A farmland water anomaly diagnosis method based on information carrier response change detection

Abstract

The application discloses a farmland water anomaly diagnosis method based on information carrier response change detection, and comprises the following steps: acquiring multi-source remote sensing time series data in the crop growth period, dividing remote sensing features into multiple information carrier groups according to physical sensing mechanisms; calculating the class separability measure of each carrier group on the water state at each time step, constructing a time series curve, and identifying the carrier dominant response change node through a change point detection; according to the response change node, combining the crop water requirement characteristics and the phenology process, adaptively dividing multiple carrier dominant intervals, determining the dominant carrier group and its core features in each interval, and screening supplementary features from the non-dominant group to establish an interval classification model; routing the pixel to be diagnosed to the corresponding model to obtain the water state classification result; calling the reference state of the corresponding crop and growth stage in the reference library, calculating the deviation index, and outputting the anomaly diagnosis result. The application improves the stage adaptability and classification stability of multi-source remote sensing monitoring, and is suitable for water anomaly diagnosis of various crops such as rice, corn and wheat.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing monitoring and farmland moisture status diagnosis technology, specifically to a method for diagnosing abnormal farmland moisture status based on the detection of response changes in multi-source information carriers. Background Technology

[0002] Farmland moisture status is a crucial factor influencing the growth and yield of major food crops such as maize, rice, and wheat. In recent years, multi-source remote sensing data, including optical remote sensing, multispectral remote sensing, thermal infrared remote sensing, and synthetic aperture radar (SAR), have been widely used for monitoring farmland waterlogging, waterlogging, soil moisture, and drought stress. Existing methods typically use remote sensing indicators such as water index, vegetation index, land surface temperature, and radar backscattering characteristics, combined with machine learning or threshold discrimination methods, to identify farmland moisture status, thereby serving agricultural irrigation management, disaster monitoring, and crop growth assessment.

[0003] Existing farmland moisture monitoring methods primarily focus on determining whether farmland is waterlogged, flooded, or drought-stricken at a specific moment, with less consideration given to the water requirements of different crops at different growth stages and changes in land cover. The same moisture signal may have different meanings for different crops and at different growth stages, and relying solely on fixed features or models can easily lead to misjudgments. Furthermore, while some methods employ staged modeling, the stage division often relies on human experience, fixed dates, or pre-set phenological nodes, making it difficult to adapt to differences in regions, years, and crop growth processes, resulting in insufficient model generalization ability and diagnostic stability. Summary of the Invention

[0004] To address the aforementioned issues, this patent proposes a method for diagnosing farmland moisture anomalies based on multi-source information carrier response change detection. This method utilizes multi-source remote sensing time-series data to calculate the separability of farmland moisture state changes across different remote sensing information carriers, automatically identifies nodes exhibiting changes in the dominant information carrier's response, and adaptively segments the data by incorporating different crop growth stages, water requirement patterns, and land cover growth patterns.

[0005] The technical solution adopted in this invention is as follows:

[0006] A method for diagnosing farmland moisture anomalies based on information carrier response change detection, the method comprising the following steps:

[0007] S1. Collect and divide multi-source remote sensing time-series data and information carriers to obtain multi-source remote sensing time-series data of the target farmland area during the crop growth period, and divide the remote sensing features extracted from the multi-source remote sensing time-series data into multiple information carrier groups according to the physical sensing mechanism corresponding to the remote sensing features.

[0008] S2. Detect changes in the response of information carriers, calculate the class separability measure of each information carrier group to the farmland water state at different time steps, construct a separability time series curve, and identify the dominant response change nodes of information carriers based on the separability time series curve.

[0009] S3. Divide the carrier-dominant intervals and establish interval models. Based on the information carrier-dominant response change nodes and combined with the water requirement characteristics and surface cover growth characteristics of the target crop at different growth stages, divide the crop growth period into multiple carrier-dominant intervals. In each carrier-dominant interval, determine the core features and supplementary features, and establish the corresponding interval classification model.

[0010] S4. Perform routing classification of the objects to be diagnosed. Route the farmland pixels or regions to be diagnosed to the corresponding interval classification models according to their time step, growth stage or carrier-dominated interval to obtain the farmland water status classification results.

[0011] S5. Perform water state deviation calculation and anomaly diagnosis output, call the baseline water state under the corresponding crop type and growth stage, compare the farmland water state classification result with the baseline water state, calculate the water state deviation index, and output the farmland water state anomaly diagnosis result based on the water state deviation index.

[0012] Further, in step S1, the remote sensing data collection and information carrier division includes the following steps:

[0013] S11. Determine the target farmland area, target crop type, and monitoring time range, wherein the target crop type includes at least one of corn, rice, and wheat;

[0014] S12. Acquire at least two of the following data for the target farmland area during the crop growth period: optical remote sensing data, multispectral remote sensing data, thermal infrared remote sensing data, synthetic aperture radar (SAR) data, and soil moisture-related data.

[0015] S13. Preprocess the multi-source remote sensing time-series data and extract one or more of the following: water index, vegetation index, multispectral band features, surface temperature features, SAR backscattering features, soil moisture correlation features, and crop growth features.

[0016] S14. Based on the physical sensing mechanism corresponding to the remote sensing features, the extracted remote sensing features are divided into at least two information carrier groups from the following groups: optical water index feature group, multispectral band feature group, vegetation growth feature group, surface temperature feature group, SAR feature group, and soil moisture related feature group.

[0017] The farmland moisture status includes one or more of the following: normal moisture, excess moisture, flooding, waterlogging, poor drainage, insufficient water supply, and drought stress.

[0018] Further, in step S2, the detection and segmentation of response change nodes in the information carrier response change includes the following steps:

[0019] S21. For each time step, calculate the class separability measure of each information carrier group for farmland moisture status, and construct the separability time series curve corresponding to each information carrier group.

[0020] S22. The change point detection adopts a change point search algorithm based on Bayesian information criterion or the PELT (PrunedExact Linear Time) algorithm, which determines the optimal change point position in the separable time series curve where the statistical characteristics change significantly by minimizing the penalty objective function, wherein the penalty term coefficient of the penalty objective function is adaptively determined according to the data length and variance of the separable time series curve.

[0021] S23. The dominant response change node of the information carrier is taken as the time node when the dominant remote sensing information source changes during the crop growth period, and the response change node is output in Julian day form.

[0022] The class separability measure is used to characterize the ability of a corresponding information carrier group to distinguish different farmland moisture states at the current time step. The class separability measure includes Fisher's discriminant ratio, the ratio of inter-class distance to intra-class distance, mutual information, classification boundary margin, area under the receiver operating characteristic curve, or feature importance score.

[0023] For the feature data of the g-th information carrier group at time step t, its separability value Sg(t) for different farmland water state categories is calculated, and a separability time series curve Sg={Sg(t1), Sg(t2), …,Sg(tn)} is formed in chronological order. When the separability curves of different information carrier groups show a significant increase, a significant decrease, crossover, or abrupt change in statistical characteristics, it is determined that the dominant remote sensing information source has undergone a response change during the crop growth period.

