Tilled land change detection and identification method and device based on time sequence feature vector

Abstract

The invention relates to a tilled land change detection and identification method and device based on a time sequence feature vector, belongs to the technical field of agricultural tilled land monitoring, and solves the problems that a tilled land change time node cannot be accurately determined and a part of tilled land change remote sensing weak signals are difficult to judge in an existing remote sensing technology. According to the tilled land change detection and identification method based on the time sequence feature vector, after remote sensing image data and multi-source land utilization data are obtained, multi-period harmonic fitting of the remote sensing image data is carried out based on a time domain sliding window, tilled land change vectors are constructed, and relative change vectors of the tilled land change vectors in adjacent periods are calculated; determining a potential cultivated land change area; dividing the potential cultivated land change area into a cultivated land area and a non-cultivated land area based on the training sample set, and generating a corresponding cultivated land classification result and a cultivated land probability; and combining cultivated land classification results to form a cultivated land change result of a continuous time sequence, and correcting the cultivated land change result in combination with the cultivated land probability. The detection accuracy of the cultivated land change dynamic state is improved.

Description

Technical Field

[0001] The present application relates to the technical field of agricultural cultivated land monitoring, and particularly to a method and device for detecting and identifying cultivated land changes based on time series feature vectors. Background Art

[0002] Cultivated land is the land formed by the development of natural soil and capable of growing crops, with a natural environment suitable for the growth, development, and maturity of crops, and the cultivated land will change accordingly with the changes in the natural environment. Therefore, accurate and timely information on cultivated land changes and change types is of great significance for cultivated land monitoring, food security, and regional sustainable development.

[0003] Traditional cultivated land monitoring is mostly carried out in administrative regions, combined with manual visual interpretation and field recheck and verification. Although relatively high monitoring accuracy can be obtained, this method is time-consuming and laborious and cannot meet the needs of large-scale, rapid, and timely agricultural monitoring. Remote sensing technology can efficiently and quickly obtain information on large-area cultivated land changes and is an important means for identifying large-area cultivated land changes and cultivated land spatial distribution. At present, existing methods for identifying cultivated land changes mainly include the post-classification comparison method, image differencing method, and time series change detection method. The first method is to separately generate annual land use classification maps and detect changes by comparing the results. The second method is to detect changes by first extracting remote sensing information in the time series without classification and then classifying the changed areas, including the threshold method, segmentation method, and statistical boundary method. The third method is to use time series data for change monitoring analysis to judge cultivated land changes and capture the time of change occurrence, and based on statistical hypotheses, realize change detection by constructing a spatio-temporal change model, such as the LandTrendr algorithm, BFAST algorithm, and CCDC algorithm.

[0004] However, existing methods for identifying cultivated land changes focus on confirming the occurrence of changes through the deviation of the phenological trajectories of predicted values and observed values, and are insufficient in extracting cultivated land changes with weak signals such as small change amplitudes or similar phenology before and after changes; in addition, the response delay of the spectral change peak caused by the transition period of land cover change results in the lag of change detection time; cultivated land changes are jointly affected by natural elements and human activities, and cultivated land with different crop categories, planting systems, and climate environments has complex spectral and time characteristics, making it difficult to effectively distinguish cultivated land changes from environmental noise and pseudo-changes. Therefore, existing methods for identifying cultivated land changes cannot accurately determine the time nodes of cultivated land changes and are difficult to distinguish some weak remote sensing signals of cultivated land changes. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a method and device for detecting and identifying cultivated land changes based on time series feature vectors, so as to solve the technical problems that the existing cultivated land change identification methods cannot accurately determine the time nodes of cultivated land changes and it is difficult to distinguish weak remote sensing signals of some cultivated land changes.

[0006] An embodiment of the present application provides a method for detecting and identifying cultivated land changes based on time series feature vectors, including the steps of: Obtain remote sensing image data and multi-source land use data; among them, fuse the multi-source land use data to obtain a training sample set of cultivated land and non-cultivated land; Perform multi-period harmonic fitting on the remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, calculate the relative change vector of the cultivated land change vectors in adjacent time periods, determine the potential cultivated land change area based on the comparison result between the relative change vector and a preset recognition threshold, and determine the year when the potential cultivated land change occurs based on the peak value of the relative change vector; Classify the potential cultivated land change area into cultivated land areas and non-cultivated land areas based on the training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities; Synthesize the cultivated land classification results to form a cultivated land change result in a continuous time series, and correct the cultivated land change result in combination with the cultivated land probability.

[0007] For the method for detecting and identifying cultivated land changes based on time series feature vectors in the embodiment of the present application, after obtaining the remote sensing image data and multi-source land use data, perform multi-period harmonic fitting on the remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, calculate the relative change vector of the cultivated land change vectors in adjacent time periods, determine the potential cultivated land change area based on the comparison result between the relative change vector and a preset recognition threshold, and determine the year when the potential cultivated land change occurs based on the peak value of the relative change vector. Classify the potential cultivated land change area into cultivated land areas and non-cultivated land areas based on the training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities; synthesize the cultivated land classification results to form a cultivated land change result in a continuous time series, and correct the cultivated land change result in combination with the cultivated land probability. Realize data fusion of long time series based on the time-domain sliding window, solve problems such as missed recognition of weak signals and time deviation of change detection, and improve the detection accuracy of cultivated land change dynamics.

