the water ruler in the image plane is obtained by detecting the vertical edge line of the water ruler; the directional angle value of the water ruler in the image plane is mapped and transformed with the pre-calibrated reference scale points, and the spatial transformation matrix of the water ruler is obtained by establishing a correspondence between the image space and the physical space; the spatial transformation matrix is ​​combined with the image resolution scale factor for calculation, and the three-dimensional coordinates and posture angle data of the water ruler are obtained through the principle of triangulation.

[0057] The temporally preprocessed image data is fed into the first stage of a two-stage feature pyramid network. This network, using ResNet50 as its backbone, extracts multi-level features from the image and then fuses feature maps of different resolutions using a feature pyramid structure. The feature pyramid structure consists of five levels, P2 to P6, corresponding to scales of 1 / 4, 1 / 8, 1 / 16, 1 / 32, and 1 / 64 of the original image, respectively. The network sets anchor boxes of varying sizes on these feature maps, optimizing for the elongated features of the water gauge, with aspect ratios set to 1:5, 1:8, and 1:10. A region proposal network calculates the confidence score and bounding box coordinate offset for each anchor box containing the water gauge, generating a set of candidate region coordinates for the water gauge.

[0058] When performing non-maximum suppression on the coordinate set of water gauge candidate regions, all candidate regions are first sorted in descending order by confidence score. The Intersection over Union (IoU) between the candidate regions is then calculated. When the IoU value of two candidate regions exceeds 0.5, the region with the higher confidence score is retained, and the region with the lower confidence score is suppressed. This process is repeated until all regions have been processed. Finally, a confidence threshold of 0.85 is set to filter out the bounding boxes of high-confidence water gauge regions, ensuring the accuracy of the detection results.

[0059] After cropping the image within the high-confidence water gauge region bounding box, it is fed into the second-stage network for fine-grained feature extraction. The second-stage network utilizes a convolutional neural network with an attention mechanism, specifically optimized for water gauge scale features. This network includes spatial and channel attention modules, adaptively focusing on the scale lines and numerals, effectively extracting the fine structural features of the water gauge. The network outputs a pixel-level segmentation mask of the water gauge, forming fine-scale contour data and accurately locating the edges and scale positions of the water gauge.

[0060] When performing a Hough transform on fine-scale water gauge contour data, edge detection is first performed on the contour data to extract a set of edge points. These edge points are then mapped into Hough space, and line parameters are calculated using an accumulator. Angular constraints are set to prioritize near-vertical edge lines based on the vertical characteristics of the water gauge. Peak detection identifies the primary vertical edge lines, and their orientation and angle values ​​in the image plane are calculated to accurately characterize the water gauge's tilt.

[0061] When mapping the orientation and angle values ​​of the water gauge in the image plane to pre-calibrated reference scale points, the pre-marked reference scale points on the water gauge are used, which have known physical coordinates. By establishing a correspondence between the pixel coordinates of the reference points in the image and their physical coordinates, the perspective transformation parameters are calculated to form a spatial transformation matrix. This matrix describes the transformation from image space to physical space. When combining the spatial transformation matrix with the image resolution scale factor, the known correspondence between image resolution and actual physical dimensions is used to determine the conversion ratio from pixels to physical units. Then, based on the principles of triangulation, the water gauge's precise coordinates and attitude angles in three-dimensional space are calculated, combining its position, orientation angle, and spatial transformation relationship in the image. The calculated three-dimensional coordinates of the water gauge are expressed in a field coordinate system. The attitude angles include the water gauge's inclination relative to the vertical and its azimuth relative to north.

[0062] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0063] Perform image enhancement processing on the water gauge image area according to the water gauge three-dimensional coordinates and attitude angle data to obtain a high-contrast water gauge image;

[0064] Adaptively binarize the high-contrast water gauge image to obtain a scale feature binary image;

[0065] The scale feature binary image is input into the convolutional neural network for feature extraction, and the scale reference system data of the water gauge is obtained by identifying the main scale lines and numbers on the water gauge.

[0066] Based on the water gauge scale reference system data, edge detection operation is performed to obtain the water surface boundary line coordinate point set by locating the boundary line between the water surface and the water gauge;

[0067] Perform robust analysis on the water surface boundary line coordinate point set to obtain the water surface line position data;

[0068] The water surface line position data and the water gauge scale reference system data are linearly interpolated and the water gauge scale reading and water surface boundary intersection data are obtained by calibrating the water gauge zero point elevation.

[0069] Specifically, the water gauge image region is enhanced based on the three-dimensional coordinates and attitude angle data to obtain a high-contrast water gauge image. The high-contrast water gauge image is then adaptively binarized to obtain a scale feature binary image. This scale feature binary image is then input into a convolutional neural network for feature extraction. The scale reference data is then obtained by identifying the main scale lines and numbers on the water gauge. Edge detection is then performed based on the scale reference data to locate the intersection of the water surface and the water gauge, obtaining a set of water surface boundary coordinate points. Robustness analysis is then performed on this set of water surface boundary coordinate points to obtain water surface line position data. Linear interpolation is then performed between the water surface line position data and the scale reference data. The zero-point elevation of the water gauge is then calibrated to obtain the scale reading and water surface boundary intersection data. Based on the acquired three-dimensional coordinates and attitude angle data, the water gauge region in the time-series preprocessed image is accurately located and extracted. The extracted water gauge image region often suffers from insufficient contrast, especially in poor lighting conditions or when the water surface is highly reflective. To address this issue, an adaptive contrast enhancement algorithm was used to process the water gauge image region. This algorithm dynamically adjusts contrast parameters based on the brightness distribution characteristics of the local image region, focusing on enhancing the contrast between the water gauge scale and the background. During this process, the water gauge region image undergoes wavelet decomposition, enhancing different frequency components, particularly the mid- and high-frequency components corresponding to the scale lines, while suppressing the low-frequency components corresponding to background noise. The resulting composite image is a high-contrast water gauge image.

