Rice nutrient precision monitoring and fertilization decision system based on multi-source data fusion

Abstract

The application discloses a rice nutrient precision monitoring and fertilization decision system based on multi-source data fusion, and particularly relates to the field of precision agriculture, and is used for solving the problems of nutrient monitoring deviation and inaccurate fertilization caused by spectral saturation during the rice vigorous growth period, and is achieved by the following steps: collecting red edge and near-infrared reflection from multiple angles, synchronously acquiring texture and structure images and fusing to generate an anti-saturation feature cube, outputting a vigorous growth period feature package by spatiotemporal segmentation at the phenological transition point, constructing a separable nitrogen trace spectrum by eliminating saturated response, outputting a prescription front graph by adjusting the balance coefficient under the calculation of energy concentration and water layer confluence indicators in a context constraint graph, and finally outputting an execution prescription graph by combining historical trajectories and micro-terrain to check the landing point, so that the vigorous growth period diagnosis precision is improved, over-fertilization or under-fertilization is avoided, and efficient nutrient management is realized.

Description

Technical Field

[0001] This invention relates to the field of precision agriculture, and more specifically, to a rice nutrient precision monitoring and fertilization decision system based on multi-source data fusion. Background Technology

[0002] Nutrient monitoring and fertilization decisions in rice fields rely on integrated air-ground observation and coordinated execution by machinery, with information sources covering canopy reflectance, soil physicochemical properties, and meteorological factors. After the rice enters its vigorous growth stage, leaf layer superposition, chlorophyll accumulation, and structural scattering become dominant, leading to a plateau in the spectral response. If common canopy indices are still used as the primary diagnostic tool, the discriminative power of input features will decrease, consequently causing overall deviations in predictions and prescriptions. Prescription maps may show excessive or insufficient amounts in vigorous growth areas, and subsequent staging will amplify these biases by using compressed spectral cues. Monitoring and prescription generation during the vigorous growth stage require joint constraints from source-level anti-saturation, robust model learning, and closed-loop validation at the prescription level to avoid chain-like distortions from perception to execution.

[0003] The published patent application CN120355524A, entitled "An Intelligent Decision-Making System and Method for Fertilization and Topdressing Based on Rice Aerial and Ground Sensing Data", proposes to use UAV multispectral data, satellite remote sensing, ground sensor network and weather station as the data acquisition end. After spatiotemporal fusion, it enters the temporal prediction of convolutional network and memory network, outputs growth assessment and maturity judgment, and generates fertilization prescription and operation path from rule base. It is then sent to agricultural machinery for execution through IoT interface, forming a closed process from observation to operation.

[0004] The reliance on dual-band canopy indexes (red and near-infrared) for nitrogen fertilization during periods of high fertility suffers from inherent saturation: strong absorption in the red band and high reflectivity in the near-infrared band plateau under high coverage, compressing the exponential gain. This triggers a chain reaction where exponential saturation weakens feature separability, causing the discrimination boundary formed by the time-series model based on the training distribution to shift. Rule thresholds mistake seemingly consistent saturation responses for field consistency, leading to a shift in prescription magnitude and spatial boundaries. After prescriptions are issued, variable fertilization creates mismatches in fertile plots, accumulating errors in subsequent stages. The root cause of this problem lies in the dual effects of spectral-structural coupling and rule-model linkage: the former compresses available information at the source, while the latter amplifies downstream decision-making biases. Only by simultaneously observing spectral plateauing, model boundary drift, and prescription shifts in the context of long-term fertility can these technical problems be fully identified and explained.

[0005] To address the aforementioned problems, a technical solution is provided. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion. It acquires red-edge and near-infrared reflections from multiple angles, simultaneously obtains texture and structural images, and fuses them to generate anti-saturation feature cubes. It segments the temporal and spatial dimensions using phenological transition points and aggregates plots to generate peak-season feature packages. It removes saturation responses to form separable nitrogen trace spectra, calculates red-edge energy concentration and water layer confluence indicators, and combines historical data to generate a calibration coefficient to adjust the pre-prescription map boundary. It then combines historical trajectories, time delays, and topographical data to verify the landing point and update the water surface reference output execution pre-prescription map. This addresses the monitoring bias and fertilization inaccuracies caused by peak-season spectral saturation in the prior art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] Data acquisition and fusion unit: During the peak period, multi-angle reflections of red edge and near-infrared are acquired and texture and structural images are acquired simultaneously. After completing the radiation and geometry consistency, the images are fused in a unified manner to generate anti-saturation feature cubes.

[0009] Spatiotemporal segmentation and aggregation unit: Using phenological transition points as time anchors, spatiotemporal segmentation is performed on the anti-saturation feature cube and aggregated by plot, outputting a peak period feature package for peak period identification.

[0010] Saturated response elimination unit: Based on the long-term feature package, a nitrogen-sensitive discrimination sub-graph is constructed. The saturated response is eliminated by combining the red edge morphological difference with the local orientation consistency to obtain the separable nitrogen trace spectrum;

[0011] Contextual adjustment unit: First, calculate the red edge energy concentration and water layer flow indicator within the contextual constraint diagram and output the peak season contextual adjustment coefficient using a Bayesian additive model. Then, use the peak season contextual adjustment coefficient to participate in the threshold interval mapping and prescription boundary contraction or relaxation to generate the prescription pre-map.

[0012] Landing point verification output unit: The landing point is jointly verified by the pre-prescription map, historical operation trajectory, controller response delay and micro-topography. The prescription map is output for execution after the water surface reference is updated by the near-windless period or by the quick-capture image.