[0008] In an embodiment of the present application, when performing multi-period harmonic fitting on the remote sensing image data based on a time-domain sliding window, the following formula is used:

[0009] Wherein, represents any pixel within the maximum cultivated land range; Denote any Julian day with remote sensing observations, including all Julian days of remote sensing observations from year t to year t + w of the observation year; w is the time window for function fitting. Denote The predicted value of the NDVI index of the pixel at time ; Is the time The remote sensing observation value of the NDVI index; Is the number of days in a year; Is the overall amplitude or intercept, reflecting the average growth status of vegetation during the period from year t to year t + w; Is the slope, reflecting the inter-annual change of vegetation from year t to year t + w; Is the order of the harmonic component, reflecting the intra-annual change of vegetation at different frequencies, where And , And , And Reflect the intra-annual changes once a year, twice a year, and three times a year respectively.

[0010] In an embodiment of the present application, the relative change vector of the cultivated land change vector in adjacent time periods is calculated by the following formula:

[0011] Wherein, Is The standardized relative change vector of the pixel in year t, Represents the standardized vector difference of the i-th pixel in year t, And Are respectively the harmonic function fitting vectors of the i-th pixel in the two time periods from year t to year t + w and From year to year t; The calculated parameters include the intercept of the harmonic function from year t to year t + w and 7 harmonic fitting parameters of the first, second, and third orders, namely ; The calculated parameters include The intercept of the harmonic function from year to year t and 7 harmonic fitting parameters of the first, second, and third orders, namely Are respectively the intercept and root mean square error of the harmonic function fitting in the two time periods from year t to And From year to year t.

[0012] In an embodiment of the present application, the following two formulas are used to determine the year when potential cultivated land changes occur:

[0013] Among them, CT represents a preset recognition threshold, and the area that satisfies the conditions of the above two formulas simultaneously is the potential cultivated land change area. They are the standardized relative change vectors of year t-2, t-1, t, t+1, and t+2 respectively. When the above two formulas are met, the standardized relative change vector in year t is greater than that in the other years t-2, t-1, t+1, and t+2; then year t is the year when potential cultivated land changes occur.

[0014] In an embodiment of the present application, the multi-source land use data includes CLCD data and GLAD data; The process of fusing multi-source land use data to obtain the training sample sets of cultivated land and non-cultivated land includes the steps of: Reclassify the CLCD data into land use types; Taking the four years of each GLAD product as the interval time period, superimposing the CLCD data of the same time period to obtain the area of the same land use type within this period, and then superimposing it with the cultivated land and non-cultivated land ranges of the corresponding time period of the GLAD product to generate the areas where the cultivated land and non-cultivated land are completely consistent, and correspondingly construct the training sample sets of cultivated land and non-cultivated land.

[0015] In an embodiment of the present application, before the process of constructing the training sample sets of cultivated land and non-cultivated land, it further includes the steps of: Perform resolution resampling on the areas where the cultivated land and non-cultivated land are completely consistent to construct a cell grid; Construct and optimize the remote sensing image features based on the multi-source land use data, and perform training sample migration.

[0016] In an embodiment of the present application, the process of correcting the cultivated land change result by combining the cultivated land probability includes the steps of: If there is a significant difference between the cultivated land probability and other cultivated land probability segments in the time series, then retain the cultivated land change result, otherwise eliminate it; among them, the significant difference is judged based on the T-test of the shapelet algorithm.

[0017] The embodiment of the present application also provides a device for detecting and identifying cultivated land changes based on a time series feature vector, including: A data acquisition module, used to acquire remote sensing image data and multi-source land use data; among them, fuse the multi-source land use data to obtain the training sample sets of cultivated land and non-cultivated land; A data fusion module, used to perform multi-period harmonic fitting on the remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, calculate the relative change vector of the cultivated land change vectors in adjacent time periods, determine the potential cultivated land change area based on the comparison result of the relative change vector and the preset recognition threshold, and discriminate the year when potential cultivated land changes occur based on the peak value of the relative change vector; A data classification module, configured to classify potential cultivated land change regions into cultivated land regions and non-cultivated land regions based on a training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities; A data detection module, configured to synthesize the cultivated land classification results to form cultivated land change results of a continuous time series, and correct the cultivated land change results in combination with the cultivated land probabilities.

[0018] The cultivated land change detection and recognition device based on a time series feature vector according to an embodiment of the present application, after obtaining remote sensing image data and multi-source land use data, performs multi-period harmonic fitting on the remote sensing image data based on a time domain sliding window to construct a cultivated land change vector, and calculates a relative change vector of the cultivated land change vectors in adjacent time periods, determines a potential cultivated land change region based on the comparison result between the relative change vector and a preset recognition threshold, and discriminates the year when the potential cultivated land change occurs based on the peak value of the relative change vector. Classify the potential cultivated land change regions into cultivated land regions and non-cultivated land regions based on a training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities; synthesize the cultivated land classification results to form cultivated land change results of a continuous time series, and correct the cultivated land change results in combination with the cultivated land probabilities. Realize data fusion of a long time series based on a time domain sliding window, solve problems such as missed recognition of weak signals and time deviation of change detection, and improve the detection accuracy of the cultivated land change dynamics.