[0070] When adaptively binarizing high-contrast water gauge images, an adaptive thresholding algorithm is used instead of a global thresholding method, taking into account the uneven lighting conditions and localized reflections present in the water gauge image. This algorithm segments the water gauge image into multiple small windows, calculates the optimal local threshold for each window, and then smoothly connects the thresholds of each window to form a threshold map. This threshold map is applied to the water gauge image, effectively separating the scale lines, digital markings, and background, resulting in a clear binary image of the scale features.

[0071] The scale feature binary image is fed into a specially trained convolutional neural network for feature extraction. This network employs a multi-branch architecture, consisting of branches for detecting scale lines and recognizing numerals. The scale line detection branch employs a densely connected convolutional network, with inter-layer feature reuse improving the detection of slender scale lines. The numeral recognition branch employs a convolutional architecture with an attention mechanism, improving the recognition accuracy of small numerals. The network establishes a scale reference system by comprehensively analyzing the spacing and arrangement of scale lines, as well as the position of numerals. It identifies major scale points and their corresponding values, generating the water gauge scale reference system data. Edge detection is performed based on this water gauge scale reference system data, focusing on locating the boundary between the water surface and the water gauge. At this boundary, the water surface often exhibits significant changes in brightness and texture. Using an improved Canny edge detection algorithm, with appropriate gradient thresholds and edge connectivity parameters, the water surface edge contour is extracted. Taking into account the verticality of the water gauge, the search range is limited, focusing on detecting edges close to the horizontal direction and eliminating interfering edges. Ultimately, the boundary between the water surface and the water gauge is determined, generating a set of coordinate points for the water surface boundary. Robustness analysis is performed on the coordinate point set of the water surface boundary to eliminate false detections caused by water surface fluctuations and partial occlusions. The RANSAC (Random Sample Consensus) algorithm is used to remove outliers from the coordinate point set and fit the optimal horizontal line. In the case of multiple candidate water surface lines, the most reasonable water surface line is selected by combining time series information and water level data from adjacent frames, using temporal continuity constraints. This processing results in stable and reliable water surface line position data that accurately reflects the current water surface position.

[0072] Linear interpolation is performed between the water surface line position data and the water gauge scale reference system data to determine the precise water level reading. Based on the identified main scale point positions and corresponding values, and combined with the relative positional relationship between the water surface line and the scale, a linear interpolation algorithm is used to calculate the precise scale value corresponding to the water surface line. Simultaneously, relative readings are converted to absolute elevation values ​​based on the pre-calibrated water gauge zero point elevation data to ensure consistency and comparability of measurement results across the entire irrigation area. The resulting water gauge scale readings and water surface boundary intersection data contain complete information such as measurement time, water gauge number, relative water level reading, and absolute elevation value.

[0073] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0074] Data preprocessing is performed on the high-density discrete point cloud of the field obtained through measurement, and the terrain feature point set is obtained through sliding window outlier detection and correction;

[0075] The terrain feature point set is input into the variational adaptive thin plate spline interpolation algorithm, and the continuous field surface digital elevation model is obtained by setting the 0.5m×0.5m resolution parameter.

[0076] Perform regional analysis and processing on the continuous field surface digital elevation model to obtain terrain structure data;

[0077] The water gauge scale readings and water surface boundary intersection data are spatially aligned with the terrain structure data to obtain the water surface-terrain correspondence relationship;

[0078] Based on the water surface-topography correspondence, raster calculation is performed and initial water depth distribution data is obtained by comparing water level elevation with field surface elevation.

[0079] The initial water depth distribution data is input into the hydraulic model for flow field analysis. By considering the water flow direction and velocity parameters to correct the water surface imbalance factor, a dynamic water depth distribution grid model is obtained.

[0080] Specifically, data preprocessing was performed on the high-density discrete point cloud obtained from the field survey. A sliding window was used to detect and correct outliers to obtain a set of terrain feature points. This set of terrain feature points was then fed into a variational adaptive thin plate spline interpolation algorithm, and a continuous field surface digital elevation model (DEM) was generated by setting a resolution of 0.5 m × 0.5 m. Regional analysis and processing were performed on the DEM to obtain terrain structure data. Water gauge readings and water surface boundary intersection data were spatially aligned with the terrain structure data to obtain a water surface-terrain correspondence. Based on this water surface-terrain correspondence, raster calculations were performed, and initial water depth distribution data was obtained by comparing water level elevation with field surface elevation. This initial water depth distribution data was then fed into a hydraulic model for flow field analysis. By considering flow direction and velocity parameters to correct for water surface imbalance, a dynamic water depth distribution grid model was generated. High-density field surveys were performed using RTK-GPS technology to obtain discrete point cloud data containing three-dimensional coordinate information. These data points are evenly distributed within the field, and each point records precise longitude, latitude, and elevation values, with coordinate accuracy reaching the centimeter level. There may be abnormal points in the collected raw point cloud data, such as measurement errors, temporary obstacles, or non-ground target points. To address these problems, a sliding window outlier detection method is used for data preprocessing. This method sets a fixed-size spatial window on the point cloud data with a window size of 3 meters × 3 meters, and calculates the statistical characteristics of the elevation value within the window, including the mean, median, and standard deviation. When the elevation value of a point deviates from the median value in the window by more than three times the standard deviation, it is judged as an abnormal point and is corrected or eliminated by interpolation of neighboring points. This process is iteratively performed on the entire point cloud data, and ultimately a clean and accurate set of terrain feature points is obtained.

