A farmland soil compaction diagnosis method based on multi-source sensor data

Abstract

The present application relates to the field of data analysis, and more particularly to a farmland soil compaction diagnosis method based on multi-source sensor data, which comprises: obtaining a gradient energy phase locking factor by gradient energy feature construction and phase alignment cross-correlation evaluation on soil resistance depth curve data; obtaining a rheological morphology coupling factor by rheological consistency coupling evaluation on soil resistance depth curve morphology features and soil volume water content; obtaining an adaptive similarity distance by joint correction on the basic Euclidean distance and performing clustering iteration; obtaining a soil compaction accurate diagnosis map and generating a differentiated deep loosening operation instruction by spatial modeling and prescription mapping on the clustering results and sampling point position information, so as to solve the problem that the existing Euclidean distance is easy to cause compaction layer feature misplacement and rheological strength distortion under the conditions of farmland micro-terrain fluctuation and water content difference, resulting in clustering partition distortion.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and in particular to a method for diagnosing farmland soil compaction based on multi-source sensor data. Background Technology

[0002] In modern precision agriculture, repeated tillage operations by machinery can easily lead to varying degrees of soil compaction in the topsoil and plow pan. Soil compaction reduces porosity, aeration, and permeability, hindering root development, restricting nutrient and water transport, and further causing problems such as waterlogging after rain, water shortage during droughts, and decreased plant resistance, ultimately affecting crop yield and quality. Therefore, in field management, spatial and zonal diagnosis of soil compaction status, and the development of differentiated deep tillage or tillage prescriptions accordingly, is a crucial foundation for improving topsoil structure, reducing energy consumption from ineffective operations, and achieving refined land preparation and management. Current technologies typically employ hydraulic cone penetrators equipped with GPS, depth measurement devices, and pressure measurement devices to obtain objective data on soil compaction. These penetrators sample the soil along a grid path or equidistant route within the field, obtaining curves showing the change in soil cone penetration resistance with depth at each sampling point, while simultaneously collecting environmental data such as soil volumetric moisture content. Due to the large number of field sampling points and the high dimensionality of profile curves, traditional manual interpretation methods are inefficient and highly subjective. Existing technologies often introduce unsupervised clustering algorithms to automatically partition and manage the sampling data. Among them, clustering methods based on distance metrics are more common. For example, clustering algorithms are used to calculate the similarity of soil resistance depth curves at different sampling points, and similar curves are grouped into the same cluster, thereby forming different compaction levels or management zones. Then, combined with the geographical location information of the sampling points, a compaction diagnostic map is generated to guide deep tillage operations.

[0003] In the aforementioned clustering process, the method of calculating the similarity between sampling points directly determines the reliability of the partitioning results. Existing technologies generally use Euclidean distance as the basic distance metric, which involves summing the squares of the resistance values ​​of two soil resistance depth curves at the same sampling depth location and taking the square root to measure the degree of curve difference. This distance is then used as the core basis for sample assignment and clustering iteration. However, farmland environments exhibit significant unstructured characteristics. Soil resistance depth curves are affected by both topographic relief and moisture content variations, leading to simultaneous drift and distortion on both the depth and amplitude axes. In real-world scenarios, surface micro-topography such as ruts and furrows can cause inconsistencies in the zero-point reference for depth at different sampling points, resulting in misaligned depth positions for the same physically compacted layer in the curves at different sampling points. Simultaneously, as a porous medium, soil exhibits significant rheological characteristics in its mechanical response with varying moisture content. The same compacted structure often shows a high and sharp resistance peak under dry conditions, while under moist conditions, the peak value may decrease but the shape broadens. Because Euclidean distance requires strict alignment of features at depth and measures only the absolute difference in numerical intensity, it can easily misclassify samples with "only depth misalignment but consistent structure" as having significant differences. It can also easily classify samples with "variable amplitude and morphology due to moisture influence but consistent compaction structure" into different categories. This leads to deviations between the generated compaction zoning map and the actual physical state in the field, resulting in problems such as missed tillage or over-tillage. Therefore, how to construct a similarity measurement mechanism more adapted to the unstructured environmental characteristics of farmland under multi-source sensor data conditions to improve the accuracy and stability of clustering and zoning remains a pressing technical problem to be solved in this field. Summary of the Invention

[0004] In view of this, the present invention aims to propose a farmland soil compaction diagnosis method based on multi-source sensor data, in order to solve the problem that the existing Euclidean distance is prone to causing misalignment of compaction layer features and distortion of rheological intensity under farmland micro-topographic undulation and moisture content differences, resulting in clustering and partitioning distortion.

[0005] To achieve the above objectives, the technical solution of the present invention is implemented as follows:

[0006] A method for diagnosing farmland soil compaction based on multi-source sensor data, the method comprising:

[0007] Step S1: Obtain a multidimensional feature dataset by collecting and standardizing the raw data from multiple sources of sensors in farmland;

[0008] Step S2: Obtain the gradient energy phase locking factor by constructing gradient energy characteristics and evaluating phase alignment cross-correlation of soil resistance depth curve data;

[0009] Step S3: Obtain the rheological morphology coupling factor by performing a rheological consistency coupling evaluation on the morphological characteristics of the soil resistance depth curve and the soil volumetric water content;

[0010] Step S4: By jointly correcting the basic Euclidean distance using the gradient energy phase locking factor and the rheological morphology coupling factor, an adaptive similarity distance is obtained and clustering iteration is performed;

[0011] Step S5: By performing spatial modeling and prescription mapping on the clustering results and sampling point location information, a precise diagnosis map of soil compaction is obtained and differentiated deep loosening operation instructions are generated.