[0019] At least one embodiment of the present application further provides a data control device, including: One or more memories, non-transiently storing computer-executable instructions; One or more processors, configured to run the computer-executable instructions, wherein when the computer-executable instructions are run by the one or more processors, the cultivated land change detection and recognition method based on a time series feature vector according to any embodiment of the present application is implemented.

[0020] The above data control device, after obtaining remote sensing image data and multi-source land use data, performs multi-period harmonic fitting on the remote sensing image data based on a time domain sliding window to construct a cultivated land change vector, and calculates a relative change vector of the cultivated land change vectors in adjacent time periods, determines a potential cultivated land change region based on the comparison result between the relative change vector and a preset recognition threshold, and discriminates the year when the potential cultivated land change occurs based on the peak value of the relative change vector. Classify the potential cultivated land change regions into cultivated land regions and non-cultivated land regions based on a training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities; synthesize the cultivated land classification results to form cultivated land change results of a continuous time series, and correct the cultivated land change results in combination with the cultivated land probabilities. Realize data fusion of a long time series based on a time domain sliding window, solve problems such as missed recognition of weak signals and time deviation of change detection, and improve the detection accuracy of the cultivated land change dynamics.

[0021] At least one embodiment of the present application further provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, a cultivated land change detection and recognition method based on a temporal feature vector according to any embodiment of the present application is implemented.

[0022] For the above non-transitory computer-readable storage medium, after obtaining remote sensing image data and multi-source land use data, multi-temporal harmonic fitting of the remote sensing image data is performed based on a time-domain sliding window to construct a cultivated land change vector, and a relative change vector of the cultivated land change vectors in adjacent time periods is calculated. The potential cultivated land change area is determined based on the comparison result between the relative change vector and a preset recognition threshold, and the year when the potential cultivated land change occurs is discriminated based on the peak value of the relative change vector. The potential cultivated land change area is classified into a cultivated land area and a non-cultivated land area based on a training sample set, and corresponding cultivated land classification results and cultivated land probabilities are generated; the cultivated land classification results are synthesized to form a cultivated land change result in a continuous time series, and the cultivated land change result is corrected in combination with the cultivated land probability. Long-time series data fusion is realized based on the time-domain sliding window, problems such as missed recognition of weak signals and time deviation in change detection are solved, and the detection accuracy of the cultivated land change dynamics is improved. Description of the Drawings

[0023] Figure 1 It is a flowchart of a cultivated land change detection and recognition method based on a temporal feature vector according to an embodiment of the application; Figure 2 It is a module structure diagram of a cultivated land change detection and recognition device based on a temporal feature vector according to an embodiment of the application; Figure 3 It is a schematic block diagram of a data control device provided by the present invention; Figure 4 It is a schematic diagram of a non-transitory computer-readable storage medium provided by the present invention. Detailed Embodiments

[0024] In order to make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the embodiments of the present application will be clearly and completely described below in conjunction with the drawings of the embodiments of the present application. Obviously, the described embodiments are some, rather than all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the described embodiments of the present application without creative efforts shall fall within the scope of protection of the present application.

[0025] Unless otherwise defined, technical terms or scientific terms used in this application shall have the ordinary meanings as understood by those with ordinary skills in the field to which this application pertains. The terms "first", "second" and similar words used in this application do not denote any order, quantity or importance, but are only used to distinguish different components. Words such as "comprising" or "including" mean that the elements or objects appearing before this word cover the elements or objects listed after this word and their equivalents, without excluding other elements or objects. Words such as "connected" or "linked" are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Upper", "lower", "left", "right", etc. are only used to indicate relative positional relationships, and when the absolute position of the object being described changes, the relative positional relationship may also change accordingly.

[0026] To keep the following description of the embodiments of this application clear and concise, detailed descriptions of some known functions and known components are omitted in this application.

[0027] Figure 1 It is a flowchart of a cultivated land change detection and recognition method based on a time-series feature vector according to an embodiment of an application, as Figure 1 shown. A cultivated land change detection and recognition method based on a time-series feature vector according to an embodiment of an application includes steps S100 to S103: S100, obtaining remote sensing image data and multi-source land use data; wherein, fusing multi-source land use data to obtain a training sample set of cultivated land and non-cultivated land; S101, performing multi-period harmonic fitting on remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, calculating a relative change vector of the cultivated land change vectors in adjacent time periods, determining a potential cultivated land change area based on the comparison result between the relative change vector and a preset recognition threshold, and discriminating the year when potential cultivated land changes occur based on the peak value of the relative change vector; S102, classifying the potential cultivated land change area into a cultivated land area and a non-cultivated land area based on the training sample set, and generating corresponding cultivated land classification results and cultivated land probabilities; S103, synthesizing the cultivated land classification results to form a cultivated land change result in a continuous time series, and correcting the cultivated land change result in combination with the cultivated land probability.

[0028] In the embodiment of this application, the remote sensing image data is the remote sensing images of the study area in a long time series. The cultivated land recognition scope of this application does not exceed the study area.

