Remote sensing adaptive sampling and weighted representation method of multi-plot recording unit

Abstract

The application discloses a kind of remote sensing self-adaptive sampling and weighted representation method of multiple land block recording units, comprising: obtaining the remote sensing image data of multiple land blocks associated with target recording unit;Grouping based on the spatial distance between land blocks obtains spatial cluster;Determine the number of sampling points based on land area and layout, generate adaptive sampling range according to the distance from sampling point to land boundary;Filter pixels in sampling range and eliminate anomaly, calculate the remote sensing characteristic value of sampling point and the proportion of land comprehensive effective pixel;Determine land correction weight based on land area, land comprehensive effective pixel proportion and temporal stability;Based on correction weight, remote sensing characteristic value is aggregated into the remote sensing representation value of target recording unit step by step.The application solves the distortion problem of remote sensing representation in multiple land block scene, and improves the feature accuracy.

Description

Technical Field

[0001] This invention relates to the field of remote sensing data processing and spatial information extraction technology, specifically to a remote sensing adaptive sampling and weighted representation method for multi-plot recording units. Background Technology

[0002] In applications such as agricultural monitoring and ecological surveys, remote sensing imagery is often used to acquire spectral information of crop-growing areas and needs to be matched and analyzed with operational data such as production records and transaction records. In practical operational scenarios, a single operational record object (i.e., a record unit) often corresponds to multiple spatially dispersed plots of land with varying areas. Current technologies for extracting remote sensing features from these multi-plot record units typically employ methods such as directly merging all pixels and calculating the average, or performing an equal-weighted average of the average values ​​from each plot. However, this approach has significant drawbacks: firstly, the area differences between different plots mean that simple averaging cannot reasonably reflect the actual contribution of each plot to the record unit, resulting in distorted contributions; secondly, fixed sampling or full-pixel statistics easily introduce mixed pixels at plot edges and anomalous pixels affected by cloud shadows, and ignore the spatial proximity structure between plots, making it difficult for the final generated remote sensing representation value to accurately correspond to the true state of the operational record, thus affecting the accuracy of downstream applications such as phenological identification and yield prediction. Summary of the Invention

[0003] The purpose of this invention is to provide a remote sensing adaptive sampling and weighted representation method for multi-plot recording units. This invention solves the problem of remote sensing representation distortion in multi-plot scenarios, improves feature accuracy, and achieves high-fidelity matching between remote sensing features and operational records.

[0004] The technical solution of this invention: a remote sensing adaptive sampling and weighted representation method for multi-plot recording units, comprising the following steps:

[0005] Step 1: Acquire remote sensing image data of multiple plots associated with the target recording unit;

[0006] Step 2: Group the multiple land parcels based on the spatial distance between them to obtain at least one spatial cluster;

[0007] Step 3: For each plot of land, determine the number of target sampling points based on the plot area and deploy the sampling points. Generate an adaptive sampling range based on the distance from the sampling points to the boundary of the plot.

[0008] Step 4: Perform quality screening on remote sensing pixels within the adaptive sampling range, collect candidate pixels from each sampling range within the same spatial cluster, perform multi-index joint anomaly removal based on the local statistical distribution of the spatial cluster, determine the effective pixel set for each sampling range, and calculate the remote sensing feature value of the sampling point and the comprehensive effective pixel ratio of the plot.

[0009] Step 5: Determine the correction weight for each plot based on the plot area, the proportion of effective pixels in the plot, and the temporal stability.

[0010] Step 6: Based on the corrected weights, the remote sensing feature values ​​are aggregated step by step into the remote sensing representation values ​​of the target recording unit.

[0011] The above method, which generates an adaptive sampling range based on the distance from the sampling point to the land parcel boundary, includes the following specific implementation methods:

[0012] Calculate the shortest distance from the sampling point to the boundary of the plot;

[0013] The adaptive sampling radius is determined based on the nearest distance and a preset radius coefficient, wherein the adaptive sampling radius is not greater than the nearest distance. The formula for calculating the adaptive sampling radius is as follows:

[0014] ;

[0015] in, For the first The first record unit Plot No. The adaptive sampling circle radius for each sampling point For the maximum sampling radius, For radius coefficient, This is the shortest distance from the sampling point to the boundary of the plot;

[0016] A sampling circle is generated with the sampling point as the center and the adaptive sampling radius as the radius, serving as the adaptive sampling range.

[0017] The aforementioned method, wherein the quality screening, anomaly removal, and determination of the effective pixel set of remote sensing pixels include:

[0018] For each sampling circle, pixels whose center point falls within the sampling circle and are located within the boundary of the corresponding plot are taken as a geometric pixel set. Based on the remote sensing image quality label, pixels with clouds, shadows, missing measurements, and invalid values ​​are removed to obtain qualified candidate pixels.

[0019] Quality-qualified candidate pixels from each sampling circle within the same spatial cluster are aggregated to form a local statistical pixel pool. Based on this pool, the local mean and local standard deviation of each remote sensing feature are calculated, and the standardized anomaly index of the candidate pixels is also calculated.

[0020] The number of remote sensing features whose standardized anomaly index absolute value is greater than a preset anomaly threshold among the candidate pixels is counted. Candidate pixels whose number reaches the preset anomaly feature number threshold are judged as anomaly pixels, and the remaining candidate pixels constitute the set of valid pixels. When the number of candidate pixels in the local statistical pixel pool is less than 2, the standardized anomaly judgment is not performed, and the qualified candidate pixels are taken as the set of valid pixels.

[0021] The aforementioned method calculates the comprehensive effective pixel ratio and remote sensing feature value of the sampling points as follows:

[0022] Calculate the first The effective pixel ratio of each sampling point to the sampling circle:

[0023] ;

[0024] In the formula, For the first The effective pixel ratio of each sampling point to the sampling circle; For the first Each sampling point in time phase The corresponding set of geometric pixels; For the first The first record unit Plot No. Each sampling point in time phase The corresponding set of valid pixels;

[0025] The comprehensive effective pixel ratio of the land parcel is calculated using the geometric pixels and effective pixels of all actually selected sampling circles as the statistical range:

[0026] ;

[0027] In the formula, The overall effective pixel ratio of the plot. For the plot of land in time phase The actual set of selected sampling point indices;

[0028] Calculate the remote sensing feature values ​​for the sampling points that belong to the valid sampling point index set:

[0029] ;

[0030] In the formula, For remote sensing feature values, For pixels In time phase Next The values ​​of each remote sensing feature; For the first Each sampling point in time phase The corresponding set of valid sampling point indices.

[0031] The aforementioned method determines the adjustment weights for each parcel, including:

[0032] Using the plot area as the basic contribution weight, and correcting it using the plot's comprehensive effective pixel ratio and temporal stability score, the unnormalized weight of the plot is obtained:

[0033] ;

[0034] In the formula, The area of ​​the land parcel. The overall effective pixel ratio of the plot. The score is for phase stability.

