Risk assessment method and system for agricultural products in chromium-nickel high geological background area

Abstract

The invention discloses a chromium-nickel high geological background area agricultural product risk assessment method and system, and relates to the technical field of chromium-nickel pollution monitoring, and the method comprises the steps: firstly, synchronously obtaining laser echo, binocular parallax and illumination change in the same time window, carrying out the triple alignment, and generating a multi-mode original ranging signature; calculating a real-time reliability weight according to the echo integrity and the texture complexity, and implementing kernel attenuation weighting on the depth candidates; fusing target short-time motion coherence information, and outputting a steady-state depth sequence by means of iterative correction of a residual learning network and a Lyapunov criterion; and finally, a pixel-level uncertainty boundary is generated based on a Bayesian residual error and a heteroscedastic Gaussian process, and the laser power, the exposure time and the frame rate are adjusted on line by using a reliability weight, so that distance measurement, evaluation and acquisition parameter adaptive closed loop are realized, and data support is provided for rapid, fine and credible chromium-nickel pollution risk judgment.

Description

Technical Field

[0001] This invention relates to the field of chromium and nickel pollution monitoring technology, specifically to a method and system for risk assessment of agricultural products in areas with high chromium and nickel geological background. Background Technology

[0002] High background levels of heavy metals in soil have become a research hotspot in recent years, involving disciplines such as geology, geochemistry, soil science, and environmental science. Chromium-nickel composite pollution has become a common safety hazard faced by key grain and vegetable producing areas in South and Southwest my country. Due to industrial and mining emissions, leakage from chromium-containing refractory material stockpiles, and irrigation with nickel-containing electroplating wastewater, field soil-irrigation water-aerosols exhibit multi-media and multi-scale enrichment characteristics. Furthermore, the root types and transpiration intensities of economic crops such as rice, leafy vegetables, and melons vary, resulting in significant plant-temporal dual differences in heavy metals within the same geomorphological unit.

[0003] Current risk assessment methods largely rely on manual sampling or single-path sensors: while laboratory chemical analysis is precise, its sampling density is sparse and time lag is large, failing to reflect rapid changes in light, wind field, and plant posture throughout the day; single-sensor laser or vision systems often experience ranging drift due to canopy shading, light dips, and ground reflections, thus amplifying the calculation error of the soil-crop enrichment coefficient. Meanwhile, acquisition parameters such as laser power, camera exposure, and frame rate are usually set once before flight, making it difficult to automatically adjust to real-time signal-to-noise changes. This results in a lack of interpretable confidence in subsequent model inputs, and makes it difficult for regulatory authorities to quickly determine high-risk areas and their priority for remediation.

[0004] In the context of chromium and nickel contamination in the field, there is an urgent need for an agricultural product risk assessment technology that can continuously output high spatial resolution, clear confidence boundaries, and automatically adjust the collection strategy according to the environment within a minute-scale time scale. The core challenge lies in how to make the system maintain low sensitivity to external drastic changes such as light, wind speed, and plant swaying in the same ranging task, while suppressing the cascading misjudgments triggered by sensor drift within the algorithm link.

[0005] Specifically, laser echoes can be truncated due to multipath truncation caused by leaf pores, and stereo matching can become mismatched due to a surge in texture complexity. The coupling of these two factors causes depth estimation to be instantaneously impacted by high-energy noise. Without dual quantification of echo integrity and texture complexity, outlier depths cannot be eliminated in real time, and residuals will accumulate along the time axis, leading to misjudgments of the location and intensity of high-risk heavy metal areas. This, in turn, exacerbates the risks to the soil-crop-food chain due to misaligned control measures. Furthermore, if real-time reliability information cannot be promptly written back to the acquisition end, the next sampling cycle will still enter with unbalanced signal-noise, creating a vicious cycle of measurement-assessment disconnect. How to simultaneously complete adaptive fusion of multimodal data, steady-state depth correction, uncertainty quantification, and data acquisition parameter writing back in a dynamic field environment has become a key technical bottleneck restricting the engineering implementation of existing risk assessment methods.

[0006] Therefore, this invention provides a method and system for risk assessment of agricultural products in areas with high chromium and nickel geological background. Summary of the Invention

[0007] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a method and system for risk assessment of agricultural products in areas with high chromium and nickel geological backgrounds. It simultaneously acquires laser echo, binocular parallax, and illumination changes within the same time window and aligns them triple-wise to generate a multimodal raw ranging signature. Then, it calculates real-time reliability weights based on echo integrity and texture complexity, applying kernel attenuation weighting to depth candidates. Subsequently, it integrates short-term motion coherence information of the target, iteratively corrects it using a residual learning network and Lyapunov criterion, and outputs a steady-state depth sequence. Finally, it generates pixel-level uncertainty boundaries based on Bayesian residuals and heteroscedastic Gaussian processes, and uses reliability weights to adjust laser power, exposure time, and frame rate online, achieving an adaptive closed loop for ranging, assessment, and acquisition parameters, thereby solving the technical problems described in the background art.

[0008] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: Risk assessment methods for agricultural products in areas with high chromium and nickel geological backgrounds include: The laser echo sequence and the binocular parallax image are acquired synchronously within the same time window, and the illumination change characteristics are recorded. Triple alignment is achieved based on timestamp, radiance and spatial reprojection. Multimodal original ranging signature is generated and output after being marked with frame number. The real-time reliability weight is calculated based on the echo integrity score and the texture complexity score. Kernel attenuation weighting is performed on multiple depth candidates to construct a weighted depth candidate set and output the corresponding index matrix to form a weight mapping table for subsequent use. The weighted depth candidate set and the short-term motion coherence information of the target are jointly input into the residual learning network. The network is iteratively corrected by gated residual accumulation and Lyapunov relaxation mapping to output a steady-state depth sequence and generate a residual variance index simultaneously. Pixel uncertainty boundaries are generated based on the steady-state depth sequence and the residual variance. The distance-uncertainty pair is input into the heavy metal risk model, and the laser power, exposure time and frame rate are adjusted according to the real-time reliability weight to update the environmental adaptation parameters for the next sampling period.

[0009] Furthermore, the observation window length is taken as twice the reciprocal of the target plant's trunk swaying frequency, and the laser emission and left and right camera exposure are triggered simultaneously through a high-precision timing unit; The time consistency function is calculated in real time to monitor the difference between the laser sampling time and the visual sampling time. When the difference exceeds the threshold, the corresponding frame is discarded to ensure the time calibration consistency of the multimodal original ranging signature.

