A knowledge-guided large-scale long-time-series crop regularized filling method for sparse period of remote sensing data

Abstract

The application discloses a kind of knowledge guiding type large-scale long time series crop regularized filling method for remote sensing data sparse period, comprising the following steps: step 1: define relevant terms and input-output parameters;Step 2: build leading planting mode automatic identification algorithm;Step 3;Step 4: output filling result;The present application deeply integrates planting mode and other agricultural knowledge, establishes a multi-scale leading planting mode automatic identification and coding system, builds a knowledge-guided crop regularized filling framework, realizes the identifiable, traceable and reusable of large-scale regional crops in historical years.Coupling crop temporal characteristics and spatial proximity characteristics before and after, it gets rid of the high dependence on remote sensing data, and the hierarchical multi-level regularized filling method based on the consistency of regional long-term planting mode effectively improves the spatio-temporal consistency and agricultural interpretability of crop filling results under the background of remote sensing sparse data.

Description

Technical Field

[0001] This invention relates to the field of crop spatiotemporal distribution reconstruction technology, specifically a knowledge-guided method for regularizing and filling in large-scale, long-term crop data. Background Technology

[0002] Long-term, high-precision spatiotemporal succession characteristics of crops are closely related to land security, food security, and ecological security. However, due to limitations such as intergenerational differences in sensors during the early stages of Earth observation and insufficient digitization of agricultural archives, it is often difficult to accurately obtain spatiotemporal distribution information of crops in sparse historical images. Data interpolation is a key step in high-precision crop distribution reconstruction under sparse image constraints. Existing technologies typically focus on image reconstruction, filling in missing pixels by constructing correlations between missing pixels and other pixels in specific spatial, spectral, and temporal domains. Examples include using Kriging interpolation to repair Landsat ETM+ stripes, local least squares fitting to repair missing MODIS spectral information, and Savitzky-Golay filtering to reconstruct time-series images of vegetation indices.

[0003] However, these traditional methods perform poorly in complex surface environments and scenarios with large-scale data gaps, easily producing noise artifacts, distorted spectral features, and unreliable spatial textures. Furthermore, they fail to accurately characterize spatiotemporal attributes such as crop planting structure and patterns, leading to significant uncertainty in the reconstruction results. Regional planting patterns are driven by multiple factors, including natural conditions, agricultural policies, and market factors. Their changes typically exhibit temporal continuity, providing a highly reliable and valuable constraint reference for large-scale crop data imputation. In the early stages of remote sensing development, my country's agricultural production was constrained by economic conditions and policy guidance, resulting in relatively stable overall planting areas for major crops. Changes in planting patterns exhibited a time lag, making crop types highly repetitive and predictable on an interannual scale. Explicitly introducing agronomical knowledge, such as typical planting patterns, into the data imputation process can constrain the rational direction of crop succession in the temporal dimension and strengthen the continuity and consistency of crop distribution in the spatial dimension, thereby effectively reducing the uncertainty in crop identification in data-gap areas. However, transforming planting pattern characteristics such as continuous cropping and crop rotation into an expressible and quantifiable imputation rule system is a crucial technical bottleneck in overcoming large-scale crop data imputation. Summary of the Invention

[0004] The purpose of this invention is to provide a knowledge-guided method for regularizing and filling in large-scale, long-term crop data, in order to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a knowledge-guided method for large-scale, long-term crop regularization and filling, comprising the following steps:

[0006] Step 1: Define relevant terms and input / output parameters. The terms include crop rotation, continuous cropping and corresponding quantification rules. The input / output parameters include annual classification raster set, cell set, crop category code set, annual mask to be filled, cell dominant pattern raster, pattern lookup table, village-level data and output parameters.

[0007] Step 2: Construct an automatic identification algorithm for dominant planting patterns, including pixel long-term series construction, continuous cropping pattern identification, crop rotation pattern identification, and dominant pattern determination;

[0008] Step 3: Construct a rule-based filling algorithm based on the dominant planting pattern, including the selection of pixels to be filled, pixel-scale filling, neighborhood-scale dominant planting pattern constraint, village-scale dominant planting pattern fallback filling, and result uncertainty labeling;

[0009] Step 4: Output the filling result.