[0013] Furthermore, the data acquisition and fusion unit acquires red-edge and near-infrared multi-angle reflections according to the key angle set, simultaneously acquires texture and structural images, and completes radiometric consistency with the help of a ground reference plate and airborne radiometric correction. Radiometric consistency includes completing atmospheric and sensor link reflectivity inversion with a ground reflectivity reference plate, establishing angular kernel functions according to solar geometry and apparent geometry to solve for angular normalized reflectivity to standard geometry, and performing spectral resampling according to the band response curve.

[0014] Furthermore, the data acquisition and fusion unit completes geometric consistency, which includes constructing an orthoreflectivity image by matching sparse high-precision control points with dense ones, eliminating stitching seams and local stretching, generating a direction histogram on a single scale to calculate direction entropy, and expressing direction consistency by the eigenvalue ratio of the structure tensor.

[0015] Furthermore, the data acquisition and fusion unit performs homogeneous fusion, strictly registers the data according to pixel coordinates and a unified timestamp, and stacks the red-edge main band, red-edge adjacent band, near-infrared main band, directional entropy, and directional consistency to form an anti-saturation feature cube.

[0016] Furthermore, the spatiotemporal segmentation and aggregation unit uses phenological transition points as time anchors. The location of phenological transition points includes performing piecewise polynomial fitting on the red-edge integral time series, calculating the difference between adjacent segments based on the model selection criteria, and recording the position of the largest difference as the transition point. If there is a gap in optical observation, the same time anchor is achieved by compensating for the difference in piecewise fitting of the radar amplitude sequence.

[0017] Furthermore, the spatiotemporal segmentation and aggregation unit performs spatiotemporal segmentation, extracting peak-season observation segments and excluding non-peak-season samples based on time anchors, and completing partitioned aggregation and outputting plot-level data blocks using plot boundary vectors. The peak-season feature package consists of a cubic subset of anti-saturation features for each plot, time anchor labels, and observation angle set labels.

[0018] Furthermore, the saturated response elimination unit calculates the red edge morphology quantity from the long-term feature package. The red edge morphology quantity is obtained by calculating the half-width and quarter-width of the red edge main band curve, taking the ratio of the two as the morphology expansion ratio, and multiplying the distance from the spectral center position to the lower edge of the red edge by the morphology expansion ratio to obtain the morphology expansion distance.

[0019] Furthermore, the saturated response elimination unit calculates the directional organization quantity, constructs the directional organization index by combining directional entropy and directional consistency, and constructs a two-dimensional dominant discrimination surface by dividing directional consistency by one and adding directional entropy. If the morphological extension distance is in the low variation region and the directional organization index is in the low order region, it is marked as a saturated response candidate. The candidate is verified by the difference dominance relationship between the two frames before and after the time anchor, and non-candidate pixels are retained to form a nitrogen-sensitive discrimination sub-map.

[0020] Furthermore, the context-adjustment unit calculates the red-edge energy concentration, calculates the reflection integral in the red-edge main band, calculates the reflection integral in the red-edge adjacent band, and obtains the red-edge energy concentration by dividing the main band integral by the adjacent band integral. The intersection integral of the upper and lower quantile envelopes of the angular sequence is taken and then the ratio is calculated to calculate the water layer confluence indicator. The free water surface coverage ratio is estimated by the specular candidate set and polarization observation. The depression skeleton length density is extracted by the digital surface model. The water layer confluence indicator is obtained by hyperbolic tangent transformation of the product of the coverage ratio and the length density.

[0021] Furthermore, the situational adjustment unit uses a Bayesian additive model to combine historical prescription effects and inspection records to output peak season situational adjustment coefficients. A threshold interval mapping function is defined on the separable nitrogen trace spectrum. The deviation of the peak season situational adjustment coefficient from the midpoint determines the boundary adjustment range. The boundary adjustment factor is equal to the sign function multiplied by twice the absolute deviation. When the peak season situational adjustment coefficient is less than the midpoint, boundary contraction is performed; when it is greater than the midpoint, boundary widening is performed, forming the pre-prescription map.

[0022] The technical effects and advantages of this invention, a rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion, are as follows:

[0023] This invention achieves closed-loop optimization from perception to execution in rice nutrient monitoring and fertilization decision-making through a holistic technical logic of multi-source data fusion. The anti-saturation feature cube generated by multi-angle data acquisition and fusion provides a robust input foundation for subsequent spatiotemporal segmentation, ensuring the spatiotemporal continuity of the peak-season feature package under optical gap compensation. Furthermore, based on the saturation response elimination and nitrogen trace spectrum construction of this package, the parameters of the coordinated context constraint graph are calculated and the calibration coefficient mapping is performed, dynamically adjusting the prescription boundary to adapt to field confluence risks. Finally, the combined historical trajectory, time delay, and micro-topography landing point verification further corrects water surface disturbances, achieving accurate output of the prescription map. This mutually reinforcing chain mechanism not only overcomes the decrease in feature discrimination and decision bias caused by spectral saturation in traditional methods but also incorporates historical feedback through a Bayesian framework, forming an adaptive learning path that improves the anti-interference capability of monitoring and the accuracy of fertilization variables. This significantly reduces the risk of accumulated peak-season bias, promoting the sustainability and economic benefits of rice nutrient management. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the structure of the rice nutrient precision monitoring and fertilization decision system based on multi-source data fusion according to the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Example 1: Figure 1 This invention presents a rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion, comprising:

[0027] Data acquisition and fusion unit: During the peak period, multi-angle reflections of red edge and near-infrared are acquired and texture and structural images are acquired simultaneously. After completing the radiation and geometry consistency, the images are fused in a unified manner to generate anti-saturation feature cubes.

[0028] Spatiotemporal segmentation and aggregation unit: Using phenological transition points as time anchors, spatiotemporal segmentation is performed on the anti-saturation feature cube and aggregated by plot, outputting a peak period feature package for peak period identification.