[0024] Further, in step S3, establishing the interval classification model includes the following steps:

[0025] S31. Based on the information carrier dominant response change nodes, the water requirement characteristics of different growth stages of the target crop, the surface cover growth characteristics, and the dynamics of phenological processes, the crop growth period is divided into multiple carrier dominant intervals.

[0026] S32. Within each carrier dominance interval, compare the category separability measure of each information carrier group, and determine the information carrier group with the highest separability or that meets the preset dominance condition as the dominant information carrier group.

[0027] S33. The features of the dominant information carrier group are determined as core features, and the Recursive Feature Elimination (RFE) or SHAP (SHapley Additive exPlanations) importance analysis method is used to screen features from the non-dominant information carrier group whose contribution meets the preset threshold as supplementary features.

[0028] S34. Using the core features and supplementary features, combined with the ground real labels, train an interval classification model for the dominant interval of the corresponding carrier. The interval classification model is any one of the random forest model, support vector machine model, or gradient boosting decision tree model.

[0029] The dominant region of the carrier is not divided according to fixed calendar dates, but rather dynamically based on the change nodes of the dominant response of the remote sensing information carrier, combined with the water requirement patterns of the target crop at different growth stages, changes in land cover growth, and phenological processes. This avoids fixed zoning errors caused by differences in different years, regions, planting systems, and crop growth processes.

[0030] Further, in step S4, based on the temporal location, crop type, growth stage, and information carrier dominant response change node of the farmland pixel or region to be diagnosed, the carrier dominant interval to which it belongs is determined, and the interval classification model corresponding to the carrier dominant interval is invoked. The interval classification model outputs the water status classification result of the farmland pixel or region to be diagnosed, and the water status classification result includes one or more of the following: normal water, excess water, flooding, waterlogging, poor drainage, insufficient water supply, or drought stress.

[0031] Furthermore, in step S5, the diagnosis and output of abnormal moisture status includes the following steps:

[0032] S51. Based on the crop type, growth stage and regional attributes of the farmland pixel or region to be diagnosed, call the corresponding baseline moisture status in the crop moisture status baseline library.

[0033] S52. Compare the currently obtained farmland moisture status classification results with the baseline moisture status, and calculate the moisture status deviation index.

[0034] S53. Determine whether the farmland water status is abnormal based on the relationship between the water status deviation index and the anomaly discrimination threshold.

[0035] S54. Output one or more of the following: farmland water status classification results, anomaly type, anomaly level, information carrier dominant response change node, carrier dominant interval, and water status deviation index.

[0036] The moisture state deviation index is the standardized deviation index (FDI), which is calculated as follows:

[0037] FDI = (F obs - F base ) / σ base ;

[0038] Where Fobs represents the observed water status index of the farmland pixel, region, or statistical unit to be diagnosed at the current time step or current growth stage, F base σ represents the baseline water state index for the corresponding crop type and growth stage in the crop water state baseline database. base It represents the fluctuation scale of the baseline water state index for the corresponding crop type and growth stage;

[0039] When FDI is within the preset normal range, the moisture content is considered normal.

[0040] When FDI is greater than a preset positive threshold, it is determined that there is an abnormality of excessive moisture, flooding, waterlogging or poor drainage.

[0041] When FDI is less than the preset negative threshold, it is determined that there is insufficient water supply or drought stress.

[0042] The crop moisture status benchmark database is used to store one or more of the following: dominant information carrier groups, benchmark moisture status, water requirement thresholds, land cover growth indicators, and anomaly detection thresholds for different crop types, growth stages, planting systems, and regional types. The crop moisture status benchmark database is constructed using one or more of the following: historical remote sensing time-series data, measured moisture status data, agricultural management data, soil moisture monitoring data, disaster survey data, or manually imported data, and is updated based on newly added monitoring data.

[0043] Through the above technical solution, this invention, based on multi-source remote sensing time-series data, divides optical, multispectral, thermal infrared, SAR, and soil moisture-related features into multiple information carrier groups according to physical sensing mechanisms. By calculating the class separability measure of farmland water status for each information carrier group at different time steps, it automatically identifies the nodes of dominant response changes of information carriers during crop growth periods where the dominant remote sensing information source changes. Furthermore, by combining the water requirement characteristics, land cover growth characteristics, and phenological processes of different crops at different growth stages, it adaptively divides carrier dominance intervals and constructs corresponding classification models within each interval. Finally, by calling the corresponding benchmark water status through the crop water status benchmark library, it calculates water status deviation indicators and outputs anomaly diagnosis results.

[0044] The beneficial results achieved by the present invention using the above solution are as follows:

[0045] (1) This invention can improve the stage adaptability of multi-source remote sensing farmland moisture status identification. By calculating the class separability measure of different information carrier groups at different time steps and detecting the dominant response change nodes of information carriers based on the separability time series curve, this invention can automatically determine the most suitable remote sensing information source for characterizing farmland moisture status at different stages, thereby avoiding the problem of fixed feature combinations failing or declining recognition ability at some growth stages. At the same time, this invention does not depend on a single crop type and can be applied to a variety of major food crops such as corn, rice, and wheat. It can also adapt to the changes in growth stages of different crops under crop rotation conditions, improving the applicability of the method under complex planting structures.

[0046] (2) This invention can improve the accuracy and stability of farmland water status classification models. Based on the response change nodes of remote sensing information carriers, and combined with the water requirement characteristics of target crops at different growth stages, land cover growth characteristics, and phenological processes, this invention adaptively divides carrier-dominant intervals. Within each carrier-dominant interval, the features of the dominant information carrier group are used as core features, and supplementary features are selected from the non-dominant information carrier groups to establish an interval classification model, enabling the model to dynamically adapt to different crops, different stages, and different regional conditions.

[0047] (3) This invention can improve the agricultural interpretability and model reusability of farmland water anomaly diagnosis. This invention stores the dominant information carrier groups, baseline water status, water requirement thresholds, land cover growth indicators, and anomaly discrimination thresholds for different crop types, growth stages, planting systems, and regional types through a crop water status benchmark library. During diagnosis, the model compares the current water status classification result with the corresponding baseline water status, calculates the water status deviation index, and thus determines whether the water status is abnormal. This method can not only identify the current water status but also determine whether it meets the current crop growth requirements, and supports cross-year benchmark calls and dynamic updates. Simultaneously, the benchmark library can be expanded and updated for different crops and their rotation combinations, enabling the model to have good versatility and long-term application capabilities under multi-crop, multi-season, and rotation planting conditions. Attached Figure Description

[0048] Figure 1 This is a flowchart illustrating the method for diagnosing abnormal farmland moisture conditions provided by the present invention.