[0010] Furthermore, the quantification rule for crop rotation in step 1 is as follows: a certain subsequence of a certain set of time series data repeats at least twice consecutively, and this subsequence is the crop rotation pattern of the set of time series data. The set of all consecutive subsequences is the longest subsequence applied to this pattern. The quantification rule for continuous cropping is as follows: a certain crop type of a certain set of time series data appears three or more consecutively, and the set of time series data is identified as having a continuous cropping pattern. The years in which the same crop is planted consecutively correspond to the longest continuous cropping subsequence of this set of time series data.

[0011] Furthermore, the expression for the annual classification raster set mentioned in step 1 is: ,in For a set of years, This represents the spatial distribution of crops in the image coverage area for any given year, where n represents the year span; the expression for the crop category coding set is: Where 1 = corn, 2 = soybean, 3 = rice, 4 = spring wheat, 0 represents other agricultural land cover, and NoData represents no data.

[0012] Furthermore, the expression for the annual mask to be filled in step 1 is: ,in This indicates that the cell needs to be filled; the expression for the dominant mode raster of the cell is:

[0013]

[0014] in, This indicates that the pixel has been identified as a valid planting pattern. This indicates that no valid planting pattern was identified for this pixel; the expression in the Pattern lookup table is:

[0015]

[0016] Among them, PatternTypeCode=1 indicates continuous operation, and PatternTypeCode=2 indicates rotation.

[0017] Furthermore, the expression for the pixel long-time sequence construction described in step 2 is as follows: ,in For a single pixel, the time sequence during the study period is used. Invalid values ​​and non-arable land are uniformly treated as 0. In the crop rotation pattern recognition, a sliding time window of 2-5 years is set to traverse the data and search for the smallest crop rotation unit with a time span of 2-5 years. When the crop rotation unit has at least 2 cycles, the smallest cycle unit is set as a candidate crop rotation planting pattern and the corresponding code is recorded.

[0018] Furthermore, the priority of the dominant mode determined in step 2 is "the mode with the longest duration and the most complete rotation and continuous cropping unit structure". When there are multiple candidate modes in a set of time series, the optimal mode is selected as the dominant planting mode of the cell according to this priority. If no continuous cropping or rotation structure that meets the conditions is identified, it is marked as "unidentified mode" with a value of -1.

[0019] Furthermore, the threshold for the village-scale dominant planting pattern constraint mentioned in step 3... The threshold is set to 0.5. When the proportion of the rotational cropping pattern in the administrative unit exceeds this threshold, the planting pattern with the highest proportion will be used to fill the gap.

[0020] Furthermore, step 3 includes neighborhood-scale constraint filling and village-scale catch-all filling, which include: defining spatial neighborhood-scale and village-scale dominant planting pattern rules; firstly, constructing a neighborhood-scale dominant planting pattern; and when no effective planting pattern is identified at the neighborhood scale, using a village-scale dominant planting pattern as a catch-all.

[0021] Furthermore, the uncertainty of the result mentioned in step 3 is marked as follows: the filling result is written into the target year raster, and the filling source type and confidence level are recorded for each cell. The filling source type includes cell-dominant mode filling, neighborhood-scale dominant mode filling, and village-scale dominant mode filling. The confidence level is divided into five levels, with the highest confidence level being 5 and the lowest being 1.

[0022] Compared with the prior art, the beneficial effects of the present invention are:

[0023] This invention deeply integrates agronomical knowledge, such as strong spatiotemporal constraints on planting patterns, to construct a multi-level automatic identification and coding system for dominant planting patterns. This enables the traceability, verifiability, and reusability of dominant planting patterns across large-scale regions. By coupling the temporal and spatial proximity characteristics of crops, it eliminates the heavy reliance on remote sensing data. Based on planting pattern characteristics such as regional crop rotation, it constructs a progressive, rule-based filling method, effectively improving the spatiotemporal consistency and agronomic interpretability of the crop filling results.