[0081] The terrain feature point set is processed using a variational adaptive thin plate spline interpolation algorithm to construct a continuous digital elevation model (DEM) of the field surface. Based on the principle of minimum curvature, this algorithm uses thin plate spline functions to fit discrete point cloud data, generating a smooth and continuous terrain surface. A variational adaptive mechanism is incorporated into the algorithm to increase the density of control points in areas with significant terrain variation to improve interpolation accuracy, while reducing the number of control points in areas with flat terrain to improve computational efficiency. The interpolation process uses a fixed grid resolution of 0.5 m x 0.5 m, meaning that every 25 square meters contains one elevation sample point, ensuring that the microtopographic variations of the field surface are captured. After interpolation, a continuous DEM covering the entire field is generated, with precise elevation values ​​recorded for each grid cell. Regional analysis is performed on the DEM to identify key terrain structures within the field. A slope analysis algorithm is used to calculate the slope and aspect values ​​for each grid cell, identifying areas of significant terrain variation. Combined with a gradient threshold segmentation method, linear features such as field boundaries, ridges, and drainage ditches are extracted. These features often represent areas of sudden elevation changes. Furthermore, a region growing algorithm is used to partition the elevation model, dividing the fields into sub-regions with similar elevation characteristics. Through these analyses, complete terrain structure data is generated, including information such as field boundaries, elevation partitions, and topographic structure lines, providing a topographic foundation for water depth calculations.

[0082] The water gauge scale readings and water surface boundary intersection data are spatially aligned with the terrain structure data to establish a unified spatial reference system. First, based on the GPS coordinate system or the local coordinate system, the water gauge position is georeferenced with the field digital elevation model to determine the precise position of the water gauge in the digital elevation model. Then, based on the absolute elevation reading of the water gauge and the water surface boundary intersection data, the absolute elevation value of the current water surface is determined. Considering that the water surface may not be completely level during irrigation, a water surface elevation model is constructed through spatial interpolation methods based on the readings of multiple water gauges within the field to accurately describe the water surface morphology within the entire field. Ultimately, a corresponding relationship between the water surface elevation and the terrain elevation in the same coordinate system is formed, namely the water surface-terrain correspondence.

[0083] Raster calculation is performed based on the water surface-terrain correspondence to calculate the water depth value of each grid cell in the field. For each grid cell in the digital elevation model, its terrain elevation value is obtained, and the water surface elevation value at the same position is extracted from the water surface elevation model. The water depth value of the point is calculated by the difference between the two, and water depth distribution data covering the entire field is generated. The calculation formula is: water depth = water surface elevation - terrain elevation. When the calculation result is a positive value, it means that the point is covered by water, and the value is the water depth; when the calculation result is a negative value, it means that the point is above the water surface and is not covered by water, and the water depth is recorded as zero. In this way, the initial water depth distribution data is formed, which intuitively reflects the water distribution status in the field.

[0084] The initial water depth distribution data is input into the hydraulic model for flow field analysis, taking into account the influence of hydrodynamic characteristics on water depth distribution. The shallow water equations are used to describe the movement of irrigation water flow, taking into account the influence of factors such as gravity, friction, and inertia on water velocity and water surface morphology. The model input parameters include field slope, soil roughness, inlet flow rate, and outlet conditions. The water flow field distribution is calculated using a numerical solution method to obtain the flow velocity vector and water surface slope of each grid unit. Based on the calculation results, the initial water depth distribution data is corrected, taking into account the water surface imbalance factors caused by hydrodynamics, such as the rise of water level near the water inlet and the decrease of water level in areas with higher flow rates. After correction, a dynamic water depth distribution grid model is formed.

[0085] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0086] The dynamic water depth distribution grid model is compared with the crop irrigation water demand database to obtain the water depth deviation distribution data by extracting the ideal irrigation water depth parameters of the current crops.

[0087] The uniformity index of water depth deviation distribution data is calculated, and the irrigation uniformity evaluation index is obtained by quantitatively evaluating the consistency of water depth distribution;

[0088] The irrigation uniformity evaluation index is combined with the irrigation deficiency rate and irrigation excess rate to obtain the comprehensive irrigation efficiency index through a comprehensive evaluation method.

[0089] Based on the comprehensive irrigation efficiency index, spatial cluster analysis was conducted to identify the distribution pattern of abnormal water depth areas and obtain the characteristic map of abnormal water depth areas.

[0090] The characteristic map of the water depth abnormal area is input into the Markov decision model, and the classification result of the irrigation abnormality cause is obtained by matching it with the expert knowledge base pattern.

[0091] The classification results of irrigation anomaly causes are combined with the characteristic map of water depth anomaly areas. By establishing the corresponding relationship between irrigation problems and regional distribution, the irrigation heterogeneity distribution map and irrigation system fault diagnosis matrix are obtained.

[0092] Specifically, the dynamic water depth distribution grid model was compared and analyzed with a crop irrigation water demand database. By extracting the ideal irrigation water depth parameters for the current crop, water depth deviation distribution data was obtained. A uniformity index was calculated for the water depth deviation distribution data, and the irrigation uniformity evaluation index was obtained by quantitatively evaluating the consistency of water depth distribution. The irrigation uniformity evaluation index was combined with the under-irrigation rate and over-irrigation rate to obtain a comprehensive irrigation efficiency index through a comprehensive assessment method. Spatial cluster analysis based on the comprehensive irrigation efficiency index identified the distribution patterns of abnormal water depth areas and generated a characteristic map of abnormal water depth areas. The characteristic map of abnormal water depth areas was input into a Markov decision model, and pattern matching with an expert knowledge base was used to obtain a classification of irrigation anomaly causes. The irrigation anomaly cause classification results were combined with the characteristic map of abnormal water depth areas to establish a correspondence between irrigation problems and regional distributions, resulting in an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix. The dynamic water depth distribution grid model was compared and analyzed with a crop irrigation water demand database. The crop irrigation water demand database contains ideal irrigation water depth parameters for different crops at various growth stages. These parameters are determined based on crop species, growth stage, soil type, and climatic conditions. The system extracts the ideal irrigation water depth from a database based on the crop type and growth stage currently planted in the field. This ideal value is then compared point by point with the actual water depth in the dynamic water depth distribution grid model, calculating the water depth deviation for each grid cell. A positive deviation indicates over-irrigation, while a negative deviation indicates under-irrigation. This method generates water depth deviation data covering the entire field, providing a visual representation of irrigation imbalances.