[0049] Figure 2 A schematic diagram of the farmland moisture state anomaly diagnosis model based on information carrier response change detection provided by the present invention;

[0050] Figure 3 This is a flowchart illustrating step S2;

[0051] Figure 4 This is a flowchart illustrating step S3;

[0052] Figure 5 This is a summary of the multi-source response contribution, FDI changes, and abnormal moisture state diagnosis results at typical locations in a specific embodiment of the present invention. Detailed Implementation

[0053] This embodiment provides an adaptive method for diagnosing abnormal water status in paddy fields based on the detection of responses to changes in multi-source information carriers. For example... Figure 1As shown, this method generally includes steps such as crop type identification, multi-source data acquisition and information carrier segmentation, information carrier response change detection, response dominance interval segmentation, interval classification model construction, routing of objects to be diagnosed, and calculation of moisture state deviation and anomaly diagnosis output. This embodiment was implemented in a field demonstration in the middle reaches of the Lianshui River in Xiangtan City, Hunan Province. The implementation targets were contiguous paddy fields, and the monitoring period covered the rice transplanting and greening stage, tillering stage, jointing and booting stage, heading and flowering stage, and grain filling and ripening stage. During implementation, the crop type in the target area is first identified using UAV images and fixed field camera images. Then, remote sensing data, ground monitoring data, meteorological data, and agricultural management data are collected to construct a multi-source information carrier group. Next, the separability measure and response contribution of different information carrier groups to the paddy field water status at different time steps are calculated to identify the response change nodes of the multi-source information carriers. Subsequently, the rice growth period is adaptively segmented by combining phenological processes, effective accumulated temperature, precipitation and irrigation / drainage events, and changes in land cover growth, and the interval classification model corresponding to each dominant response interval is trained. Finally, the object to be diagnosed is routed to the corresponding interval classification model to obtain the water status classification result, which is compared with the benchmark water status in the crop water status benchmark library to output the anomaly diagnosis result.

[0054] This embodiment selected 18 representative paddy field plots for the field demonstration area, covering low-lying fields near rivers, fields adjacent to ditches, conventionally irrigated and drained fields, and fields with poor drainage. Each plot's boundary was used as the basic statistical unit, and a rice growth period time series sample was constructed using 5 days as a fixed time step. Monitoring times were uniformly converted to Julian days. This embodiment generated a total of paddy field-time step samples, some of which were manually labeled as normal moisture state, excessive moisture state, poor drainage state, or insufficient water supply state. These manual survey labels, along with field water depth, soil volumetric water content, irrigation and drainage records, and crop growth records, served as the data basis for model training and benchmark database construction.

[0055] 1. Identify crop types, collect multi-source data, and classify information carriers.

[0056] In this embodiment, crop images are jointly acquired by a drone platform and a fixed field camera. The drone platform acquires visible light and multispectral images of the target farmland area during the rice transplanting and greening stage, tillering stage, jointing and booting stage, heading and flowering stage, and grain-filling and ripening stage, respectively, with the flight altitude set to obtain canopy texture and water surface exposure information at the plot scale. The fixed field camera acquires rice canopy images on a daily scale to record canopy coverage, leaf color changes, and growth changes. After acquisition, the crop images are sequentially processed by image cropping, distortion correction, geometric correction, brightness normalization, scale normalization, and sample slicing, and the slicing results are spatially correlated with the plot boundaries.

[0057] This embodiment uses the DeepLabV3+ semantic segmentation model to identify crop types. Before implementation, samples of rice, field ridges, water surfaces, bare soil, weeds, and ditches from UAV images and fixed field camera images are manually labeled to form crop identification training samples. After inputting the preprocessed crop images into the DeepLabV3+ semantic segmentation model, the model outputs the crop type label corresponding to each spatial location. When there are multiple plots within the target farmland area, the target farmland area is divided into several spatial units, and crop type identification is performed on each spatial unit to generate a crop type spatial distribution map. In this embodiment, the spatial unit identified as rice and consistent with the boundary of the measured plot is used as the object to be diagnosed, and the corresponding growth period rules, crop baseline temperature, baseline moisture status, and anomaly discrimination threshold for rice are called.

[0058] After crop type identification, multi-source time-series data of the target farmland area during the rice growth period are acquired. The multi-source time-series data actually used in this embodiment includes remote sensing data, ground monitoring data, meteorological data, and agricultural management data.

[0059] Remote sensing data includes multispectral remote sensing imagery, thermal infrared land surface temperature products, SAR imagery, and microwave soil moisture products. Multispectral remote sensing imagery is used to extract water indices, vegetation indices, and multispectral reflectance; thermal infrared land surface temperature products are used to extract land surface temperature and temperature anomaly characteristics; SAR imagery is used to extract VV polarization backscattering coefficient, VH polarization backscattering coefficient, VH / VV ratio, VV-VH difference, and radar vegetation index; microwave soil moisture products are used to characterize regional-scale soil moisture changes.

[0060] Ground monitoring data is obtained from soil moisture sensors, water level recorders, and manual surveys deployed in the field. In this embodiment, soil moisture sensors with a depth of 10cm are deployed in typical plots to collect the volumetric water content of soil at different soil layers; water level gauges and water level recorders are deployed in low-lying plots, plots adjacent to ditches, and conventionally irrigated and drained plots to collect data on field water depth, drainage ditch water level, and groundwater depth; manual surveys are used to verify the water status of paddy fields and record phenomena such as surface water, water receding rate, leaf curling, leaf color lightening, water accumulation in low-lying areas, and waterlogging in drainage ditches.

[0061] Meteorological data were obtained from agricultural meteorological stations near the target farmland area, including daily maximum and minimum temperatures, precipitation, relative humidity, wind speed, solar radiation, and reference crop evapotranspiration. Agricultural management data were obtained from field trial records, including transplanting date, irrigation date, irrigation volume, drainage date, drainage volume, fertilization date, and field management records.

[0062] This embodiment performs radiometric calibration, atmospheric correction, cloud and shadow removal, geometric correction, and spatial registration on remote sensing images in sequence; orbit correction, radiometric calibration, speckle filtering, topographic correction, and decibel conversion on SAR images; outlier removal, sensor drift checking, and time consistency processing on ground monitoring data; spatial matching of meteorological data according to the location of target farmland areas; and conversion of agricultural management records into time-series event features according to date. All data are unified to the same coordinate system, the same time step, and the same plot boundary. For continuous missing time steps, this embodiment uses Savitzky-Golay filtering and linear interpolation to complete time-series reconstruction; for event-type variables such as irrigation, drainage, and heavy precipitation, their event attributes are preserved without smoothing; for manually verified labels, labels are not generated through interpolation. For data with inconsistent spatial scales, this embodiment performs mean, quantile, or area-weighted aggregation according to plot boundaries to obtain multi-source feature sequences at the plot scale.

[0063] After preprocessing, features related to farmland moisture status are extracted from multi-source time-series data. For example... Figure 2 As shown, this embodiment uses remote sensing data, ground monitoring data, meteorological data, agricultural management data, and crop image data as multi-source inputs. A complete diagnostic process is formed through information carrier response capability assessment, carrier dominant region division, model and knowledge base construction, and an anomaly diagnosis module. For optical and multispectral remote sensing data, this embodiment extracts water body indices, vegetation indices, multispectral band reflectance, and crop growth characteristics. Water body indices include the Normalized Difference Water Index (NDWI), the Modified Normalized Difference Water Index (MNDWI), and the Land Surface Water Index (LSWI), which are calculated as follows:

[0064]

[0065]

[0066]

[0067] Wherein, G represents the reflectivity of the green light band, NIR represents the reflectivity of the near-infrared band, and SWIR represents the reflectivity of the short-wave infrared band.