[0029] In a preferred embodiment, the remote sensing image data includes orthorectified surface reflectance images of Landsat-5 TM, Landsat-7 ETM+, and Landsat-8 OLI with a long time series. By way of example, the image time range of the selected remote sensing image data is from 1986 to 2023. To ensure pixel quality, the QA band from the CFMask algorithm is used to remove observation data of poor quality such as clouds, shadows, snow / ice, etc.; for pixels with missing valid observations for four consecutive months, the nearest valid observations before and after are searched, and the linear interpolation algorithm is used to supplement the pixels with missing valid observations.

[0030] In a preferred embodiment, the multi-source land use data includes GLAD data (cultivated land products, with data produced once every 4 years based on the decision tree algorithm) and CLCD data (cultivated land products, annual data generated based on the random forest algorithm and the GEE cloud computing platform). By extracting the cultivated land distribution range of the CLCD data, the maximum cultivated land range is obtained by overlaying the cultivated land distributions of the CLCD data from 1985 to 2022 and the GLAD data from 2000 to 2019.

[0031] In the embodiment of the present application, in step S101, the process of performing multi-temporal harmonic fitting on the remote sensing image data based on the time-domain sliding window to construct the cultivated land change vector is as follows:

[0032] Where, represents any pixel within the maximum cultivated land range; represents any Julian day with remote sensing observations, and the range includes the Julian days of all remote sensing observations from year t to year t + w of the observation year; w is the time window for function fitting; represents the predicted value of the NDVI index of the pixel at time ; is the remote sensing observation value of the NDVI index at time ; is the number of days in a year; is the overall amplitude or intercept, reflecting the average growth status of the vegetation in the time period from year t to year t + w; is the slope, reflecting the inter-annual change of the vegetation from year t to year t + w; is the order of the harmonic component, reflecting the intra-annual change of the vegetation at different frequencies, where and , and , and reflect the intra-annual changes once a year, twice a year, and three times a year respectively.

[0033] Among them, multi-temporal harmonic fitting of remote sensing image data is performed based on a time-domain sliding window. Within the delineated maximum cultivated land area, long-term Landsat satellite image data is used to calculate the NDVI index. A harmonic function including intercept, slope, and first-, second-, and third-order harmonic parameters is used to fit the NDVI index, and 8 parameters of the harmonic function are calculated by pixel-wise fitting using LASSO regression.

[0034] Preferably, w is the time window for function fitting. A smaller window width is restricted by the number of effective observations and affected by environmental noise and cannot obtain a stable harmonic fitting model. While a larger window width will smooth out cultivated land changes with a short duration of change (such as abandonment). In the embodiments of the present application, a window width of "w" years ("w = 5") is defined.

[0035] Preferably, is the number of days in a year, that is, 365.25. Since the harmonic function contains a total of 8 parameters, at least 12 effective observations are required to start model fitting. Considering that there are less than 15 effective observation data within the 5-year window for some pixels, the stability of function fitting decreases significantly at this time. Therefore, the predicted values are calculated by forward filling with the harmonic fitting results of adjacent time periods.

[0036] In the embodiments of the present application, the process of calculating the relative change vector of the cultivated land change vector in adjacent time periods in step S101 is as follows:

[0037] Among them, is the standardized relative change vector of pixel t in the year, represents the standardized vector difference of pixel i in the year t, and are the two harmonic function fitting vectors of pixel i from year t to year t + w and from year to year t respectively; The calculated parameters include the intercept of the harmonic function from year t to year t + w and 7 harmonic fitting parameters of the first, second, and third orders, that is ; The calculated parameters include the intercept of the harmonic function from year to year t and 7 harmonic fitting parameters of the first, second, and third orders, that is and are the intercept and root mean square error of the harmonic function fitting of the two time periods from year t to

[0038] Among them, the construction of the cultivated land change vector based on phenological stability. On an annual scale, the standardized relative change vector (SRCV) is calculated for adjacent time periods to detect whether cultivated land has changed.

[0039] In the embodiment of the present application, the process of determining the potential cultivated land change area based on the comparison result between the relative change vector and the preset recognition threshold, and determining the year when the potential cultivated land change occurs by discriminating the peak value of the relative change vector is as follows:

[0040] Among them, CT represents the preset recognition threshold, and satisfying both of the above conditions is the potential cultivated land change area. They are the standardized relative change vectors for years t - 2, t - 1, t, t + 1, and t + 2 respectively. When the above formula is met, the standardized relative change vector for year t is greater than those for years t - 2, t - 1, t + 1, and t + 2. Therefore, year t is the year when the potential cultivated land change occurs.

[0041] Preferably, the harmonic vectors of the two 5 - year windows before and after the occurrence of the cultivated land change should have the largest difference, that is, the SRCV increases significantly, and at the same time, the SRCV for this year is greater than the SRCVs for the two years before and after. To accurately determine the year when the change occurs, after testing multiple sets of model parameters in different agricultural regions and performing threshold determination on the results, the preset recognition threshold CT for cultivated land change recognition is determined. Theoretically, different preset recognition thresholds should be adopted for different agricultural sub - regions, but it is found through analysis in multiple case areas that CT = 0.5 is a robust threshold for detecting changes. Therefore, in the embodiment of the present application, CT is set to 0.5.