[0093] The uniformity index of the water depth deviation distribution data is calculated to quantitatively evaluate the uniformity of irrigation. The standard Crystal uniformity coefficient (CU) evaluation index is used, which reflects the consistency of the irrigation water depth distribution. During the calculation process, the water depth values ​​of all valid grid cells in the field are first obtained, the average water depth value is calculated, and then the absolute deviation of the water depth value of each grid cell from the average value is calculated. Finally, the uniformity coefficient is calculated based on these deviation values. The closer the uniformity coefficient is to 100%, the better the irrigation uniformity. In addition, statistical indicators such as the standard deviation of water depth and the coefficient of variation are calculated to comprehensively evaluate the distribution characteristics of irrigation water depth and form a set of irrigation uniformity evaluation indicators.

[0094] The irrigation uniformity evaluation index is combined with the under-irrigation rate and over-irrigation rate to perform a comprehensive efficiency assessment. The under-irrigation rate measures the percentage of the total field area where water depth is below the ideal value, reflecting under-irrigation conditions. The over-irrigation rate measures the percentage of the total field area where water depth exceeds the ideal value, reflecting over-irrigation conditions. A weighted average method is used to combine the uniformity coefficient, under-irrigation rate, and over-irrigation rate to calculate a comprehensive irrigation efficiency index. Weighting parameters are determined based on the crop's sensitivity to water stress and local water resources. Generally, under-irrigation rate is given a higher weight than over-irrigation rate. The comprehensive irrigation efficiency index is expressed as a percentage on a scale of 0 to 100, with higher values ​​indicating better irrigation efficiency. Spatial cluster analysis is performed based on the comprehensive irrigation efficiency index to identify areas of irrigation anomalies. A density clustering algorithm is used to process water depth deviation distribution data, grouping spatially adjacent raster cells with similar deviation characteristics into a single cluster, forming several irrigation anomaly areas. This algorithm does not preset the number of clusters; instead, it automatically determines the number and range of clusters based on the density distribution characteristics of the data. For each identified abnormal area, characteristic parameters such as area, average deviation value, and deviation direction (excess or deficiency) are calculated to construct a feature vector for the abnormal area. The spatial distribution pattern of the abnormal area is also analyzed, such as whether it is strip-like, block-like, or gradient-like, and whether it is related to the topographic characteristics of the field. This creates a characteristic map of the abnormal water depth area.

[0095] The characteristic maps of areas with abnormal water depths are input into a Markov decision model to diagnose the causes of irrigation anomalies. This model, built on an expert knowledge base, encompasses common irrigation anomalies and their characteristic patterns, such as those typically manifested by uneven terrain, leaking channels, and uneven water distribution. Using a Markov decision process framework, the model represents the irrigation system status as a feature vector of the anomaly area. By calculating similarity with patterns in the knowledge base, the model identifies the most likely cause of the anomaly. For complex situations, the model can perform multi-step reasoning, considering the combined influence of multiple factors to improve diagnostic accuracy. The diagnostic results include the anomaly cause type, confidence score, and problem severity rating, forming a classification of the cause of the irrigation anomaly.

[0096] The classification results of irrigation anomaly causes are combined with the characteristic map of water depth anomaly areas to establish a correspondence between problems and regions. First, within the spatial representation of the characteristic map of water depth anomaly areas, each anomaly area is labeled with its corresponding anomaly cause type to form a problem distribution map. A correlation matrix between anomaly types and regional characteristics is then constructed. The rows of the matrix represent different anomaly cause types, the columns represent different anomaly areas, and the cell values ​​indicate the confidence level that the area belongs to that anomaly type. Furthermore, for each anomaly type, typical areas are extracted as representative examples, and their characteristic parameters and spatial extent are recorded. Through these processes, an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix are formed, which intuitively reflect the spatial distribution patterns and cause analysis results of field irrigation problems.

[0097] In a specific embodiment, the process of performing step S106 can specifically include the following steps:

[0098] Data integration is performed on the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix, and irrigation control demand data is obtained by establishing a field water dynamic relationship;

[0099] The irrigation control demand data is input into a multi-objective optimization algorithm, and irrigation control candidate schemes are obtained by simultaneously considering irrigation uniformity, water resource utilization rate, and operation cost;

[0100] Based on the irrigation control candidate schemes, scheme classification processing is performed, and differentiated irrigation optimization schemes are obtained by distinguishing different irrigation problem types;

[0101] The differentiated irrigation optimization schemes are prioritized, and an irrigation execution task list is obtained by comprehensively considering the implementation difficulty and the expected effect;

[0102] The irrigation execution task list is matched with field characteristic parameters, and regional irrigation parameters are obtained by calculating the optimal irrigation time and flow;

[0103] The regional irrigation parameters are processed in time sequence, and a regional irrigation control parameter set is obtained by generating an operation instruction sequence and an execution schedule.