[0012] Furthermore, the process of collecting and standardizing the raw data from multiple sources of farmland sensors to obtain a multi-dimensional feature dataset includes:

[0013] Using an agricultural machinery platform equipped with a high-precision global positioning system, a hydraulic cone penetrator, and a time-domain reflectometer, continuous sampling is performed within the target farmland area according to a preset grid path. For each sampling point, the location information of the sampling point, soil cone penetration resistance sequence data varying with depth, sampling depth sequence data corresponding to the soil cone penetration resistance sequence data, and soil volumetric moisture content data at the sampling point location are simultaneously acquired. The soil cone penetration resistance sequence data covers a vertical profile from the surface to the preset depth. After completing the raw data acquisition from the multi-source sensors in the farmland, the soil... Outlier removal was performed on the soil cone penetration resistance sequence data and soil volumetric moisture content data. The outliers included abrupt outliers caused by poor sensor contact or stone collisions. The soil cone penetration resistance sequence data and soil volumetric moisture content data after outlier removal were then standardized and mapped to eliminate the influence of different sensor dimensions. The standardized and mapped soil cone penetration resistance sequence data, sampling depth sequence data, soil volumetric moisture content data, and location information were combined to obtain a multidimensional feature dataset containing location information, soil resistance depth curves, and soil volumetric moisture content.

[0014] Furthermore, the step of constructing gradient energy characteristics and evaluating phase alignment cross-correlation on soil resistance depth curve data to obtain gradient energy phase locking factors includes:

[0015] By performing differential gradient calculation and gradient squared weighted coupling processing on the soil resistance depth curve data of the sampling points, gradient energy operator sequence data and feature layer energy centroid depth data are obtained.

[0016] The gradient energy phase locking factor is obtained by performing displacement alignment cross-correlation and normalization mapping on the gradient energy operator sequence data between sampling points based on the offset of the energy centroid of the feature layer.

[0017] Furthermore, the step of obtaining gradient energy operator sequence data and feature layer energy centroid depth data by performing differential gradient calculation and gradient squared weighted coupling processing on the soil resistance depth curve data of the sampling points includes:

[0018] For any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions at each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data. First-order difference calculations are performed on the soil cone penetration resistance sequence data corresponding to adjacent depth sampling positions of the target sampling point to obtain the resistance differential gradient data at each depth sampling position of the target sampling point. Multiplicative coupling operations are performed on the squares of the soil cone penetration resistance sequence data and the resistance differential gradient data at each depth sampling position of the target sampling point to obtain the gradient energy operator sequence data corresponding to the target sampling point. The gradient energy operator sequence data and the corresponding sampling depth sequence data are weighted and summed, and the ratio of the weighted summation result to the cumulative summation result of the gradient energy operator sequence data is normalized to obtain the feature layer energy centroid depth data corresponding to the target sampling point.

[0019] Furthermore, the step of obtaining the gradient energy phase locking factor by performing displacement-aligned cross-correlation and normalization mapping processing based on the feature layer energy centroid offset on the gradient energy operator sequence data between sampling points includes:

[0020] For any first sampling point and second sampling point, extract the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the first sampling point, and the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the second sampling point from the multidimensional feature dataset, respectively.

[0021] The absolute value of the difference between the feature layer energy centroid depth data of the first sampling point and the feature layer energy centroid depth data of the second sampling point is calculated to obtain the feature layer energy centroid offset data corresponding to the first sampling point and the second sampling point.

[0022] For each depth sampling position in the sampled depth sequence data, the gradient energy operator value of the first sampling point at that depth sampling position is multiplied by the gradient energy operator value of the second sampling point at the depth sampling position corresponding to the depth sampling position minus the feature layer energy centroid offset data. The product results of all depth sampling positions are then summed to obtain the cross-correlation multiplication accumulation evaluation value. The gradient energy operator values ​​of the gradient energy operator sequence data of the first sampling point at all depth sampling positions are then summed by squares to obtain the first autocorrelation energy evaluation value. The gradient energy operator values ​​of the gradient energy operator sequence data of the second sampling point at all depth sampling positions are then summed by squares to obtain the second autocorrelation energy evaluation value.

[0023] The normalized cross-correlation evaluation value is obtained by dividing the sum of the cross-correlation multiplication evaluation values ​​by the square root of the product of the first autocorrelation energy evaluation value and the second autocorrelation energy evaluation value; the difference between the constant and the normalized cross-correlation evaluation value is used as the gradient energy phase locking factor corresponding to the first sampling point and the second sampling point.

[0024] Furthermore, the method of obtaining the rheological morphology coupling factor by performing a rheological consistency coupling evaluation on the morphological characteristics of the soil resistance depth curve and the soil volumetric water content includes:

[0025] By performing curve integration energy accumulation and peak intensity extraction on the soil resistance depth curve data at the sampling points, the integrated energy data and peak intensity data are obtained.

[0026] Plasticity index data is obtained by normalizing the ratio of integral energy data to peak intensity data;

[0027] By calculating the consistency difference of morphological contribution per unit water content between plastic morphological index data and soil volumetric water content data and processing it with nonlinear mapping constraints, the rheological morphology coupling factor is obtained.

[0028] Furthermore, the step of obtaining integrated energy data and peak intensity data by performing curve integration energy accumulation and peak intensity extraction processing on the soil resistance depth curve data at the sampling points includes:

[0029] For any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions at each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data. The soil cone penetration resistance sequence data of the target sampling point at all depth sampling positions are accumulated and summed to obtain the curve integral energy accumulation evaluation value corresponding to the target sampling point, and the curve integral energy accumulation evaluation value is used as the integral energy data corresponding to the target sampling point. Peak search and maximum value extraction operations are performed on the soil cone penetration resistance sequence data of the target sampling point at all depth sampling positions to obtain the resistance peak intensity evaluation value corresponding to the target sampling point, and the resistance peak intensity evaluation value is used as the peak intensity data corresponding to the target sampling point.

[0030] Furthermore, the process of obtaining plasticity index data by normalizing the ratio of integrated energy data to peak intensity data includes:

[0031] For any target sampling point, extract the integral energy data and peak intensity data corresponding to the target sampling point from the multidimensional feature dataset; use the integral energy data of the target sampling point as the numerator and the peak intensity data of the target sampling point as the denominator, and use the corresponding fraction calculation result as the plasticity index data corresponding to the target sampling point.