[0024] Compared to traditional image reconstruction methods, the algorithm proposed in this invention has stronger spatiotemporal generalization capabilities, effectively addressing the need for crop distribution reconstruction during periods of sparse imagery. Furthermore, it does not rely on a large number of ground truth samples and can be applied across regions. Simultaneously, by synchronously recording the source type and confidence level of the infilling, the infilling results possess traceability and interpretability, providing a reliable, implementable, and scalable common key technology for efficient crop infilling under sparse data constraints. Based on the constructed infilling framework, it supports the reconstruction of the temporal distribution of crops during this period, thereby providing a long-term benchmark reference for the establishment of my country's "40+ years" agricultural information database, grain yield prediction, and agricultural land system resilience assessment. Attached Figure Description

[0025] Figure 1 This is a schematic diagram showing the filling results and comparison results of a typical area to be filled in according to the present invention;

[0026] Figure 2 This is a flowchart of the method of the present invention. Detailed Implementation

[0027] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1-2 This invention provides a knowledge-guided method for large-scale, long-term crop regularization and filling, comprising the following steps:

[0029] Step 1: Define relevant terminology and input / output parameters

[0030] 1) Definition of Rotation and Continuous Crop Planting Pattern

[0031] Crop rotation: planting different crops in a certain order on the same field in different years. The rotation cycle is the number of years required to implement one cycle. 2-5 years is a short cycle, and more than 6 years is a long cycle.

[0032] Continuous cropping: planting the same crop on the same piece of land one after another;

[0033] Rotation quantization rule: If a certain subsequence of a set of time series data repeats at least twice consecutively, the subsequence is identified as the rotation pattern of the set of time series data, and the set of all consecutive subsequences is the longest subsequence applied to the pattern;

[0034] Continuous cropping quantification rule: If a certain crop type appears three or more times consecutively in a set of time series data, that is, the same crop is planted for three or more consecutive years, the set of time series data is identified as having a continuous cropping pattern. All the years in which the same crop is planted consecutively correspond to the longest subsequence of continuous cropping in this set of time series data.

[0035] 2) Data structure definition

[0036] Annual category grid collection:

[0037] (1)

[0038] in, For a set of years, This represents the spatial distribution of crops in the area covered by the image in any year, where n represents the year span.

[0039] Pixel set: The pixel set of the study area is denoted as Ω, and any pixel is Ω. ;

[0040] Crop category code set:

[0041] (2)

[0042] Where 1 = corn, 2 = soybean, 3 = rice, 4 = spring wheat; 0 represents other agricultural land cover; NoData indicates no data.

[0043] Annual mask to be filled:

[0044]

[0045] A value of 1 indicates that the cell needs to be filled.

[0046] Pixel-dominant pattern raster (PatternID):

[0047] (4)

[0048] In the above formula, a PatternID greater than zero indicates that the cell has been identified as a valid planting pattern, and a value of -1 indicates that the cell has not been identified as a valid planting pattern.

[0049] Pattern lookup table: Given a PatternID, output the minimum repeating unit (MinimalUnit) and the pattern type code (PatternTypeCode):

[0050] (5)

[0051] 3) Definition of neighborhood scale

[0052] Pixel-scale dominant planting patterns may exhibit significant uncertainty in areas with sparse remote sensing data or unstable institutional systems. To enhance the spatial consistency and institutional rationality of crop infilling results, this study introduces a spatial neighborhood-scale dominant planting pattern identification rule. This rule statistically summarizes pixel-scale dominant planting patterns within a certain spatial range, extracting regionally representative dominant institutional types for institutional constraints and corrections to low-confidence pixel infilling results. Following the principle of "from near to far, from small to large, and gradually increasing constraints," this rule identifies dominant planting patterns within multi-scale spatial neighborhoods and participates in subsequent infilling decisions through coarse classification.

[0053] 4) Village-level data definition

[0054] Village boundary vector: , pixel The village to which it belongs is ;

[0055] Village-scale dominant planting pattern table: for villages Given a village-led planting model Village model type Village rotation / continuous cropping area percentage .