[0042] The potential cultivated land change area identified based on the above method includes pseudo - changes where the phenological trajectory changes due to changes in the cultivated land planting system but the land cover type remains unchanged. Therefore, it is called the potential cultivated land change area.

[0043] In the embodiment of the present application, the process of fusing multi - source land use data in step S100 to obtain the training sample sets for cultivated land and non - cultivated land includes the following steps: Re - classify the CLCD data into land use types; Taking the four years of each GLAD product as the interval time period, superimpose the CLCD data and the CLCD data for the same time period to obtain the areas of the same land use type within this time period, and then superimpose them with the cultivated land and non - cultivated land ranges corresponding to the GLAD product for this time period to generate areas where the cultivated land and non - cultivated land are completely consistent, and correspondingly construct the training sample sets for cultivated land and non - cultivated land.

[0044] Preferably, multi-source land use data are integrated to obtain a training sample set of cultivated land and non-cultivated land. CLCD products are reclassified into six land use types: cultivated land, forest land, grassland, water area, construction land and unused land. Taking the four years of each GLAD data as the interval, the CLCD data of the same time period are superimposed to obtain the area with the same land use type for four years, and then superimposed on the cultivated land and non-cultivated land range of the corresponding period of the GLAD data to generate a completely consistent area of ​​cultivated land and non-cultivated land for a period of four years; at the same time, for the years not covered by the GLAD data, only the CLCD data of four years and four periods are used to obtain the completely consistent area of ​​cultivated land and non-cultivated land. Finally, we obtained completely consistent areas of cultivated land and non-cultivated land in nine time periods: 1986-1991, 1992-1995, 1996-1999, 2000-2003, 2004-2007, 2008-2011, 2012-2015, 2016-2019 and 2020-2023.

[0045] In the embodiment of the present application, before constructing the training sample set of cultivated land and non-cultivated land, the following steps are also included: Resample the resolution of the areas with completely consistent cultivated land and non-cultivated land to construct a cell network; Remote sensing image features are constructed and optimized based on multi-source land use data, and training sample migration is performed.

[0046] Preferably, in order to avoid the influence of mixed pixels and edge effects, the embodiment of the present application resamples the resolution of the areas where the cultivated land and non-cultivated land of CLCD and GLAD are completely consistent from 30m to 90m, extracts pixels with 100% cultivated land or non-cultivated land in the 3×3 range of 30m pixels, and performs random sampling of 1°×1° grid partitions, i.e., initial training samples every 4 years. 1000 cultivated land samples and 200 samples of other land use types are obtained in each 1°×1° grid; if the number of available training samples in the 1°×1° grid is less than the target number, all available training samples are used.

[0047] Preferably, in the process of constructing remote sensing image features, in addition to the blue (Blue), green (Green), red (Red), near infrared (NIR), short wave infrared (SWIR-1, SWIR-2) bands of Landsat satellite images, the normalized vegetation index (NDVI), enhanced vegetation index (EVI), chlorophyll index (GCVI), normalized burn index (NBR), normalized water index (NDWI), normalized building index (NDBI), bare soil index (BSI), modified soil conditioned vegetation index (MSAVI2), normalized tillage index (NDTI) and spectral variability vegetation index (SVVI) are also calculated, a total of 16 indicators, as shown in the following formula:

[0048] By performing multi-temporal synthesis on all available Landsat satellite images of the year, 250 indicators including statistical indicators, time-series indicators, harmonic function coefficient indicators, and terrain factor indicators were constructed to fully reflect the differences in remote sensing image features of cultivated land and non-cultivated land at different times, as follows: ① Statistical indicators, namely the minimum (min), maximum (max), quantiles (1 / 4 quantile (Q1), median (Q2), 3 / 4 quantile (Q3)), range (d max-min ), interquartile range (d Q3-Q1 ), and standard deviation (std) of 16 spectral bands and indices, totaling 128 indicators; ② Time-series indicators, namely 16 spectral bands and indices that form an annual time series sorted from smallest to largest by date, and extracting the minimum (min), maximum (max), quantiles (1 / 4 quantile (Q1), median (Q2), 3 / 4 quantile (Q3)), range (d max-min ), interquartile range (d Q3-Q1 ) and other data of the spectral bands and indices corresponding to 7 dates, totaling 112 indicators; ③ Harmonic function coefficient indicators, namely 7 indicators such as the intercept, slope, first-order, second-order, and third-order fitting coefficients obtained by performing harmonic function fitting on the NDVI index time series; ④ Terrain factors, namely 3 such as altitude, slope, and aspect.

[0049] Preferably, in the process of optimizing remote sensing image features, a random forest model is used for feature optimization and subsequent model training. First, integrate the training samples for each 4-year period obtained and based on the set of remote sensing image feature indicators obtained. Second, use the random forest model for training and evaluate the feature importance based on the Gini impurity index. Then, continuously remove the 10% least important features using the recursive feature elimination method of cross-validation in backward feature elimination, and use 5-fold cross-validation to test the model accuracy until the training model accuracy reaches the highest, and take the feature set corresponding to the highest model accuracy as the final optimized remote sensing image feature set.