[0029] Saturated response elimination unit: Based on the long-term feature package, a nitrogen-sensitive discrimination sub-graph is constructed. The saturated response is eliminated by combining the red edge morphological difference with the local orientation consistency to obtain the separable nitrogen trace spectrum;

[0030] Contextual adjustment unit: First, calculate the red edge energy concentration and water layer flow indicator within the contextual constraint diagram and output the peak season contextual adjustment coefficient using a Bayesian additive model. Then, use the peak season contextual adjustment coefficient to participate in the threshold interval mapping and prescription boundary contraction or relaxation to generate the prescription pre-map.

[0031] Landing point verification output unit: The landing point is jointly verified by the pre-prescription map, historical operation trajectory, controller response delay and micro-topography. The prescription map is output for execution after the water surface reference is updated by the near-windless period or by the quick-capture image.

[0032] In the field of rice nutrient monitoring, traditional methods relying on a single canopy index are prone to spectral saturation during the vigorous growth phase, leading to prediction bias and fertilizer prescription deviations. This problem stems from information compression dominated by leaf layer superposition and structural scattering, affecting the accuracy of the entire process from perception to execution. However, by employing a multi-angle, multi-source data acquisition and fusion strategy, feature discrimination is enhanced at the source end, establishing a foundation for anti-saturation observation, providing robust input for vigorous growth phase diagnosis, and ensuring closed-loop constraints in the decision chain.

[0033] 1-1: Multi-angle reflection and image acquisition.

[0034] Observing the rice canopy during its peak growing season presents a challenge of spectral saturation, necessitating the acquisition of information from multiple sources to enhance feature robustness. To address the sensitivity in the red-edge and near-infrared bands, key angle sets are used for directional acquisition, ensuring coverage of representative combinations of solar azimuth and observation zenith angles.

[0035] The peak growth canopy refers specifically to the dense layer of top branches and leaves that forms after rice plants enter their vigorous growth period. This stage usually corresponds to the jointing to heading stage. At this time, the plant grows rapidly, and the overlapping of leaves and accumulation of chlorophyll are significant, leading to a more complex canopy structure and dominating structural scattering, which in turn affects the spectral response and the accuracy of nutrient monitoring.

[0036] In practice, the key angle set is first defined as eight points uniformly distributed between 0 and 360 degrees of solar azimuth and five gradient points within the range of 0 to 60 degrees of observed zenith angle, forming 40 acquisition angle pairs. Airborne multispectral sensors are used to simultaneously acquire red-edge reflectance data, covering the main red-edge band (680 nm to 750 nm) and adjacent red-edge bands (750 nm to 780 nm); simultaneously, near-infrared main band reflectance (780 nm to 900 nm) is acquired. Homogeneous texture images are acquired in parallel, using a high-resolution camera to capture canopy surface texture details with a resolution of at least 0.05 meters; homogeneous structural images are also acquired, and a three-dimensional point cloud of the canopy is reconstructed using a stereo imaging module, with a point density of at least 100 points per square meter. All acquisitions are performed under a unified timestamp to ensure spatiotemporal alignment of the data. After this acquisition process is completed, an original acquisition set is formed, including red-edge reflectance, near-infrared reflectance, texture images, and structural images, providing a multimodal basis for consistent processing.

[0037] 1-2: Radiation and geometric uniformity treatment.

[0038] Based on the original collected reflection data, due to atmospheric interference and sensor link effects, it is necessary to unify the radiation dimensions and align the geometry to eliminate external biases. Under the background of high canopy reflectivity during peak season, uniformization processing is required to prevent saturation amplification.

[0039] In practice, the first step is to perform radiation uniformity: multiple ground reflectance reference plates with known reflectance are placed in the field, and after the sensors collect their reflection signals, atmospheric path radiation is simulated using an atmospheric transmission model. After deducting scattering and absorption effects, the surface reflectance is obtained by inversion. An angular kernel function is established according to solar geometry and apparent geometry, and a semi-empirical model is used to solve for the angular normalized reflectance, thus standardizing the multi-angle data to the nadir observation geometry. Spectral resampling is performed according to the band response curve, and the Gaussian response function is used to interpolate the red-edge main band, red-edge adjacent band, and near-infrared main band to ensure a smooth transition of band boundaries.

[0040] Next, geometrical homogenization is performed: fixed field features are selected as sparse, high-precision control points, with at least 20 points having centimeter-level coordinates obtained via GPS; dense matching is performed using a structure-from-motion algorithm to generate digital orthophotos, eliminating parallax and projection distortion caused by multi-angle acquisition; a multinomial transformation model is applied band-by-band to the radiometrically homogenized reflectance dataset, with parameters derived from least-squares fitting of the control points; texture and structure images are processed simultaneously to ensure all images share a unified coordinate reference system. After processing, an orthoreflectance image set is output, with all channels' reflectance dimensions unified to a unitless percentage form, supporting unbiased input for subsequent derived quantity extraction.

[0041] 1-3: Extraction and fusion of texture and structural derivatives.

[0042] Orthoreflectivity image sets provide a unified foundation, but structural information of the canopy during peak season needs to be quantified to supplement spectral saturation discrimination clues and fused with spectral data to construct a high-dimensional space. Directional features are extracted at a single scale to avoid multi-scale interference.

[0043] In practice, for texture images, firstly, an orientation histogram is generated, and the gradient direction is calculated using a gradient operator to statistically analyze the distribution from 0 degrees to 180 degrees; then, the orientation entropy is calculated, and the randomness of the texture is quantified by summing the natural logarithm of the orientation histogram.