[0010] Furthermore, the construction of the multimodal original ranging signature includes: performing joint principal component analysis on the laser amplitude sequence after radiometric normalization and the disparity depth matrix; retaining the principal components with a cumulative contribution rate of not less than 80% to generate a compressed vector; and then linearly fusing the illumination feature vector obtained by time-series filtering according to the illumination sensitivity weight to output the signature.

[0011] Furthermore, the echo integrity score is calculated using the spectral entropy inverse scaling of the normalized spectral energy, and the texture complexity score is obtained by normalizing the local fractal dimension after performing box counting on the Laplacian response of the corrected parallax depth matrix, and finally output in the range of zero to one.

[0012] Furthermore, the calculation of the real-time reliability weight includes: inputting the echo integrity score and the texture complexity score into the Dirichlet prior model, adaptively obtaining the single-source weight system based on the mother sample variance, and then using an exponential kernel decay function to weight the outlier degree of the initial depth candidates to obtain the weighted depth candidate set.

[0013] Furthermore, the generation of the target's short-term motion coherence information includes: estimating the coarse motion vector field using the pyramid optical flow algorithm, and employing... The motion coherence field vector is estimated by eliminating false matches, and the motion embedding matrix is ​​generated by performing an exponential decay mapping through the Euclidean distance between it and the average vector of the whole field. After normalization, the coherence information is output.

[0014] Furthermore, the residual learning network consists of two layers of gated recurrent units and one fully connected layer. It uses a gated residual accumulation method to update the depth in multiple rounds, and stops iterating when the Lyapunov candidate function converges to a preset ratio compared to the previous round, and outputs the steady-state depth sequence.

[0015] Furthermore, the generation of the pixel uncertainty boundary first establishes a conjugate Bayesian model with unknown mean and unknown scale based on the residual vector to obtain the Student-t posterior confidence interval, then models the depth error surface with a spatial heteroscedastic Gaussian process, and finally fuses the residual uncertainty and spatial uncertainty according to the maximum entropy criterion.

[0016] Furthermore, the adjustment of the acquisition parameters first uses a calibrated linear gain matrix to map the real-time reliability weight to the reference values ​​of laser power, exposure time, and acquisition frame rate. Then, during the sampling process, the time consistency error and energy compensation factor are monitored, and the parameters are incrementally corrected according to the sliding mode adjustment law.

[0017] Risk assessment system for agricultural products in areas with high chromium and nickel geological background, including: The data acquisition module synchronously acquires the laser echo sequence and the binocular parallax image within the same time window, records the illumination change characteristics, achieves triple alignment based on timestamp, radiance and spatial reprojection, generates a multimodal original ranging signature and marks the frame number before outputting it; The weight fusion module calculates the real-time reliability weight based on the echo integrity score and the texture complexity score, performs kernel attenuation weighting on multiple depth candidates, constructs a weighted depth candidate set and outputs the corresponding index matrix to form a weight mapping table for subsequent use. The deep steady-state module inputs the weighted depth candidate set and the short-term motion coherence information of the target into the residual learning network, and iteratively corrects it through gated residual accumulation and Lyapunov relaxation mapping, outputting a steady-state depth sequence and simultaneously generating a residual variance index. The adaptive feedback module generates pixel uncertainty boundaries based on the steady-state depth sequence and the residual variance, inputs the distance-uncertainty pair into the heavy metal risk model, and adjusts the laser power, exposure time and frame rate according to the real-time reliability weight to update the environmental adaptation parameters for the next sampling period.

[0018] (III) Beneficial Effects This invention provides a method and system for risk assessment of agricultural products in areas with high chromium and nickel geological backgrounds, which has the following beneficial effects: By simultaneously acquiring laser echo, binocular parallax, and illumination changes and generating multimodal raw ranging signatures, radiometric, geometric, and temporal triple alignment is completed in the same time slot. A consistent depth starting point can be obtained without relying on an additional calibration board, directly reducing the interference of plant shading and sudden changes in illumination on subsequent links. By constructing reliability weights based on echo integrity and texture complexity, and then combining kernel attenuation to generate a weighted depth candidate set, a collaborative mechanism for adaptively identifying reliable information from both the signal layer and the scene layer is realized. This enables depth estimation to prioritize the retention of high-confidence pixels, significantly alleviating the contradiction that traditional single-threshold filtering cannot take into account both sparse high-reflectivity spots and dense dark areas. By introducing coherent short-term motion information of the target to redistribute the weighted depth candidate set, and progressively outputting the steady-state depth sequence through a residual learning network, the project innovatively combines temporal consistency constraints and gated residual accumulation, so that the drift error is pushed into a controllable decay channel, solving the problem of depth jump caused by field wind. By superimposing Bayesian residual modeling with spatial heteroscedastic Gaussian processes, pixel-level uncertainty boundaries are generated, providing interpretable confidence bandwidth for subsequent chromium-nickel enrichment and health risk models. At the same time, reliability weights are mapped to acquisition parameters such as laser power, camera exposure, and frame rate for write-back, forming a three-in-one collaborative closed chain of measurement-evaluation-adjustment, ensuring that the next sampling cycle is still in the optimal signal-to-noise range. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the process for risk assessment of agricultural products in areas with high chromium and nickel geological backgrounds according to the present invention; Figure 2 This is a schematic diagram of the structure of the risk assessment system for agricultural products in areas with high chromium and nickel geological backgrounds according to the present invention; Figure 3 Rice yields under different treatments; Figure 4 The potential non-carcinogenic risk of rice consumption in typical basalt weathering areas of the Leizhou Peninsula is shown in (a) the average level of rice in basalt weathering pollution areas, and (b) the average level of rice with excessive Cr. Figure 5 The study investigated the heavy metal content in rice grains and sugarcane stalks to assess crop planting structure and pollution status. Detailed Implementation

[0020] 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.

[0021] Please see Figure 1 This invention provides a method for risk assessment of agricultural products in areas with high chromium and nickel geological backgrounds, including, In conducting risk assessments of agricultural products in farmland contaminated with chromium and nickel, the primary step is to collect raw observations that reflect the coupling state of soil, crops, and the environment. Traditional methods often rely on single sensors or discrete sampling, making it difficult to maintain observational consistency under conditions of significantly fluctuating natural light, vegetation canopy structure, and differences in surface texture. This results in subsequent risk calculations being plagued by noise amplification, spatial interpolation drift, and temporal distortion.