[0050] Preferably, during the training sample migration process, based on the obtained initial training samples of each land use type for a four-year period, the obtained initial training samples are applied to the remote sensing images of any target year within these four years. Principal component analysis is performed on the preferred remote sensing image feature set, and the obtained principal component variables are used to calculate the robust Mahalanobis distance from the initial training samples to the sample center. A 3σ criterion is established to eliminate the outliers in the initial training sample set, effectively identifying the outliers and obtaining the training samples for the target year. For example, in the embodiments of the present application, the initial training samples from 2000 to 2003 are respectively migrated to 2000, 2001, 2002, and 2003, and finally the training samples for the target years 2000, 2001, 2002, and 2003 are obtained.

[0051] Based on this, in step S102, based on the generated training sample set, the extracted potential cultivated land change areas are classified using a local random forest classifier, and the potential cultivated land change areas are classified into cultivated land areas and non-cultivated land areas, generating the cultivated land classification results and cultivated land probabilities within each 1°×1° cell grid year by year.

[0052] Among them, in the embodiments of the present application, in order to balance the accuracy and operation time of the random forest algorithm, three parameters of the random forest algorithm are adjusted: 1) the number of decision trees is 500; 2) the sample ratio used for training each tree is set to 2 / 3, that is, each tree contains 2 / 3 of the training samples; 3) the minimum number of samples in the leaf nodes, which is used to control the minimum number of samples in the leaf nodes of each tree to prevent over-segmentation and overfitting, and the minimum number of samples in the leaf nodes is set to 10.

[0053] In the embodiments of the present application, the process of correcting the cultivated land change results in combination with the cultivated land probability includes the steps: If there is a significant difference between the cultivated land probability and other cultivated land probability segments in the time series, the cultivated land change results are retained, otherwise they are eliminated; among them, the significant difference is judged based on the T-test of the shapelet algorithm.

[0054] Preferably, in order to maximize the description of the annual dynamic information of cultivated land, in the embodiments of the present application, cultivated land that has not been planted with field crops for at least three consecutive years is regarded as abandoned. Cultivated land that has not been planted for less than three years is regarded as fallow activities in this study and is not considered in the cultivated land change. The cultivated land classification results of the potential change areas for each year obtained by synthesis form a continuous time series of potential cultivated land changes. Using the cultivated land probability in the three-year window and the shapelet method, compare whether there is a significant difference (T-test) between the cultivated land probability in the window and other cultivated land probability segments in the time series to determine whether the change is reliable. If there is a significant difference, retain the change; otherwise, eliminate the change. As a supplement to the shapelet method, use 0.5 as the threshold for cultivated land and non-cultivated land. Set the transition period window to two years. Define that the average cultivated land probability for three consecutive years in the first five-year window > 0.5 and the average cultivated land probability for three consecutive years in the last five-year window < 0.5 as a reduction in cultivated land; conversely, if the average cultivated land probability for three consecutive years in the first five-year window < 0.5 and the average cultivated land probability for three consecutive years in the last five-year window > 0.5, it is newly added cultivated land.

[0055] Based on this, in the preferred embodiment, the method for realizing long-time series multi-period harmonic fitting and cultivated land change recognition based on a time-domain sliding window can amplify mutations, gradual changes, and phenological differences through the stability of segmented fitting trajectories, effectively solving problems such as weak signal leakage recognition and change detection time deviation existing in existing cultivated land change recognition methods. At the same time, based on key steps such as multi-source land use data fusion, feature optimization, sample migration, machine learning classification, and pseudo-change discrimination and correction, it can not only expand the scope of cultivated land change detection but also provide a more accurate cultivated land change detection location. Compared with traditional methods, it improves the accuracy and robustness of cultivated land change detection.

[0056] The embodiments of the present application also provide a cultivated land change detection and recognition device based on a time-series feature vector.

[0057] Figure 2 For the module structure diagram of the cultivated land change detection and recognition device based on the time-series feature vector in an embodiment of the application, as Figure 2 shown, a cultivated land change detection and recognition device based on a time-series feature vector in one embodiment includes: A data acquisition module 100, configured to acquire remote sensing image data and multi-source land use data; wherein, multi-source land use data is fused to obtain a training sample set of cultivated land and non-cultivated land; A data fusion module 101, configured to perform multi-period harmonic fitting on remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, calculate a relative change vector of the cultivated land change vector in adjacent time periods, determine a potential cultivated land change area based on the comparison result of the relative change vector and a preset recognition threshold, and determine the year when potential cultivated land change occurs based on the peak value of the relative change vector; The data classification module 102 is configured to classify potential cultivated land change areas into cultivated land areas and non-cultivated land areas based on a training sample set, and generate corresponding cultivated land classification results and cultivated land probabilities. The data detection module 103 is configured to synthesize the cultivated land classification results to form cultivated land change results in a continuous time series, and correct the cultivated land change results in combination with the cultivated land probabilities.

[0058] After obtaining remote sensing image data and multi-source land use data, the cultivated land change detection and recognition device based on the time series feature vector in the embodiment of the present application performs multi-period harmonic fitting on the remote sensing image data based on a time-domain sliding window to construct a cultivated land change vector, and calculates a relative change vector of the cultivated land change vectors in adjacent time periods. The potential cultivated land change areas are determined based on the comparison result between the relative change vector and a preset recognition threshold, and the year when the potential cultivated land change occurs is discriminated based on the peak value of the relative change vector. The potential cultivated land change areas are classified into cultivated land areas and non-cultivated land areas based on a training sample set, and corresponding cultivated land classification results and cultivated land probabilities are generated; the cultivated land classification results are synthesized to form cultivated land change results in a continuous time series, and the cultivated land change results are corrected in combination with the cultivated land probabilities. The data fusion of a long time series is realized based on the time-domain sliding window, problems such as missed recognition of weak signals and time deviation of change detection are solved, and the detection accuracy of the cultivated land change dynamics is improved.