[0056] Output parameter definition for each year Construct the set of cells that need to be filled:

[0057] (6)

[0058] Step 2: Construct an automatic identification algorithm for dominant planting patterns

[0059] S1. Pixel Long-Term Series Construction: For each pixel, extract its crop classification results for each year within the study period (e.g., 1990-2000) to construct a time series of length T.

[0060] (7)

[0061] Invalid values ​​and non-arable land are uniformly treated as 0.

[0062] S2. Continuous cropping pattern recognition: Based on time series data, identify crops that have appeared continuously for no less than 3 years for the same crop type. If they exist, record the crop type as a candidate continuous cropping pattern, record the continuous cropping pattern code, and record the longest duration at the same time.

[0063] S3. Crop Rotation Pattern Recognition: Based on time series data, a sliding time window of 2-5 years is set, and the data is traversed sequentially to search for the smallest crop rotation unit with a time span of 2-5 years. It is determined whether a certain crop rotation unit has at least 2 cycles (meeting the crop rotation condition). If it does, the smallest cycle unit is set as a candidate crop rotation planting pattern, and the corresponding planting pattern code is recorded.

[0064] S4. Dominant mode determination: When there are multiple candidate modes in a time series, they are sorted according to the priority of "longest mode duration and most complete rotation and crop rotation unit structure", and the optimal mode is selected as the dominant planting mode of the cell. If no rotation or crop rotation structure that meets the conditions is identified, it is marked as "unidentified mode" with a value of -1.

[0065] Step 3: Construct a rule-based filling algorithm based on dominant patterns

[0066] S1. Filtering of pixels to be filled: targeting a specific year Based on the data to be filled, the set of pixels to be filled is extracted on a yearly basis.

[0067] S2, Pixel-scale filling: For each pixel in the set, read its dominant planting pattern. If the pattern is an effective rotation or continuous cropping structure, then based on the pattern cycle and under the constraint of the observations of known years, determine the corresponding crop type in the pattern for the year to be filled.

[0068] S3. Neighborhood Scale Filling: For pixels where no dominant planting pattern is identified at the pixel scale, constraint filling is performed using the dominant planting pattern at the neighborhood scale. Square neighborhood windows of different scales (3*3, 5*5, 11*11, 17*17) are constructed, and the dominant planting patterns at different scales are statistically analyzed for constraint filling. The main principles are the minimum finite scale principle and the institutional stability constraint principle, to preserve spatial details to the greatest extent possible.

[0069] S4. Village-level Dominant Planting Pattern Constraints: When high-confidence crop imputation cannot be carried out at the pixel and neighborhood scales, this study introduces village-level dominant planting pattern identification rules as the final catch-all constraint level in the rule-based imputation framework. The area proportion of different planting systems within the administrative unit is statistically analyzed. If the proportion of rotational cropping exceeds a threshold... If the value is set to 0.5, then the planting mode with the highest percentage will be used to fill the gap.

[0070] S5. Uncertainty Marking of Results: The filling results are written into the target year raster, and the filling source type and confidence level are recorded for each cell for subsequent uncertainty analysis. The filling source type includes cell-dominant mode filling, neighborhood-scale dominant mode filling, and village-scale dominant mode filling. The confidence level is divided into five levels, with the highest confidence level being 5 and the lowest being 1.

[0071] Step 4: Output Results

[0072] The constructed infilling algorithm was applied to carry out crop infilling experiments, using crop distribution in the image-covered area of ​​the study area as the basic data and the missing areas in the image as the data to be filled. The results of the spatiotemporal distribution of crops after infilling were output.

[0073] Example:

[0074] The Northeast region was selected as the study area from 1990 to 2000. Based on the crop distribution in the image-covered areas of this region, and with missing areas in the images as data to be filled, a crop filling experiment was conducted. The specific steps are as follows:

[0075] Step 1: Define relevant terminology and input / output parameters

[0076] The definition of crop rotation and continuous cropping patterns: Rotation means planting different crops in a certain order on the same field in different years, and continuous cropping means planting the same crop continuously on the same piece of land. The quantification rule for crop rotation is that a certain subsequence of a certain time series data is repeated at least twice, and that subsequence is a crop rotation pattern. The quantification rule for continuous cropping is that a certain crop type of a certain time series data appears three or more times consecutively.