[0032] Furthermore, the process of calculating the consistency difference in rheological contribution per unit water content between plastic morphology index data and soil volumetric water content data, and performing nonlinear mapping constraint processing, to obtain the rheological morphology coupling factor includes:

[0033] For any first sampling point and second sampling point, extract the plasticity index data and soil volumetric water content data corresponding to the first sampling point and the plasticity index data and soil volumetric water content data corresponding to the second sampling point from the multidimensional feature dataset, respectively.

[0034] The plastic morphology index data of the first sampling point is compared with the soil volumetric moisture content data of the first sampling point to obtain the morphological contribution assessment value of the first unit moisture content. The plastic morphology index data of the second sampling point is compared with the soil volumetric moisture content data of the second sampling point to obtain the morphological contribution assessment value of the second unit moisture content.

[0035] The absolute value of the difference between the first unit moisture content form contribution assessment value and the second unit moisture content form contribution assessment value is calculated to obtain the consistency difference assessment value of the unit moisture content form contribution; the soil volume moisture content data of the first sampling point and the soil volume moisture content data of the second sampling point are summed to obtain the moisture content summed weight assessment value; the consistency difference assessment value of the unit moisture content form contribution assessment value and the moisture content summed weight assessment value are multiplicatively coupled to obtain the rheological non-uniform potential energy assessment value.

[0036] Set the sensitivity adjustment coefficient and the minimum distance maintenance constant; take the rheological non-uniform potential energy assessment value as the numerator, the sum of the rheological non-uniform potential energy assessment value and the sensitivity adjustment coefficient as the denominator, and the calculation result of adding the corresponding fraction with the minimum distance maintenance constant as the rheological morphology coupling factor corresponding to the first sampling point and the second sampling point.

[0037] Furthermore, the process of obtaining adaptive similarity distance and performing clustering iterations by jointly correcting the basic Euclidean distance using gradient energy phase locking factors and rheological morphology coupling factors includes:

[0038] For any sampling point and any cluster center in the multidimensional feature dataset, for each depth sampling location in the sampling depth sequence data, the difference in soil cone penetration resistance between the sampling point and the cluster center at the corresponding depth sampling location is calculated. The squares of the soil cone penetration resistance differences at each depth sampling location are summed and then square-rooted to obtain the basic Euclidean distance between the sampling point and the cluster center. The basic Euclidean distance is then multiplied and weighted by the gradient energy phase locking factor and rheological morphology coupling factor corresponding to the sampling point and the cluster center to obtain the adaptive similarity distance between the sampling point and the cluster center. During the clustering operation, multiple cluster centers are initialized. For each sampling point in the multidimensional feature dataset, the adaptive similarity distance between the sampling point and each cluster center is calculated. The cluster center with the smallest adaptive similarity distance is determined as the cluster center to which the sampling point belongs. After completing one sample allocation, the cluster centers are updated. The sample allocation and cluster center update are repeated until the preset convergence condition is met, and the category label corresponding to each sampling point is output.

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] This invention discloses a farmland soil compaction diagnosis method based on multi-source sensor data. By introducing a feature layer localization and phase alignment mechanism based on resistance gradient energy distribution, similar compacted structures can still be identified as belonging to the same category even when there is a shift in the depth axis. This significantly reduces distance misalignment and clustering errors caused by depth misalignment. The resulting compaction zone boundaries more closely resemble the continuous distribution of actual compacted zones, avoiding the artificial division of the same compacted zone into multiple fragmented small blocks, thus improving the interpretability and usability of the diagnostic map in field operations. Simultaneously, differences in soil moisture content induce nonlinear rheological changes in the mechanical response, causing the same compacted structure to exhibit a high-amplitude peak under dry conditions and a low-amplitude broad peak under wet conditions. Traditional measurement methods based solely on numerical amplitude differences can easily misclassify these structures as different compaction levels. This technical solution constructs a curve morphology index that can characterize the "peak-area" relationship and establishes consistency verification and nonlinear compensation mapping with moisture content. This effectively resolves the intensity distortion and morphological linkage changes caused by moisture, keeping compacted areas with different wet and dry conditions but similar structures compact in the cluster space. This reduces the risk of missed tillage and over-tillage, improves the stability and energy-saving effect of differentiated deep tillage prescriptions, and ultimately improves the precision level of farmland soil management. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0042] Figure 1This is a flowchart illustrating a method for diagnosing farmland soil compaction based on multi-source sensor data, as described in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0044] See Figure 1 This is a flowchart of a method for diagnosing farmland soil compaction based on multi-source sensor data, as provided in Embodiment 1 of the present invention. Figure 1 As shown, a method for diagnosing farmland soil compaction based on multi-source sensor data may include:

[0045] Step S1: Obtain a multidimensional feature dataset by collecting and standardizing the raw data from multiple sources of sensors in the farmland.

[0046] First, using an agricultural machinery platform equipped with a high-precision GPS positioning device, a hydraulic cone penetrator, and a time-domain reflectometer, continuous sampling is conducted within the target farmland area according to a preset grid path. For each sampling point, the location information of the sampling point, soil cone penetration resistance sequence data varying with depth, sampling depth sequence data corresponding to the soil cone penetration resistance sequence data, and soil volumetric water content data at the sampling point location are simultaneously acquired. The soil cone penetration resistance sequence data covers a vertical profile from the surface to the preset depth. After completing the raw data acquisition from the multi-source sensors in the farmland, the data is then processed... Outlier removal was performed on the soil cone penetration resistance sequence data and soil volumetric moisture content data. Outliers included abrupt changes caused by poor sensor contact or stone collisions. The soil cone penetration resistance sequence data and soil volumetric moisture content data after outlier removal were then standardized and mapped to eliminate the influence of different sensor dimensions. The standardized and mapped soil cone penetration resistance sequence data, sampling depth sequence data, soil volumetric moisture content data, and location information were combined to obtain a multidimensional feature dataset containing location information, soil resistance depth curves, and soil volumetric moisture content.

[0047] This completes the process of acquiring and standardizing the raw data from multiple sources of sensors in farmland to obtain a multidimensional feature dataset.