[0059] At least one embodiment of the present application further provides a data control device. Figure 3 It is a schematic block diagram of a data control device provided by at least one embodiment of the present application. For example, as Figure 3 shown, the data control device 20 may include one or more memories 200 and one or more processors 201. The memory 200 is used to non-transiently store computer-executable instructions; the processor 201 is used to run the computer-executable instructions, and when the computer-executable instructions are run by the processor 201, the processor 201 can be made to execute one or more steps in the cultivated land change detection and recognition method based on the time series feature vector according to any embodiment of the present application.

[0060] For the specific implementation of each step of the cultivated land change detection and recognition method based on the time series feature vector and the related explanatory content, reference may be made to the relevant content in the embodiments of the cultivated land change detection and recognition method based on the time series feature vector described above, which will not be elaborated here. It should be noted that Figure 3 the components of the data control device 20 shown are exemplary and not restrictive. According to actual application needs, the data control device 20 may also have other components.

[0061] In one embodiment, the processor 201 and the memory 200 can communicate with each other directly or indirectly. For example, the processor 201 and the memory 200 can communicate through a network connection. The network can include a wireless network, a wired network, and / or any combination of a wireless network and a wired network. The type and function of the network are not limited herein. For another example, the processor 201 and the memory 200 can also communicate through a bus connection. The bus can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. For example, the processor 201 and the memory 200 can be set at a remote data server side (cloud) or a distributed energy system side (local side), or can also be set at a client side (such as a mobile device like a mobile phone). For example, the processor 201 can be a Central Processing Unit (CPU), a Tensor Processing Unit (TPU), or a Graphics Processing Unit (GPU), etc., which has data processing capabilities and / or instruction execution capabilities, and can control other components in the data control device 20 to perform desired functions. The Central Processing Unit (CPU) can be of an X86 or ARM architecture, etc.

[0062] In one embodiment, the memory 200 can include any combination of one or more computer program products. The computer program products can include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. Volatile memory can include, for example, Random Access Memory (RAM) and / or cache memory, etc. Non-volatile memory can include, for example, Read-Only Memory (ROM), a hard disk, Erasable Programmable Read-Only Memory (EPROM), a Portable Compact Disc Read-Only Memory (CD-ROM), a USB memory, a flash memory, etc. One or more computer-executable instructions can be stored on the computer-readable storage media. The processor 201 can run the computer-executable instructions to implement various functions of the data control device 20. Various application programs and various data can also be stored in the memory 200, as well as various data used and / or generated by the application programs, etc.

[0063] It should be noted that the data control device 20 can achieve similar technical effects as the aforementioned cultivated land change detection and recognition method based on temporal feature vectors. The repeated parts will not be elaborated here.

[0064] At least one embodiment of the present application also provides a non-transitory computer-readable storage medium. Figure 4 It is a schematic diagram of a non-transitory computer-readable storage medium provided by at least one embodiment of the present application. For example, as Figure 4As shown, one or more computer-executable instructions 301 can be non-transiently stored on a non-transitory computer-readable storage medium 30. For example, when the computer-executable instructions 301 are executed by a computer, the computer can be caused to execute one or more steps in the cultivated land change detection and recognition method based on temporal feature vectors according to any embodiment of the present application.

[0065] In one embodiment, the non-transitory computer-readable storage medium 30 can be applied to the above data control device 20. For example, it can be the memory 200 in the data control device 20.

[0066] In one embodiment, the description of the non-transitory computer-readable storage medium 30 can refer to the description of the memory 200 in the embodiment of the data control device 20, and repeated parts will not be elaborated.

[0067] It should be noted that when the memory 200 stores different non-transitory computer-executable instructions, the data control device 20 correspondingly serves as a firmware upgrade device. When the computer-executable instructions are run by the processor 201, the processor 201 can be caused to execute one or more steps in the cultivated land change detection and recognition method based on temporal feature vectors according to any embodiment of the present application.

[0068] For the present application, the following points also need to be explained: (1) The accompanying drawings of the embodiments of the present application only relate to the structures involved in the embodiments of the present application, and other structures can refer to the general design.

[0069] (2) Without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other to obtain new embodiments. The above are only the specific implementation manners of the present application, but the protection scope of the present application is not limited thereto. The protection scope of the present application shall be subject to the protection scope of the claims.

[0070] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combinations of these technical features do not conflict, they should all be considered as the scope recorded in this specification.

[0071] The above embodiments only express several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the patent application. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the appended claims.