[0044] For structural images, the eigenvalue ratio of the structural tensor is calculated: a 2x2 tensor matrix is ​​constructed, where eigenvalues ​​represent the intensity of the principal and secondary directions, and the ratio expresses directional consistency. After extraction, strict registration is performed according to pixel coordinates and a unified timestamp, using affine transformation to ensure sub-pixel-level alignment. The red-edge main band, red-edge adjacent band, near-infrared main band, directional entropy, and directional consistency are stacked to form a multi-channel cube, with all channels sharing the same coordinate reference and reflectivity dimension. Texture and structural channels are treated as dimensionless quantities, and unit labels are added to avoid dimensional drift. After processing, an anti-saturation feature cube is generated as the core data carrier for peak season monitoring, supporting seamless integration with subsequent spatiotemporal segmentation and nitrogen-based diagnostics.

[0045] By constructing an anti-saturation feature cube through data acquisition and fusion unit processing, multi-source data is fused at the source end to improve the discrimination of canopy features during peak season, avoid chain distortion caused by spectral saturation, and provide a robust foundation for subsequent processing.

[0046] The anti-saturation feature cube enhances the discriminative power of peak-season canopy information at the source end through multi-source fusion, avoiding the plateau bias of traditional spectral indices and laying a robust foundation for time-series analysis. This data structure retains the multi-channel characteristics of the red-edge main band, red-edge adjacent band, near-infrared main band, directional entropy, and directional consistency, ensuring that subsequent processing can directly call upon unified dimensions and coordinate references. However, in peak-season monitoring, temporal discontinuities and spatial heterogeneity can amplify observation gaps, necessitating the introduction of phenological transition points as an anchoring mechanism to achieve precise spatiotemporal segmentation and compensation, thereby outputting highly targeted feature packages to support the clarification of nitrogen diagnosis boundaries and decision-making closure.

[0047] 2-1: Location of phenological transition points.

[0048] The anti-saturation feature cube provides multi-temporal observation sequences, but the onset of the peak growth period needs to be accurately anchored to isolate the saturation effect; otherwise, the segmentation bias will accumulate in the downstream discrimination. Analysis of the red-edge integral time series can capture growth inflection points.

[0049] The peak growth period specifically refers to the stage when plants enter a period of vigorous growth, corresponding to the critical developmental window from the jointing stage to the heading stage. At this time, the plant growth rate accelerates, leaves and stems expand rapidly, chlorophyll content accumulates significantly, leading to an increase in canopy density and dominating the structural scattering effect, thereby affecting the spectral response and the accuracy of nutrient diagnosis.

[0050] In practice, the reflection integral sequences of the red-edge main band and adjacent red-edge bands are first extracted from the anti-saturation feature cube and sorted by timestamp to form a continuous time series. A piecewise polynomial fitting algorithm is then used to process this sequence, with a third-order polynomial piecewise fitted to each potential segment. Next, the difference between adjacent segments is calculated using model selection criteria such as the Akaike Information Criterion, and the location of the largest difference is selected as the phenological transition point and marked as a time anchor.

[0051] If gaps occur in optical observations during the fitting period, such as data interruption due to cloud cover, the process switches to radar amplitude sequences for compensation: the radar echo intensity time series is extracted, and the difference is calculated using a piecewise polynomial fitting algorithm to achieve continuous alignment of the time anchor. After this positioning process is completed, a unique time anchor label is output to accurately define the start of the peak period, supporting subsequent unbiased spatiotemporal reference segments.

[0052] 2-2: Spatiotemporal segmentation and land parcel aggregation.

[0053] While time anchor labels have established peak season boundaries, it is necessary to perform segmentation against saturation feature cubes to remove irrelevant segments and consider spatial partitioning to match field heterogeneity. This allows us to focus on the peak season saturation response and avoid non-target samples interfering with aggregation accuracy.

[0054] In practice, the peak observation period is extracted from the anti-saturation feature cube, which is the sequence from the time anchor to the end of the current observation. All samples before the time anchor are removed channel by channel to ensure that only the peak period data is retained.

[0055] Next, using the parcel boundary vector as a spatial mask, the segmented sequence is partitioned and aggregated: a predefined parcel boundary vector file containing polygon coordinates is loaded; average aggregation is performed on pixels within each parcel boundary to generate parcel-level data blocks, preserving all dimensions of the original channels such as the red-edge main band. After this segmentation and aggregation, a sequence of parcel-level data blocks is output, with channel naming and dimensions consistent with the anti-saturation feature cube, providing partitioned input for feature package assembly.

[0056] Average aggregation refers to data dimensionality reduction by calculating the simple arithmetic mean of pixel values ​​within each channel. For example, for a plot containing N pixels, on the red-edge main band channel, the aggregation value is the sum of the reflectance of all N pixels divided by N, thus generating a plot-level data block. This data block is a multidimensional array whose dimensions are consistent with the original anti-saturation feature cube, but the spatial resolution is reduced to the plot cell level. This method helps reduce the impact of noise, improve computational efficiency, and maintain the integrity of information between channels, facilitating the subsequent construction and use of nitrogen-sensitive sub-maps without losing any original dimensions or coordinate references.

[0057] 2-3: Generation of long-term feature packages.

[0058] The plot-level data block sequence is ready, but tags need to be integrated to form a complete input carrier, supporting continuous invocation of nitrogen-sensitive analysis. In peak season scenarios, this packet structure can embed time anchor and observation angle information, enhancing the spatiotemporal robustness of the discrimination.

[0059] In practice, a time anchor label is attached to each plot-level data block, directly copying the self-localization results; simultaneously, an observation angle set label is attached, recording the combination of solar azimuth and observation zenith angle of the key angle set, such as 8 azimuth point locations and 5 zenith angle gradients. After integration, a peak-season feature package is formed, including an anti-saturation feature cube subset (i.e., plot-level data blocks), time anchor labels, and observation angle set labels, with all elements sharing a unified coordinate reference. After this generation process, the peak-season feature package is output as the input carrier for constructing the nitrogen-sensitive discrimination submap.