[0035] when empty or At that time, Set it to 0 and mark the plot as having no valid observations. The remaining plots are treated as valid plots and participate in the weight normalization within the spatial cluster.

[0036] The aforementioned method aggregates remote sensing feature values ​​into remote sensing representation values ​​for target recording units step by step, including:

[0037] Based on the effective pixel ratio of effective sampling points within the plot, the remote sensing feature values ​​of the sampling points are aggregated into plot-level remote sensing feature values;

[0038] Based on the unnormalized weights of each effective block within the spatial cluster, the block-level remote sensing feature values ​​are aggregated into spatial cluster-level remote sensing feature values.

[0039] The sum of the unnormalized weights of the effective blocks within a spatial cluster is used as the total corrected weight of the spatial cluster. Based on the total corrected weight of each effective spatial cluster, the spatial cluster-level remote sensing feature values ​​are aggregated into the remote sensing representation values ​​of the target recording unit.

[0040] When a spatial cluster does not contain any valid blocks or its total correction weight is 0, the spatial cluster is excluded from inter-cluster aggregation; when there is no valid spatial cluster in the target recording unit, missing values ​​and a record unit without valid observations are output.

[0041] The aforementioned method, wherein the determination of the phase stability includes:

[0042] The current observation phase and its adjacent valid observation phases are obtained to form a time window;

[0043] The coefficients of variation of each remote sensing feature of the land parcel within the time window are calculated and averaged to obtain the temporal coefficient of variation. The formula for calculating the temporal coefficient of variation is as follows:

[0044] ;

[0045] in, The phase variation coefficient is denoted as . The number of remote sensing features, For remote sensing feature set, For a certain remote sensing feature, This represents the function for calculating the standard deviation. This represents the mean function. For the plot of land in terms of characteristics Upper time window The set of characterization values ​​for all valid observation phases within the scope. To prevent constants with a denominator of zero;

[0046] A phase stability score is calculated based on the phase variation coefficient, wherein the phase stability score is negatively correlated with the phase variation coefficient, and the formula for calculating the phase stability score is as follows:

[0047] ;

[0048] in, The score represents the phase stability.

[0049] The aforementioned method groups multiple land parcels based on the spatial distance between them to obtain at least one spatial cluster, including:

[0050] Calculate the shortest distance between the boundaries of any two plots of land. The formula for calculating the shortest distance is:

[0051] ;

[0052] in, For the plot of land With the plot of land The shortest distance between the boundaries and These represent the boundaries of the two plots of land. and These are points on the boundaries of the two plots of land. This represents the Euclidean distance between two points;

[0053] When the shortest distance is not greater than a preset distance threshold, an adjacency relationship is established between the corresponding two plots. The criteria for determining the adjacency relationship are as follows:

[0054] ;

[0055] in, The shortest distance to the boundary. The preset distance threshold;

[0056] Based on the adjacency relationship, directly or indirectly connected land parcels are grouped into the same spatial cluster.

[0057] The aforementioned method, for each plot of land, determines the number of target sampling points based on the plot area and deploys the sampling points, including:

[0058] The number of sampling points is converted to an integer using a rounding function and then limited to a preset range of minimum and maximum sampling points.

[0059] ;

[0060] In the formula, It is the number of sampling points. This is the maximum number of sampling points allowed per plot of land. This is the minimum allowed number of sampling points for a single plot of land. This refers to the actual area of ​​the land parcel. The area of ​​a single remote sensing pixel. The sampling point density coefficient;

[0061] Let the set of indexes of the actually selected sampling points be denoted as The actual number of sampling points is:

[0062] ;

[0063] when At that time, the corresponding land parcel will be in the time phase Marked as having no valid observations.

[0064] A remote sensing adaptive sampling and weighted representation device for multi-plot recording units includes:

[0065] Memory, used to store computer programs;

[0066] The processor is used to implement the remote sensing adaptive sampling and weighted representation method for multi-plot recording units as described above when executing the computer program.

[0067] Compared with existing technologies, this invention improves the accuracy and robustness of remote sensing representation of multiple record units by adaptive sampling with boundary distance constraints, local statistical anomaly detection of spatial clusters, and hierarchical reliability weighting. The effective pixel ratio of the sampling circle uses the total number of geometric pixels within the sampling circle as the denominator, reflecting the loss of effective information caused by clouds, shadows, missing measurements, and anomalies. The comprehensive effective pixel ratio of the plot further characterizes the overall effective coverage at the plot scale. The core function of the spatial cluster is to provide a local statistical benchmark for anomaly detection, enabling plots with large spatial distances and significant differences in background reflectance conditions to calculate statistical parameters separately. In the aggregation stage, the original contribution weights are determined by plot area, comprehensive effective pixel ratio of the plot, and temporal stability, and the same set of original contribution relationships is transmitted at the plot, spatial cluster, and record unit levels, thereby avoiding contribution distortion caused by equal-weighted averaging and repeated normalization. For cases where there are no effective observations for sampling points, plots, spatial clusters, or record units, exclusion and missing measurement output rules are set respectively, ensuring that the entire calculation process remains closed under both normal and insufficient data conditions. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0069] Figure 2 This is a schematic diagram of the target recording unit in an embodiment of the present invention;

[0070] Figure 3 This is a schematic diagram of land parcel grouping according to an embodiment of the present invention;

[0071] Figure 4 This is a schematic diagram of the sampling point layout according to an embodiment of the present invention;

[0072] Figure 5 This is a schematic diagram of remote sensing pixel quality screening and anomaly removal;

[0073] Figure 6 The comparison of standardized means of various remote sensing indices under normal mean sampling conditions is presented.

[0074] Figure 7 This invention presents a comparison chart of the grouped mean values ​​of multiple remote sensing indices under adaptive sampling.

[0075] Figure 8 This displays the box plot of the best single remote sensing descriptor feature under normal mean sampling conditions;

[0076] Figure 9 The invention demonstrates the optimal single remote sensing descriptor feature boxplot under adaptive sampling.

[0077] Figure 10 This displays a box plot of the integrated discriminant fractional features based solely on remote sensing under normal mean sampling conditions.

[0078] Figure 11 The invention presents a box plot of remote sensing integrated discrimination score features under adaptive sampling.

[0079] Figure 12 The figure shows the ROC curve of binary cross-validation for early and late sampling under normal mean sampling conditions with only remote sensing feature input;

[0080] Figure 13 The histogram of the AUC distribution under the normal mean sampling condition is shown.

[0081] Figure 14 The diagram shows the ROC curve of early and late sampling binary classification cross-validation with only remote sensing feature input in this invention.