[0022] This scheme collaboratively acquires three types of information—laser echo amplitude, binocular parallax depth, and concurrent illumination variations—within the same time slot, and uses this to construct a multimodal raw ranging signature. This provides a unified spatiotemporal benchmark for subsequent reliability assessment, depth steady-state correction, and pollution-enrichment-risk linkage simulation. This approach captures the combined effects of soil surface undulations, canopy height differences, and abrupt changes in illumination on laser scattering and visual matching accuracy without complicating the testing process. This allows for continuous observation and characterization of the spatial occurrence and crop enrichment mechanisms of heavy metals on a three-dimensional scale, further providing high-confidence and traceable data support for chromium and nickel pollution risk zoning and remediation priority ranking.

[0023] Step 1: By implementing time synchronization, radiation normalization, and spatial reprojection alignment of laser echo, binocular parallax, and illumination changes within the same time window, a multimodal original ranging signature under a unified reference frame is finally generated.

[0024] Step 101: Multi-source synchronous acquisition and spatiotemporal alignment To ensure that subsequent algorithmic steps can process data within the same reference frame, a time-gating and spatial calibration mechanism is used to unify the laser echo amplitude sequence, binocular parallax image, and illumination variation characteristics into a complete and seamless multimodal raw ranging signature within a single observation window. Only such consistent input can prevent data mismatch caused by acquisition timing drift or field-of-view misalignment, ensuring that the three-dimensional quantification of the impact of chromium-nickel contamination remains accurate at the source.

[0025] First, set an observation window length. The high-precision timing unit simultaneously sends trigger pulses to the laser ranging subsystem and the binocular vision subsystem, ensuring that the laser pulse emission time is strictly aligned with the shutter time of the left and right cameras.

[0026] Next, within the observation window length The instantaneous illuminance sequence output by the internal light sensor is continuously collected and denoted as the illuminance change vector. To ensure that the three data streams have a corresponding relationship during subsequent fusion, a time consistency function is introduced. : ; Wherein: laser sampling time With visual sampling time Representing the first Timestamps for frame laser and visual data.

[0027] Observation window length The time span of the data collection period can be set according to the wind frequency of the plants and the cruising speed of the drone. Time Consistency Function : A scalar measure of the synchronization error between laser and visual acquisition. When the time consistency function Synchronization is considered successful if the value is less than a set threshold; real-time monitoring of the time consistency function is performed, and if the time consistency function... If the limit is exceeded, the system immediately discards the corresponding frame to avoid the spread of subsequent depth estimation error chains. Based on time alignment, it is necessary to ensure that the laser beam scanning area completely overlaps with the field of view of the binocular camera.

[0028] Therefore, the echo intensity map of the laser ranging system Point cloud reprojection is performed using a laser intrinsic parameter matrix. With laser-camera extrinsic matrix Calculate laser point At the left camera pixel coordinates Mapping: ; in, For the camera intrinsic parameter matrix, is the scaling factor, and is the homogeneous coordinate scaling factor; Laser intrinsic parameter matrix : A matrix containing the laser scanning angular resolution and ranging linearity coefficients; Laser-camera extrinsic matrix : A four-by-four matrix describing the rigid transformation of the coordinate system between the laser emitter and the camera; Camera intrinsic matrix A 3x3 matrix represented by focal length, principal point coordinates, and pixel size; scaling factor : Ensures the homogenization ratio of pixel coordinates after reprojection, used to calculate the actual pixel position.

[0029] After reprojection, the laser intensity and visual texture can be matched one by one at the pixel level to provide a precise location correspondence for subsequent confidence assessment. Due to the difference between surface reflectivity and canopy leaf orientation, the dynamic range of laser echo intensity and visual brightness will be inconsistent, so bidirectional normalization needs to be performed during the acquisition stage.

[0030] Laser energy compensation factor is used on the laser side. Correction for distance attenuation; correction factor used on the visual side. The corrected nonlinear response, after normalization, yields the corrected laser amplitude sequence. With disparity depth matrix : ; Where: laser energy compensation factor The proportionality coefficient calculated based on the ranging and energy attenuation model is greater than 0; correction factor. : A real number that adjusts the linearity of image grayscale, generally greater than 0; ; Corrects the laser amplitude sequence Laser echo intensity sequence after distance attenuation compensation; parallax depth matrix :Finish The corrected depth matrix.

[0031] After implementing bidirectional normalization, the three data streams have a unified benchmark in terms of radiometry and geometric scale, providing comparable dimensions for the generation of confidence weights in the next step. Through a collaborative acquisition mechanism integrating time synchronization, radiometric correction, and spatial reprojection, this step integrates laser echo, binocular parallax, and illumination variations into a unified reference frame. This significantly reduces source mismatches caused by plant shading, terrain undulations, and solar flicker, achieving a stable starting point for subsequent deep fusion and providing a true and reliable spatiotemporal benchmark for reliability assessment.

[0032] Step 102: Construction of Multimodal Raw Ranging Signature After completing the synchronous acquisition and alignment, the three data streams still need to be processed into a unified signature that can be directly called by subsequent algorithms in order to track the coupling effect of heavy metal migration and highly dynamic crops in the risk assessment chain.

[0033] Based on the reprojected point-to-pixel correspondence, the corrected laser amplitude sequence With disparity depth matrix Perform joint principal component analysis (J-PCA) to project high-dimensional intensity-depth features into feature compression vectors. The formula is as follows: ; Where, mapping matrix It is the transformation matrix learned by J-PCA based on the cumulative contribution rate threshold; the mapping matrix It contains several columns of feature vectors, each column corresponding to a principal component direction; feature compression vector. A vector set that simultaneously represents laser amplitude and depth information in a low-dimensional space; Cumulative contribution rate threshold: This threshold determines the percentage of principal components retained, generally greater than 50%; it is achieved through feature compression vectors. The structure of the data not only compresses the amount of data but also retains the dominant information that affects ranging drift.