[0060] By forming a peak-season feature package through spatiotemporal segmentation and aggregation units, the anti-saturation feature cube is optimized in the spatiotemporal dimension, compensating for observation gaps and aggregating plot heterogeneity, improving the accuracy of peak-season nitrogen monitoring, avoiding saturation accumulation bias, and providing a clear boundary basis for nitrogen sensitivity analysis.

[0061] The peak season feature package has optimized the anti-saturation feature cube through spatiotemporal segmentation and compensation. Time anchor labels and observation angle set labels are embedded in the plot-level aggregation to ensure the spatiotemporal robustness and channel consistency of the input carrier, injecting precise peak season clues into nitrogen diagnosis. It directly supports the calculation of discriminant quantities, avoiding the confounding interference of saturation responses. However, in the context of high-density canopies during peak season, differences in red edge morphology and directional consistency need to be collaboratively eliminated to remove saturation responses, in order to construct clear sub-graphs and generate separable nitrogen trace spectra. This improves the boundary accuracy of nitrogen-sensitive discrimination, prevents the amplification of biases in the decision chain, and achieves closed-loop constraints from the source to the prescription.

[0062] 3-1: Determining the red-edge morphological quantity.

[0063] The long-term feature package provides multi-channel data at the plot level, but the shape of the red-edge curve needs to be quantified to capture morphological differences; otherwise, the saturation response will obscure nitrogen indications. This method can highlight the sensitivity to changes in the main red-edge band.

[0064] In practice, the red-edge main band curve is extracted from the long-term feature packet and sorted by wavelength to form a spectral sequence. A peak detection algorithm is used to locate the maximum value of the curve, and the half-width at half-maximum (HWHM), i.e., the wavelength distance between the left and right intersections at half the height of the peak, is calculated. Similarly, the quarter-width at half-maximum (WWHM), i.e., the wavelength distance between the left and right intersections at one-quarter of the height of the peak, is calculated. Next, the ratio of HWHM to WWHM is taken as the morphological expansion ratio to quantify the curve broadening degree. The distance from the spectral center (the wavelength of the curve's centroid) to the lower edge of the red edge (680 nm) is calculated and multiplied by the morphological expansion ratio to obtain the morphological extension distance, in nanometers. The red-edge morphological quantity is output, including the morphological expansion ratio and the morphological extension distance, supporting dimensional input for saturation discrimination.

[0065] 3-2: Determination of directional tissue quantity.

[0066] The red-edge morphology has quantified spectral characteristics, but it needs to be combined with structural information to construct an organization index to enhance directional discrimination; otherwise, low-order regions are easily dominated by saturation. This combination can balance entropy and consistency, avoiding bias from a single indicator.

[0067] In practice, directional entropy and directional consistency channels are extracted from the long-term feature package. A directional organization index is constructed and calculated using the formula of directional consistency divided by directional entropy, outputting a dimensionless value that emphasizes the weight of ordered structures. After this calculation process, the directional organization quantity, i.e., the directional organization index, is output as an auxiliary dimension for saturation removal, ensuring seamless matching with the red-edge morphological quantity.

[0068] The directional organization index characterizes the degree of directional order in canopy texture. It is calculated by dividing directional consistency by directional entropy. The larger the value, the more directional consistency and low randomness the texture has, i.e., the canopy structure is highly organized. The smaller the value, the more random and less consistent the texture has, i.e., the canopy exhibits disorder or saturation response.

[0069] 3-3: Saturation response elimination and nitrogen-sensitive discrimination subgraph formation:

[0070] The directional organization quantity and red-edge morphology quantity are complete, but a discriminant surface needs to be constructed to remove saturated candidates to purify the data; otherwise, peak-period response platformization will interfere with nitrogen trace boundaries. Precise filtering is achieved through two-dimensional dominance and temporal difference.

[0071] In the specific implementation, a two-dimensional dominance discrimination surface is constructed, using morphological extension distance and directional organization index as coordinate axes. Low-variance regions are defined as those with morphological extension distances below a preset threshold, and low-order regions are those with directional organization indices below a preset threshold. If a pixel is simultaneously in both regions, it is marked as a saturation response candidate. Next, the difference dominance relationship is calculated between frames before and after the time anchor label. The inter-frame absolute difference is used to verify the candidates, retaining only non-candidate pixels with differences exceeding a preset threshold. The retained pixels are integrated to generate a nitrogen-sensitive discrimination sub-map, preserving the original coordinate reference. This nitrogen-sensitive discrimination sub-map serves as the basic carrier for nitrogen trace spectral mapping.

[0072] 3-4: Separable nitrogen trace spectrum generation.

[0073] The nitrogen-sensitive discrimination submap has been saturated, but energy kernels need to be mapped and boundaries defined to output a clear spectrum; otherwise, prescription generation is susceptible to residual bias. In peak-period diagnosis, this mapping can integrate energy and tissue quantity, improving separability.

[0074] In practice, red-edge morphological quantities are extracted from the nitrogen-sensitive discrimination sub-image and mapped to red-edge energy kernels. These kernels are calculated by dividing the red-edge principal band integral by the red-edge adjacent band integral, resulting in a dimensionless energy kernel. Together with directional organization quantities, they define the spatial boundary. Boundary tracing algorithms, such as the Moore neighborhood method, are used to scan the sub-image edges to generate closed boundary contours. After integration, a separable nitrogen trace spectrum is formed. The output separable nitrogen trace spectrum supports direct access to the context constraint graph.

[0075] By generating separable nitrogen trace spectra through the saturation response removal unit, saturation responses are removed and boundary clarity is enhanced in the nitrogen sensitivity analysis during the peak season, thereby improving the discrimination of nutrient monitoring, avoiding chain distortion, and providing robust input for contextual calibration.

