A method and system for identifying abandoned land plots based on crop type changes

Abstract

The embodiment of the application provides a kind of abandoned land identification method and system based on crop type change, belong to image processing technical field.The method comprises: the pre-processing is executed to multiple time phase remote sensing image to obtain pre-processing image, and the cadastral vector boundary is based on generation and each land corresponding land time sequence sample;Extract multi-dimensional feature vector;Identify crop type label, construct crop type change sequence, and obtain management state preliminary judgment result based on land discrimination model;Based on the confusion resolution rule set of pre-established, the management state preliminary judgment result is executed correction, and the abandoned land identification result is output.The scheme of the present application realizes high-precision identification of abandoned land by constructing crop type change sequence and combining multi-dimensional feature vector, which is jointly processed by discrimination model and confusion resolution rule, effectively avoids misjudgment of fallow, multiple cropping and disaster damage.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image processing, in particular to a method for identifying abandoned land based on crop type change and a system for identifying abandoned land based on crop type change. BACKGROUND

[0002] With the development of remote sensing monitoring means, using multi-temporal satellite images to identify the utilization status of cultivated land has become an important way for land resource management and agricultural production supervision. In the prior art, the vegetation index is usually calculated or the time series change curve is constructed by using remote sensing data obtained at different agricultural times to identify whether the cultivated land is in normal planting state. However, the method based on single spectral index or simple threshold judgment has obvious deficiencies in practical application.

[0003] The existing method has limitations in distinguishing abandoned land and policy-induced fallow land. Abandoned land usually presents low coverage or bare land characteristics due to abandonment, and policy-induced fallow land also presents low vegetation coverage for a period of time. The two are highly similar in spectral characteristics and time series change curve. If only the vegetation index threshold is relied on for judgment, the fallow land is easy to be misidentified as abandoned, thereby affecting the accuracy of land use statistics and may cause deviation in the implementation of agricultural management policies.

[0004] The existing method also has the risk of misjudgment in dealing with abnormalities caused by single-season disasters. For example, natural disasters such as floods or droughts will cause the crops to reduce yield or even die in the current season, which is manifested as a sub-peak or declining vegetation index in remote sensing images. In the traditional time series monitoring method, such disaster-damaged land is often directly classified as abandoned land, ignoring the fact that it may recover normal planting in the next season. This not only reduces the accuracy of the identification result, but also may mislead the assessment of agricultural production potential and post-disaster recovery capacity.

[0005] Therefore, how to effectively distinguish abandoned land from fallow land in remote sensing image monitoring and avoid misjudging single-season disaster-damaged land as abandoned land is a technical problem to be solved in the field of cultivated land use status identification. SUMMARY

[0006] The purpose of the embodiments of the present application is to provide a method and system for identifying abandoned land based on crop type change, to at least solve the problems that it is difficult to distinguish abandoned land from fallow land in the prior art, and single-season disaster-damaged land is easy to be misjudged as abandoned land.

[0007] To achieve the above object, the application provides a method for identifying abandoned land plots based on crop type changes, which comprises the following steps: acquiring multi-temporal remote sensing images and cadastral vector boundaries of a target area; performing preprocessing on the multi-temporal remote sensing images to obtain preprocessed images; performing spatial registration and vector mask cropping on the preprocessed images based on the cadastral vector boundaries to generate land plot time series samples corresponding to each land plot; extracting land plot multi-dimensional feature vectors including spectral features, time series features and spatial structure features from the land plot time series samples within a preset phenological window; identifying crop type labels of each time phase in the land plot time series samples based on a time series segmentation network, constructing a crop type change sequence, and inputting the crop type change sequence and the land plot multi-dimensional feature vectors into a land plot discrimination model to obtain an initial management state judgment result; and performing correction on the initial management state judgment result based on a preset confusion resolution rule set to output an abandoned land plot identification result.