Claims

1. A method for detecting and identifying cultivated land changes based on time series feature vectors, characterized in that: Includes steps: Acquire remote sensing image data and multi-source land use data; wherein, fuse the multi-source land use data to obtain a training sample set of cultivated land and non-cultivated land; Perform multi-period harmonic fitting of remote sensing image data based on a time domain sliding window to construct a cultivated land change vector, and calculate the relative change vector of the cultivated land change vector in adjacent time periods, determine the potential cultivated land change area by comparing the relative change vector with a preset recognition threshold, and identify the year when the potential cultivated land change occurs by the relative change vector peak value; Classifying the potential cultivated land change area into cultivated land area and non-cultivated land area based on the training sample set, and generating corresponding cultivated land classification results and cultivated land probabilities; The cultivated land classification results are synthesized to form a continuous time series of cultivated land change results, and the cultivated land change results are corrected in combination with the cultivated land probability.

2. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 1 is characterized in that: The multi-period harmonic fitting of remote sensing image data based on the time domain sliding window adopts the following formula: ； ； in, Represents any pixel within the maximum cultivated land area; represents any Julian day with remote sensing observations, ranging from all Julian days with remote sensing observations from observation year t to observation year t+w; w is the time window for function fitting; express Pixel in time The predicted value of NDVI index; For time Remote sensing observation value of NDVI index; is the number of days in a year; is the overall amplitude or intercept, reflecting the average growth status of vegetation in the period from year t to year t + w; is the slope, reflecting the interannual variation of vegetation from year t to year t+w; is the order of the harmonic component, reflecting the annual changes in different frequencies of vegetation, where and , and , and It reflects the annual changes once a year, twice a year and thrice a year respectively.

3. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 2 is characterized in that: The relative change vector of the cultivated land change vector in adjacent time periods is calculated using the following formula: in, for The normalized relative change vector of pixel t in year, represents the standardized vector difference of pixel i in year t, and They are respectively the i pixel from year t to year t+w and The harmonic function fitting vector of the two time periods from year to year t; The calculated parameters include the intercept of the harmonic function from year t to year t+w and seven harmonic fitting parameters of the first, second and third order, namely ; The calculated parameters include The intercept of the harmonic function from year to year t and the first-order, second-order and third-order harmonic fitting parameters, which are ; From t to and The intercept and RMS error of the harmonic function fit for the two time periods from year to year t.

4. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 3 is characterized in that: The following two formulas are used to identify the year in which potential cultivated land change occurs: Among them, CT represents the preset identification threshold, and the area that meets the above two equations is the potential cultivated land change area; are the standardized relative change vectors of t-2, t-1, t, t+1, and t+2 respectively. When the above two formulas are met, the standardized relative change vector of year t is greater than that of the other years t-2, t-1, t+1, and t+2; Then year t is the year when potential cultivated land change occurs.

5. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 1 is characterized in that: The multi-source land use data includes CLCD data and GLAD data; The process of fusing the multi-source land use data to obtain a training sample set of cultivated land and non-cultivated land includes the following steps: Reclassify CLCD data into land use types; The CLCD data and CLCD data of the same time period are superimposed to obtain the same land use type area, and then superimposed on the cultivated land and non-cultivated land range of the corresponding time period of the GLAD product to generate a completely consistent area of ​​cultivated land and non-cultivated land, and correspondingly construct a training sample set of cultivated land and non-cultivated land.

6. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 5 is characterized in that: Before constructing the training sample set of cultivated land and non-cultivated land, the following steps are also included: Resampling the areas where the cultivated land and the non-cultivated land are completely consistent with each other in resolution to construct a cell network; Based on the multi-source land use data, remote sensing image features are constructed and optimized, and training sample migration is performed.

7. The method for detecting and identifying cultivated land changes based on time series feature vectors according to claim 1 is characterized in that: The process of correcting the cultivated land change result in combination with the cultivated land probability comprises the steps of: If there is a significant difference between the cultivated land probability and other cultivated land probability segments in the time series, the cultivated land change result is retained, otherwise it is eliminated; wherein the significant difference is judged based on the shapelet algorithm.

8. A device for detecting and identifying cultivated land changes based on time series feature vectors, characterized in that: include: A data acquisition module is used to acquire remote sensing image data and multi-source land use data; wherein the multi-source land use data is integrated to obtain a training sample set of cultivated land and non-cultivated land; A data fusion module is used to perform multi-period harmonic fitting of remote sensing image data based on a time domain sliding window to construct a cultivated land change vector, and calculate the relative change vector of the cultivated land change vector in adjacent time periods, determine the potential cultivated land change area by comparing the relative change vector with a preset recognition threshold, and identify the year when the potential cultivated land change occurs by the relative change vector peak value; A data classification module, for classifying the potential cultivated land change area into cultivated land area and non-cultivated land area based on the training sample set, and generating corresponding cultivated land classification results and cultivated land probability; The data detection module is used to synthesize the cultivated land classification results to form a continuous time series of cultivated land change results, and to correct the cultivated land change results in combination with the cultivated land probability.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer-executable instructions, and when the computer-executable instructions are executed by a processor, the method for detecting and identifying cultivated land changes based on time series feature vectors as described in any one of claims 1 to 7 is implemented.

10. A data control device, characterized in that: include: one or more memories non-transitorily storing computer-executable instructions; One or more processors are configured to run computer executable instructions, wherein the computer executable instructions, when executed by the one or more processors, implement the method for detecting and identifying cultivated land changes based on time series feature vectors as described in any one of claims 1 to 7.

Images