[0076] Separable nitrogen trace spectra have enhanced the discriminative power of nitrogen cues during peak fertilization periods through saturation elimination and boundary constraints. At the plot level, they retain a clear mapping between red-edge energy cores and directional tissue quantities, ensuring the separability and channel consistency of the input carrier. This provides a precise diagnostic basis for situational adjustments and avoids the neglect of peak fertilization confluence risks. However, in the complex field environment during peak fertilization periods, it is necessary to integrate red-edge energy concentration and water layer confluence indicators to construct a constraint map. Furthermore, a Bayesian framework should be used to fuse historical data and output a calibration coefficient, thereby dynamically adjusting prescription boundaries. This achieves closed-loop optimization by shrinking high-risk areas and widening low-risk areas, preventing the accumulation of fertilization biases and improving decision-making accuracy.

[0077] The scenario constraint diagram is a two-dimensional parameter framework for monitoring nutrient levels during the peak rice growing season. It consists of red-edge energy concentration as the first parameter, water layer runoff indicator as the second parameter, and a plot scenario field. The red-edge energy concentration is calculated by the ratio of the reflectance integral of the main zone to the adjacent zone to quantify the spectral energy distribution. The water layer runoff indicator is obtained by taking the product of the free water surface coverage ratio and the length density of the depression skeleton through hyperbolic tangent transformation to assess runoff risk. The diagram combines the two parameters into a two-dimensional observation vector and integrates historical prescription effects and inspection records based on the Bayesian likelihood ratio accumulation method to output a peak season scenario calibration coefficient to dynamically adjust the fertilization prescription boundary and ensure that the decision-making adapts to complex field scenarios.

[0078] 4-1: Calculation of red-edge energy concentration.

[0079] Separable nitrogen trace spectra provide an integral basis for the red-edge main band and the red-edge adjacent band, but the energy distribution needs to be quantified to suppress angular differences; otherwise, multi-angle responses during peak periods will introduce instability, thus avoiding plateau amplification bias.

[0080] In practice, the red-edge main band reflection integral sequence is extracted from the separable nitrogen trace spectrum, and the main band integral value is obtained by integrating along wavelength. Similarly, the red-edge adjacent band reflection integral sequence is extracted, and the adjacent band integral value is obtained by integrating along wavelength. The preliminary ratio is obtained by dividing the main band integral by the adjacent band integral, which serves as the basis for energy concentration. To suppress angular differences, the sequence is sorted according to the observation angle set labels, and the intersection integration region of the upper and lower quartiles (i.e., the upper quartile and the lower quartile) is calculated. Then, the ratio is calculated for this region, and the dimensionless red-edge energy concentration is output. The red-edge energy concentration is used as the first parameter of the scenario constraint graph, supporting the assembly of two-dimensional vectors.

[0081] 4-2: Calculation of water layer confluence indicator.

[0082] The red-edge energy concentration is sufficient, but it needs to be combined with water surface and topographic information to quantify the confluence risk; otherwise, water layer interference during peak season will blur the prescription boundary. In non-polarized data scenarios, specular detection can estimate the free water surface, improving the robustness of the indication.

[0083] In practice, the existence of polarization observation is first checked. If it exists, the polarization observation is used directly to estimate the proportion of free water surface coverage, and the proportion value is calculated by polarization degree threshold segmentation. If there is no polarization data, a speckle detection algorithm such as threshold segmentation is used to identify the speckle candidate set, and a speckle removal algorithm such as connected component analysis is applied to remove noise spots and estimate the proportion of free water surface coverage.

[0084] Next, the depression skeleton is extracted from the digital surface model, and the length density, i.e., the total length of the skeleton per unit area, is calculated using skeletonization algorithms such as Zhang's refinement method. The product of the cover percentage and the length density is used as input, and a dimensionless water layer confluence indicator is obtained through hyperbolic tangent transformation. After this calculation process, the water layer confluence indicator is output as the second parameter of the context constraint graph, which is combined with the red-edge energy concentration to form a two-dimensional observation vector.

[0085] 4-3: Mapping of the balancing coefficient during peak season.

[0086] Two-dimensional observation vectors have been formed, but historical data needs to be integrated to output coefficients in a Bayesian manner; otherwise, prescription adjustments are easily dominated by current biases. The likelihood ratio accumulation method can balance fit and bias class labels, improving the objectivity of peak-season adjustments.

[0087] In practice, a small number of field inspection samples and historical prescription effects are loaded as a labeled dataset. Quantile equidistant interpolation and anisotropic monotonic spline fitting algorithms are applied to the edge densities of the three classes (fit, over-, and under-fit) on the red-edge energy concentration and water layer runoff indicator, respectively, to obtain unweighted density estimates. The likelihood ratio is calculated, where the numerator is the joint density of the fit class and the denominator is the supremum of the joint density of the over- and under-fit classes. After taking the ratio, it is mapped to the zero-to-one interval using the formula κ = ratio divided by one, outputting the dimensionless peak season scenario balancing coefficient κ, which serves as the control parameter for threshold interval adjustment.

[0088] The peak season scenario balance coefficient characterizes the balance factor of the field scenario during the peak season of rice. It is calculated based on historical prescription effects and inspection records using the likelihood ratio accumulation method. The larger the value, the closer it is to 1, the more suitable the scenario is and the lower the risk, thus relaxing the fertilization boundary in the low-risk area; the smaller the value, the closer it is to 0, the higher the risk of over-fertilization or under-fertilization, thus narrowing the fertilization boundary in the high confluence risk area.

[0089] 4-4: Prescription image generation.

[0090] The peak season scenario correction coefficient has been output, but it needs to be applied to the separable nitrogen trace spectrum to dynamically map the boundary; otherwise, the boundary of the high confluence risk zone is prone to over-expansion. In peak season decision-making, this adjustment can shrink or widen the risk zone to ensure that the prescription matches the field scenario.