[0008] Optionally, the pre-processing rule for the multi-temporal remote sensing images is: sequentially performing radiation calibration, atmospheric correction and geometric correction on the multi-temporal remote sensing images to obtain preprocessed images with spectral comparability and spatial consistency between different time images; wherein, the radiation calibration is used to convert the original pixel value into physical reflectivity; the atmospheric correction is used to eliminate the interference of atmospheric scattering and absorption on the spectral curve; and the geometric correction is used to spatially align each time phase image with the cadastral vector boundaries.

[0009] Optionally, the spatial registration and vector mask cropping of the preprocessed images based on the cadastral vector boundaries to generate land plot time series samples corresponding to each land plot comprises the following steps: performing coordinate system and geometric registration of the cadastral vector boundaries and the preprocessed images to ensure one-to-one correspondence between land plot boundaries and image pixels; performing vector mask operation on the registered images using the cadastral vector boundaries to remove non-land plot area pixels and only keep the pixel set consistent with the land plot boundary range; and grouping the pixel set according to the block identifier of the cadastral vector boundaries to form land plot time series samples corresponding to each land plot.

[0010] Optionally, the multi-dimensional feature vector of the plot is extracted from the plot time series sample in a preset phenology window, including spectral features, time series features, and spatial structure features, comprising: calculating a spectral index curve based on the pixel reflectance sequence of the plot time series sample in the phenology window, and taking the spectral index curve as the spectral feature; extracting trend information and key node information of the crop growth stage based on the time series variation of the plot time series sample in the phenology window, and taking the trend information and the node information as the time series feature; extracting boundary tightness and internal spatial texture parameters based on the image texture distribution and cadastral vector boundary morphology of the plot time series sample, and taking the boundary tightness and the internal spatial texture parameters as the spatial structure feature; and fusing the spectral feature, the time series feature, and the spatial structure feature to obtain a multi-dimensional feature vector.

[0011] Optionally, the multi-dimensional feature vector is obtained by fusing the spectral feature, the time series feature, and the spatial structure feature, comprising: performing normalization processing on the spectral feature, the time series feature, and the spatial structure feature respectively; splicing or weighted summing the normalized features in the feature dimension according to a preset weight to form a fusion feature matrix; and extracting key feature components on the fusion feature matrix by a dimension reduction algorithm or a feature selection method to obtain a final multi-dimensional feature vector.

[0012] Optionally, the crop type label of each time phase in the plot time series sample is identified based on the time series segmentation network, and a crop type change sequence is constructed, comprising: inputting the plot time series sample into the trained time series segmentation network, performing pixel-by-pixel classification on the pixels of each time phase image, and outputting the corresponding crop type label; aggregating the crop type labels according to the cadastral vector boundary to generate a crop type result corresponding to each plot; and encoding the crop type results of each time phase in time sequence to form a crop type change sequence for representing the crop evolution process.

[0013] Optionally, the crop type change sequence and the multi-dimensional feature vector of the plot are jointly input into a plot discrimination model to obtain an initial management state judgment result, comprising: performing time series encoding on the crop type change sequence to generate a time series embedding vector representing the crop evolution process; performing feature mapping on the multi-dimensional feature vector of the plot to generate a comprehensive feature embedding representing the spectral information, time series information, and spatial structure information of the plot; fusing the time series embedding vector and the comprehensive feature embedding in a unified feature space, and inputting them into the plot discrimination model; and the plot discrimination model outputs the initial management state judgment result based on the fused input.

[0014] Optionally, the training rule of the field discrimination model is that: a historical field time sequence sample with a field management state label is collected, the historical field time sequence sample contains a corresponding crop type change sequence and a field multi-dimensional feature vector; the crop type change sequence and the field multi-dimensional feature vector are taken as inputs, the field management state label is taken as a supervision signal, and supervised learning training is performed on the field discrimination model; a boundary constraint loss and a time sequence consistency loss are introduced in the training process, and the field discrimination model is obtained after training convergence.