[0034] Illumination Variation Vector The impact on ranging reliability is significant. To quantify the perturbation of depth inference caused by illumination fluctuations, an illumination sensitivity weight is introduced. Define the formula: ; Among them, the fluctuation of illuminance inside the window , representing the extreme difference in illuminance within the observation window, and the light modulation coefficient. Control the steepness of the weighting curve and the fluctuation of illuminance within the window. The difference between the maximum and minimum illumination values ​​in the current observation window, a non-negative real number; Light modulation coefficient : A positive real number used to adjust the slope of the S-shaped weight curve; light sensitivity weight : A coefficient that quantifies the impact of illumination fluctuations on ranging reliability, with a value between 0 and 1.

[0035] Large fluctuations in illuminance inside the window This will result in a higher light sensitivity weight. In subsequent confidence assessments, the initial weights of laser and visual features are reduced to decrease the risk of drift. This is achieved by compressing the comprehensive feature vector. Weighted by light sensitivity Multimodal raw ranging signatures are generated using a linear weighting method. : ; Among them, the illumination feature vector Light change vector The stable illumination feature vector obtained by the time-series filter, the illumination feature vector A low-dimensional vector describing the trend of light change; Multimodal raw ranging signature : A compressed observation vector that integrates laser, depth, and illumination information; linear weighting coefficients : Ensure that the three information streams are adaptively adjusted according to the light sensitivity.

[0036] When the light intensity fluctuates drastically, the light sensitivity weighting Increase the illumination feature vector Increasing the weight of the feature vector can explicitly remind subsequent algorithms to pay attention to the uncertainty of illumination in ranging; conversely, decreasing the weight of the feature vector can compress the feature vector. leading.

[0037] After completing the time-space-radial triple alignment in step 101, step 102 fuses the three observations into a multimodal raw ranging signature using J-PCA compression and illumination sensitivity weighting mechanisms. .

[0038] By utilizing joint principal component compression and illumination sensitivity-guided linear fusion, multiple high-dimensional observations are reduced into a compact ranging signature. This reduces storage and transmission burden while retaining the feature vectors that are most decisive for depth drift, enabling subsequent algorithms to complete real-time processing under conditions of abundant information without overloading.

[0039] Step one employs an interdisciplinary synchronous acquisition and fusion strategy to rigorously integrate three types of heterogeneous data: laser echo, binocular parallax, and illumination variation, forming a compressed yet informative multimodal raw ranging signature. This signature semantically links the microscopic geometric features of the contaminated land surface with the light disturbance under vegetation cover, providing an input with a light sensitivity label for subsequent confidence weight generation and depth steady-state correction, so that the ranging chain can still maintain low error gain diffusion even when the environment changes drastically.

[0040] For risk assessment of farmland contaminated with chromium and nickel, any inference of depth-enrichment relationships depends on a precise grasp of the reliability of distance measurements; the aforementioned steps have generated multimodal raw ranging signatures. However, it still contains unstable components caused by laser scattering, canopy transmission, and light flicker. If these unstable signals are not identified and their effects are not quantified beforehand, subsequent deep steady-state corrections will accumulate errors step by step, ultimately weakening the spatial resolution of high-risk areas for heavy metals.

[0041] Step 2: Signing the original multimodal ranging measurement Based on this, real-time reliability weights are output through two-dimensional reliability quantization and adaptive fusion. And based on this, a weighted depth candidate set is generated. This provides weighted priors for subsequent deep steady-state correction processes.

[0042] Step 201: Confidence Measurement in Two Dimensions Multipath scattering of laser signals in vegetation gaps leads to echo truncation, while high-frequency canopy texture disrupts stereo matching consistency. By separately assessing echo integrity and texture complexity, the two main causes of ranging drift can be analyzed at the information level. Therefore, entropy spectral analysis is used to quantify the integrity score of laser echoes. Visual texture complexity scores are extracted using fractal Laplacian dimensions. By mapping the two to the zero-to-one range using a unified scale, a comparable baseline is established for weight fusion.

[0043] In an aligned and radiation-normalized laser amplitude sequence In laser beams, truncation and multipath reflection lead to non-uniform energy distribution in the frequency domain. First, consider the laser amplitude sequence... The amplitude spectrum is obtained by performing a discrete Fourier transform. Calculate the normalized spectral energy Then, spectral entropy is used to characterize signal uncertainty, and subsequently, integrity is measured in reverse: ; Where: Normalized spectral energy : No. The ratio of the energy of each frequency component to the total energy. ; Total number of frequency components The number of sampling points in the Discrete Fourier Transform is an integer power of two; the echo integrity score. The upper limit of 1 represents concentrated spectral energy and a complete echo, while the lower limit of zero represents dispersed energy and a truncated echo.

[0044] Low spectral entropy → Approaching zero → A value close to 1 indicates a continuous echo waveform; conversely, a value far from 1 suggests truncation, requiring a reduction in subsequent weights. This applies to the radiometrically corrected disparity depth matrix. Extracting the Laplace response Then, the local fractal dimension is estimated using box counting. Texture complexity score Defined as a dimension normalization value: ; Where: Local fractal dimension The dimension estimated using box counting slope ranges from 2 to 3; Minimum Dimension Maximum dimension The theoretical extreme value of visual texture complexity determined by modeling; Texture complexity score Zero indicates a simple texture that is easy to register, while one indicates an extremely complex texture that is easy to mismatch.

[0045] Texture complexity score Corresponding to visual error sensitivity, high texture complexity score The region needs to rely more on laser ranging, with low texture complexity scores. The visual depth contribution from a region is more reliable.

[0046] Step 201: Output echo integrity score Texture complexity score Both have been mapped to 0 to 1, providing dimensionally consistent input for the next step of weight fusion.

[0047] By measuring echo integrity inversely using spectral entropy and characterizing texture complexity using fractal dimensions, this step transforms the two error sources, laser and vision, into dimensional fractions, truly quantifying ranging reliability from both the signal and scene layers. This lays a comparable and continuous foundation for subsequent weight fusion and outlier suppression.

[0048] Step 202: Reliability weight fusion and weighted depth candidate generation A single score is insufficient to fully reflect the reliability of ranging; the echo integrity score must also be considered. Texture complexity score Adaptive fusion based on scene characteristics generates reliable weights that can directly drive deep candidate filtering. .

[0049] First, the information source weight system is automatically estimated using the Dirichlet-Gamma Transform. Subsequently, the reliability weights were adjusted using kernel decay mapping. Applied to the initial depth candidate set Output weighted depth candidate set .