[0048] Step S2: Obtain the gradient energy phase locking factor by constructing gradient energy characteristics and evaluating phase alignment cross-correlation of soil resistance depth curve data.

[0049] When diagnosing soil compaction, the first step is to address the misalignment of characteristic layer depths caused by micro-topographical undulations in farmland. In real-world field operations, the repeated compaction by large agricultural machinery creates ruts, and tillage creates furrows, resulting in significant vertical elevation differences on the farmland surface. However, when using penetrometers to collect data, the moment of contact with the surface is typically taken as the zero-depth point. This leads to a situation where, although the underground plow pan or compacted layer is usually parallel to the mean sea level or historical tillage layer, the inconsistency in the zero-depth point causes compacted layers with the same physical properties to appear at different depths in sensor data from different sampling points. For example, data collected within ruts may show a peak compacted layer depth several centimeters shallower than data collected on ridges. In this case, if the traditional Euclidean distance is used to calculate the similarity between two resistance depth curves, the algorithm will compare the high resistance value of one sampling point with the low resistance value of another at the same depth, calculating a significant difference and classifying these two sampling points, which essentially belong to the same compaction type, as different categories. To address this issue, a simple overall depth translation is insufficient because the soil profile also contains noise from uncompacted layers. A method is needed that can automatically pinpoint the core location of the compacted layer and elastically correct the distance metric based on the relative deviation of that location. Specifically, the region with the most dramatic resistance variation (i.e., the compacted layer boundary) should be used as a feature anchor point. The phase difference between two sampling points on the feature layer should be calculated, and a correction factor that tolerates depth translation should be constructed. This allows sampling points that are physically similar but have different depth locations to be correctly identified as belonging to the same category.

[0050] In summary, this invention first obtains gradient energy operator sequence data and feature layer energy centroid depth data by performing differential gradient calculation and gradient square weighted coupling processing on the soil resistance depth curve data of sampling points. Specifically, for any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions of each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data; first-order difference calculation is performed on the soil cone penetration resistance sequence data corresponding to the adjacent depth sampling positions of the target sampling point to obtain the resistance differential gradient data of the target sampling point at each depth sampling position; multiplicative coupling operation is performed on the squares of the soil cone penetration resistance sequence data and the resistance differential gradient data at each depth sampling position of the target sampling point to obtain the gradient energy operator sequence data corresponding to the target sampling point; weighted summation operation is performed on the gradient energy operator sequence data and the corresponding sampling depth sequence data, and the ratio of the weighted summation operation result to the cumulative summation operation result of the gradient energy operator sequence data is normalized to obtain the feature layer energy centroid depth data corresponding to the target sampling point.

[0051] After obtaining the gradient energy operator sequence data and the feature layer energy centroid depth data, the gradient energy phase locking factor is obtained by performing displacement alignment cross-correlation and normalization mapping on the gradient energy operator sequence data between sampling points based on the feature layer energy centroid offset. Specifically, for any first sampling point and second sampling point, the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the first sampling point and the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the second sampling point are extracted from the multidimensional feature dataset respectively; the absolute value of the difference between the feature layer energy centroid depth data of the first sampling point and the feature layer energy centroid depth data of the second sampling point is calculated to obtain the feature layer energy centroid offset data corresponding to the first sampling point and the second sampling point; for each depth sampling position in the sampling depth sequence data, the gradient energy operator sequence data of the first sampling point at the depth sampling position is calculated. The gradient energy operator value is multiplied by the gradient energy operator value at the depth sampling position corresponding to the second sampling point minus the feature layer energy centroid offset data at the depth sampling position, and the product results are summed over all depth sampling positions to obtain the cross-correlation multiplication accumulation evaluation value; the gradient energy operator values ​​of the gradient energy operator sequence data of the first sampling point at all depth sampling positions are summed by squares to obtain the first autocorrelation energy evaluation value, and the gradient energy operator values ​​of the gradient energy operator sequence data of the second sampling point at all depth sampling positions are summed by squares to obtain the second autocorrelation energy evaluation value; the cross-correlation multiplication accumulation evaluation value is divided by the square root of the product of the first autocorrelation energy evaluation value and the second autocorrelation energy evaluation value to obtain the normalized cross-correlation evaluation value; the difference between the constant and the normalized cross-correlation evaluation value is used as the gradient energy phase locking factor corresponding to the first sampling point and the second sampling point.

[0052] In one implementation, assume the first Each sampling point at depth The gradient energy operator at is ;No. The sampling point and the first The offset of the feature layer energy center between each sampling point is Then the first The sampling point and the first The expression for calculating the gradient energy phase-locking factor corresponding to each sampling point is:

[0053]

[0054] in, Indicates the first The sampling point and the first Gradient energy phase locking factor corresponding to each sampling point; Indicates the first Each sampling point at depth Gradient energy operator at the location; Indicates the first The sampling point and the first The offset of the feature layer energy center between each sampling point; Indicates the first Each sampling point at depth The gradient energy operator at the location.