[0015] Optionally, the management state preliminary judgment result is modified based on a preset confusion resolution rule set, and a abandoned field recognition result is output, including: the confusion resolution rule set is pre-defined with a multiple cropping discrimination rule, a fallow discrimination rule and a disaster loss inspection rule, and is stored in a rule library; the management state preliminary judgment result is compared with the crop type change sequence and the field multi-dimensional feature vector, and the confusion resolution rule set is called one by one; if the multiple cropping discrimination rule is met, the management state preliminary judgment result is modified to a multiple cropping state; if the fallow discrimination rule is met, the management state preliminary judgment result is modified to a fallow state; if the disaster loss inspection rule is met, the management state preliminary judgment result is modified to a disaster loss state; if none of the multiple cropping discrimination rule, the fallow discrimination rule and the disaster loss inspection rule is met, the management state preliminary judgment result is confirmed to be a abandoned state, and the abandoned field recognition result is output.

[0016] The second aspect of the application provides a abandoned field recognition system based on crop type change, the system includes: an acquisition unit for obtaining multi-temporal remote sensing images and cadastral vector boundaries of a target area, performing preprocessing on the multi-temporal remote sensing images to obtain preprocessed images, and performing spatial registration and vector mask cropping on the preprocessed images based on the cadastral vector boundaries to generate field time sequence samples corresponding to each field; a feature extraction unit for extracting a field multi-dimensional feature vector including spectral features, time sequence features and spatial structure features from the field time sequence samples within a preset phenology window; a judgment unit for identifying crop type labels of each time phase in the field time sequence sample based on a time sequence segmentation network, constructing a crop type change sequence, and inputting the crop type change sequence and the field multi-dimensional feature vector into a field discrimination model to obtain a management state preliminary judgment result; and an output unit for modifying the management state preliminary judgment result based on a preset confusion resolution rule set and outputting a abandoned field recognition result.

[0017] Through the technical solution, the application achieves accurate acquisition of land block level time sequence samples through registration and cutting of multi-temporal remote sensing images and cadastral vector boundaries, and ensures consistency of data spatial positioning. Comprehensive extraction of spectral, time sequence and spatial structure features within a phenology window can comprehensively depict land block growth states and provide multi-dimensional basis for subsequent analysis. In combination with crop type labels output by a time sequence segmentation network, a crop type change sequence is formed, and is input into a land block discrimination model together with a multi-dimensional feature vector, thereby effectively improving accuracy of land block management state recognition. Finally, through a confusion resolution rule set, a management state preliminary judgment result is corrected, which can avoid misjudgment of fallow, re-planting or disaster damage as abandoned land, thereby realizing accurate recognition of abandoned land blocks and ensuring reliability and practicality of monitoring results.

[0018] Other features and advantages of the embodiments of the present application will be described in detail in the following detailed description of the embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings are included to provide a further understanding of the embodiments of the application, and constitute a part of the specification, and are used together with the following detailed description of the embodiments to explain the embodiments of the application, but do not constitute a limitation of the embodiments of the application. In the drawings:

[0020] Figure 1 is a step flow chart of a method for abandoned land block recognition based on crop type change provided by an embodiment of the application;

[0021] Figure 2 is a schematic diagram of generation of a land block pixel set through cutting of cadastral vector boundaries provided by an embodiment of the application;

[0022] Figure 3 is a system structure diagram of an abandoned land block recognition system based on crop type change provided by an embodiment of the application. DETAILED DESCRIPTION

[0023] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. It should be understood that the specific embodiments described herein are only used to illustrate and explain the embodiments of the present application, and are not used to limit the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] It should be noted that if the embodiments of the present application involve directionality indication (such as up, down, left, right, front, back, etc.), the directionality indication is only used to explain the relative position relationship, motion condition, etc. between components in a certain posture (as shown in the drawings), and if the certain posture changes, the directionality indication also changes accordingly.

[0025] In addition, if the embodiments of the present application involve descriptions such as "first", "second", etc., the descriptions of "first", "second", etc. are only for description purposes, and cannot be understood as indicating or implying the relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include at least one of the features. In addition, the technical solutions of each embodiment can be combined with each other, but it must be based on the realization of ordinary skilled in the art, when the combination of technical solutions appears contradictory or cannot be realized, it should be considered that the combination of technical solutions does not exist, and is not within the protection scope required by the present application.