[0050] Based on the multi-source fusion theory, echo integrity scores are used in an unsupervised manner. Texture complexity score Post-test weighting of mother samples Its negative log-likelihood is determined by the gamma function. Combining the logarithms of the two, setting the gradient to zero, we obtain a closed-form solution: ; ; Where: Dirichlet's prior total strength Initial hyperparameters, positive real numbers; second-order trigonometric functions The second logarithmic derivative of the Dirichlet distribution (Digamma function); inverse function : The range of values ​​is positive; single-source weight system : Corresponds to the importance of laser and visual information sources, respectively, positive real numbers; reliability weight : Real numbers that combine two source fractions, normalized to 0 to 1.

[0051] When echo integrity score Or texture complexity score In extremely poor cases, the corresponding single-source weight system Rapidly reduce the reliability weight Automatically biased towards highly reliable sources.

[0052] Initial depth candidate set Self-multimodal raw ranging signature Decoding yields multiple depth estimates. Reliability weighting Construct a kernel decay function for the core. This forms a weighted depth candidate set. : ; Where: weighted depth candidate set The depth set after being weighted by both reliability weight and kernel decay function; nuclear decay function A function that exponentially suppresses the weights of candidate depths based on their outlier tendencies; average depth : Mid-depth mean; single depth candidate value The first one obtained by multimodal ranging signature decoding Original depth estimate; distance attenuation coefficient : A positive real number that controls the decay rate of the kernel function; power exponent : A positive real number that determines the shape of the decay curve.

[0053] Nuclear decay mapping in reliability weights Based on this, the outlier depth is further suppressed so that the final candidate set both adapts to the reliability of the information source and remains sensitive to extreme values.

[0054] Step 202 Output Reliability Weights and weighted depth candidate set Reliability weight The weighted depth candidate set will be dynamically updated in real time based on lighting, echoes, and textures. It provides weighted inputs for deep residual learning, enabling a smooth transfer of parameters and data between steps.

[0055] By leveraging the Dirichlet adaptive weight system and exponential kernel decay, this step maps double scores to real-time reliability weights and dynamically rearranges all depth candidates, prioritizing high-confidence pixels for subsequent learning processes while subjecting low-confidence pixels to flexible suppression, thereby suppressing the perturbation of network gradients by outliers from the source.

[0056] Through the two-layer processing in step two, the multimodal raw ranging signature has been obtained. Echo integrity score was extracted Texture complexity score The reliability weights are generated through adaptive fusion of the Dirichlet weight system. Then, a weighted depth candidate set is constructed using kernel decay mapping. Reliability weight The reliability gradient of ranging data in the current light-vegetation hybrid scene was quantified, and the weighted depth candidate set was... Using this weight as an anchor, depth candidates are filtered and rearranged, significantly reducing the interference of outlier depths on the residual learning network. Therefore, the weighted depth candidate set can be directly used in subsequent step three. As a weighted input, and with reliability weights To enhance the steady-state correction capability in dynamic environments, the network loss function is adaptively modulated based on the prior strength. This step seamlessly integrates with the previous synchronous data acquisition and signature construction, realizing a complete chain from data generation to reliability measurement and candidate selection, thus laying a solid foundation for the depth, accuracy, and stability of the risk assessment system for agricultural products in areas with high chromium and nickel geological backgrounds.

[0057] In rice paddies and vegetable fields contaminated with chromium and nickel, laser-visual collaborative ranging is often interfered with by multiple dynamic factors such as plant swaying, wind-induced shadows, and drone vibration. These disturbances exhibit a complex spectrum of short-period high-frequency and medium-period low-frequency overlaps over time. If only a static weighted depth candidate set is used... A one-time selection cannot fully absorb the continuous motion constraints of the crop canopy between adjacent frames, nor can it gradually reduce the accumulated potential energy of outlier depth on the time axis, resulting in the output depth still jumping when there is a sudden drop in light or local reflection.

[0058] Step 3 aims to utilize short-term motion coherence information With residual learning networks The synergistic effect on the weighted depth candidate set Temporal consistency filtering and gradual steady-state correction are performed to finally output a steady-state depth sequence. .

[0059] Step 301: Motion Consistency-Driven Candidate Depth Reassignment Weighted depth candidate set Reliability weights have been applied. The initial rearrangement using outlier distance is completed; however, the set still lacks temporal constraints on the small translational and rotational displacements caused by the plants swaying in the wind. Without introducing motion consistency correction, high-confidence depths may exhibit abrupt jumps in adjacent frames, thus misleading the gradient direction of the residual learning network.

[0060] In two consecutive disparity-amplitude aligned images, the coarse motion vector field is first estimated using the sparse feature point pyramid-optical flow method. Subsequently, based on minimizing the robust kernel The estimation criterion eliminates spurious matches caused by illumination flicker, yielding a coherent motion field vector. In order to convert the motion coherence field vector Translate into depth domain displacement and construct the motion embedding matrix. : ; And the motion embedding matrix Normalization is performed to form short-term motion coherence information. Motion embedding matrix : A matrix that performs an exponential decay mapping on the degree of motion coherence; Where: motion vector field : The set of pixel displacement vectors predicted by pyramid-optical flow; motion coherence field vectors : Robust core Estimate the effective motion vectors after filtering out false matches; Mean motion vector : Spatial average; attenuation coefficient : Positive real number that modulates the sensitivity to coherence; short-term motion coherence information : Normalized time-consistent scalar field, with values ​​ranging from zero to one.

[0061] Motion Embedding Matrix A larger value indicates a smaller difference between local motion and the overall coherent field, and the corresponding depth should obtain higher temporal reliability in subsequent weighting. This yields short-term motion coherence information. Then, it is combined with the reliability weight. Coupled to form a time-series weighting factor :

[0062] And update the candidate depth set accordingly. : ; Where: time-series weighting factor The weights for integrating reliability and motion coherence range from zero to one. Numerical stability term : Prevent small positive numbers with a denominator of zero; information on motion continuity The short-time motion coherent field obtained by optical flow-robust estimation is mapped to the first... The closer the value after a pixel is to 1, the more consistent the movement direction is with the surrounding pixels. Updated candidate depth set : The deep set of objects into which motion consistency weights have been injected.

[0063] When a region has complex lighting but continuous motion, the time-weighted factor... Due to short-term continuous motion information A higher weight and a larger value will prevent the depth of the region from being excessively decayed; conversely, strong outlier movements will reduce the weights and reduce interference with the residual network.