[0104] Specifically, data integration is performed on the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix, and irrigation control demand data is obtained by establishing a field water dynamic relationship; the irrigation control demand data is input into a multi-objective optimization algorithm, and irrigation control candidate schemes are obtained by simultaneously considering irrigation uniformity, water resource utilization rate, and operation cost; based on the irrigation control candidate schemes, scheme classification processing is performed, and differentiated irrigation optimization schemes are obtained by distinguishing different irrigation problem types; the differentiated irrigation optimization schemes are prioritized, and an irrigation execution task list is obtained by comprehensively considering the implementation difficulty and the expected effect; the irrigation execution task list is matched with field characteristic parameters, and regional irrigation parameters are obtained by calculating the optimal irrigation time and flow; the regional irrigation parameters are processed in time sequence, and a regional irrigation control parameter set is obtained by generating an operation instruction sequence and an execution schedule.

[0105] During the implementation process, data from the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix were first integrated to establish a dynamic relationship model for field water content. During the integration process, the spatial information in the irrigation heterogeneity distribution map was linked to the problem type information in the fault diagnosis matrix, forming a three-dimensional correlation data structure: problem-region-cause. For each irrigation problem type, key parameters such as impact range, severity, and duration were extracted to quantitatively assess its impact on crop growth. Furthermore, the ideal water replenishment amount and replenishment rate for each region were calculated based on the field soil type, crop water requirements, and current growth stage. The integrated results formed structured irrigation control demand data, which included spatial partitioning information, problem type, control target value, and time window constraints. This irrigation control demand data was then input into a multi-objective optimization algorithm to optimize the irrigation control plan. The non-dominated sorting genetic algorithm (NSGA-II) was used as the optimization framework, simultaneously considering three objective functions: maximizing irrigation uniformity, maximizing water resource utilization, and minimizing operating costs. The decision variables set in the algorithm included the flow allocation ratio of each water inlet, the total irrigation duration, the staged irrigation schedule, and the treatment of special areas. Constraints include total water volume limits, irrigation time windows, and equipment capacity limitations. The algorithm uses multiple generations of evolutionary computation to generate a series of non-dominated solutions, known as the Pareto optimal solution set. Each solution represents a feasible candidate irrigation control scheme. These solutions, each with its own emphasis, form an optimal balance surface between uniformity, efficiency, and cost.

[0106] Based on candidate irrigation control solutions, we categorize them and develop differentiated strategies for different types of irrigation problems. First, we classify candidate solutions into several categories based on the cause of irrigation anomalies: uneven terrain, uneven water flow distribution, ridge leakage, and time control. For problems caused by uneven terrain, we select solutions that can compensate for elevation differences, such as adjusting the inlet flow distribution and extending the drainage time in low-lying areas. For uneven water flow distribution, we focus on optimizing the inlet flow distribution ratio and control timing. For ridge leakage, we adopt a zoned irrigation strategy to control water level differences between adjacent areas. Through this categorization process, we select differentiated irrigation optimization solutions from the candidate solution set that address the specific issues facing the current field.

[0107] Differentiated irrigation optimization plans are prioritized and the order of implementation is determined. The ranking process comprehensively considers four factors: implementation difficulty, expected improvement effect, resource consumption, and time urgency. These four factors are quantitatively scored for each plan, and then a comprehensive priority score is calculated through weighted summation. The weight parameters are dynamically adjusted based on the current crop sensitivity to water stress and water resource supply conditions. The ranking results form a hierarchical irrigation execution task list. Each task item in the list contains information such as the operation object, operation type, control parameter range, and execution time window. It is arranged from high to low priority to provide guidance for specific implementation.

[0108] The irrigation execution task list is matched with the field characteristic parameters to calculate precise irrigation control parameters. Field characteristic parameters include soil infiltration characteristics, terrain slope, crop canopy cover, etc. For each task item, the characteristic parameters of the area of ​​influence are combined to perform detailed irrigation parameter calculations. For irrigation flow, the optimal irrigation intensity per unit area is determined by considering soil infiltration rate and surface runoff risk. For irrigation time, the precise duration required to reach the target water depth is calculated based on the soil water storage capacity and water depth required. These calculation results form regional irrigation parameters, with each area equipped with independent irrigation flow and duration values.

[0109] Regional irrigation parameters are time-sequenced to form an implementation plan. Considering the hydraulic constraints of the irrigation system, such as network pressure variations and total flow constraints, irrigation operations in different regions need to be rationally sequenced and combined. Using a time series optimization algorithm, irrigation tasks for each region are sorted according to priority and hydraulic constraints to generate an optimal operational sequence. At the same time, detailed control instructions are generated for each operation step, including valve opening, pump station flow, duration, and switching conditions. These instructions are organized into an operational instruction sequence in the order of execution and accompanied by a detailed execution schedule, forming a regional irrigation control parameter set that provides direct input for the automated control of the irrigation system.

[0110] It should be noted that when matching the irrigation execution task list with the field characteristic parameters, the irrigation execution task list can also be input into the computer-aided decision-making system, and the water infiltration curve under different parameter combinations can be simulated to obtain the field regional irrigation control plan; the soil moisture content of the field regional irrigation control plan is predicted and analyzed, and the regional optimized irrigation parameters are obtained by considering the soil infiltration characteristics and crop water absorption dynamics; the regional optimized irrigation parameters are matched with the hydraulic condition constraints, and the irrigation system water distribution model is obtained by solving the flow balance equation; water distribution calculation is performed based on the irrigation system water distribution model, and the flow value of each control unit is obtained by minimizing the energy loss function; the flow value of each control unit is subjected to field verification simulation, and the flow compensation coefficient is obtained by calculating the moisture uniformity index; the flow compensation coefficient is combined with the control unit flow value, and the irrigation sequence is divided by generating time periods to obtain the regional irrigation parameters.