[0082] Figure 15 The histogram of the AUC distribution under the adaptive sampling of this invention is shown. Detailed Implementation

[0083] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0084] Example 1: This example provides a remote sensing adaptive sampling and weighted representation method for multi-plot recording units, such as... Figure 1 As shown, it includes the following steps:

[0085] Step 1: Obtain remote sensing image data of multiple plots associated with the target recording unit.

[0086] In this step, let the set of record units be... Recording unit The associated set of crop plots is ,in Indicates the first The first record unit under the record unit One plot of land, Indicates its area. Observation phase. This refers to a set of remote sensing images with a definite imaging date. Representation of record unit Available set of observation time phases. Spatial attribution is determined by whether the pixel center point falls within the plot boundary. In time phase The initial full set of pixels is defined as follows:

[0087] ;

[0088] In the formula: For the first The first record unit The plot of land in the time phase The initial full set of pixels, Represents remote sensing pixels, Indicates the center point of a pixel. Indicates the first The first record unit under the record unit A crop plot Indicates the observation phase with a definite imaging date. Representation of record unit The available set of observation phases.

[0089] A target recording unit refers to an object associated with business data. In practical applications, it can be an agricultural business entity, a production cooperative, a sample number, or a transaction account, etc. Since a single recording unit often corresponds to multiple geographically dispersed plots of land with varying shapes, it is necessary to establish a hierarchical mapping relationship between recording units, plots, and remote sensing pixels. For example, ... Figure 2 As shown, a recording unit is associated with multiple land parcels, each of which is covered by several remote sensing pixels. This step aims to acquire multispectral or hyperspectral image data of these land parcels at a specific observation time phase, serving as the basic input for subsequent processing. It should be understood that although this embodiment uses optical remote sensing imagery as an example, other types of remote sensing data, such as radar and thermal infrared data, can also be used in other embodiments, as long as they can provide characteristic information reflecting the land cover status.

[0090] Step 2: Group the multiple land parcels based on the spatial distance between them to obtain at least one spatial cluster. The spatial cluster is also used to determine the local statistical range for remote sensing pixel anomaly detection.

[0091] In this step, land parcels within the same recording unit may exhibit spatial clustering or discrete distribution. Directly merging all parcels would ignore the spatial heterogeneity between them. This step calculates the spatial distance between parcels, grouping spatially adjacent parcels into the same spatial cluster, such as... Figure 3 As shown, the recording unit is thus obtained. Set of spatial clusters:

[0092] ;

[0093] Clustering results are used not only to organize subsequent hierarchical aggregation at the parcel, spatial cluster, and record unit levels, but also to determine the local statistical range for remote sensing pixel anomaly detection. For the same spatial cluster, candidate pixels from all qualified sampling circles of each parcel within the cluster are aggregated to establish a spatial cluster-level local remote sensing feature distribution, and the Z-score of the candidate pixels is calculated based on this local distribution. Local statistical parameters are calculated separately for different spatial clusters, thereby reducing misjudgments caused by parcels that are spatially distant and have significant differences in terrain and background reflection conditions participating in anomaly detection.

[0094] Step 3: For each plot of land, determine the number of target sampling points based on the plot area and deploy the sampling points. Generate an adaptive sampling range based on the distance from the sampling points to the boundary of the plot.

[0095] In this step, a differentiated sampling strategy is adopted for plots of different sizes. On the one hand, the number of sampling points is dynamically determined based on the plot size, so that larger plots have more sampling points to capture internal variations, while smaller plots have fewer sampling points to avoid redundancy. On the other hand, for each sampling point, its nearest distance to the plot boundary is calculated, and an adaptive sampling range is generated accordingly. For example... Figure 4 As shown, the sampling point density varies within plots of different sizes, and the radius of the circular sampling range around each sampling point varies with its distance from the boundary. This method ensures that the sampling range is controlled within the physical boundaries of the plot, solving the problem that fixed-radius sampling can easily cross the boundary and introduce non-target features. It also ensures that sampling points located inside the plot can make full use of the surrounding effective information.

[0096] Step 4: Perform quality screening on remote sensing pixels within the adaptive sampling range, collect candidate pixels from each sampling range within the same spatial cluster, perform multi-index joint anomaly removal based on the local statistical distribution of the spatial cluster, determine the effective pixel set for each sampling range, and calculate the remote sensing feature value of the sampling point and the comprehensive effective pixel ratio of the plot.

[0097] In this step, within each adaptive sampling range, firstly, clouds, shadows, and invalid pixels are removed based on image quality indicators. Then, local abnormal pixels are identified and removed based on multi-feature statistical distribution. On this basis, the mean feature value of the retained valid pixels is calculated as the remote sensing feature value of that sampling point. Simultaneously, the ratio of the number of valid pixels to the total number of pixels within the sampling range is calculated to obtain the overall valid pixel ratio of the land parcel. For example... Figure 5 As shown, some pixels are marked as abnormal and removed, while the remaining blank squares represent valid pixels.

[0098] Step 5: Determine the correction weight for each plot based on the plot area, the proportion of effective pixels in the plot, and the temporal stability.

[0099] In this step, to reasonably represent the contribution of each plot to the recording unit, this invention abandons simple equal-weighted averaging or weighting solely by area, and instead constructs a multi-dimensional positive weighting system. Plot area serves as the basic contribution weight, reflecting the objective size proportion of the plot within the recording unit; the comprehensive effective pixel ratio of the plot serves as a quality correction factor, reflecting the usability of the current observation data; and temporal stability serves as a time-series correction factor, reflecting the fluctuation of plot characteristics between adjacent time phases. These three factors work together to give higher weights to plots with large areas, good data quality, and stable temporal sequences, while correspondingly reducing the weights of small areas, severely disturbed plots, or plots with drastic fluctuations.

[0100] Step 6: Based on the corrected weights, the remote sensing feature values ​​are aggregated step by step into the remote sensing representation values ​​of the target recording unit.

[0101] In this step, firstly, within the set of valid sampling points, the features of the sampling points are aggregated into plot-level features according to the proportion of valid pixels in the sampling circle. Then, only plots with valid observations and unnormalized weights greater than 0 are aggregated within spatial clusters. Finally, only spatial clusters with a total corrected weight greater than 0 are aggregated at the record unit level. Spatial clusters influence the set of valid pixels, plot features, and plot weights through the aforementioned local statistical anomalies. The two-level aggregation itself is used to maintain consistent transmission of the same original contribution weight between the plot, spatial cluster, and record unit levels, and to provide interpretable intermediate representations at the spatial cluster level. When no valid object exists at a certain level, division or weight normalization is not performed; instead, the object is excluded. When all spatial clusters of a record unit have no valid observations, missing values ​​and a no-valid-observation flag are output.

[0102] Through steps 1 to 6 above, this embodiment constructs a complete remote sensing adaptive sampling and weighted representation framework. This framework not only solves the problem of inaccurate representation caused by scale inconsistency, boundary mixing, and contribution distortion in fragmented multi-plot scenarios using traditional methods, but also achieves deep linkage between sampling quality control and weight aggregation through the comprehensive effective pixel ratio of plots, thereby improving the matching accuracy between remote sensing features and operational records.