[0026] Figure 1 is a method flowchart of a crop type change based abandoned land block identification method provided by an embodiment of the present application. As shown in Figure 1 , the present application provides a crop type change based abandoned land block identification method, which comprises the following steps:

[0027] Step S1: acquiring multi-temporal remote sensing images of a target area and cadastral vector boundaries, performing preprocessing on the multi-temporal remote sensing images to obtain preprocessed images, and performing spatial registration and vector mask cropping on the preprocessed images based on the cadastral vector boundaries to generate land block time sequence samples corresponding to each land block.

[0028] Specifically, the rule for performing preprocessing on the multi-temporal remote sensing images is: sequentially performing radiation calibration, atmospheric correction and geometric correction on the multi-temporal remote sensing images to obtain preprocessed images with spectral comparability and spatial consistency between different time images; wherein, the radiation calibration is used to convert the original pixel value into physical reflectivity; the atmospheric correction is used to eliminate the interference of atmospheric scattering and absorption on the spectral curve; the geometric correction is used to spatially align each time phase image with the cadastral vector boundaries.

[0029] Further, based on the cadastral vector boundary, the pre-processed image is spatially registered and vector mask cropped to generate land parcel time series samples corresponding to each land parcel, including: performing coordinate system and geometric registration of the cadastral vector boundary and the pre-processed image to ensure one-to-one correspondence between land parcel boundaries and image pixels; performing vector mask operation on the registered image to remove non-land parcel area pixels and only keep pixel sets consistent with land parcel boundary ranges; grouping the pixel sets according to the block identifiers of the cadastral vector boundary to form land parcel time series samples corresponding to each land parcel.

[0030] In the implementation process, first, multi-temporal remote sensing images of the target area and corresponding cadastral vector boundaries need to be obtained. Multi-temporal remote sensing images refer to remote sensing data covering the same area obtained at different time nodes, which can reflect the dynamic changes of land parcels with seasons and farming times. Cadastral vector boundaries are vectorized boundary data that accurately delineate the ranges of each land parcel in the target area. In order to ensure consistency between time series images and land parcel boundaries in the subsequent analysis process, these two types of data must be combined to form time series samples that can be directly corresponded to each land parcel.

[0031] Before converting the original remote sensing image into usable data, it needs to be pre-processed. The rules for pre-processing are: performing radiation calibration, atmospheric correction, and geometric correction on the multi-temporal remote sensing images in sequence to obtain pre-processed images with spectral comparability and spatial consistency. Among them, the role of radiation calibration is to convert the pixel values in the original image from the digital quantization values recorded by the sensor into reflectance values with physical meaning. After such processing, the images obtained at different times have spectral comparability, otherwise the differences in sensor imaging conditions will cause deviations in vegetation index and spectral curve. Atmospheric correction is used to eliminate the interference of atmospheric scattering and absorption on the spectral curve, especially in the shortwave and near-infrared bands, atmospheric conditions have a great impact on remote sensing data. If there is no atmospheric correction, images of the same area obtained at different dates may have spectral curve drift due to different water vapor content or aerosol concentration, which seriously affects subsequent time-based change analysis. Geometric correction is a key link to ensure that the images of each time phase are strictly aligned in space. By referring to the cadastral vector boundary and control points for geometric registration, the pixel positions of images at different times can be corrected to a unified geographic coordinate system, avoiding misalignment of land parcel boundaries and pixel sets due to position deviation.

[0032] After obtaining the pre-processed images, it is necessary to perform spatial registration and vector mask cropping with the cadastral vector boundary to generate parcel time series samples corresponding to each parcel. Specifically, first, the cadastral vector boundary and the pre-processed images are coordinate system unified and geometrically registered. Coordinate system unification here refers to unifying data from different sources to the same spatial reference system, ensuring that the boundary and the image can be overlaid in the same coordinate system; geometric registration refers to accurately aligning the cadastral boundary and the pre-processed images through control point matching and geometric transformation, ensuring that the parcel boundary and the image pixels can correspond one by one.