[0055] It should be noted that, firstly, to address the difficulty in feature extraction caused by the mixing of compacted and ordinary soil layers in soil profiles, the gradient energy operator not only considers the magnitude of the resistance value itself but also introduces the square of the gradient of resistance change with depth as a weight. Since the physical nature of the compacted layer is a sudden change in soil compaction, its resistance curve inevitably has sharp rising or falling edges, while the gradient change in ordinary soil layers or noisy areas is smaller. Therefore, by multiplying the square of the gradient with the original value, the energy proportion of the compacted layer in the signal can be amplified, and the interference of background noise from the uncompacted layer can be suppressed, thereby achieving accurate feature enhancement of the physical location of the compacted layer. Based on the extracted gradient energy operator, to quantify the degree of depth misalignment caused by surface undulations, the feature layer energy centroid and centroid offset are introduced. Unlike a simple geometric center, the energy centroid uses the gradient energy distribution as a density function to calculate its first moment, which can accurately reflect the physical location of the core area of ​​the compacted layer on the vertical profile. By calculating the difference between the centroids of two sampling points, the relative depth translation caused by the difference in surface elevation between the two sampling points is actually obtained. Finally, the main structure of the gradient energy phase-locking factor adopts a displacement-corrected sliding cross-correlation form. In the numerator of the formula, the sampling points... The gradient energy sequence is shifted according to the calculated offset so that its centroid is aligned with the sampling point. Align the samples and then calculate their inner product. The core logic of this design is that if the two sampling points do indeed belong to the same compaction type, then after eliminating the depth deviation caused by surface undulations, their waveforms should be highly overlapping. In this case, the translated inner product value will approach the root of the autocorrelation energy product in the denominator, making the ratio approach... This ultimately leads to the gradient energy phase-locking factor approaching [value]. Conversely, if the two soils are essentially different types of soil structures (e.g., one with a compacted layer and the other without), even with centroid alignment, their waveforms will not match, resulting in a smaller inner product value and a gradient energy phase-locking factor that approaches [value missing]. By applying this factor as a multiplier to the original Euclidean distance, distance compression was achieved for samples that are similar in shape but displaced, thus effectively solving the clustering misjudgment problem caused by micro-topographic undulations and ensuring that the diagnostic results reflect the true physical compaction state of the soil rather than topographic differences.

[0056] Thus, the gradient energy phase locking factor was obtained by constructing gradient energy characteristics and evaluating phase alignment cross-correlation on soil resistance depth curve data.

[0057] Step S3: By performing a rheological consistency coupling evaluation on the morphological characteristics of the soil resistance depth curve and the soil volumetric water content, the rheological morphology coupling factor is obtained.

[0058] After addressing the spatial misalignment of feature layers using the gradient energy phase-locking factor, clustering analysis still faces a deeper problem: the distortion of rheological intensity caused by changes in soil mechanical properties with water content. Although step S2 achieves depth alignment of the compacted layers, the two aligned sampling points may still exhibit significant differences in resistance values. This is because soil, as a typical porous medium, exhibits a significant nonlinear negative correlation between its shear strength and water content. In real-world scenarios, within the same severely compacted field, if local areas are relatively dry, the soil exhibits brittle hardening, with the penetration resistance curve showing an extremely high peak and a sharp waveform (narrow peak width); while if local areas are relatively moist, water acts as a lubricant between soil particles, resulting in plastic softening. The peak penetration resistance value decreases significantly, but due to the viscosity effect, the waveform exhibits obvious broadening characteristics (wider peak width, greater energy dissipation). In this case, if the algorithm judges solely based on resistance values, it will classify dry, high-resistance samples as severely compacted and moist, low-resistance samples as lightly compacted, leading to misjudgments and missed cultivation of the same type of severely compacted areas.

[0059] Therefore, simply normalizing the values ​​is insufficient to solve the problem, because moisture not only alters the magnitude of resistance but also changes the waveform of the curve. A mechanism that deeply couples the relationship between moisture, morphology, and intensity is needed. This mechanism should identify sampling points with significantly different values ​​but identical essential structures by quantifying whether the soil's plasticity characteristics conform to the rheological laws under specific moisture conditions. Specifically, the constant ratio of the integral energy corresponding to the peak resistance per unit area to moisture should be used as a physical criterion to construct a correction factor capable of compensating for differences in moisture content.

[0060] In summary, this invention first processes the soil resistance depth curve data at sampling points by performing curve integral energy accumulation and peak intensity extraction to obtain integral energy data and peak intensity data. Specifically, for any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions at each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data. The soil cone penetration resistance sequence data at all depth sampling positions of the target sampling point are accumulated and summed to obtain the curve integral energy accumulation evaluation value corresponding to the target sampling point, and the curve integral energy accumulation evaluation value is used as the integral energy data corresponding to the target sampling point. Peak search and maximum value extraction operations are performed on the soil cone penetration resistance sequence data at all depth sampling positions of the target sampling point to obtain the resistance peak intensity evaluation value corresponding to the target sampling point, and the resistance peak intensity evaluation value is used as the peak intensity data corresponding to the target sampling point.

[0061] After obtaining the peak intensity data, the plasticity index data is obtained by normalizing the ratio of the integral energy data and the peak intensity data. Specifically, for any target sampling point, the integral energy data and peak intensity data corresponding to the target sampling point are extracted from the multidimensional feature dataset. The integral energy data of the target sampling point is used as the numerator, the peak intensity data of the target sampling point is used as the denominator, and the result of the corresponding fraction calculation is used as the plasticity index data corresponding to the target sampling point.

[0062] After obtaining the plastic morphology index data, the rheological morphology coupling factor is obtained by calculating the consistency difference of morphological contribution per unit water content and processing it with nonlinear mapping constraints between the plastic morphology index data and the soil volumetric water content data. Specifically, for any first sampling point and second sampling point, the plastic morphology index data and soil volumetric water content data corresponding to the first sampling point and the corresponding soil volumetric water content data of the second sampling point are extracted from the multidimensional feature dataset, respectively. The ratio of the plastic morphology index data and the soil volumetric water content data of the first sampling point is calculated to obtain the first unit water content morphological contribution evaluation value, and the ratio of the plastic morphology index data and the soil volumetric water content data of the second sampling point is calculated to obtain the second unit water content morphological contribution evaluation value. The difference between the first unit water content morphological contribution evaluation value and the second unit water content morphological contribution evaluation value is then calculated. The absolute value of the value is calculated to obtain the consistency difference assessment value of the morphological contribution per unit water content; the soil volumetric water content data of the first sampling point and the soil volumetric water content data of the second sampling point are summed to obtain the water content summed weight assessment value; the consistency difference assessment value of the morphological contribution per unit water content is multiplicatively coupled with the water content summed weight assessment value to obtain the rheological non-consistency potential energy assessment value; a sensitivity adjustment coefficient and a minimum distance maintenance constant are set. In this embodiment of the invention, the sensitivity adjustment coefficient is set to 0.5 and the minimum distance maintenance constant is set to 0.05 to prevent the distance from becoming too zero when the two samples are completely consistent, which would lead to matrix singularity; the rheological non-consistency potential energy assessment value is used as the numerator, the sum of the rheological non-consistency potential energy assessment value and the sensitivity adjustment coefficient is used as the denominator, and the calculation result of adding the corresponding fraction with the minimum distance maintenance constant is used as the rheological morphological coupling factor corresponding to the first sampling point and the second sampling point.