[0077] Data structure definition: Annual categorized raster set n=10, corresponding to the crop category code set from 1990 to 2000. (1=Corn, 2=Soybean, 3=Rice, 4=Spring Wheat), Annual Masks to be Filled Pixel-dominated mode raster In the Pattern lookup table, PatternTypeCode=1 corresponds to consecutive operations, and PatternTypeCode=2 corresponds to rotation operations.

[0078] Village-level data definition: Village boundary vector Statistics on the percentage of rotational / continuous cropping area in each village .

[0079] Output parameter definition: Construct the output results of the pixels to be filled in each year according to formula (6). .

[0080] Step 2: Construct an automatic identification algorithm for dominant planting patterns

[0081] Pixel long-time series construction: for each pixel in the study area Extract the crop classification results for each year from 1990 to 2000, and construct a time series according to formula (7). Invalid values ​​and non-arable land are uniformly treated as 0.

[0082] Continuous cropping pattern recognition: Traverse the time sequence of each cell to identify cases where the same crop type appears continuously for ≥3 years. For example, if a cell was continuously planted with corn from 1990 to 1993 (coded 1), then corn is recorded as a candidate continuous cropping pattern for that cell, with the longest duration being 4 years.

[0083] Crop rotation pattern recognition: Set a sliding time window of 2-5 years, traverse the time series, search for the smallest crop rotation unit. If the "soybean-corn" (code 2-1) subsequence in a certain pixel time series appears in two consecutive cycles in 1990-1993 and 1994-1997, then "soybean-corn" is set as a candidate crop rotation planting pattern.

[0084] Dominant mode determination: The dominant mode is selected based on the priority of "longest duration of the mode and most complete rotation and continuous cropping unit structure". If a cell has both continuous cropping and rotation candidate modes, and the rotation mode has a longer duration, then the rotation mode is determined as the dominant planting mode for that cell; cells where no valid mode is identified are marked as... =−1.

[0085] Step 3: Construct a rule-based filling algorithm based on dominant patterns

[0086] Pixels to be filled: For each year t from 1990 to 2000, extract... The set of pixels with a value of 1 is the set of pixels to be filled.

[0087] Pixel-scale filling (one round): For each pixel to be filled, its dominant cropping pattern is read. If it is an effective crop rotation or continuous cropping structure, the crop type for the year to be filled is determined according to the pattern cycle. For example, if the dominant crop rotation pattern of a pixel is "soybean-corn" (2-year cycle), and 1990 was soybean and 1991 was corn, then the value to be filled in 1992 would be soybean and 1993 would be corn, and so on.

[0088] Neighborhood-scale dominant planting pattern constraint filling (second round): To For pixels with a value of -1, the dominant planting pattern type in the neighborhood at different scales is statistically analyzed. The scale is determined step by step from small to large. If it is a continuous cropping type, it is a strong constraint and is directly filled according to this type. If it is a stable crop rotation type (e.g., corn-soybean), it is filled according to the rotation. If it is a multi-cropping rotation or is not determined, it proceeds to the next round of filling.

[0089] Village-scale dominant cropping pattern fallback (three-round): For pixels where fallback cannot be implemented at the neighborhood scale, fallback is provided based on the village-scale dominant cropping pattern. The percentage of rotational cropping area within the village is statistically analyzed. ,like If the percentage is greater than 0.5, then the village's most prevalent planting pattern will be used to fill in the missing pixels. For example, if a village has a corn rotation area of ​​0.6%, which exceeds the threshold of 0.5, then the pixels in that village that do not have a dominant pattern will be filled in as corn.

[0090] Uncertainty marker for results: The filling results are written to the raster of the corresponding year, and the filling source type and confidence level of each cell are recorded. The source types include: cell-dominant mode filling (code 1), neighborhood-scale filling (code 2), and village-scale dominant mode filling (code 3). The confidence level is divided into 5 levels from high to low, with the highest confidence level being 5 and the lowest being 1.