[0111] Specifically, the irrigation execution task list is input into a computer-aided decision system for processing. The system is built based on a physical model, including three sub-modules of water movement, heat transfer and crop growth. The information in the irrigation execution task list, such as operation objects, time windows and target water depth, is parsed and converted into model input parameters by the system. The system simulates the soil-plant-atmosphere continuum under different irrigation parameter combinations using the finite element method, calculating the water infiltration curve in the vertical profile. Various boundary conditions are considered in the simulation, including surface irrigation intensity, bottom drainage conditions and lateral water movement. By comparing the water infiltration performance under different irrigation parameters, the system selects the parameter combination that meets the crop water requirement and has the highest water use efficiency, forming a field sub-area irrigation control scheme. When predicting and analyzing the soil moisture content of the field sub-area irrigation control scheme, the system combines the soil texture and structure characteristics of the field, and applies the Richards equation to simulate the water movement process in porous media. In the prediction and analysis, the soil profile is divided into multiple calculation layers, each layer setting different hydraulic parameters, including saturated hydraulic conductivity, water holding capacity and water characteristic curve. Meanwhile, considering the differences in crop root distribution characteristics and growth stages, the root water uptake in each layer of soil is calculated. According to the predicted soil moisture content distribution changes, the system evaluates the water use efficiency of different irrigation schemes, optimizes the irrigation intensity and duration parameters, and obtains the regional optimized irrigation parameters. When matching the regional optimized irrigation parameters with the hydraulic conditions, the system establishes a field irrigation water network model, including inlet channels, water outlets and field ditch systems. In the matching process, the Bernoulli equation and continuity equation are used to describe the water flow motion law in the channel, calculating the water head loss and flow velocity distribution under different flow conditions. The system considers physical parameters such as channel cross-section shape, roughness coefficient and slope, solves the nonlinear flow balance equation set, and determines the hydraulic conditions that meet the irrigation demand of each sub-area. Through repeated iterative calculation, the irrigation system water distribution model that balances the hydraulic conditions of each region is obtained. Based on the irrigation system water distribution model, the system sets a multi-objective optimization function, with minimizing energy loss as the main optimization target. The energy loss function considers factors such as lift loss, pipe friction loss and local head loss, and comprehensively evaluates the operation efficiency of the irrigation system. The optimization calculation uses the Lagrange multiplier method combined with the gradient descent algorithm to solve the optimal water distribution scheme under the constraint condition of meeting the minimum flow demand of each region. The calculation results give the accurate flow value of each control unit, which not only meets the crop irrigation demand, but also minimizes energy consumption. When verifying the flow value of each control unit in the field, the system uses numerical simulation methods to predict the water distribution uniformity under actual irrigation conditions. In the verification simulation, factors such as field micro-topography changes, spatial heterogeneity of soil permeability and water flow boundary effects are considered, and the two-dimensional diffusion wave equation is used to describe the surface water flow motion process.The system calculates water uniformity indicators such as the Crystal uniformity coefficient and coefficient of variation to evaluate the deviation between actual irrigation results and the theoretical model. Based on the deviation analysis results, it calculates the flow compensation coefficient for each control unit, which is used to adjust the theoretical flow value and improve actual irrigation uniformity. When the flow compensation coefficient is combined with the control unit flow value, the system applies the corresponding compensation coefficient to the theoretical flow value of each control unit to obtain the corrected actual controlled flow rate. At the same time, the system divides the entire irrigation process into multiple time periods, arranging different irrigation intensities and durations based on the irrigation needs of different growth stages. The system generates a detailed time-based irrigation sequence, including the opening and closing times and flow setpoints for each control unit in each time period. These parameters are combined to form the regional irrigation parameters.

[0112] The above describes the field surface water depth measurement method based on image intelligent recognition in the embodiment of the present application. The following describes the field surface water depth measurement system based on image intelligent recognition in the embodiment of the present application. Figure 2 In the embodiment of the present application, an embodiment of a field surface water depth measurement system based on image intelligent recognition includes:

[0113] The processing module 201 is used to collect images of the field with the water gauge in multiple time periods and perform time series processing to obtain time series pre-processed image data;

[0114] Matching module 202, configured to input the temporal preprocessed image data into a two-stage feature pyramid network to perform multi-scale anchor frame matching to obtain water level three-dimensional coordinates and posture angle data;

[0115] A segmentation module 203 is used to perform instance segmentation based on the water gauge three-dimensional coordinates and posture angle data to obtain water gauge scale readings and water surface boundary intersection data;

[0116] The correction module 204 is used to combine the water gauge scale readings and water surface boundary intersection data with the measured and sampled field high-density discrete point cloud, and obtain a dynamic water depth distribution grid model through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing;

[0117] An analysis module 205 is configured to perform Markov decision analysis based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix;

[0118] The input module 206 is used to input the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix into the deep reinforcement learning framework to perform multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set.

[0119] Through the cooperation of the above-mentioned components, by applying image intelligent recognition technology to the field water depth measurement field, the whole process automation from image acquisition to irrigation control is realized, effectively solving the technical problems of low efficiency and poor precision of traditional water depth measurement methods, and the combination of deep reinforcement learning framework and Markov decision analysis provides intelligent decision support for irrigation optimization. Specifically, the multi-time period image acquisition and timing processing technology ensures the continuity and reliability of the original data, providing a high-quality data foundation for subsequent analysis; the introduction of the two-stage feature pyramid network greatly improves the accuracy and robustness of the water gauge positioning, especially under complex lighting conditions and partial occlusion, through the multi-scale anchor box matching mechanism to effectively extract the water gauge feature, accurately obtain its three-dimensional coordinate and attitude angle data; the instance segmentation method based on the water gauge coordinates solves the problem of water gauge scale recognition, even under the interference of water surface fluctuation, the scale value can be accurately read; the method combining the variational adaptive thin plate spline interpolation algorithm with the hydraulic simulation correction realizes the accurate conversion from discrete points to continuous water depth distribution, overcoming the limitation that traditional point measurement cannot express spatial distribution; Markov decision analysis introduces timing data processing capability, through the comprehensive analysis of historical state and current observation, accurately diagnoses the irrigation abnormal phenomenon; the deep reinforcement learning framework optimizes the multi-step decision-making under the constraint condition, and adaptively generates the optimal irrigation control strategy.