[0063] In one implementation, assume the first The plasticity index of each sampling point is ;No. The plasticity index of each sampling point is ;No. The soil volumetric moisture content at each sampling point was [missing information]. ;No. The soil volumetric moisture content at each sampling point was [missing information]. The sensitivity adjustment coefficient is The minimum distance preservation constant is Then the first The sampling point and the first The formula for calculating the rheological coupling factor corresponding to each sampling point is:

[0064]

[0065] in, Indicates the first The sampling point and the first The rheological coupling factor corresponding to each sampling point; Indicates the first Plastic morphology index of each sampling point; Indicates the first Plastic morphology index of each sampling point; Indicates the first Soil volumetric moisture content at each sampling point; Indicates the first Soil volumetric moisture content at each sampling point; This represents the sensitivity adjustment coefficient; This indicates that the minimum distance remains constant.

[0066] It should be noted that, in response to the morphological difference between the "sharp peaks" of dry soil and the "wide, shallow peaks" of moist soil, this invention first constructs a plastic morphology index. This index captures the rheological characteristics of soil by calculating the ratio of the total integral area of ​​the resistance curve (representing the total energy consumed during penetration) to the peak resistance (representing the maximum instantaneous intensity). For dry, brittle soils, although the peak value is large, the waveform is narrow, and the total area is relatively small, leading to... The values ​​are relatively small; however, for moist plastic soils, although the peak value is small, the waveform is wide and severely sluggish, and the total area is relatively large, resulting in... The value increased significantly. Based on this, to determine whether the two sampling points belonged to the same soil structure, rheological non-uniform potential energy was introduced. The core logic of this formula is based on the rheological assumptions of soil mechanics: for soils of the same texture, their plasticity index... With moisture content There is usually a positive correlation between them, that is, the plastic deformation capacity contributed by unit moisture content. It should tend to a relatively stable constant. This can be achieved through calculation. In essence, this verifies whether two sampling points follow the same rheological evolution trajectory. If the ratio between the two is extremely small, it indicates that the samples... The low resistance is solely due to the high moisture content; its physical nature is unrelated to the sample. Consistent. At the same time, the product term... This serves as an amplitude weighting mechanism, ensuring that differences are appropriately amplified even when there is very little data variation (high data uncertainty), thus avoiding misjudgment. Finally, By employing a nonlinear mapping function, the unbounded potential energy difference is transformed into... Distance weighting of intervals. When two sampling points conform to the same rheological law (i.e., one is dry compacted soil and the other is wet compacted soil). Approaching ,at this time Approaching the minimum value This significantly compresses the originally huge Euclidean distance, allowing the algorithm to classify them as similar, thus solving the problem of false outliers caused by differences in moisture content. Conversely, if the ratios of two samples differ greatly (for example, one is moist compacted soil, and the other is moist stone, the latter's ratio...), the algorithm will not recognize them as similar. (extremely small) Get bigger Approaching The distance remains constant, thus ensuring the algorithm's ability to distinguish real outlier samples.

[0067] Thus, the rheological morphology coupling factor was obtained by conducting a rheological consistency coupling evaluation of the morphological characteristics of the soil resistance depth curve and the soil volumetric water content.

[0068] Step S4: By jointly correcting the basic Euclidean distance using gradient energy phase locking factor and rheological morphology coupling factor, an adaptive similarity distance is obtained and clustering iteration is performed.

[0069] After obtaining the gradient energy phase-locking factor and the rheological morphology coupling factor, this invention further obtains an adaptive similarity distance and performs clustering iteration by jointly correcting the basic Euclidean distance using the gradient energy phase-locking factor and the rheological morphology coupling factor. Specifically, for any sampling point and any cluster center in the multidimensional feature dataset, for each depth sampling position in the sampling depth sequence data, the difference in soil cone penetration resistance between the sampling point and the cluster center at the corresponding depth sampling position is calculated. The squares of the soil cone penetration resistance differences at each depth sampling position are then summed and the square root is taken to obtain the basic Euclidean distance corresponding to the sampling point and the cluster center. The distance is calculated by multiplying the basic Euclidean distance with the gradient energy phase-locking factor and rheological coupling factor corresponding to the sampling point and the cluster center, and obtaining the adaptive similarity distance between the sampling point and the cluster center. When performing clustering operation, multiple cluster centers are initialized. For each sampling point in the multidimensional feature dataset, the adaptive similarity distance between the sampling point and each cluster center is calculated. The cluster center with the smallest adaptive similarity distance is determined as the cluster center to which the sampling point belongs. After completing one sample allocation, the cluster centers are updated. The sample allocation and cluster center update are repeated until the preset convergence condition is met, and the category label corresponding to each sampling point is output.

[0070] It should be noted that the present invention uses the elbow rule to select the number of cluster centers.

[0071] Thus, the adaptive similarity distance was obtained and clustering iteration was performed by jointly correcting the basic Euclidean distance using gradient energy phase locking factor and rheological morphology coupling factor.

[0072] Step S5 involves spatial modeling and prescription mapping of the clustering results and sampling point location information to obtain a precise diagnostic map of soil compaction and generate differentiated deep loosening operation instructions.

[0073] Based on the final clustering result labels output from the steps, combined with the original labels of each sampling point... Location information was used to construct a spatial distribution model of soil compaction across the entire field. Because an adaptive distance metric incorporating gradient energy phase-locking and rheological morphology coupling factors was employed during clustering, the classification results effectively eliminated interference introduced by micro-topographic undulations and differences in soil moisture content, ensuring that areas classified into the same cluster have a highly consistent physical compaction structure.

[0074] Based on the above classification results, the following specific diagnostic and work instruction procedures will be implemented.