[0091] In practice, a threshold interval mapping function is defined from the separable nitrogen trace spectrum, i.e., an initial threshold boundary is set based on a piecewise linear function of nitrogen trace values. The magnitude of κ's deviation from the midpoint, i.e., 0.5, is calculated, and the boundary adjustment factor is equal to the sign function multiplied by twice the absolute deviation, where the sign function is sgn(κ-0.5). If κ is less than 0.5, boundary contraction is performed, i.e., the boundary of the high confluence risk zone is reduced to decrease the amount of fertilizer applied; if κ is greater than 0.5, boundary widening is performed, i.e., the boundary of the low-risk zone is expanded to increase the flexibility of fertilization, forming a pre-prescription map, which serves as the input carrier for subsequent verification.

[0092] The pre-prescription map refers to the preliminary fertilization prescription mapping structure generated based on separable nitrogen trace spectra. It dynamically adjusts the threshold interval boundaries using peak season scenario calibration coefficients, forming a pre-processed version of fertilization intensity and spatial zoning. Its function is to integrate scenario parameters such as red-edge energy concentration and water layer runoff indicators, achieving boundary contraction in high-runoff-risk areas and widening in low-risk areas. This provides an intermediate decision-making basis adapted to the field environment during peak season, facilitating subsequent verification by combining historical operation trajectories, controller response delays, and micro-topography, avoiding the accumulation of fertilization deviations and improving overall decision-making accuracy.

[0093] The pre-prescription map is obtained through the treatment of the situational adjustment unit. Energy and flow parameters are integrated in the peak season situational constraints, and the boundary is dynamically adjusted through the adjustment coefficient to improve the adaptability of fertilization decision-making, avoid the accumulation of deviations, and provide a closed-loop basis for prescription execution.

[0094] The pre-prescription map dynamically adjusts the threshold boundary through a scenario-based adjustment coefficient, integrating red-edge energy concentration and water layer confluence indicators at the plot level to ensure the risk adaptability and coordinate consistency of the initial prescription, providing a closed-loop foundation for execution optimization. This map structure directly supports landing point verification, avoiding spatiotemporal deviations in peak fertilization periods. However, in field machinery collaboration scenarios, multiple corrections are needed, combining historical operation trajectories, controller response delays, and micro-topography models, and updating water surface references to address wind field disturbances. This results in a refined prescription map, minimizing deviations from diagnosis to execution and improving the reliability and accuracy of overall decision-making.

[0095] 5-1: Track alignment processing.

[0096] The pre-prescription map provides initial boundaries, but it needs to be aligned with historical trajectories to verify the landing point; otherwise, peak season offsets will amplify execution errors. Alignment captures temporal relationships, ensuring correction accuracy.

[0097] In practice, spraying trigger points and positioning records are extracted from historical operation trajectories and sorted by timestamps to construct time-position offset curves. Dynamic time warping is used to solve the offset function along the way, that is, the offset parameters are calculated by minimizing the distance between curves, a continuous offset mapping is generated, and the time-position offset curve and offset function are output as input parameters for response delay verification, supporting spatiotemporal adjustment of nozzle control points.

[0098] 5-2: Response delay verification processing.

[0099] The time-position offset curves are ready, but they need to be adjusted backwards to incorporate controller characteristics; otherwise, the delay deviation will affect the accuracy of the landing point. In peak-season variable fertilization, this verification can quantify fixed and variable offsets, improving the consistency of machine response.

[0100] In practice, pulse records are extracted from the controller response delay characteristics, and the fixed offset (constant time difference) from triggering the spray imaging signal and the variable offset (dynamic time difference that varies with the input) are calculated. The above offsets are directly applied to the forward adjustment of the nozzle control point, that is, the landing point coordinates in the back-tracking prescription map are mapped back through the offset function to achieve delay compensation, and the adjusted landing point set is output as the reference input for micro-topography correction.

[0101] 5-3: Micro-topography correction processing.

[0102] The adjusted set of landing points provides a basis for verification, but boundary operations must still be performed according to the terrain to match the flow direction; otherwise, confluence in low-potential areas will lead to uneven fertilization. Corrections can dynamically compress or widen the boundaries to adapt to surface disturbances during peak seasons.

[0103] In practice, a surface flow network is loaded from the micro-topography model, and flow analysis algorithms such as the D8 method are used to extract low-potential and high-potential areas. Morphological erosion is performed on the pre-prescription map in low-potential areas to compress the fertilization boundary, and morphological dilation is performed in high-potential areas to widen the fertilization boundary. The operation radius is determined by multiplying the absolute value of the flow gradient by a preset scaling factor. For example, when the flow gradient is 0.05, the radius is set to 3 meters to accommodate boundary adjustments under mild slopes. If wind causes wave surface disturbance, a quick-capture image is acquired during a near-windless period (when the wind speed is below a threshold). Threshold segmentation is used to update the current free water surface reference, and the erosion and dilation processes are repeated to refine the boundary. After this correction process, a revised pre-prescription map draft is output, along with an updated water surface reference, supporting vector integration in the final output.

[0104] 5-4: Prescription chart output.

[0105] A revised draft prescription map has been completed, but it needs to integrate zoning and intensity tables for traceability; otherwise, the execution chain is prone to incompleteness. In the peak season decision-making loop, this output can be embedded with tags to improve traceability.

[0106] In practice, vector partitions are generated from the revised draft prescription map, and the partition outlines are defined using a boundary polygon algorithm; simultaneously, a target fertilization intensity table is generated, and intensity values ​​are calculated by partition based on nitrogen trace spectrum mapping; records are kept consistent with the pre-prescription... Figure 1 The coordinate reference and units are specified, along with the offset function version number (function iteration identifier) ​​and time anchor label. After this output process is completed, a prescription map for execution is generated, serving as the final carrier for the distribution of agricultural machinery.