[0103] As a preferred embodiment, the specific implementation of generating the adaptive sampling range based on the distance from the sampling point to the land parcel boundary includes: calculating the shortest distance from the sampling point to the land parcel boundary; determining the adaptive sampling radius based on the shortest distance and a preset radius coefficient, wherein the adaptive sampling radius is not greater than the shortest distance; and generating a sampling circle as the adaptive sampling range with the sampling point as the center and the adaptive sampling radius as the radius.

[0104] Specifically, setting up land parcels Inner Each sampling point is First, calculate the shortest distance from the sampling point to the boundary of the plot:

[0105] ;

[0106] In the formula, For the first The first record unit Within the plot of land, the first The shortest distance from each sampling point to the boundary of the plot. For the plot of land The first One sampling point, For the plot of land The boundary.

[0107] The radius of the sampling circle is adaptively determined based on the boundary distance:

[0108] ;

[0109] In the formula, For the first The first record unit Within the plot of land, the first The adaptive sampling circle radius for each sampling point For the maximum sampling radius, The radius coefficient (the range of values ​​is...) (preferably 0.8). For the first The shortest distance from each sampling point to the plot boundary. Because Not greater than The geometric range of the sampling circle does not exceed the boundary of the plot. Let... To ensure a minimum radius with sufficient pixels within the circular region; when the calculated... Less than If the sampling point is not selected, discard that point and choose the next sampling point from the candidate point sequence. The sampling circle is defined as follows:

[0110] ;

[0111] In the formula, For the set of sampling circles, For coordinates of a point in space, For the first The first record unit The first plot of land The center location of each sampling point Describes the Euclidean norm. This is the radius of the adaptive sampling circle corresponding to this sampling point.

[0112] Furthermore, after determining the adaptive sampling range, a geometric pixel set, a quality-qualified candidate pixel set, and an anomaly-removed valid pixel set are established respectively. For each sampling circle, a geometric pixel set is formed by pixels whose center point falls within the sampling circle and are located within the corresponding plot boundary:

[0113] ;

[0114] In the formula, For the first The first record unit Plot No. Each sampling point in time phase The corresponding set of geometric pixels (i.e., the original remote sensing surface observation feature values); For the first The first record unit Plots of land in the time phase The corresponding original remote sensing surface observation feature values; For a single remote sensing pixel, For pixels Corresponding planar geographic coordinates No. Recording Unit 1 The first plot of land An adaptive sampling circular region;

[0115] Based on the remote sensing image quality label, clouds, shadows, missing measurements, and invalid values ​​are removed from the geometric pixel set to obtain a set of candidate pixels with acceptable quality:

[0116] ;

[0117] in, This is an indication of pixel quality compliance.

[0118] For each spatial cluster, candidate pixels from each sampling circle of each patch within the cluster are collected to form a local statistical pixel pool:

[0119] ;

[0120] In the formula, For the plot of land in time phase The actual set of sample point indices selected; Indicates the first The first record unit A spatial cluster;

[0121] Obtain the number of valid observation sampling points:

[0122] ;

[0123] when At that time, based on the local statistical pixel pool, the local mean, local standard deviation, and standardized anomaly index of each remote sensing feature are calculated:

[0124] ;

[0125] ;

[0126] ;

[0127] in, For space clusters Inner Local mean of a remote sensing feature Space clusters Inner Local standard deviation of a remote sensing feature. For pixels In time phase Next The values ​​of each remote sensing feature; To prevent constants with a denominator of zero;

[0128] Let the standardized outlier threshold be... Pixel in the The anomaly indication and the number of anomalies on each remote sensing feature are as follows:

[0129] ;

[0130] ;

[0131] Let the threshold for the number of abnormal features be . ,when When the number of candidate pixels with abnormal features is less than the threshold, retain the candidate pixels; when If, without performing anomaly detection based on standard deviation, qualified candidate pixels are retained and a local statistical sample deficiency flag is added, then the effective pixel set is:

[0132] ;

[0133] No. The effective pixel ratio of each sampling circle is calculated with the geometric pixel set as the denominator:

[0134] ;

[0135] Define the land parcel in time phase The set of valid sampling point indices is:

[0136] ;

[0137] In the formula, The minimum sampling radius threshold is preset;

[0138] Remote sensing feature values ​​are calculated only for sampling points belonging to the set of valid sampling point indices:

[0139] ;

[0140] when When empty, the corresponding land parcel will be in the time phase. Mark as having no valid observations and exclude them from subsequent weight calculations.

[0141] The comprehensive effective pixel ratio of the land parcel is calculated using the geometric pixels and effective pixels of all actually selected sampling circles as the statistical range:

[0142] .

[0143] As a preferred option, the adjustment weights for each parcel are determined, including:

[0144] Using the plot area as the basic contribution weight, and correcting it using the plot's comprehensive effective pixel ratio and temporal stability score, the unnormalized weight of the plot is obtained:

[0145] ;

[0146] In the formula, The area of ​​the land parcel. The overall effective pixel ratio of the plot. The score is for phase stability.

[0147] when empty or At that time, Set it to 0 and mark the plot as having no valid observations. The remaining plots are treated as valid plots and participate in the weight normalization within the spatial cluster.

[0148] For an effective block, the block-level remote sensing feature value is:

[0149] ;

[0150] Define space clusters In time phase The effective block set is:

[0151] ;

[0152] The total corrected weight of the spatial cluster and the normalized weight of the blocks within the cluster are as follows:

[0153] ;

[0154] when empty or If no observations are found, the spatial cluster is marked as having no valid observations and excluded from inter-cluster aggregation; otherwise, the spatial cluster-level remote sensing characteristic value is:

[0155] ;

[0156] Define the target recording unit in time phase The effective set of spatial clusters is:

[0157] ;

[0158] The normalized weighting coefficients of the effective spatial clusters and the remote sensing representation values ​​of the target recording units are as follows:

[0159] ;

[0160] ;

[0161] Expanding the two-level aggregation formula, we get:

[0162] ;

[0163] Therefore, the two-level aggregation does not additionally change the original contribution relationship of the plots; its role is to maintain the consistent transmission of weights across levels and output intermediate results at the spatial cluster level. The substantial impact of spatial clusters on the final result comes from the changes in the effective pixel set, plot characterization value, and comprehensive effective pixel ratio caused by the local statistical anomaly discrimination of spatial clusters. When the effective spatial cluster set is empty, the recording unit outputs missing values ​​and no effective observation identifiers, and weight normalization with a denominator of 0 is not performed.