[0068] Vegetation indices include the Normalized Difference Vegetation Index (NDVI) and the Enhanced Vegetation Index (EVI), which are calculated as follows:

[0069]

[0070]

[0071] Where R represents the reflectivity of the red light band and B represents the reflectivity of the blue light band.

[0072] For thermal infrared remote sensing data, this embodiment extracts land surface temperature, canopy temperature, temperature anomaly, temperature-vegetation drought index, and crop water stress index. For SAR data, this embodiment extracts VV polarization backscattering coefficient, VH polarization backscattering coefficient, and... ratio, Difference, polarization characteristics, coherence characteristics, and radar vegetation index (RVI). .

[0073] For meteorological data, this embodiment extracts cumulative precipitation, number of consecutive rainless days, reference crop evapotranspiration, precipitation-evapotranspiration differential, water deficit, and effective accumulated temperature. The crop reference temperature T in this embodiment... base The data were determined by a combination of completed field phenological records and daily temperature data from agricultural meteorological stations. Specifically, during the field experiment, the actual dates of the rice transplanting and greening-up stage, tillering stage, jointing and booting stage, heading and flowering stage, and grain-filling and ripening stage were recorded, and the daily maximum and minimum temperatures from agricultural meteorological stations were obtained simultaneously.

[0074] Let the candidate baseline temperature be T0. Calculate the effective accumulated temperature from the transplanting date to each measured phenological node based on the candidate baseline temperature:

[0075]

[0076] in, t represents the effective accumulated temperature corresponding to the s-th phenological node calculated at the candidate reference temperature T0. s T represents the date corresponding to the s-th phenological node. max,d T represents the highest temperature on day d. min,d This indicates the lowest temperature on day d.

[0077] This embodiment uses the stability of effective accumulated temperature between measured phenological nodes as a constraint to determine the baseline temperature for rice crops:

[0078]

[0079] Where S represents the number of measured phenological nodes, This represents the average effective accumulated temperature across multiple measured samples for the same phenological stage. In this embodiment, after performing ergonomic calculations within the candidate reference temperature range, the reference temperature T for rice crops is determined. base The value is set to 10.0℃ and stored in the crop moisture state baseline database. This Tbase is used to calculate the effective accumulated temperature during subsequent adaptive segmentation and time migration correction. Therefore, in this embodiment, the effective accumulated temperature GDD is calculated as follows:

[0080]

[0081] For ground monitoring data, this embodiment extracts soil moisture content, field water depth, groundwater depth, drainage ditch water level changes, and canopy temperature anomalies. For agricultural management data, this embodiment extracts the time since the most recent irrigation, the time since the most recent drainage, cumulative irrigation volume, cumulative drainage volume, and the intensity of irrigation and drainage events. For crop images, this embodiment extracts canopy coverage, leaf color index, texture features, plant height estimates, canopy structure features, and image phenotypic growth indices.

[0082] Based on the data sources, physical response mechanisms, and the role of water status characterization, this embodiment divides the above information into two categories: one category is the independent variable information carrier group used for model input, including the optical water body response feature group, the spectral vegetation response feature group, the multispectral band feature group, the thermal infrared water stress feature group, the radar water structure response feature group, the microwave soil moisture feature group, and the meteorological water-driven feature group; the other category is the observation variables used for water status determination and model training annotation, including ground soil moisture monitoring data, field water level monitoring data, irrigation and drainage management records, and crop image phenotype and growth observation information, which serve as the basis for water status label construction and benchmark library establishment.

[0083] In this embodiment, the primary diagnostic categories for paddy field moisture status include normal moisture status, excess moisture status, poor drainage status, and insufficient water supply status. For excess moisture status, it is further labeled as shallow water accumulation, flooding, or waterlogging; for poor drainage status, it is further labeled as high water levels in drainage ditches, shallow groundwater depth, or slow water receding; for insufficient water supply status, it is further labeled as mild water shortage or drought stress. Through this method, the classification model output categories remain consistent, and the anomaly diagnosis results can express more granular anomaly types.

[0084] 2. Detect changes in the information carrier response and identify change nodes.

[0085] This embodiment uses Fisher's discriminant ratio to measure the separability of each information carrier group to the paddy field moisture state at different time steps. Let the sample feature set corresponding to the g-th information carrier group at the t-th time step be... The number of moisture state categories is C, and the number of samples in category c is n. c The mean of the c-th class of samples is The overall sample mean is Then the intraclass discreteness matrix and inter-class discreteness matrix They are respectively:

[0086]

[0087]

[0088] In this embodiment, the category separability measure is calculated according to the following formula:

[0089]

[0090] in, Represents the trace of a matrix. To prevent positive numbers with a denominator of zero.

[0091] This embodiment normalizes the category separability measure:

[0092]

[0093] Subsequently, this embodiment calculates the response contribution based on the normalized class separability measure, model feature importance, classification boundary margin, prediction confidence, and anomaly response intensity:

[0094]

[0095] in, Indicates the importance of the normalized model features. This represents the normalized classification boundary margin. This represents the normalized prediction confidence level. This represents the normalized anomalous response strength. , , , and These are the weighting coefficients. In this embodiment, the above weights are determined through cross-validation of the training set, with w1=0.35, w2=0.25, w3=0.15, w4=0.15, and w5=0.10, so that class separability and feature importance account for the main weights in the response contribution.

[0096] This embodiment uses the PELT change point detection algorithm to identify nodes where multi-source information carriers respond and change. For example... Figure 3 As shown, this embodiment first calculates the separability measure of each information carrier group, and then, by combining the response contribution time series, identifies the nodes of response change in the information carriers using the PELT change point detection algorithm. Let the response contribution vector at time step t be:

[0097]

[0098] Let the set of changing points be The optimal set of change points B* is determined according to the following objective function:

[0099]

[0100] in, Let represent the mean of the response contribution vector within the k-th segment, β be the variable point penalty coefficient, K be the number of variable points, and n be the time series length. In this embodiment, the penalty term coefficient is determined as follows:

[0101]

[0102] Where α is the adjustment coefficient. The variance of the response contribution time series is represented by α, and n represents the length of the time series. In this embodiment, α = 1.5 is used, and the minimum segment length is set to 3 time steps to avoid spurious changes caused by short-term cloud pollution, single rainfall events, or local sensor anomalies.

[0103] When the dominant information carrier group changes before and after the change point τ, or when the ranking of response contributions changes, this embodiment determines the change point as the multi-source information carrier response change node. The dominant information carrier group is determined according to the following formula:

[0104]

[0105] The actual response structure obtained in this embodiment shows that: from the transplanting and greening stage to the early tillering stage, the optical water body response feature group contributes the most to water status monitoring; from the tillering stage to the jointing and booting stage, the contribution of the spectral vegetation response feature group gradually increases; after the jointing and booting stage, the radar water structure response feature group, the thermal infrared water stress feature group, and the meteorological water driving feature group contribute more to water identification under vegetation cover conditions. Based on the above changes in response contribution, this embodiment identifies multiple response change nodes and uses them as candidate boundaries for subsequent adaptive segmentation.