[0064] Step 301 weighted depth candidate set Information related to short-term motion Combine and output the updated candidate depth set after rebalancing. And simultaneously retain the time-series weighting factor. As an explicit attentional distribution for residual learning, it provides temporally consistent priors for the next step.

[0065] By constructing a motion coherence field through optical flow-robust estimation and then fusing reliability information through weight redistribution, this step ensures that depth candidates are not only controlled in the static quality dimension but also constrained in the temporal coherence dimension. This prevents jumps in depth caused by instantaneous wind or slight shaking of the aircraft from entering the learning channel, thereby improving sequence stability.

[0066] Step 302: Residual Learning and Steady-State Depth Asymptotic Correction Despite the updated candidate depth set While motion consistency has been incorporated, point outlier errors may still occur during sudden increases in wind speed or minor equipment vibrations. To obtain a steady-state depth sequence that can sustainably track the three-dimensional structure of vegetation-soil-water, a learning unit is needed that can use historical sequences as memory and residuals as feedback to update the candidate depth set. Multiple rounds of iterative corrections were performed.

[0067] The network input is a weight-depth pair. In the In each iteration, the prediction depth is calculated. Compared to the previous round of output depth residuals: ; And using the residual gating function: ; Adjusting the cumulative residual: ; Where: pixel index , mark Depth entries for each pixel or voxel; Iteration rounds The current update round of the residual learning network; Predicted depth : In the The depth estimate given by the first round, and the first round obtained by the network forward inference. A depth estimate; This round of updates depth : No. The depth value after the wheel residuals are accumulated, and the result after adding the gated residuals to the depth of the previous wheel; Previous output depth , No. The depth value at the end of the round iteration; Weighted residuals The difference between the predicted depth and the previous depth, multiplied by the time-series weighting factor, is used as an adjustment amount. Residual Gating Function : By the Sigmoid function The dynamic coefficients constituted range from 0 to 1; Trainable parameters : Control gating sensitivity and offset; Time-series weighting factor Mapping this onto the residuals allows reliability and motion coherence to directly influence the gradient magnitude; the gating function ensures that significant adjustments are only made when the residuals are significant and reliable, avoiding overfitting. To prevent endless iterations, theoretical constraints are needed on the convergence of the deep sequence, introducing a Lyapunov candidate function. : ; If it exists make: ; The current depth sequence is considered to have met the steady-state criterion, and the output is:

[0068] Otherwise, continue iterating until the condition is met or the maximum number of rounds is reached. .

[0069] Where: Lyapunov candidate function : Measure the depth sequence at the th The dispersion of the wheel; Then it is the first Wheel dispersion; average depth Mean depth of the current round; convergence coefficient The relaxation ratio is set, with a value between 0 and 1; Maximum round Positive integer thresholds to prevent excessive iterations; steady-state depth sequence : Depth output that satisfies the Lyapunov steady-state criterion.

[0070] Lyapunov relaxation mapping guarantees that the variance of the deep sequence converges geometrically, theoretically preventing the jitter-compensation-overcompensation cycle and making the output stable and reliable.

[0071] Step 302 uses a combination of gated residual accumulation and Lyapunov relaxation to redistribute candidates. Transform into a steady-state depth sequence At the same time, retain This provides data and indicators for the next step of inferring the uncertainty range, with consistent parameter signs throughout, thus providing data for subsequent information writing back.

[0072] The residual learning network progressively corrects the depth sequence under the dual strategies of gated accumulation and Lyapunov relaxation, ensuring that the error converges monotonically within a finite number of rounds. Compared with traditional one-time filtering, this mechanism can continuously track environmental changes and adjust the step size in real time, so that the depth output remains smooth and reliable in dynamic scenes.

[0073] Step three employs a two-stage process of motion consistency-driven redistribution and residual learning-Lyapunov steady-state correction to eliminate the temporal perturbations of the dynamic field environment on depth estimation, successfully outputting a steady-state depth sequence. Among them, the time-series weighting factor The gradient is reliably adjusted within the residual channel, while the Lyapunov criterion provides a rigorous theoretical guarantee for convergence. This steady-state depth sequence not only spatially matches the micro-topographical height differences in high-risk chromium-nickel pollution areas, but also temporally maintains a smooth response to plant swaying and light flashdown. This lays a solid foundation for uncertainty quantification and reliability weight rewriting in step four, enabling the entire risk assessment system to maintain both high accuracy and high stability even in complex scenarios.

[0074] In field measurements of farmland contaminated with chromium and nickel, the laser-vision chain has obtained a steady-state depth sequence through the first three steps. and real-time reliability weight However, to ensure that risk assessment results truly serve field-specific policy implementation, two key questions must be addressed: First, steady-state depth is still affected by random factors such as plant posture, wind speed, and machine vibration. How can a reliable uncertainty range be assigned to each pixel-level distance value so that the subsequent heavy metal enrichment and health risk model can dynamically adjust the confidence bandwidth? Second, the field environment changes rapidly on a minute-by-minute scale in terms of solar radiation angle, cloud cover thickness, and wind field distribution. If the parameters at the data acquisition end cannot be adjusted in real time according to reliability weights... The feedback will cause the entire link to fall back into signal-to-noise imbalance in the next cycle.

[0075] Step 4: Through two steps—uncertainty quantification and adaptive parameter write-back—a steady-state depth sequence is achieved. The reliability boundary is output, and the acquisition unit is driven to determine the reliability weight in real time. Dynamically adjust environmental adaptation parameters.

[0076] Step 401: Uncertainty Measurement and Credibility Boundary Generation Although the Lyapunov relaxation mapping has led to geometric convergence of the variance of the depth sequence, discrete point noise and spatial correlation errors have not been completely eliminated. If single-valued depths are directly input into the crop enrichment-health risk model, the confidence interval of the parameter estimates will be amplified, causing fluctuations in the high-risk threshold judgment.