[0164] As a preferred embodiment, the specific implementation of grouping multiple land parcels based on the spatial distance between multiple land parcels to obtain at least one spatial cluster includes: calculating the shortest distance between the boundaries of any two land parcels; when the shortest distance is not greater than a preset distance threshold, establishing an adjacency relationship between the corresponding two land parcels; and based on the adjacency relationship, classifying directly or indirectly connected land parcels into the same spatial cluster.

[0165] When calculating spatial distances, this embodiment uses the "shortest distance between boundaries" instead of the traditional "centroid distance." This is because in real-world scenarios such as agricultural monitoring, plots are often irregularly shaped, with many being long and narrow, L-shaped, or encircling. If centroid distance were used, two geographically adjacent long and narrow plots, even separated only by a field ridge, might have centroids that are far apart, leading the algorithm to incorrectly classify them as unrelated plots. Conversely, some large but oddly shaped plots might have centroids located outside the plot or far from the main planting area, resulting in distorted distance measurements with other plots.

[0166] land parcel and The distance between them is defined as:

[0167] ;

[0168] In the formula, For the plot of land With the plot of land The distance between them and Representing plots of land and The boundary, and These are points on the boundaries of the two plots of land. Point With point The Euclidean distance between them.

[0169] By calculating the minimum Euclidean distance between any two points on the polygonal boundary of two plots, the actual proximity of the plot edges can be accurately characterized, eliminating the risk of misjudgment of distance caused by geometric center offset and ensuring that the spatial grouping results conform to the true geographical adjacency relationship.

[0170] Single-link hierarchical clustering is used, with distance thresholds. Truncate the clustering tree. When the shortest distance between the boundaries of two plots is not greater than... At that time, establish an adjacency relationship between the two:

[0171] ;

[0172] In the formula, Indicates the first The first record unit under the record unit One plot of land, Indicates the first The first record unit under the record unit One plot of land, For the plot of land and The shortest distance between the boundaries This is the distance threshold for clustering. Based on the established adjacency relationships, directly or indirectly connected plots are grouped into the same spatial cluster, thus obtaining the recording unit. spatial cluster set In the Fuding white tea example, Take 60m; in other applications, the value can be determined within the range of 30 to 100m, depending on the degree of land fragmentation, operating radius, and spatial resolution of remote sensing images.

[0173] After obtaining the shortest distance between any two plots, it is compared with a preset distance threshold. When the shortest distance between a pair of plots is less than or equal to the threshold, the two plots are considered to be spatially adjacent. Subsequently, the plots are grouped according to their connectivity based on the established adjacency relationships. This grouping identifies plots that are directly or indirectly connected: if plot A is adjacent to plot B, and plot B is adjacent to plot C, they will still be grouped into the same spatial cluster, even if the direct distance between plot A and plot C exceeds the threshold. This mechanism simulates the continuity characteristics of agricultural production management in reality, where adjacent plots often share the same microclimate, have similar soil conditions, and are subject to uniform management measures. Treating them as a whole helps preserve the local spatial homogeneity structure within the recording unit.

[0174] Regarding the selection of the preset distance threshold, it is adaptively adjusted based on the degree of land fragmentation, the spatial resolution of remote sensing images, and the typical management radius in the specific application scenario. Generally, the preferred range for this threshold is 30 meters to 100 meters. For example, in hilly tea garden scenarios where plots are relatively fragmented and densely distributed, in order to accurately distinguish independent planting areas on different hillsides or slopes, the threshold can be set to a smaller value. For instance, in the application verification of early and late harvesting identification of Fuding white tea, considering the average spacing of local tea garden plots and the scale of management units, setting the distance threshold to 60 meters achieved the best grouping effect. In contrast, in plains field crop scenarios where plots are large and relatively sparsely distributed, to avoid incorrectly splitting plots belonging to the same management unit but slightly separated by roads and ditches, the threshold can be appropriately widened to 80 meters or even 100 meters.

[0175] Based on the grouping method of shortest boundary distance and adjacency connectivity, this embodiment transforms discrete geographic entities into spatial clusters with semantic relationships. This preprocessing step lays the foundation for subsequent hierarchical weighted aggregation, enabling the aggregation process to be carried out preferentially in local areas with high spatial correlation. This effectively avoids information smoothing and feature distortion caused by forcibly averaging far apart and environmentally different plots, thereby improving the ability of the final generated record unit-level remote sensing representation values ​​to restore the actual business state.

[0176] As a preferred method, for each plot of land, the specific implementation of determining the target number of sampling points and deploying sampling points based on the plot area includes: calculating the initial number of sampling points based on the plot area, the area of ​​a single pixel, and the sampling ratio adjustment coefficient, and limiting the initial number of sampling points to a preset range of minimum and maximum number of sampling points to obtain the target number of sampling points; generating candidate sampling points using a system grid method, and prioritizing the selection of pixel centers that are of qualified quality and whose distance from the plot boundary is greater than a preset boundary threshold as sampling points; wherein, the distance between any two sampling points is not less than a preset minimum point spacing.

[0177] Before setting up sampling points, a basic pre-screening is performed on the initial pixel set. Let... The boundary buffer threshold, For pixel quality indicators; This indicates that the cell was not excluded by the quality flags for cloud, shadow, missing measurement, or invalid value. The candidate cell set after boundary and quality pre-screening is as follows:

[0178] ;

[0179] In the formula, For the plot of land In time phase The initial full set of pixels, Represents remote sensing pixels, Indicates the center point of a pixel. Indicates land parcel The boundary, This represents the distance from the center of the pixel to the boundary of the plot. A boundary buffer threshold (preferably 1 to 2 times the spatial resolution of the remote sensing image) is used to exclude pixels with a high risk of edge blending. For pixel quality indicators, The set of candidate pixels after pre-screening for boundaries and quality.

[0180] When determining the target number of sampling points, an up-rounding function is used to ensure that the number of sampling points is an integer.

[0181] The number of sampling points is converted to an integer using a rounding function and then limited to a preset range of minimum and maximum sampling points.

[0182] ;

[0183] In the formula, It is the number of sampling points. This is the maximum number of sampling points allowed per plot of land. This is the minimum allowed number of sampling points for a single plot of land. This refers to the actual area of ​​the land parcel. The area of ​​a single remote sensing pixel. The sampling point density coefficient;

[0184] Candidate sampling points are generated using a system grid method. Priority is given to selecting the center of a cell that is of qualified quality and whose distance from the boundary of the plot is greater than a preset boundary threshold. Then, the selection is carried out sequentially under the condition that the distance between any two selected sampling points is not less than the preset minimum point spacing, until the target number of sampling points is reached or no candidate points that meet the conditions are found.

[0185] Let the set of indexes of the actually selected sampling points be denoted as The actual number of sampling points is:

[0186] ;

[0187] When the number of candidate points is insufficient, it is not forced to be supplemented to the target number; instead, the actual number selected is used in subsequent calculations. At that time, the corresponding land parcel will be in the time phase Marked as having no valid observations.