[0106] 3. Divide the carrier-dominant regions and establish a region classification model based on the corresponding characteristics.

[0107] After obtaining the response change nodes of the multi-source information carrier, this embodiment adaptively segments the rice growth period. For example... Figure 4 As shown, after obtaining the response change nodes of multi-source information carriers, this embodiment divides multiple carrier-dominant intervals based on the response change nodes and determines the dominant information carrier group corresponding to each interval. Simultaneously, supplementary features are selected from the non-dominant information carrier groups, and finally, a classification model corresponding to the dominant response interval is trained. Adaptive segmentation does not use a fixed calendar date as the sole boundary, but rather comprehensively considers the response change nodes of multi-source information carriers, crop type, phenological processes, effective accumulated temperature, precipitation and irrigation / drainage events, water demand sensitive periods, and changes in land cover growth to determine the dominant response interval.

[0108] This embodiment identifies rice phenological nodes based on EVI, NDVI, and crop image phenotypic time series. Let V(t) be the time series of EVI, NDVI, or image phenotypic growth indicators, then its rate of change is: When V(t) exceeds the preset growth threshold and RV(t) is positive for two or more consecutive time steps, it is determined that the plant has entered the rapid growth stage; when V(t) reaches near its peak and the absolute value of the rate of change is less than the preset threshold, it is determined that the plant has entered the high-coverage stable stage; when V(t) continuously decreases, it is determined that the plant has entered the maturity and decline stage. In this embodiment, the above phenological nodes are verified with the transplanting greening, tillering, jointing and booting, heading and flowering, and grain filling and maturity dates recorded in manual surveys to obtain the actual phenological process of rice in the current year.

[0109] Let the adaptive segmentation result be , Let L represent the dominant response interval. Then, the adaptive segmentation result is determined according to the following objective function:

[0110]

[0111] Where R(t) represents the multi-source information carrier response contribution vector at time step t. Indicates the first The mean of the response contribution vector within each dominant response interval, where L represents the number of dominant response intervals, λ is the segment number penalty coefficient, and η is the crop growth stage constraint weight. This indicates the degree of deviation between the segmented result Ω and the phenological stage, water-sensitive period, or accumulated temperature process corresponding to crop type c.

[0112] Furthermore, the set T of multi-source information carrier response change nodes is... r phenological node set T p Water demand change node set T w The set of meteorological or irrigation / drainage event nodes T m The nodes are merged to obtain a set of candidate segment nodes. When the time distance between two candidate nodes is less than a preset threshold δ, they are merged into one node.

[0113]

[0114] Among them, b i w represents a candidate node i The weight of the candidate nodes is represented by δ. In this embodiment, δ is set to 5 days, and higher weights are assigned to nodes showing changes in response to multi-source information carriers and nodes showing actual phenological phenomena. Nodes showing single precipitation or irrigation / drainage events are weighted according to the event intensity and duration. The merged nodes are then selected based on the closest actual observation time step as the final segment boundary.

[0115] In this embodiment, the rice growth period is divided into four dominant response intervals using the above method: the first dominant response interval corresponds to the period from transplanting and greening up to the early tillering stage, with the dominant response coming from the optical water body response feature group; the second dominant response interval corresponds to the tillering stage to the early jointing and booting stage, with the dominant response coming from the spectral vegetation response feature group; the third dominant response interval corresponds to the jointing and booting stage to the heading and flowering stage, with the dominant response coming from the radar water structure response feature group and the thermal infrared water stress feature group; and the fourth dominant response interval corresponds to the grain filling and maturity stage, with the dominant response coming from the thermal infrared water stress feature group, the meteorological water-driven feature group, and the radar water structure response feature group.

[0116] Within each dominant response interval, the average response contribution of each information carrier group is calculated:

[0117]

[0118] in, This indicates that the g-th information carrier group is in the k-th dominant response interval I. k Average response contribution within, Indicate interval I k The number of time steps included. Information carrier groups that meet the following conditions are identified as the dominant information carrier group for this interval:

[0119]

[0120] Or, when a certain information carrier group Greater than the preset dominant threshold θ d When this condition is met, it is determined as the dominant information carrier group. If multiple information carrier groups meet the preset dominant condition, then all multiple information carrier groups are jointly designated as the dominant information carrier group. In this embodiment, θ d The threshold is set to 0.18. When the average response contribution of a certain information carrier group is higher than this threshold, it will be included in the dominant information carrier group of the corresponding interval.

[0121] After determining the dominant information carrier group, the features within this group are identified as core features. For features in non-dominant information carrier groups, this embodiment employs recursive feature elimination (RFE) and SHAP importance analysis to screen for supplementary features. When using the SHAP importance analysis method, the average absolute SHAP value of the j-th feature to the model output is calculated:

[0122]

[0123] in, Let represent the SHAP value of the j-th feature in the i-th sample, and N represent the number of samples. If If the following conditions are met, then this feature will be considered a supplementary feature:

[0124]

[0125] in, Here, p represents the preset contribution threshold, and p is the number of candidate features. In this embodiment... Set to 0.2.

[0126] Subsequently, this embodiment utilizes the core and supplementary features within each dominant response interval, combined with ground-based real labels, historical diagnostic labels, and manually labeled samples, to train the interval classification model for the corresponding dominant response interval. This embodiment employs a random forest model as the interval classification model for each dominant response interval, with the number of model trees set to 300. The number of candidate features used during node partitioning adopts a square root strategy, and the class weights are automatically balanced according to the class frequency of the training samples. Training samples are divided into training and validation sets according to land parcels to avoid spatial-temporal leakage caused by adjacent time steps of the same land parcel simultaneously entering both the training and validation sets. After training, the first dominant response interval model, the second dominant response interval model, the third dominant response interval model, and the fourth dominant response interval model are obtained, respectively.

[0127] 4. Route the object to be diagnosed to the corresponding interval model to obtain the water state classification result.

[0128] For the object z to be diagnosed, this embodiment calculates the routing score based on the matching degree of time step or growth stage, the completeness of multi-source data, and the availability of key features:

[0129]

[0130] in, The time step or growth stage matching degree is calculated by comparing the current Julian Day, effective accumulated temperature process, and EVI or NDVI growth status of the object to be diagnosed with the time range and phenological stage corresponding to each dominant response interval. It represents the completeness of multi-source data, which is obtained by statistical normalization based on the completeness of remote sensing data, ground monitoring data, meteorological data and agricultural management data available at the current time step; This indicates the availability of key features, evaluated based on the availability of the core features required by the interval classification model in the current sample. A higher value is assigned when fewer core features are missing. (Parameter) and For routing weight coefficients, this embodiment takes... This allows the matching of growth stages to play a major role in the routing process.

[0131] This embodiment first determines the optimal dominant response interval for the object to be diagnosed: Here, k represents the index of the optimal response dominant interval corresponding to the object z to be diagnosed. Subsequently, the object z to be diagnosed is input into the interval classification model corresponding to the k-th response dominant interval. The water state classification results are as follows: , where x z y represents the feature vector of the object z to be diagnosed at the current time step. z This indicates the output classification results of farmland moisture status.