[0075] Semantic mapping of compaction level:

[0076] Calculate the average conic index of each cluster ( Peak and average compaction depth. According to agronomic standards (such as...) And the depth is Within a certain range, the subsurface is considered severely compacted. The numerical labels output by clustering are mapped to compaction levels with actual agronomical meanings, such as “severely compacted area”, “moderately compacted area”, and “lightly / uncompacted area”.

[0077] Generating accurate diagnostic thematic maps: Using ordinary kriging interpolation from statistics, based on the compaction level of discrete sampling points, spatial interpolation is performed on unsampled areas to generate a continuous thematic map covering the entire field for soil compaction diagnosis. This layer can intuitively display the spatial distribution continuity and boundary range of the compacted layer, providing a basic base map for agricultural machinery operation path planning.

[0078] Develop differentiated deep tillage operation strategies and generate corresponding intelligent agricultural machinery operation instructions for different compaction zones:

[0079] For "severely compacted areas": the deep loosening operation depth is set below the bottom boundary of the compacted layer. This is to completely break up the plow pan and restore soil permeability.

[0080] For "moderately compacted zones": Set a shallower subsoil depth (e.g.) Alternatively, an intermittent deep tillage method can be adopted to improve soil physical properties while reducing fuel consumption.

[0081] For “lightly / uncompacted areas”: adopt no-till or only shallow topsoil cultivation to avoid unnecessary mechanical compaction and energy waste.

[0082] Through the above steps, this invention transforms multi-source sensor data into executable differentiated operation prescriptions, enabling accurate diagnosis of soil compaction conditions in complex and ever-changing farmland environments. It effectively solves the problems of missed or over-cultivation caused by terrain and water interference in traditional methods, and significantly improves farmland operation efficiency and soil management level.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing soil compaction in an agricultural field based on multi-source sensor data, characterized by, The method includes: Step S1: Obtain a multidimensional feature dataset by collecting and standardizing the raw data from multiple sources of sensors in farmland; Step S2: Obtain the gradient energy phase locking factor by constructing gradient energy characteristics and evaluating phase alignment cross-correlation of soil resistance depth curve data; Step S3: Obtain the rheological morphology coupling factor by performing a rheological consistency coupling evaluation on the morphological characteristics of the soil resistance depth curve and the soil volumetric water content; Step S4: By jointly correcting the basic Euclidean distance using the gradient energy phase locking factor and the rheological morphology coupling factor, an adaptive similarity distance is obtained and clustering iteration is performed; Step S5: By performing spatial modeling and prescription mapping on the clustering results and sampling point location information, a precise diagnosis map of soil compaction is obtained and differentiated deep loosening operation instructions are generated.

2. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 1, characterized in that, The process involves collecting and standardizing raw data from multiple sources of sensors in farmland to obtain a multidimensional feature dataset, including: Using an agricultural machinery platform equipped with a high-precision global positioning system, a hydraulic cone penetrator, and a time-domain reflectometer, continuous sampling is performed within the target farmland area according to a preset grid path. For each sampling point, the location information of the sampling point, soil cone penetration resistance sequence data varying with depth, sampling depth sequence data corresponding to the soil cone penetration resistance sequence data, and soil volumetric moisture content data at the sampling point location are simultaneously acquired. The soil cone penetration resistance sequence data covers a vertical profile from the surface to the preset depth. After completing the raw data acquisition from the multi-source sensors in the farmland, the soil... Outlier removal was performed on the soil cone penetration resistance sequence data and soil volumetric moisture content data. The outliers included abrupt outliers caused by poor sensor contact or stone collisions. The soil cone penetration resistance sequence data and soil volumetric moisture content data after outlier removal were then standardized and mapped to eliminate the influence of different sensor dimensions. The standardized and mapped soil cone penetration resistance sequence data, sampling depth sequence data, soil volumetric moisture content data, and location information were combined to obtain a multidimensional feature dataset containing location information, soil resistance depth curves, and soil volumetric moisture content.

3. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 1, characterized in that, The step involves constructing gradient energy characteristics and evaluating phase-aligned cross-correlation of soil resistance depth curve data to obtain a gradient energy phase-locking factor, including: By performing differential gradient calculation and gradient squared weighted coupling processing on the soil resistance depth curve data of the sampling points, gradient energy operator sequence data and feature layer energy centroid depth data are obtained. The gradient energy phase locking factor is obtained by performing displacement alignment cross-correlation and normalization mapping on the gradient energy operator sequence data between sampling points based on the offset of the energy centroid of the feature layer.

4. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 3, characterized in that, The process involves performing differential gradient calculation and gradient squared weighted coupling on the soil resistance depth curve data at sampling points to obtain gradient energy operator sequence data and feature layer energy centroid depth data, including: For any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions at each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data. First-order difference calculations are performed on the soil cone penetration resistance sequence data corresponding to adjacent depth sampling positions of the target sampling point to obtain the resistance differential gradient data at each depth sampling position of the target sampling point. Multiplicative coupling operations are performed on the squares of the soil cone penetration resistance sequence data and the resistance differential gradient data at each depth sampling position of the target sampling point to obtain the gradient energy operator sequence data corresponding to the target sampling point. The gradient energy operator sequence data and the corresponding sampling depth sequence data are weighted and summed, and the ratio of the weighted summation result to the cumulative summation result of the gradient energy operator sequence data is normalized to obtain the feature layer energy centroid depth data corresponding to the target sampling point.

5. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 3, characterized in that, The step involves performing displacement-aligned cross-correlation and normalization mapping on the gradient energy operator sequence data between sampling points based on the feature layer energy centroid offset to obtain the gradient energy phase locking factor, including: For any first sampling point and second sampling point, extract the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the first sampling point, and the gradient energy operator sequence data and feature layer energy centroid depth data corresponding to the second sampling point from the multidimensional feature dataset, respectively. The absolute value of the difference between the feature layer energy centroid depth data of the first sampling point and the feature layer energy centroid depth data of the second sampling point is calculated to obtain the feature layer energy centroid offset data corresponding to the first sampling point and the second sampling point. For each depth sampling position in the sampled depth sequence data, the gradient energy operator value of the first sampling point at that depth sampling position is multiplied by the gradient energy operator value of the second sampling point at the depth sampling position corresponding to the depth sampling position minus the feature layer energy centroid offset data. The product results of all depth sampling positions are then summed to obtain the cross-correlation multiplication accumulation evaluation value. The gradient energy operator values ​​of the gradient energy operator sequence data of the first sampling point at all depth sampling positions are then summed by squares to obtain the first autocorrelation energy evaluation value. The gradient energy operator values ​​of the gradient energy operator sequence data of the second sampling point at all depth sampling positions are then summed by squares to obtain the second autocorrelation energy evaluation value. The normalized cross-correlation evaluation value is obtained by dividing the sum of the cross-correlation multiplication evaluation values ​​by the square root of the product of the first autocorrelation energy evaluation value and the second autocorrelation energy evaluation value; the difference between the constant and the normalized cross-correlation evaluation value is used as the gradient energy phase locking factor corresponding to the first sampling point and the second sampling point.

6. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 1, characterized in that, The method involves evaluating the rheological consistency between the morphological characteristics of the soil resistance depth curve and the soil volumetric water content to obtain the rheological morphology coupling factor, including: By performing curve integration energy accumulation and peak intensity extraction on the soil resistance depth curve data at the sampling points, the integrated energy data and peak intensity data are obtained. Plasticity index data is obtained by normalizing the ratio of integral energy data to peak intensity data. By calculating the consistency difference of morphological contribution per unit water content between plastic morphological index data and soil volumetric water content data and processing it with nonlinear mapping constraints, the rheological morphology coupling factor is obtained.

7. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 6, characterized in that, The process of obtaining integrated energy data and peak intensity data by performing curve integration energy accumulation and peak intensity extraction on the soil resistance depth curve data at sampling points includes: For any target sampling point, the soil cone penetration resistance sequence data and sampling depth sequence data corresponding to the target sampling point are extracted from the multidimensional feature dataset, and the sampling positions at each depth of the soil cone penetration resistance sequence data are determined based on the sampling depth sequence data. The soil cone penetration resistance sequence data of the target sampling point at all depth sampling positions are accumulated and summed to obtain the curve integral energy accumulation evaluation value corresponding to the target sampling point, and the curve integral energy accumulation evaluation value is used as the integral energy data corresponding to the target sampling point. Peak search and maximum value extraction operations are performed on the soil cone penetration resistance sequence data of the target sampling point at all depth sampling positions to obtain the resistance peak intensity evaluation value corresponding to the target sampling point, and the resistance peak intensity evaluation value is used as the peak intensity data corresponding to the target sampling point.

8. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 6, characterized in that, The process of obtaining plasticity index data by normalizing the ratio of integral energy data to peak intensity data includes: For any target sampling point, extract the integral energy data and peak intensity data corresponding to the target sampling point from the multidimensional feature dataset; use the integral energy data of the target sampling point as the numerator and the peak intensity data of the target sampling point as the denominator, and use the corresponding fraction calculation result as the plasticity index data corresponding to the target sampling point.

9. A method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 6, characterized in that, The process involves calculating the consistency difference in rheological contribution per unit water content between plasticity index data and soil volumetric water content data, and applying nonlinear mapping constraints to obtain the rheological coupling factor, including: For any first sampling point and second sampling point, extract the plasticity index data and soil volumetric water content data corresponding to the first sampling point and the plasticity index data and soil volumetric water content data corresponding to the second sampling point from the multidimensional feature dataset, respectively. The plastic morphology index data of the first sampling point is compared with the soil volumetric moisture content data of the first sampling point to obtain the morphological contribution assessment value of the first unit moisture content. The plastic morphology index data of the second sampling point is compared with the soil volumetric moisture content data of the second sampling point to obtain the morphological contribution assessment value of the second unit moisture content. The absolute value of the difference between the first unit moisture content form contribution assessment value and the second unit moisture content form contribution assessment value is calculated to obtain the consistency difference assessment value of the unit moisture content form contribution; the soil volume moisture content data of the first sampling point and the soil volume moisture content data of the second sampling point are summed to obtain the moisture content summed weight assessment value; the consistency difference assessment value of the unit moisture content form contribution assessment value and the moisture content summed weight assessment value are multiplicatively coupled to obtain the rheological non-uniform potential energy assessment value. Set the sensitivity adjustment coefficient and the minimum distance maintenance constant; take the rheological non-uniform potential energy assessment value as the numerator, the sum of the rheological non-uniform potential energy assessment value and the sensitivity adjustment coefficient as the denominator, and the calculation result of adding the corresponding fraction with the minimum distance maintenance constant as the rheological morphology coupling factor corresponding to the first sampling point and the second sampling point.

10. The method for diagnosing farmland soil compaction based on multi-source sensor data according to claim 1, characterized in that, The process involves jointly correcting the basic Euclidean distance using a gradient energy phase-locking factor and a rheological morphology coupling factor to obtain an adaptive similarity distance and performing clustering iterations, including: For any sampling point and any cluster center in the multidimensional feature dataset, for each depth sampling location in the sampling depth sequence data, the difference in soil cone penetration resistance between the sampling point and the cluster center at the corresponding depth sampling location is calculated. The squares of the soil cone penetration resistance differences at each depth sampling location are summed and then square-rooted to obtain the basic Euclidean distance between the sampling point and the cluster center. The basic Euclidean distance is then multiplied and weighted by the gradient energy phase locking factor and rheological morphology coupling factor corresponding to the sampling point and the cluster center to obtain the adaptive similarity distance between the sampling point and the cluster center. During the clustering operation, multiple cluster centers are initialized. For each sampling point in the multidimensional feature dataset, the adaptive similarity distance between the sampling point and each cluster center is calculated. The cluster center with the smallest adaptive similarity distance is determined as the cluster center to which the sampling point belongs. After completing one sample allocation, the cluster centers are updated. The sample allocation and cluster center update are repeated until the preset convergence condition is met, and the category label corresponding to each sampling point is output.

Images