[0091] Step 4: Output Results

[0092] Output the spatiotemporal distribution of crops in Northeast China from 1990 to 2000 after data filling. Verify the effectiveness of the method by comparing it with provincial statistical yearbook data. Typical regional filling results are shown below. Figure 1 As shown, this effectively highlights the characteristics of interannual maize rotation.

[0093] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 knowledge-guided method for large-scale, long-term crop regularization and filling, characterized in that: Includes the following steps: Step 1: Define relevant terms and input / output parameters. The terms include crop rotation, continuous cropping and corresponding quantification rules. The input / output parameters include annual classification raster set, cell set, crop category code set, annual mask to be filled, cell dominant pattern raster, pattern lookup table, village-level data and output parameters. Step 2: Construct an automatic identification algorithm for dominant planting patterns, including pixel long-term series construction, continuous cropping pattern identification, crop rotation pattern identification, and dominant pattern determination; Step 3: Construct a rule-based filling algorithm based on the dominant planting pattern, including the selection of pixels to be filled, pixel-scale filling, neighborhood-scale dominant planting pattern constraint, village-scale dominant planting pattern fallback filling, and result uncertainty labeling; Step 4: Output the filling result.

2. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The quantification rule for crop rotation in step 1 is as follows: a certain subsequence of a certain time series data is repeated at least twice, and this subsequence is the crop rotation pattern of the time series data. The set of all consecutive subsequences is the longest subsequence applied to this pattern. The quantification rule for continuous cropping is as follows: a certain crop type of a certain time series data appears three or more times consecutively, and the time series data is considered to have a continuous cropping pattern. The years in which the same crop is planted consecutively correspond to the longest continuous cropping subsequence of this time series data.

3. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The expression for the annual classification raster set mentioned in step 1 is: ,in For a set of years, This represents the spatial distribution of crops in the image coverage area for any given year, where n represents the year span; the expression for the crop category coding set is: Where 1 = corn, 2 = soybean, 3 = rice, 4 = spring wheat, 0 represents other agricultural land cover, and NoData represents no data.

4. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The expression for the annual mask to be filled in step 1 is: ,in This indicates that the cell needs to be filled; the expression for the dominant mode raster of the cell is: in, This indicates that the pixel has been identified as a valid planting pattern. This indicates that no valid planting pattern was identified for this pixel; the expression in the Pattern lookup table is: Among them, PatternTypeCode=1 indicates continuous operation, and PatternTypeCode=2 indicates rotation.

5. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The expression for the pixel long-time series construction in step 2 is: ,in For a single pixel, the time sequence during the study period is used. Invalid values ​​and non-arable land are uniformly treated as 0. In the crop rotation pattern recognition, a sliding time window of 2-5 years is set to traverse the data and search for the smallest crop rotation unit with a time span of 2-5 years. When the crop rotation unit has at least 2 cycles, the smallest cycle unit is set as a candidate crop rotation planting pattern and the corresponding code is recorded.

6. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The priority for determining the dominant mode in step 2 is "the mode with the longest duration and the most complete crop rotation unit structure". When there are multiple candidate modes in a set of time series, the optimal mode is selected as the dominant planting mode of the cell according to this priority. If no crop rotation or continuous cropping structure that meets the conditions is identified, it is marked as "unidentified mode" with a value of -1.

7. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The threshold constrained by the village-scale dominant planting pattern mentioned in step 3 The threshold is set to 0.

5. When the proportion of the rotational cropping pattern in the administrative unit exceeds this threshold, the planting pattern with the highest proportion will be used to fill the gap.

8. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: Step 3, which involves filling in neighborhood-scale constraints and village-scale catch-all conditions, includes defining the dominant planting pattern rules at the spatial neighborhood and village scales. First, a neighborhood-scale dominant planting pattern is established. When no effective planting pattern is identified at the neighborhood scale, a village-scale dominant planting pattern is adopted as a backup.

9. The knowledge-guided large-scale long-term crop regularization filling method according to claim 1, characterized in that: The uncertainty of the result mentioned in step 3 is marked as follows: the filling result is written into the target year raster, and the filling source type and confidence level are recorded for each cell. The filling source type includes cell-dominant mode filling, neighborhood-scale dominant mode filling, and village-scale dominant mode filling.

Images