[0077] First, the statistical analysis of the gated residual accumulation mechanism in the convergence rounds is performed. residual column vector Treating it as a floating distribution with unknown mean and unknown scale, we introduce a priori information: ; Through conjugate derivation, the posterior distribution is: ; and posterior mean With scale Calculate the confidence interval of pixel residuals

[0078] ; Where: prior mean Historical residual expectation; population mean of residuals Random variables to be inferred; population variance of residuals Random variables to be inferred; prior sample size : Positive real number, representing the strength of the prior; prior parameter The shape and scale of the inverse gamma prior are controlled by positive real numbers; degrees of freedom : Posterior Student-t distribution degrees of freedom, positive real numbers; posterior mean : Posterior expectation of residuals; posterior metric : Posterior standard deviation of residuals; quantiles Degrees of freedom are Confidence level is Student-t quantiles; Bayesian confidence intervals : The interval of quantization residual uncertainty.

[0079] Compared to commonly used variance estimation, Student-t posterior is naturally robust to heavy-tailed noise and can accurately capture sporadic high-energy errors. Residual distribution alone cannot reveal spatially correlated noise; Gaussian processes are needed to analyze steady-state depth sequences. Spatial error surface compensation is performed. The input position vector is defined. Depth error Modeled as a zero-mean Gaussian process: ; Wherein the kernel function is: ; And introduce heteroscedasticity modulation term ; Finally, the pixel-level spatial uncertainty is obtained: ; in To predict the covariance matrix, the covariance is calculated based on the kernel function for the test points.

[0080] Depth Spatial Residual In position The random depth error at that point is considered as a zero-mean Gaussian process sample. Zero-mean Gaussian process , with kernel function A random field model with covariance; Signal variance The overall amplitude of the process is determined and adaptively adjusted according to the light sensitivity weight; the baseline variance... : Initial signal variance when illumination is stable; Model initial signal variance; Scale matrix The diagonal elements are the square length scales in each direction, controlling the spatial correlation decay rate; position vector : The two-dimensional or three-dimensional position of the corresponding pixel in the ground coordinate system; kernel parameter vector : Controlling the smoothness and noise of the Gaussian process; distance vector The difference between the two positions in the ground coordinate system; light sensitivity weight. : Defined in step one; Kronecker symbol : 1 is taken when the positions completely overlap, otherwise 0 is taken; Spatial uncertainty : The square root of the variance predicted by a Gaussian process; By using a heteroscedasticity modulation term to directly amplify or reduce the nuclear amplitude due to illumination fluctuations, illumination-spatial error coupling compensation is achieved, and the Bayesian confidence interval is extended. Spatial uncertainty Linear fusion: ; Obtain the final trust boundary , where the fusion coefficient The residual fat-tail shape is adaptively set by the maximum entropy criterion. Final credible boundary. With steady-state depth sequence A distance-uncertainty binary output is generated to set a pixel-level confidence bandwidth for the chromium-nickel enrichment model.

[0081] By combining Bayesian residual modeling with spatial heteroscedastic Gaussian process, interpretable uncertainty boundaries are generated for each pixel. This enables subsequent chromium-nickel enrichment models to flexibly adjust thresholds based on confidence bandwidth, avoiding misjudgments of high-risk areas due to the invisibility of the underlying layer caused by depth errors, thus making regulatory decisions more transparent.

[0082] Step 402: Reliability weight-driven adaptive acquisition parameters Real-time reliability weight It is an indicator that integrates multiple factors such as ambient light, texture complexity, and echo cutoff. If it is fed back to the acquisition end in a timely manner, the laser pulse power, camera exposure, and frame rate can be finely controlled to ensure the optimal performance in the next sampling cycle. and The system remains within a controllable threshold. This step proposes two key techniques: weight-parameter mapping and feedforward-backward coupling, to achieve reliability-driven environmental adaptation.

[0083] Establish a linear-piecewise mapping: ; in: , ; And set up saturation value constraints: ; Where: laser pulse power Adjusting the laser output energy; camera exposure time. Controls visual image brightness; captures frame rate. Determines time resolution; mapping matrix Weight-parameter gain vector; , Gain coefficient; bias vector Baseline parameters; , As the reference bias; Upper and lower limits , , , , , Hardware security or quality control thresholds.

[0084] When reliability weight Decrease, automatically increase laser pulse power And shorten the camera exposure time This improves the laser signal-to-noise ratio and suppresses dynamic blur, while conversely reducing power consumption and the risk of overexposure.

[0085] To prevent sudden jumps caused by mapping, a two-layer feedforward-backward adjustment loop is designed. In the feedforward stage, parameters are updated according to the previous formula; in the backward adjustment stage, the synchronization error is monitored. With energy compensation factor If overshoot occurs, adjust according to the sliding mold adjustment law: ; Where: error vector .

[0086] Synchronization Error Reference Target time consistency; energy compensation benchmark Target energy compensation factor; gain matrix Positive definite diagonal matrix, determines the adjustment rate; sign function Element-level symbolic retrieval function; parameter increment : The fine-tuning vector applied to the acquisition parameters.

[0087] By adjusting the sliding mode, the system can quickly return to the target acquisition state area, while ensuring that parameter changes are within the range that the hardware can withstand.

[0088] Step 402 outputs updated collected parameters and together with the real-time reliability weight The data is written into the acquisition node configuration table to provide prior information for synchronous acquisition and radiometric normalization in the next cycle, thus completing the output-writeback connection. Through weight-parameter mapping and feedforward-back-adjustment coupling, this step feeds back real-time reliability to laser power, camera exposure, and frame rate, forming a three-loop adaptive update chain of ranging, evaluation, and acquisition. The acquisition end adjusts the signal-to-noise ratio and dynamic range in real time according to environmental fluctuations, continuing to provide high-quality input for data fusion in the next cycle.

[0089] Step four involves generating pixel-level reliable boundaries through Bayesian-Gaussian dual modeling. This allows the deep-enriched-risk model to obtain uncertainty constraints at the input layer; on the other hand, by leveraging weight-parameter mapping and feedforward-back-adjustment coupling, the real-time reliability weights are... The system rapidly writes back the three core acquisition parameters—laser power, camera exposure, and frame rate—ensuring the acquisition-fusion link remains in optimal signal-to-noise and synchronization condition for the next sampling cycle. This achieves reliable measurement and environmentally adaptive two-way quantization, providing interpretable depth-confidence combination data for risk assessment of agricultural products in high-chromium-nickel geological background areas. Furthermore, the parameter update mechanism maintains the long-term stability of the ranging chain, providing solid data and methodological support for subsequent regulatory decisions and agronomic interventions.