[0132] If the model outputs class probabilities, the water state class is determined as follows:

[0133]

[0134] Where 'c' represents the moisture state category. This represents the dominant interval of the k-th response.

[0135] In this embodiment, during actual diagnosis, the current Julian day, effective accumulated temperature progress, EVI or NDVI growth status, recent precipitation and irrigation events, and available data source status are first read for each plot to be diagnosed. Then, the routing score between the plot and the four dominant response intervals is calculated. For example, a plot is in the jointing-booting to heading and flowering stage on Julian day 189, the EVI is in the high-coverage stable stage, there has been continuous precipitation recently and the drainage ditch water level is high, and SAR, thermal infrared, ground moisture, and field water level data are complete. The routing result is the third dominant response interval. The plot is then input into the third dominant response interval random forest classification model. The model outputs the category probabilities of normal water status, water surplus status, poor drainage status, and insufficient water supply status, and the category with the highest probability is determined as the current water status classification result of the plot.

[0136] When some information carrier groups are missing, this embodiment employs supplementary feature substitution, homologous feature interpolation, cross-source feature mapping, or backup interval classification models to complete the classification. For example, when optical remote sensing data is missing due to cloud cover, the weights of SAR feature groups, ground moisture monitoring feature groups, and meteorological moisture-driven feature groups are increased in the route classification; when ground monitoring data is missing, remote sensing inversion soil moisture, thermal infrared moisture stress features, or meteorological moisture deficit features are used to supplement the data. For samples with severe missing data that cannot meet the minimum feature availability requirements, a low-confidence diagnostic result is output, indicating the need for supplementary observations.

[0137] 5. Calculation and anomaly diagnosis output for moisture state deviation

[0138] The crop moisture status benchmark library in this embodiment is constructed from completed field test data, historical normal year monitoring data, and agricultural meteorological station data. The data in the benchmark library includes crop type, crop benchmark temperature, growth stage, regional type, cropping system, meteorological moisture background, dominant information carrier group, benchmark moisture status, benchmark moisture status fluctuation scale, field water layer threshold, soil moisture threshold, meteorological moisture threshold, and anomaly detection threshold.

[0139] The crop baseline temperature was determined by actual phenological dates and daily temperature data from agricultural meteorological stations; the baseline moisture state was statistically derived from field moisture observations of normal samples, manual verification results, and historical moisture state diagnostic results. To avoid a circular dependency between the baseline database construction and the classification model, field water depth, soil volumetric water content, manual survey labels, and historical records from normal years were used as the main basis for constructing the baseline moisture state, while the output probability of the classification model was only used as an auxiliary consistency check or robust weighting basis.

[0140] In constructing the baseline moisture state, this embodiment first screens normal samples from field test samples and historical monitoring samples. Normal samples simultaneously meet the following conditions: First, the plot corresponding to the sample has no record of flooding, waterlogging, drought, or poor drainage disasters during this growth stage; Second, field irrigation and drainage records comply with local rice management regulations; Third, manual survey records show that the rice is growing normally and there is no obvious water stress; Fourth, the field water depth and soil volumetric water content are within the normal observation range for the corresponding rice growth stage; Fifth, there is no abnormal moisture background caused by continuous extreme precipitation or long periods without rain during the corresponding period of the sample.

[0141] For each normal sample, this embodiment calculates the water state index F based on field water depth, soil volumetric water content, classification model output probability, and manually verified labels. i For rice paddies, F i Calculate as follows:

[0142]

[0143] in, This represents the deviation of the normalized field water depth. This indicates the deviation of the normalized soil volumetric moisture content. This represents the probability that sample i is classified as being in a state of excess moisture, flooding, waterlogging, or poor drainage. This represents the probability that sample i is classified as a state of insufficient water supply or drought stress. and Let be the weighting coefficients, and satisfy:

[0144]

[0145] This embodiment takes =0.30、 =0.25、 =0.25、 =0.20. F iAs a directional water state indicator, a positive deviation indicates an increased risk of excess water, flooding, waterlogging, or poor drainage, while a negative deviation indicates an increased risk of insufficient water supply or drought stress. and Obtained by normalization of field sensor observations. and The labels are obtained from the interval classification model, and manual verification is used to remove samples with inconsistent labels or inconsistent directions.

[0146] Let N be the number of normal samples under the s-th growth stage, r-th regional type, q-th cropping system, and m-th meteorological and water background. Then, what is the baseline water state index under these conditions? Calculate as follows:

[0147]

[0148] Corresponding benchmark moisture state fluctuation scale Calculate as follows:

[0149]

[0150] in, This represents the water status index of the i-th normal sample under the s-th growth stage, r-th regional type, q-th planting system, and m-th meteorological and water background.

[0151] The benchmark database simultaneously stores the field water layer threshold, soil moisture threshold, and meteorological moisture threshold for each growth stage. The field water layer threshold is determined by the distribution of field water layer depth in normal samples from field experiments; the soil moisture threshold is determined by the distribution of soil volumetric water content in different soil layers in normal samples; and the meteorological moisture threshold is determined by statistics of daily precipitation, reference crop evapotranspiration, and water deficit from agricultural meteorological stations. For each growth stage, this embodiment uses the 5th to 95th percentiles of the normal sample distribution as the normal fluctuation range, and uses the average field water layer depth, average soil moisture content, or comprehensive moisture state index exceeding this range as candidate criteria for anomaly detection.

[0152] For the observed water state index Fobs of the object to be diagnosed z at the current time step t, it is calculated according to the weighted method of the response contribution of multiple information carriers:

[0153]

[0154]

[0155] in, This represents the moisture status characterization value output by the g-th information carrier group for the subject z to be diagnosed at the t-th time step. This represents the normalized response weight of the g-th information carrier group. ε represents the response contribution of the g-th information carrier group, G represents the number of information carrier groups, and ε is a positive number to prevent the denominator from being zero. The moisture state characterization values ​​output by each information carrier group adopt a unified direction, with positive values ​​indicating excess moisture or poor drainage, and negative values ​​indicating insufficient water supply or drought stress.

[0156] For region A to be diagnosed, within the current time step t or the current growth stage s, the observed water status index at the regional scale is calculated based on the classification results. To maintain consistency with the positive and negative directions of FDI, this embodiment differentiates the anomaly probabilities of water excess and water shortage anomalies through statistical analysis:

[0157]

[0158] Where |A| represents the number of pixels in region A, y u x represents the water state classification result of pixel u. u The feature vector representing pixel u. This represents a set of positive anomalies such as excessive moisture, flooding, waterlogging, or poor drainage. This represents a set of negative anomaly categories such as insufficient water supply or drought stress.

[0159] If only the spatial proportion of a certain type of anomaly needs to be calculated, the proportion of the target anomaly category should also be calculated as follows:

[0160]

[0161] Among them, C f Let represent the set of target anomaly categories, and I be the indicator function. When y... u Belongs to C f When I is true, I takes the value 1; otherwise, it takes the value 0. The proportion of this target anomaly category is output as the anomaly spatial distribution result, but it is not directly used for the calculation of FDI with uniform positive and negative directions.