[0033] After the above registration is completed, it is necessary to apply the cadastral vector boundary to the pre-processed images to perform vector mask operation. Vector mask operation is to use the vector boundary as a mask to remove all pixels in the image that do not belong to the boundary range, and only keep the pixel set that is completely consistent with the boundary range. The purpose of this process is to eliminate the interference of background areas or adjacent parcels, so that the spectral curves, texture features and other information extracted subsequently are strictly derived from the target parcel range.

[0034] As Figure 2 , a schematic diagram of vector mask operation is provided. The left side illustrates the state of superimposing the cadastral vector boundary on the pre-processed images, where the parcel range is defined by the cadastral vector boundary, and the images contain both the parcel area and the background area outside the parcel. By performing vector mask operation, all pixels outside the parcel can be removed, and only the pixel set inside the parcel consistent with the cadastral vector boundary is retained. The right side illustrates the result after cropping, i.e. only the pixel set within the parcel range is included, which serves as input data for subsequent parcel time series sample generation and feature extraction. This processing method can ensure that the time series image data used in the subsequent recognition process is strictly derived from the parcel interior, thereby avoiding misjudgment caused by interference from adjacent parcels or background areas.

[0035] After the mask operation is completed, the pixel set needs to be grouped according to the block identifier of the cadastral vector boundary. Usually each parcel has a unique identifier in cadastral data, and through this identifier, the pixel set belonging to the same parcel is corresponded in each time series image, thereby forming a complete parcel time series sample. The parcel time series sample generated in this way not only retains the spectral and spatial information of each parcel at different time nodes, but also ensures the consistency of the data with the parcel boundary, laying a data foundation for subsequent feature extraction and crop type change sequence construction.

[0036] The above steps address two major issues that arise when directly using existing remote sensing imagery: first, the incomparability between images from different time periods due to differences in radiometric, atmospheric, and geometric factors; and second, the impurity of samples caused by misalignment or contamination between plot boundaries and image pixels. By employing preprocessing and vector cropping rules, the resulting plot time-series samples exhibit consistent spectral curves across time phases, accurate spatial correspondence, and clear plot boundary segmentation, thus ensuring reliable data input for subsequent land abandonment assessment based on time-series characteristics.

[0037] Step S2: Extract multidimensional feature vectors of the land parcels from the temporal samples of the land parcels within a preset phenological window, including spectral features, temporal features, and spatial structure features.

[0038] Specifically, a spectral index curve is calculated based on the pixel reflectance sequence of the plot's time-series samples within the phenological window, and the spectral index curve is used as a spectral feature; trend information and key node information of crop growth stages are extracted based on the time-series changes of the plot's time-series samples within the phenological window, and the trend information and the node information are used as temporal features; boundary compactness and internal spatial texture parameters are extracted based on the image texture distribution and cadastral vector boundary morphology of the plot's time-series samples, and the boundary compactness and the internal spatial texture parameters are used as spatial structure features; a multidimensional feature vector is obtained by fusing the spectral features, the temporal features, and the spatial structure features.

[0039] Furthermore, the multidimensional feature vector is obtained by fusing the spectral features, temporal features, and spatial structure features, including: performing normalization processing on the spectral features, temporal features, and spatial structure features respectively; concatenating or weighting the normalized features according to preset weights along the feature dimension to form a fused feature matrix; and extracting key feature components from the fused feature matrix using a dimensionality reduction algorithm or feature selection method to obtain the final multidimensional feature vector.

[0040] In the specific implementation process, multi-dimensional feature vectors, including spectral features, temporal features, and spatial structure features, are extracted from the plot time series samples within a preset phenological window. This ensures that the subsequent discrimination model obtains complete and multi-dimensional plot feature information at the input end. A phenological window is a pre-defined time range based on the crop phenological patterns of the target area, such as the stages from sowing to jointing, jointing to grain filling, and grain filling to harvest. The corresponding remote sensing observation data within this time range can relatively completely reflect the crop growth trend. Therefore, limiting the plot time series samples formed by preprocessing and masking multi-temporal remote sensing images to this phenological window can minimize noise interference from redundant time points while retaining the key growth information necessary for crop type identification.