[0090] Please see Figure 2 This invention provides a risk assessment system for agricultural products in areas with high chromium and nickel geological backgrounds, including, The data acquisition module synchronously acquires the laser echo sequence and the binocular parallax image within the same time window, records the illumination change characteristics, achieves triple alignment based on timestamp, radiance and spatial reprojection, generates a multimodal original ranging signature and marks the frame number before outputting it; The weight fusion module calculates the real-time reliability weight based on the echo integrity score and the texture complexity score, performs kernel attenuation weighting on multiple depth candidates, constructs a weighted depth candidate set and outputs the corresponding index matrix to form a weight mapping table for subsequent use. The deep steady-state module inputs the weighted depth candidate set and the short-term motion coherence information of the target into the residual learning network, and iteratively corrects it through gated residual accumulation and Lyapunov relaxation mapping, outputting a steady-state depth sequence and simultaneously generating a residual variance index. The adaptive feedback module generates pixel uncertainty boundaries based on the steady-state depth sequence and the residual variance, inputs the distance-uncertainty pair into the heavy metal risk model, and adjusts the laser power, exposure time and frame rate according to the real-time reliability weight to update the environmental adaptation parameters for the next sampling period.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0094] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0095] 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 method for assessing the risk of agricultural products in a high geological background area of chromium and nickel, characterized in that: comprising, Synchronously collect the laser echo sequence and the binocular parallax image within the same time window, record the illumination change characteristics, realize triple alignment based on timestamps, radiance and spatial re-projection, generate multi-modal original ranging signatures and output after marking frame numbers; Calculate real-time reliability weights according to the echo integrity score and the texture complexity score, perform kernel decay weighting on multiple groups of depth candidates, construct a weighted depth candidate set and output a corresponding index matrix to form a weight mapping table for subsequent calling; Jointly input the weighted depth candidate set and the target short-time motion coherence information into a residual learning network, output a steady-state depth sequence and synchronously generate a residual variance index through gated residual accumulation and Lyapunov relaxation mapping iteration correction. Based on the steady-state depth sequence and the residual variance, generate a pixel uncertainty boundary, input the distance-uncertainty pair into a heavy metal risk model, and adjust the laser power, exposure time and frame rate according to the real-time reliability weight to update the environmental adaptation parameters in the next sampling period.

2. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 1, characterized in that: The length of the observation window is twice the inverse of the target plant trunk swing frequency, and the laser emission and left and right camera exposure are triggered simultaneously through a high-precision time unit; Real-time calculation of the time consistency function monitors the difference between the laser sampling time and the visual sampling time, and when the difference exceeds the threshold value, the corresponding frame is discarded to ensure that the time calibration of the multi-modal original ranging signature is consistent.

3. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 2, characterized in that: The construction of the multi-modal original ranging signature includes: performing joint principal component analysis on the laser amplitude sequence after radiance normalization and the parallax depth matrix; retaining principal components with cumulative contribution rate not less than 80% to generate compressed vectors, and then linearly fusing the illumination feature vectors obtained through time series filtering according to the illumination sensitivity weight to output the signature.

4. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 3, characterized in that: The echo integrity score is calculated by inversely scaling the spectral entropy of the normalized spectral energy, and the texture complexity score is obtained by normalizing the local fractal dimension estimated by performing box counting method on the corrected parallax depth matrix, and finally output in the range of zero to one.

5. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 4, characterized in that: The calculation of the real-time reliability weight includes: inputting the echo integrity score and the texture complexity score into a Dirichlet prior model, adaptively obtaining a single-source weight system according to the parent sample variance, and then using an exponential kernel decay function to weight the outlying degree of the initial depth candidate to obtain the weighted depth candidate set.

6. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 5, characterized in that: The generation of the target short-time motion coherence information comprises: estimating a coarse motion vector field by using a pyramid optical flow algorithm, adopting The estimated motion coherence field vector is obtained by eliminating false matching, and an exponential decay mapping is implemented by the Euclidean distance between the motion coherence field vector and the average vector of the whole field to generate a motion embedding matrix, and the coherence information is output after normalization.

7. The chromium-nickel high geological background area agricultural product risk assessment method according to claim 6, characterized in that: The residual learning network is composed of two layers of gated recurrent units and one layer of fully connected layers, adopts a gated residual accumulation method for multiple rounds of depth update, and stops iteration when a Lyapunov candidate function converges to a preset proportion compared with the previous round, and outputs the steady-state depth sequence.

8. The method according to claim 7, wherein the method is used for the risk assessment of agricultural products in a high geological background area of Cr and Ni. The generation of the pixel uncertainty boundary is based on a residual vector to establish a conjugate Bayesian model with unknown mean and unknown scale to obtain a Student-t posterior confidence interval, then a spatial heteroscedastic Gaussian process is used to model the depth error surface, and finally the residual uncertainty and spatial uncertainty are fused according to the maximum entropy criterion.

9. The method according to claim 8, wherein the method is used for the risk assessment of agricultural products in a high geological background area of Cr and Ni. The adjustment of the acquisition parameters first maps the real-time reliability weight to the reference values of laser power, exposure time and acquisition frame rate by using a calibrated linear gain matrix, then monitors the time consistency error and energy compensation factor during the sampling process, and incrementally corrects the parameters according to the sliding mode adjustment law.

10. A system for assessing the risk of agricultural products in a high geological background area of chromium-nickel, characterized by: The data acquisition module synchronously acquires the laser echo sequence and the binocular parallax image within the same time window, records the illumination change characteristics, realizes triple alignment based on timestamp, radiance and spatial re-projection, generates a multi-modal original ranging signature and marks the frame number, and then outputs; The weight fusion module calculates the real-time reliability weight according to the echo integrity score and the texture complexity score, performs kernel attenuation weighting on multiple groups of depth candidates, constructs a weighted depth candidate set and outputs the corresponding index matrix to form a weight mapping table for subsequent calling; The depth steady-state module inputs the weighted depth candidate set and the target short-time motion coherence information into a residual learning network, iteratively corrects through gated residual accumulation and Lyapunov relaxation mapping, outputs a steady-state depth sequence, and synchronously generates a residual variance index; The adaptive feedback module generates a pixel uncertainty boundary based on the steady-state depth sequence and the residual variance, inputs the distance-uncertainty pair into a heavy metal risk model, adjusts the laser power, exposure time and frame rate according to the real-time reliability weight, and updates the environmental adaptation parameters for the next sampling period. ​

Images