[0188] As a preferred embodiment, the method for determining temporal stability includes: obtaining the current observation time phase and the effective observation time phases adjacent to it to form a time window; calculating the coefficient of variation of each remote sensing feature of the land parcel within the time window and averaging them to obtain the temporal variation coefficient; and calculating the temporal stability score based on the temporal variation coefficient, wherein the temporal stability score is negatively correlated with the temporal variation coefficient.

[0189] To quantify the stability of site observations across adjacent time phases, a time window is constructed using the current time phase and the nearest valid time phase before and after it. Preferably, it includes 3 effective time phases. The average of the coefficients of variation for each feature is used to obtain the time phase variation coefficient of the plot:

[0190] ;

[0191] In the formula, The number of remote sensing indices in the feature set. For a fixed set of remote sensing indices, For a specific feature in the feature set, This represents the function for calculating the standard deviation. This represents the mean function. Indicates land parcel In features Upper time window The set of representation values ​​for all valid time phases within the range. A time window that includes the current time phase and the nearest valid time phase before and after it. To prevent small constants with a denominator of zero.

[0192] In practical remote sensing monitoring operations, observations from a single time phase are often affected by various random factors such as atmospheric conditions, sensor attitude, or the instantaneous physiological state of crops. It is difficult to distinguish between actual surface changes and observational noise based solely on single-point data. Therefore, introducing a time-dimensional stability assessment is crucial. Firstly, the system uses the current target observation time phase... A time window is constructed by searching for adjacent effective observation phases in front of and behind the center. The "effective observation phase" referred to here is the image phase whose data is confirmed to be usable after cloud detection, shadow masking, and quality labeling. In a preferred embodiment, the time window typically includes three effective phases (i.e., the current phase and one of the most recent effective phases before and after it). This setting ensures that there are sufficient time series samples to assess fluctuation trends while avoiding the inclusion of normal spectral changes caused by natural phenological evolution of crops due to excessively large time spans. This ensures that the calculated stability index mainly reflects the short-term observation noise level rather than long-term biological processes.

[0193] After constructing the time window, the coefficient of variation (CV) is calculated for each remote sensing feature (such as NDVI, EVI2, etc.) of the land parcel within that window. The CV is defined as the ratio of the standard deviation to the absolute mean of all valid temporal representations of that feature within the time window. The CV is used instead of the standard deviation because the numerical ranges and dimensions of different remote sensing indices or band reflectances vary greatly. As a dimensionless statistic, the CV can fairly measure the relative dispersion of different features. Subsequently, the CVs of all selected remote sensing features of the land parcel within the time window are arithmetically averaged to obtain the comprehensive temporal CV of the land parcel in the current time phase. This average value reflects the overall fluctuation intensity of the spectral signal of the land parcel in multiple dimensions.

[0194] Furthermore, the phase stability score is calculated based on the phase variation coefficient. This embodiment uses a reciprocal form of the mapping function to achieve the negative correlation; the specific calculation formula is as follows:

[0195] ;

[0196] In the formula, Indicates the first The first record unit The plot of land in the time phase The temporal stability score is as follows; This indicates that the land parcel was in the time phase The combined phase variation coefficient under the given conditions. The mechanism of this formula is as follows: when When the value approaches 0, it indicates that the spectral characteristics between adjacent time phases are highly consistent, and the observation is extremely stable; at this point, the score is [value missing]. It approaches the maximum value of 1; conversely, when An increase in the value indicates drastic fluctuations in the observed values, which may be due to residual clouds, shadow interference, or data anomalies. (Score) The score will decay rapidly and approach 0. This non-linear decay characteristic allows the system to impose significant penalty weights on highly volatile data while ensuring that the score always falls within the standardized range of [0, 1], which facilitates product fusion with area weights and the proportion of effective pixels in the plot.

[0197] Furthermore, considering the possibility of continuous rain or sensor malfunctions leading to scarce effective observations in practical applications, this embodiment also incorporates a degradation processing strategy to enhance the robustness of the solution. When the number of effective observation phases within a time window is less than a preset minimum threshold, the phase stability score is set to 1, and a phase insufficiency flag is added. Here, a score of 1 indicates that no phase stability penalty is applied in this weighting calculation, but does not imply that the site has the highest time series reliability.

[0198] Example 2: To verify the effectiveness of the multi-plot recording unit remote sensing adaptive sampling and weighted characterization method provided in the foregoing embodiments of the present invention, the following description uses the identification of early and late harvesting of Fuding white tea as a specific application scenario. In this scenario, the operational objective is to accurately distinguish the early and late first harvesting times of tea leaves in different recording units based on remote sensing observation data. This is of great significance for tea quality grading and market price prediction. It should be understood that although this embodiment uses Fuding white tea as an example, the technical solution of the present invention is also applicable to other remote sensing monitoring scenarios where one recording unit corresponds to multiple dispersed crop plots, such as orchards, medicinal herb planting areas, and economic forests.

[0199] In this application embodiment, the experimental data comes from high-resolution PlanetScope remote sensing imagery and corresponding tea garden plot vector data and production transaction records. Using "record unit-year" as the basic sample unit, classification labels are constructed based on the actual first-harvest date. Specifically, samples with a first-harvest date no later than the 76th day of the year are classified as the early-harvest group, containing 2435 samples; samples with a first-harvest date no earlier than the 80th day of the year are classified as the late-harvest group, containing 2745 samples; samples falling between the two thresholds are not included in this binary classification validation experiment. To eliminate the interference of different regional climate backgrounds and topographic conditions on remote sensing signals, the remote sensing descriptors within each cross-validation training layer were centralized and standardized using meteorological stations before modeling.

[0200] In contrast, a standard mean sampling method was used. This method directly averages remote sensing pixels within the same recording unit's associated plot area, or first calculates the remote sensing mean of each plot and then performs equal-weighted aggregation, without considering spatial distance between plots, area differences, the relationship between sampling location and boundaries, or outlier pixel removal based on local statistics of spatial clusters. In comparison, the method of this invention follows the process of Example 1 described above: first, spatial clustering of plots is performed based on the shortest boundary distance; second, the number of sampling points is adaptively determined and deployed according to the plot area; then, an adaptive sampling range constrained by boundary distance is generated for each sampling point, and multi-feature joint outlier removal and plot comprehensive effective pixel ratio calculation are performed within this range; finally, based on the area, plot comprehensive effective pixel ratio, and temporal stability-determined correction weights, the features are aggregated step-by-step into recording unit-level remote sensing representation values.