[0107] By operating the prescription map through the upper landing point verification output unit, the trajectory, time delay and micro-topography are combined in the peak landing point verification, and the water surface reference is updated to resist disturbance, thereby improving the accuracy of fertilization execution, avoiding chain distortion, and realizing a closed loop of the entire process from monitoring to decision-making.

[0108] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0109] It should be noted that the system of the present invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting a variety of hardware environments and usage requirements.

[0110] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0111] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely to distinguish one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0112] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion, characterized in that, include: Data acquisition and fusion unit: During the long term, multi-angle reflections of red edge and near-infrared are acquired and texture and structural images are acquired simultaneously. After completing the radiation and geometric consistency, the red edge main band, red edge adjacent band, near-infrared main band, directional entropy and directional consistency are fused from the same source to generate anti-saturation feature cube. Spatiotemporal segmentation and aggregation unit: Using phenological transition points as time anchors, spatiotemporal segmentation is performed on the anti-saturation feature cube and aggregated by plot, outputting a peak period feature package for peak period identification. Saturation response elimination unit: Based on the long-term feature packet, the red edge morphology and orientation organization are calculated. The red edge morphology is obtained by calculating the half-width and quarter-width of the red edge main band curve and taking the ratio of the two as the morphology expansion ratio. The distance from the spectral center to the lower edge of the red edge is calculated and multiplied by the morphology expansion ratio to obtain the morphology expansion distance. The orientation organization is calculated by constructing the orientation organization index by combining the orientation entropy and orientation consistency. The orientation consistency is divided by one and the orientation entropy is added to construct a two-dimensional dominant discrimination surface. If the morphology expansion distance is in the low variation region and the orientation organization index is in the low order region, it is marked as a saturation response candidate. The candidate is verified by the difference dominance relationship between the two frames before and after the time anchor. Non-candidate pixels are retained to construct a nitrogen-sensitive discrimination sub-map. The saturation response is eliminated by the red edge morphology difference combined with the local orientation consistency to obtain the separable nitrogen trace spectrum. Contextual adjustment unit: First, calculate the red edge energy concentration and water layer flow indicator within the contextual constraint diagram, and output the peak season contextual adjustment coefficient using a Bayesian additive model. Then, use the peak season contextual adjustment coefficient to participate in the threshold interval mapping and prescription boundary contraction or relaxation to generate the prescription pre-map. Landing point verification output unit: The landing point is jointly verified by the pre-prescription map, historical operation trajectory, controller response delay and micro-topography. The prescription map is output for execution after the water surface reference is updated by the near-windless period or by the quick-capture image.

2. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 1, characterized in that: The data acquisition and fusion unit acquires red-edge and near-infrared multi-angle reflections according to the key angle set, and simultaneously acquires texture and structure images. It completes radiometric homogenization with the help of a ground reference plate and airborne radiometric correction. Radiometric homogenization includes completing atmospheric and sensor link reflectivity inversion with a ground reflectivity reference plate, establishing angular kernel functions according to solar geometry and apparent geometry to solve for angular normalized reflectivity to standard geometry, and performing spectral resampling according to the band response curve.

3. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 2, characterized in that: The data acquisition and fusion unit completes geometric consistency, which includes constructing an orthoreflectivity image by matching sparse high-precision control points with dense ones, eliminating stitching seams and local stretching, generating a direction histogram on a single scale to calculate direction entropy, and expressing direction consistency by the eigenvalue ratio of the structure tensor.

4. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 3, characterized in that: When the data acquisition and fusion unit performs source fusion, it strictly registers the data according to pixel coordinates and a unified timestamp.

5. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 1, characterized in that: The spatiotemporal segmentation and aggregation unit uses phenological transition points as time anchors. The location of phenological transition points includes performing piecewise polynomial fitting on the red-edge integral time series, calculating the difference between adjacent segments based on the model selection criteria, and recording the position of the largest difference as the transition point. If there is a gap in optical observation, the same time anchor is achieved by compensating for the difference in piecewise fitting of the radar amplitude sequence.

6. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 5, characterized in that: The spatiotemporal segmentation and aggregation unit performs spatiotemporal segmentation, extracts the peak-season observation segment and removes non-peak-season samples by using the time anchor as the boundary, and completes the partitioning and aggregation by the plot boundary vector to output plot-level data blocks. The peak-season feature package consists of a cubic subset of anti-saturation features for each plot, a time anchor label, and an observation angle set label.

7. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 1, characterized in that: The context-adjustment unit calculates the red-edge energy concentration, calculates the reflection integral in the red-edge main band, and calculates the reflection integral in the red-edge adjacent band. The red-edge energy concentration is obtained by dividing the main band integral by the adjacent band integral. The intersection integral of the upper and lower quantile envelopes of the angular sequence is taken and then the ratio is calculated to calculate the water layer confluence indicator. The free water surface coverage ratio is estimated by the specular candidate set and polarization observation. The depression skeleton length density is extracted by the digital surface model. The water layer confluence indicator is obtained by hyperbolic tangent transformation of the product of the coverage ratio and the length density.

8. The rice nutrient precision monitoring and fertilization decision-making system based on multi-source data fusion according to claim 7, characterized in that: The situational adjustment unit uses a Bayesian additive model to combine historical prescription effects and inspection records to output peak season situational adjustment coefficients. A threshold interval mapping function is defined on the separable nitrogen trace spectrum. The deviation of the peak season situational adjustment coefficient from the midpoint determines the boundary adjustment range. The boundary adjustment factor is equal to the sign function multiplied by twice the absolute deviation. When the peak season situational adjustment coefficient is less than the midpoint, boundary contraction is performed; when it is greater than the midpoint, boundary widening is performed, forming the pre-prescription map.

Images