[0162] The moisture state deviation index is the standardized deviation index (FDI), which is calculated as follows:

[0163]

[0164] in, This indicates the observed water status index of the farmland pixel, plot, region, or statistical unit to be diagnosed at the current time step or current growth stage. This represents the baseline water state index in the crop water state baseline database, corresponding to the crop type, growth stage, regional type, planting system, and meteorological and water background. This indicates the fluctuation scale of the baseline moisture state index under the corresponding conditions.

[0165] Based on the relationship between FDI and the anomaly detection threshold, it is determined whether the farmland moisture status is abnormal. Preferably, when When the farmland moisture status is normal, it is determined that the farmland moisture status is normal; when At that time, it was determined that there was a slight excess of moisture or a slight flooding abnormality; when At that time, it is determined that there is a moderate or severe excess moisture, flooding, waterlogging, or poor drainage; when At that time, it was determined that there was a slight water supply shortage; when When the water supply is insufficient or drought stress is abnormal, it is determined that there is a moderate or severe water shortage or drought stress. In this embodiment, T1=1.0 and T2=2.0, and the threshold is determined based on the statistical results of historical samples.

[0166] 6. Implementation Results

[0167] This embodiment was implemented from March 15, 2024 to October 30, 2024 in a typical rice-growing area in the middle reaches of the Lianshui River in Yuhu District, Xiangtan City, Hunan Province. The entire process was completed from multi-source data acquisition, information carrier response change detection, adaptive segmentation, interval classification model diagnosis to abnormal result output. Figure 5 The results of water state anomaly diagnosis for typical diagnostic points are presented, including changes in the contribution of characteristic group responses, changes in the standardized water state deviation index (FDI), and the final anomaly diagnosis output.

[0168] like Figure 5 As shown in (a), the dominant information carriers change significantly at different growth stages. In the early growth stage of rice (DOY approximately 210–230), the optical water body response feature group has a high response contribution, effectively reflecting changes in the field surface water layer and excess water status during the transplanting and greening period. As rice enters the tillering to jointing and booting stage, the contribution of the spectral vegetation response feature group gradually increases, indicating that vegetation growth status gradually becomes an important indicator of water status changes. In the later growth stage, the response contributions of the thermal infrared water stress feature group and the radar water structure feature group increase, enhancing the ability to identify abnormal states such as poor drainage and insufficient water supply. These results verify that the proposed method can dynamically adjust the dominant information carriers according to the crop growth stage, rather than using fixed features for whole-process diagnosis.

[0169] like Figure 5As shown in (b), the FDI index at typical locations changes dynamically over time. A positive FDI value indicates an increased risk of excess water or poor drainage, while a negative value indicates an increased risk of insufficient water supply or drought stress. During DOY 220–235, the FDI at locations P4 and P6 continuously increased, indicating a risk of water accumulation due to precipitation or drainage conditions. Near DOY 250, location P5 showed a significant negative deviation, indicating a trend of insufficient water supply or water stress in a localized area. The FDI change process continuously reflects the degree of deviation of the water state from the normal baseline state.

[0170] like Figure 5 As shown in (c), based on the FDI threshold discrimination rule, the diagnostic results were divided into normal, mildly abnormal, and moderately to severely abnormal states. The green area indicates that the FDI is within the normal fluctuation range, the blue area indicates excessive water or poor drainage, and the red area indicates insufficient water supply or drought stress. The diagnostic results are consistent with the field surface water and plant water stress phenomena recorded in the field survey.

[0171] This method integrates remote sensing data, ground monitoring data, meteorological data, and agricultural management data to achieve multi-source collaborative analysis of farmland water status. It adaptively identifies dominant information carriers at different growth stages, improving the targeting and accuracy of water status diagnosis. By introducing response change detection and adaptive segmentation mechanisms, errors caused by fixed time divisions are effectively avoided, making the model more closely reflect the actual growth process of crops. Furthermore, by constructing a crop water status benchmark database and conducting comparative analysis with deviation indicators, anomaly detection exhibits strong stability and interpretability, more accurately reflecting abnormal farmland water conditions.

[0172] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for diagnosing farmland moisture anomalies based on information carrier response change detection, characterized in that, The process includes the following steps: S1, acquiring crop images, identifying crop types and constructing multi-source information carriers, acquiring crop images of the target farmland area, identifying the crop type corresponding to the object to be diagnosed based on the crop images, and acquiring multi-source time-series data of the target farmland area during the crop growth period; Feature extraction is performed on the multi-source time-series data, and the extracted features are divided into multiple information carrier groups according to data source, physical response mechanism, and water state characterization function; S2, multi-source information carrier response change detection is performed, the category separability measure or response contribution of each information carrier group to farmland water state at different time steps is calculated, the response time series curve corresponding to each information carrier group is constructed, and the multi-source information carrier response change nodes are identified based on the response time series curves; S3, adaptive segmentation and interval model establishment is performed, and the crop growth period is adaptively divided into multiple response-dominant intervals according to the multi-source information carrier response change nodes, combined with the crop type, crop phenological process, water requirement characteristics of different growth stages, and surface cover growth characteristics. Within each dominant response interval, core and supplementary features are identified, and corresponding interval classification models are established. S4. Perform routing classification of the objects to be diagnosed. Route the farmland pixels, plots, or regions to be diagnosed to the corresponding interval classification models according to their crop type, time step, growth stage, dominant response interval, and multi-source data availability to obtain farmland water status classification results. S5. Perform water status deviation calculation and anomaly diagnosis output. Call the baseline water status under the corresponding crop type and growth stage, compare the farmland water status classification results with the baseline water status, calculate the water status deviation index, and output the farmland water status anomaly diagnosis results based on the water status deviation index.

2. The method according to claim 1, characterized in that, In step S1, the crop images include one or more of the following: UAV crop images, ground-captured crop images, field-fixed camera images, images acquired by mobile terminals, or high-resolution remote sensing images. The step of identifying the crop type corresponding to the object to be diagnosed based on the crop image includes: performing cropping, noise reduction, radiometric correction, geometric correction, scale normalization or image enhancement processing on the crop image, and using an image classification model, object detection model or semantic segmentation model to identify the crop type. When multiple crops exist within the target farmland area, a spatial distribution map of crop types is generated based on the crop image recognition results. This spatial distribution map of crop types is then used as input for subsequent information carrier division, adaptive segmentation, routing classification, and reference moisture status retrieval.

3. The method according to claim 1, characterized in that, In step S1, the multi-source time-series data includes at least two of the following: remote sensing data, ground monitoring data, meteorological data, and agricultural management data. The remote sensing data includes two or more of the following: optical remote sensing data, multispectral remote sensing data, hyperspectral remote sensing data, thermal infrared remote sensing data, synthetic aperture radar (SAR) data, microwave soil moisture data, or topographic data. The ground monitoring data includes one or more of the following: soil volumetric water content, soil moisture in different layers, field water layer depth, groundwater depth, drainage ditch water level, soil temperature, soil conductivity, crop canopy temperature, or ground-measured moisture status labels. The meteorological data includes one or more of the following: precipitation, air temperature, relative humidity, wind speed, solar radiation, evapotranspiration, reference crop evapotranspiration, continuous precipitation, effective accumulated temperature, drought index, or water deficit. The agricultural management data includes one or more of the following: irrigation time, irrigation amount, drainage time, drainage amount, cropping system, sowing date, transplanting date, fertilization records, or field management records.