[0201] The two sampling strategies exhibit significant differences in three aspects: remote sensing descriptor differences, inter-group separation, and classification performance. First, under normal mean sampling conditions, the standardized means of most remote sensing descriptors from the early and late sampling groups are close, and the mean differences and their confidence intervals are mainly distributed near zero, indicating low distinguishability between the two types of samples on individual remote sensing descriptors. After applying the method of this invention, the means of multiple remote sensing descriptors from the early and late sampling groups are located on either side of zero, and the inter-group differences increase, indicating that the differences related to the early and late sampling times in remote sensing features are preserved after local circular domain screening, abnormal pixel removal, and hierarchical aggregation. Figure 6 This shows a comparison of the standardized means of various remote sensing indices under normal mean sampling conditions. Figure 7 The diagram shows a comparison of the grouped mean values ​​of multiple remote sensing indices under adaptive sampling according to the present invention.

[0202] Secondly, considering the characteristic distributions of the early and late sampling groups, under normal mean sampling, the inter-group effect size of the optimal single remote sensing descriptor is 0.238, while the inter-group effect size of the integrated remote sensing discrimination score alone is 0.619, indicating significant overlap between the two groups. After adopting the method of this invention, the inter-group effect size of the optimal single descriptor increases to 1.546, and the inter-group effect size of the integrated remote sensing discrimination score alone increases to 1.933. Furthermore, the discrimination scores of the early sampling group are mainly distributed in the lower range, while those of the late sampling group are mainly distributed in the higher range, significantly reducing the overlap between the two groups. This result demonstrates that the record unit-level representation formed by integrating multiple remote sensing descriptors has a higher ability to distinguish between early and late sampling than a single descriptor. Figure 8 This displays the box plot of the best single remote sensing descriptor feature under normal mean sampling conditions; Figure 9 The invention demonstrates the optimal single remote sensing descriptor feature boxplot under adaptive sampling. Figure 10 This displays a box plot of the integrated discriminant fractional features based solely on remote sensing under normal mean sampling conditions. Figure 11 The box plot of the remote sensing integrated discrimination score feature under adaptive sampling of the present invention is shown.

[0203] at last, Figure 12 The figure shows the ROC curve of binary cross-validation for early and late sampling under normal mean sampling conditions with only remote sensing feature input; Figure 13 The histogram of the AUC distribution under the normal mean sampling condition is shown. Figure 14 The diagram shows the ROC curve of early and late sampling binary classification cross-validation with only remote sensing feature input in this invention. Figure 15 The histogram of AUC distribution under the adaptive sampling of this invention with 5-fold cross-validation is shown. The 5-fold cross-validation results show that the AUC (area under the ROC curve) of the ordinary mean sampling method is 0.663, and the accuracy is 0.610; the AUC of the method of this invention is 0.888, and the accuracy is 0.824. Compared with ordinary mean sampling, the AUC of the method of this invention is improved by 0.225, and the accuracy is improved by 0.214. The ROC curve under the ordinary mean sampling condition is only slightly above the random classification baseline, while the ROC curve of the method of this invention is significantly shifted to the upper left, indicating that it has higher sorting ability for early and late sampling under different classification thresholds.

[0204] The random label permutation results further show that the AUCs obtained by the two methods under random label conditions are mainly distributed around 0.5, while the AUC of the method of the present invention reaches 0.888 under true label conditions, which is significantly higher than the random permutation results. Therefore, the classification performance obtained by the method of the present invention is not caused by random label combinations, but rather by the preservation of effective canopy spectral information by the remote sensing sampling and aggregation methods.

[0205] In summary, this invention reduces the impact of boundary mixed pixels, anomalous observations, and differences in contributions from multiple plots on remote sensing features by employing plot spatial clustering, area-adaptive point placement, adaptive sampling circles, outlier pixel removal based on local statistics of spatial clusters, calculation of the proportion of effective pixels across plots, and hierarchical weighted aggregation at the record unit level. While maintaining consistency in the original imagery, samples, labels, and classification process, this invention improves the remote sensing distinguishability between early and late samples of Fuding white tea, validating the effectiveness of the constructed record unit-level remote sensing characterization method.

[0206] Example 3: This example also provides a remote sensing adaptive sampling and weighted characterization device for multiple land parcel recording units. The device includes a memory and a processor. The memory stores a computer program; the processor executes the computer program to implement the remote sensing adaptive sampling and weighted characterization method for multiple land parcel recording units as described in the foregoing examples.

[0207] Specifically, the memory may include high-speed random access memory and non-volatile memory, such as one or more disk storage devices, flash memory devices, or other volatile solid-state storage devices. The processor may be a central processing unit, a general-purpose processor, a digital signal processor, an application-specific integrated circuit, a field-programmable gate array, or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. In some embodiments of this invention, the various functional modules in the device may be integrated into a processing unit, or each module may exist physically separately, or two or more modules may be integrated into a unit. The integrated unit described above can be implemented in hardware or as a software functional unit. When implemented as a software functional unit, the device can be independently deployed on a local server, edge computing node, or cloud service platform. By calling program instructions stored in the memory, it can automatically complete the entire process from acquiring image data of the target recording unit associated with the plot, spatial grouping, adaptive sampling quality control, to multi-maintenance positive weighted aggregation, thereby providing standardized remote sensing characterization data services for agricultural monitoring, phenological identification, and other businesses.

[0208] In addition, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the remote sensing adaptive sampling and weighted characterization method for multi-plot recording units as described in the foregoing embodiments.

[0209] Specifically, a computer-readable storage medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. This medium can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination thereof. For example, this medium includes, but is not limited to, portable computer disk enclosures, hard disks, random access memory, read-only memory, erasable programmable read-only memory, flash memory, portable compact disk read-only memory, optical data storage devices, etc. It should be understood that the computer-readable storage medium referred to in this invention does not include the transient propagation signal itself, but rather refers to a physical entity capable of persistently storing program instructions. In practical applications, this storage medium can be distributed as a software product or as a firmware update package in a cloud service provider's backend system, enabling various computing devices with the program installed to execute the aforementioned adaptive sampling and weighted representation method, thereby achieving comprehensive protection of the core technical solution in various business models such as software sales, cloud API calls, and embedded device integration.