[0120] The above Figure 2 The image intelligent recognition based field water depth measurement system in the embodiment of the application is described in detail from the perspective of modular functional entities, and the image intelligent recognition based field water depth measurement device in the embodiment of the application is described in detail from the perspective of hardware processing.

[0121] Figure 3is a structural schematic view of a field surface water depth measuring device based on image intelligent identification provided by an embodiment of the present application. The field surface water depth measuring device based on image intelligent identification 300 can have great differences due to different configurations or performances, and can include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, one or more storage media 330 (for example, one or more mass storage device ends) storing application programs 333 or data 332. The memory 320 and the storage medium 330 can be temporary storage or persistent storage. The programs stored in the storage medium 330 can include one or more modules (not shown in the figure), and each module can include a series of instruction operations in the field surface water depth measuring device based on image intelligent identification 300. Further, the processor 310 can be configured to communicate with the storage medium 330 and execute a series of instruction operations in the storage medium 330 on the field surface water depth measuring device based on image intelligent identification 300 to realize the steps of the above-mentioned field surface water depth measuring method based on image intelligent identification.

[0122] The field surface water depth measuring device based on image intelligent identification 300 can also include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, etc. Those skilled in the art can understand that, Figure 3 The field surface water depth measuring device based on image intelligent identification shown in the structure does not constitute a limitation on the field surface water depth measuring device based on image intelligent identification provided by the present application, and can include more or fewer components than shown, or combine certain components, or different component arrangements.

[0123] The present application also provides a computer readable storage medium, which can be a non-volatile computer readable storage medium, and can also be a volatile computer readable storage medium. The computer readable storage medium has instructions stored therein, and when the instructions are run on a computer, the computer executes the steps of the field surface water depth measuring method based on image intelligent identification.

[0124] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described system, system and unit can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0125] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application or the entire or part of the technical solutions that essentially contribute to the prior art can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a field surface water depth measuring device based on image intelligent identification (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0126] The above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalents; and these modifications or replacements do not make the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for measuring field surface water depth based on image intelligent recognition, characterized in that: The method comprises: Multi-period image acquisition and time series processing are performed on the fields with water gauges to obtain time series pre-processed image data; Inputting the temporal preprocessed image data into a two-stage feature pyramid network to perform multi-scale anchor frame matching to obtain the three-dimensional coordinates and posture angle data of the water ruler; Perform instance segmentation based on the water gauge three-dimensional coordinates and posture angle data to obtain water gauge scale readings and water surface boundary intersection data; The water gauge scale readings and water surface boundary intersection data are combined with the high-density discrete point cloud of the field sampled by actual measurement, and a dynamic water depth distribution grid model is obtained through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing; Performing Markov decision analysis based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix; The irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix are input into a deep reinforcement learning framework for multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set.

2. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The method of collecting images of the field with the water gauge in multiple time periods and performing time series processing to obtain time series preprocessed image data includes: Adjust the lighting parameters of the camera fixed on the edge of the field, and control the shooting frequency to once per hour to obtain a sequence of original images including the water gauge and water surface. Performing noise filtering on the original image sequence to obtain a field water surface image by eliminating water surface ripples and weed interference; Inputting the field water surface image into a histogram equalization processing unit to obtain standardized image data by balancing the image brightness under different lighting conditions; Performing perspective correction based on the standardized image data and obtaining an orthophoto image by performing coordinate transformation using reference points at the four corners of the field; Performing detail enhancement processing on the orthophoto image to obtain an enhanced image by sharpening edge features of the water ruler scale; The enhanced image is combined with the acquisition time, field number and geographical location information to obtain time series preprocessed image data.

3. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The step of inputting the temporal preprocessed image data into a two-stage feature pyramid network to perform multi-scale anchor frame matching to obtain water level three-dimensional coordinates and posture angle data includes: Inputting the temporal preprocessed image data into a first-stage network including a backbone network and a feature pyramid structure, and obtaining a coordinate set of a water gauge candidate area by extracting feature maps of different scales; Performing a non-maximum suppression operation on the coordinate set of the water gauge candidate area, and screening by setting a confidence threshold of 0.85 to obtain a high-confidence water gauge area bounding box; The image within the high-confidence water gauge area boundary box is cropped and input into the second-stage network, and fine-grained feature extraction is performed to obtain the water gauge fine contour data; Performing Hough transform calculation based on the water ruler fine contour data, and obtaining the direction angle value of the water ruler in the image plane by detecting the vertical edge line of the water ruler; Mapping the direction angle values ​​of the water ruler on the image plane to pre-calibrated reference scale points, and obtaining a spatial transformation matrix of the water ruler by establishing a corresponding relationship between the image space and the physical space; The spatial transformation matrix is ​​combined with the image resolution scale factor to calculate and obtain the three-dimensional coordinates and attitude angle data of the water ruler through the principle of triangulation.

4. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The example segmentation is performed based on the three-dimensional coordinates and attitude angle data of the water gauge to obtain the water gauge scale reading and water surface boundary intersection data, including: Performing image enhancement processing on the water gauge image area according to the water gauge three-dimensional coordinates and posture angle data to obtain a high-contrast water gauge image; Adaptively binarizing the high-contrast water gauge image to obtain a scale feature binary image; The scale feature binary image is input into a convolutional neural network for feature extraction, and the scale reference system data of the water gauge is obtained by identifying the main scale lines and numbers on the water gauge; Performing edge detection calculation based on the water gauge scale reference system data, and obtaining a water surface boundary line coordinate point set by locating the boundary line between the water surface and the water gauge; Performing robustness analysis on the water surface boundary line coordinate point set to obtain water surface line position data; The water surface line position data and the water gauge scale reference system data are linearly interpolated and calculated, and the water gauge scale reading and the water surface boundary intersection data are obtained through water gauge zero point elevation calibration.

5. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The water gauge scale readings and water surface boundary intersection data are combined with the field high-density discrete point cloud of the field sampled by actual measurement, and a dynamic water depth distribution grid model is obtained through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing, including: Data preprocessing is performed on the high-density discrete point cloud of the field obtained through measurement, and the terrain feature point set is obtained through sliding window outlier detection and correction; Inputting the terrain feature point set into a variational adaptive thin plate spline interpolation algorithm, and obtaining a continuous field surface digital elevation model by setting a 0.5 meter × 0.5 meter resolution parameter; Performing regional analysis on the continuous field surface digital elevation model to obtain terrain structure data; Performing spatial registration on the water gauge scale readings and the water surface boundary intersection data with the terrain structure data to obtain a water surface-terrain correspondence; Performing grid calculation based on the water surface-topography correspondence, and obtaining initial water depth distribution data by comparing water level elevation with field surface elevation; The initial water depth distribution data is input into a hydraulic model for flow field analysis, and the water surface imbalance factor is corrected by considering water flow direction and flow velocity parameters to obtain a dynamic water depth distribution grid model.

6. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The Markov decision analysis based on the dynamic water depth distribution grid model is performed to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix, including: Comparing and analyzing the dynamic water depth distribution grid model with the crop irrigation water demand database, and extracting the ideal irrigation water depth parameters of the current crops to obtain water depth deviation distribution data; Calculating the uniformity index of the water depth deviation distribution data, and obtaining an irrigation uniformity evaluation index by quantitatively evaluating the consistency of the water depth distribution; The irrigation uniformity evaluation index is combined with the irrigation deficiency rate and the irrigation excess rate to obtain a comprehensive irrigation efficiency index through a comprehensive evaluation method; Performing spatial cluster analysis based on the comprehensive irrigation efficiency index to identify the distribution pattern of abnormal water depth areas and obtain a characteristic map of abnormal water depth areas; Inputting the characteristic map of the abnormal water depth area into the Markov decision model, and obtaining the classification result of the cause of the irrigation abnormality by matching it with the expert knowledge base pattern; The classification results of the irrigation anomaly causes are combined with the characteristic map of the water depth anomaly area, and the corresponding relationship between irrigation problems and regional distribution is established to obtain the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix.

7. The method for measuring field surface water depth based on image intelligent recognition according to claim 1, characterized in that: The irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix are input into the deep reinforcement learning framework for multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set, including: Integrating the data of the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix, and obtaining irrigation control demand data by establishing a dynamic relationship of field water content; Inputting the irrigation control demand data into a multi-objective optimization algorithm to obtain a candidate irrigation control plan by simultaneously considering irrigation uniformity, water resource utilization and operating cost; Classifying the candidate irrigation control schemes, and obtaining differentiated irrigation optimization schemes by distinguishing different types of irrigation problems; Prioritizing the differentiated irrigation optimization plans, and obtaining an irrigation execution task list by comprehensively considering implementation difficulty and expected effects; Matching the irrigation execution task list with the field characteristic parameters, and obtaining regional irrigation parameters by calculating the optimal irrigation time and flow rate; The sub-region irrigation parameters are time-sequenced and processed, and an operation instruction sequence and an execution schedule are generated to obtain a sub-region irrigation control parameter set.

8. A field surface water depth measurement system based on image intelligent recognition, characterized in that: A method for measuring field surface water depth based on image intelligent recognition according to any one of claims 1 to 7, wherein the system for measuring field surface water depth based on image intelligent recognition comprises: A processing module is used to collect images of the field with the water gauge in multiple time periods and perform time series processing to obtain time series pre-processed image data; A matching module is used to input the temporal preprocessed image data into a two-stage feature pyramid network to perform multi-scale anchor frame matching to obtain the three-dimensional coordinates and posture angle data of the water ruler; A segmentation module is used to perform instance segmentation based on the three-dimensional coordinates and posture angle data of the water gauge to obtain water gauge scale readings and water surface boundary intersection data; A correction module is used to combine the water gauge scale readings and water surface boundary intersection data with the high-density discrete point cloud of the field sampled by actual measurement, and obtain a dynamic water depth distribution grid model through variational adaptive thin plate spline interpolation and hydraulic simulation correction processing; an analysis module for performing Markov decision analysis based on the dynamic water depth distribution grid model to obtain an irrigation heterogeneity distribution map and an irrigation system fault diagnosis matrix; The input module is used to input the irrigation heterogeneity distribution map and the irrigation system fault diagnosis matrix into the deep reinforcement learning framework to perform multi-step decision optimization processing under constraints to obtain a regional irrigation control parameter set.

9. A field surface water depth measurement device based on image intelligent recognition, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the method for measuring field surface water depth based on image intelligent recognition according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is enabled to execute the field surface water depth measurement method based on image intelligent recognition according to any one of claims 1 to 7.

Images