4. The method according to claim 1, characterized in that, In step S1, the multi-source time-series data undergoes time registration, spatial registration, outlier removal, missing value imputation, scale unification, and standardization. One or more of the following are extracted: water body index, vegetation index, multispectral band features, hyperspectral narrow band features, surface temperature features, temperature-vegetation-drought index features, SAR backscattering features, polarization features, interferometric coherence features, microwave soil moisture features, ground moisture features, field water level features, meteorological driving features, irrigation and drainage management features, crop image phenotypic features, and crop growth features. The information carrier group includes at least two of the following: optical water body response feature group, spectral vegetation response feature group, thermal infrared water stress feature group, radar water structure response feature group, microwave soil moisture feature group, ground moisture monitoring feature group, field water level monitoring feature group, meteorological water driving feature group, irrigation and drainage management feature group, crop image phenotypic feature group, and crop growth feature group.

5. The method according to claim 1, characterized in that, Step S2 includes: S21, for each time step, calculating the class separability measure of each information carrier group for farmland moisture status, and constructing the class separability time series curve corresponding to each information carrier group; wherein, let X be the sample feature set corresponding to the g-th information carrier group at the t-th time step. g (t), where C is the number of farmland moisture state categories and n is the number of samples in the c-th category. c The mean of the c-th class of samples is μ c,g (t), the population sample mean is μ g (t), then the intra-class scatter matrix S of the g-th information carrier group at time step t. w,g (t) and the inter-class scatter matrix S b,g (t) are respectively: ; The corresponding category separability measure J g (t) is calculated as follows: Where tr() represents the matrix trace, and ε is a positive number to prevent the denominator from being zero; S22. Based on one or more of the following: category separability measure, model feature importance, classification boundary interval, prediction confidence or abnormal response intensity, calculate the response contribution of each information carrier group at the corresponding time step, and construct the response contribution time series curve. S23. Perform change point detection on the time series curves of category separability and / or response contribution, and identify time nodes where statistical characteristics change significantly, the dominant information carrier changes, or the response contribution ranking changes, as the response change nodes of the multi-source information carrier; wherein, the change point detection adopts any one or more of the following algorithms: change point search algorithm based on Bayesian information criterion, PELT algorithm, Bayesian online change point detection algorithm, sliding window statistical test algorithm, or dynamic programming segmentation algorithm.

6. The method according to claim 5, characterized in that, The category separability measure is used to characterize the ability of the corresponding information carrier group to distinguish different farmland water status categories at the current time step. The category separability measure includes one or more of the following: Fisher discriminant ratio, inter-class distance to intra-class distance ratio, mutual information, Jeffries-Matusita distance, Bach distance, classification boundary interval, area under the receiver operating characteristic curve, average precision, model feature importance score, or SHAP feature contribution value. The multi-source information carrier response change node is output in one or more of the following forms: Julian day, calendar date, growth stage number, or time step number.

7. The method according to claim 1, characterized in that, Step S3 includes: S31, using the response change node of the multi-source information carrier as the candidate segmentation boundary, and according to the growth stage, water demand sensitive period, accumulated temperature process, precipitation irrigation and drainage events and surface cover growth characteristics corresponding to the crop type, the candidate segmentation boundary is processed by advancing, delaying, merging or splitting to obtain adaptive segmentation results; S32. Within each dominant response interval, compare the class separability measure and / or response contribution of each information carrier group, and determine the information carrier group with the highest separability, the highest response contribution, or that meets the preset dominant conditions as the dominant information carrier group. S33. The features in the dominant information carrier group are identified as core features, and recursive feature elimination (RFE), minimum redundancy maximum correlation (mRMR), LASSO, permutation importance analysis, or SHAP importance analysis methods are used to select features from the non-dominant information carrier group whose contribution meets a preset threshold as supplementary features. S34. Using the core features and the supplementary features, combined with ground real labels, historical diagnostic labels or manually labeled samples, train an interval classification model corresponding to the dominant response interval. The interval classification model is any one or more of the following: random forest model, support vector machine model, gradient boosting decision tree model, extreme gradient boosting model, lightweight gradient boosting model, neural network model or ensemble learning model.

8. The method according to claim 1, characterized in that, In step S4, the routing classification includes: determining the corresponding routing rules based on the crop type, spatial location, current time step, growth stage, dominant response interval, multi-source data completeness, and key feature availability of the farmland pixels, plots, or regions to be diagnosed; when the farmland pixels, plots, or regions to be diagnosed meet the routing conditions of a certain dominant response interval, they are input into the interval classification model corresponding to that dominant response interval; when there are multiple optional interval classification models, the final interval classification model is determined based on crop type matching degree, time distance, feature completeness, model confidence, or historical classification stability; when some information carrier groups are missing, supplementary feature substitution, homologous feature interpolation, cross-source feature mapping, or backup interval classification models are used to complete the classification.

9. The method according to claim 1, characterized in that, Step S5 includes: S51, obtaining the farmland moisture state classification results output by the interval classification model and the predicted probability corresponding to each moisture state category, and obtaining the moisture state characterization value and response contribution of each information carrier group participating in the current diagnosis; S52, constructing the observed moisture state index of the farmland pixel, plot, or region to be diagnosed at the current time step or current growth stage based on the moisture state characterization value, response contribution, and / or the predicted probability of each moisture state category of each information carrier group. The observed water state index adopts a unified directional numerical representation method, with positive values ​​indicating a tendency for excess water, flooding, waterlogging, or poor drainage, and negative values ​​indicating a tendency for insufficient water supply or drought stress; S53, based on the crop type, growth stage, regional attributes, planting system, and meteorological water background corresponding to the farmland pixel, plot, or region to be diagnosed, the corresponding benchmark water state index is retrieved from the crop water state benchmark database. Scale of fluctuation in baseline moisture state and anomaly detection threshold; S54, based on the observed moisture state index The aforementioned benchmark moisture state index and the benchmark moisture state fluctuation scale Calculate the standardized moisture state deviation index (FDI); S55. Based on the relationship between the standardized water state deviation index (FDI) and the anomaly discrimination threshold, determine whether the water state of the farmland pixel, plot, or area to be diagnosed is abnormal, and output one or more of the following: farmland water state classification result, anomaly type, anomaly level, standardized water state deviation index, and diagnostic confidence level; when the FDI is within the preset normal range, the water state is determined to be normal; when the FDI is greater than the preset positive threshold, it is determined that there is an anomaly of excess water, flooding, waterlogging, or poor drainage; when the FDI is less than the preset negative threshold, it is determined that there is an anomaly of insufficient water supply or drought stress.

Images