[0210] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A remote sensing adaptive sampling and weighted representation method for multi-plot recording units, characterized in that, Includes the following steps: Step 1: Acquire remote sensing image data of multiple plots associated with the target recording unit; Step 2: Group the multiple land parcels based on the spatial distance between them to obtain at least one spatial cluster; Step 3: For each plot of land, determine the number of target sampling points based on the plot area and deploy the sampling points. Generate an adaptive sampling range based on the distance from the sampling points to the boundary of the plot. Step 4: Perform quality screening on remote sensing pixels within the adaptive sampling range, collect candidate pixels from each sampling range within the same spatial cluster, perform multi-index joint anomaly removal based on the local statistical distribution of the spatial cluster, determine the effective pixel set for each sampling range, and calculate the remote sensing feature value of the sampling point and the comprehensive effective pixel ratio of the plot. Step 5: Determine the correction weight for each plot based on the plot area, the proportion of effective pixels in the plot, and the temporal stability. Step 6: Based on the corrected weights, the remote sensing feature values ​​are aggregated step by step into the remote sensing representation values ​​of the target recording unit; Specific implementation methods for generating adaptive sampling range based on the distance from sampling points to land parcel boundaries include: Calculate the shortest distance from the sampling point to the boundary of the plot; The adaptive sampling radius is determined based on the nearest distance and a preset radius coefficient, wherein the adaptive sampling radius is not greater than the nearest distance. The formula for calculating the adaptive sampling radius is as follows: ； in, For the first The first record unit Plot No. The adaptive sampling circle radius for each sampling point For the maximum sampling radius, For radius coefficient, This is the shortest distance from the sampling point to the boundary of the plot; A sampling circle is generated with the sampling point as the center and the adaptive sampling radius as the radius, serving as the adaptive sampling range. The quality screening, anomaly removal, and determination of the effective pixel set for remote sensing pixels include: For each sampling circle, pixels whose center point falls within the sampling circle and are located within the boundary of the corresponding plot are taken as a geometric pixel set. Based on the remote sensing image quality label, pixels with clouds, shadows, missing measurements, and invalid values ​​are removed to obtain qualified candidate pixels. Quality-qualified candidate pixels from each sampling circle within the same spatial cluster are aggregated to form a local statistical pixel pool. Based on this pool, the local mean and local standard deviation of each remote sensing feature are calculated, and the standardized anomaly index of the candidate pixels is also calculated. The number of remote sensing features whose standardized anomaly index absolute value is greater than a preset anomaly threshold among the candidate pixels is counted. Candidate pixels whose number reaches the preset anomaly feature number threshold are judged as anomaly pixels, and the remaining candidate pixels constitute the effective pixel set. When the number of candidate pixels in the local statistical pixel pool is less than 2, the standardized anomaly judgment is not performed, and the qualified candidate pixels are taken as the effective pixel set. The comprehensive effective pixel ratio and remote sensing feature value of the sampling points are calculated as follows: Calculate the first The effective pixel ratio of each sampling point to the sampling circle: ； In the formula, For the first The effective pixel ratio of each sampling point to the sampling circle; For the first Each sampling point in time phase The corresponding set of geometric pixels; For the first The first record unit Plot No. Each sampling point in time phase The corresponding set of valid pixels; The comprehensive effective pixel ratio of the land parcel is calculated using the geometric pixels and effective pixels of all actually selected sampling circles as the statistical range: ； In the formula, The overall effective pixel ratio of the plot. For the plot of land in time phase The actual set of selected sampling point indices; Calculate the remote sensing feature values ​​for the sampling points that belong to the valid sampling point index set: ； In the formula, For remote sensing feature values, For pixels In time phase Next The values ​​of each remote sensing feature; For the first Each sampling point in time phase The corresponding set of valid sampling point indices.

2. The method according to claim 1, characterized in that, Determine the adjustment weights for each parcel, including: Using the plot area as the basic contribution weight, and correcting it using the plot's comprehensive effective pixel ratio and temporal stability score, the unnormalized weight of the plot is obtained: ； In the formula, The area of ​​the land parcel. The overall effective pixel ratio of the plot. The score is for phase stability. when empty or At that time, Set it to 0 and mark the plot as having no valid observations. The remaining plots are treated as valid plots and participate in the weight normalization within the spatial cluster.

3. The method according to claim 2, characterized in that, Remote sensing feature values ​​are aggregated step by step into remote sensing representation values ​​for the target recording unit, including: Based on the effective pixel ratio of effective sampling points within the plot, the remote sensing feature values ​​of the sampling points are aggregated into plot-level remote sensing feature values. Based on the unnormalized weights of each effective block within the spatial cluster, the block-level remote sensing feature values ​​are aggregated into spatial cluster-level remote sensing feature values. The sum of the unnormalized weights of the effective blocks within a spatial cluster is used as the total corrected weight of the spatial cluster. Based on the total corrected weight of each effective spatial cluster, the spatial cluster-level remote sensing feature values ​​are aggregated into the remote sensing representation values ​​of the target recording unit. When a spatial cluster does not contain any valid blocks or its total correction weight is 0, the spatial cluster is excluded from inter-cluster aggregation; when there is no valid spatial cluster in the target recording unit, the missing measurement value and the recording unit has no valid observation identifier are output.

4. The method according to claim 2, characterized in that, The methods for determining the phase stability include: The current observation phase and its adjacent valid observation phases are obtained to form a time window; The coefficients of variation of each remote sensing feature of the land parcel within the time window are calculated and averaged to obtain the temporal coefficient of variation. The formula for calculating the temporal coefficient of variation is as follows: ； in, The phase variation coefficient is . The number of remote sensing features, For remote sensing feature set, For a certain remote sensing feature, This represents the function for calculating the standard deviation. This represents the mean function. For the plot of land in terms of characteristics Upper time window The set of characterization values ​​for all valid observation phases within the scope. To prevent constants with a denominator of zero; A phase stability score is calculated based on the phase variation coefficient, wherein the phase stability score is negatively correlated with the phase variation coefficient, and the formula for calculating the phase stability score is as follows: ； in, The score represents the phase stability.

5. The method according to claim 1, characterized in that, Multiple land parcels are grouped based on their spatial distances to obtain at least one spatial cluster, including: Calculate the shortest distance between the boundaries of any two plots of land. The formula for calculating the shortest distance is: ； in, For the plot of land With the plot of land The shortest distance between the boundaries and These represent the boundaries of the two plots of land. and These are points on the boundaries of the two plots of land. This represents the Euclidean distance between two points; When the shortest distance is not greater than a preset distance threshold, an adjacency relationship is established between the corresponding two plots. The criteria for determining the adjacency relationship are as follows: ； in, The shortest distance to the boundary. The preset distance threshold; Based on the adjacency relationship, directly or indirectly connected land parcels are grouped into the same spatial cluster.

6. The method according to claim 1, characterized in that, For each plot of land, the number of target sampling points is determined based on the plot area, and the sampling points are deployed, including: The number of sampling points is converted to an integer using a rounding function and then limited to a preset range of minimum and maximum sampling points. ； In the formula, It is the number of sampling points. This is the maximum number of sampling points allowed per plot of land. This is the minimum allowed number of sampling points for a single plot of land. This refers to the actual area of ​​the land parcel. The area of ​​a single remote sensing pixel. The sampling point density coefficient; Let the set of indexes of the actually selected sampling points be denoted as The actual number of sampling points is: ； when At that time, the corresponding land parcel will be in the time phase Marked as having no valid observations.

7. A remote sensing adaptive sampling and weighted representation device for multi-plot recording units, characterized in that, include: Memory, used to store computer programs; A processor is configured to implement the remote sensing adaptive sampling and weighted representation method for multi-plot recording units as described in any one of claims 1 to 6 when executing the computer program.

Images