[0041] Specifically, the first step is to calculate the spectral index curve based on the pixel reflectance sequence of the plot's time-series samples within the phenological window. Spectral indices are indicators composed of combinations of different bands, such as the common Normalized Difference Vegetation Index (NDVI), Soil-Adjusted Vegetation Index (SDI), and Moisture Index, all constructed through band operations to amplify the differences between vegetation and soil, and between vegetation and water bodies. During the calculation, the reflectance sequence of pixels within each plot within the phenological window is combined temporally to form a continuous spectral index time curve. This curve is then smoothed or interpolated to remove gaps and abnormal fluctuations caused by clouds and fog, thus obtaining a continuous spectral index curve. This spectral index curve serves as the spectral feature input to the plot's multidimensional feature vector, characterizing features such as vegetation cover, photosynthetic activity, and moisture status within the phenological window. Those skilled in the art will understand that the shape of this spectral curve is closely related to crop type; different crops will exhibit different curve shapes within the same phenological window. For example, wheat shows a significant rapid rise during the greening stage, while corn reaches its peak after jointing.

[0042] Furthermore, it is necessary to extract trend information and key node information of crop growth stages based on the time series changes of the plot's time series samples within the phenological window. The extraction rules for trend information can be defined as follows: by fitting the spectral index curve or directly calculating the slope of change based on the reflectance sequence, characteristic parameters of upward trend, plateau trend, or downward trend are obtained. For example, in the early stage, a slope close to zero indicates a dormant period, while a significantly positive slope corresponds to the greening or rapid growth period. The extraction rules for key node information include determining the inflection point position, peak position, and the starting point of decline. By comparing the numerical changes of continuous time phases, when the growth rate changes from positive to zero, it can be identified as an inflection point; when the value reaches a maximum value, it can be identified as a peak value; and when the value begins to decline, it is the starting point of decline. The temporal positions of these nodes directly correspond to the phenological stages of the crop and are important bases for determining crop type and plot management status. The extracted trend information and key node information are collectively referred to as temporal features and are input as independent features into the plot's multidimensional feature vector.

[0043] Furthermore, it is necessary to extract boundary compactness and internal spatial texture parameters based on the image texture distribution and cadastral vector boundary morphology of the plot's time-series samples. The extraction rule for texture distribution is as follows: select typical temporal images within the phenological window, calculate texture statistics such as the gray-level co-occurrence matrix, local binary pattern, or directional gradient histogram of pixels within the plot, thereby quantifying the uniformity and directionality of the spatial distribution within the plot. The extraction rule for boundary compactness is as follows: calculate the continuity and closure of boundary pixels based on the matching degree between the cadastral vector boundary and the actual image boundary. If there are obvious intrusions or gaps in the boundary extension area, the compactness is low. The extraction of internal spatial texture parameters is used to reflect the neatness and consistency of crop planting within the plot, for example, by statistically analyzing the internal texture entropy value of the block to characterize whether the crop rows are uniform. The aforementioned boundary compactness and texture parameters, as spatial structural features, are input into the plot's multi-dimensional feature vector alongside spectral and temporal features, ensuring that the model relies not only on spectral and temporal dimensions but also utilizes spatial dimension information.

[0044] After obtaining spectral features, temporal features, and spatial structure features, these three types of features need to be fused to obtain a complete multidimensional feature vector. The specific rules include three steps:

[0045] 1. Normalize the spectral features, temporal features, and spatial structure features respectively. Since different features have different value ranges and dimensions—for example, spectral indices are between 0 and 1, while texture entropy values ​​may be between 0 and 10—direct splicing would lead to excessively large differences in feature scales. Normalization transforms various features into the same numerical range, improving the effectiveness of subsequent fusion.

[0046] 2. After normalization, the various features are concatenated or weighted and summed according to preset weights along the feature dimension to form a fused feature matrix. The concatenation rule directly combines different feature vectors in dimensional order to obtain a higher-dimensional comprehensive feature vector. The weighted summation rule assigns a weight coefficient to each type of feature; for example, spectral features account for 0.5, temporal features for 0.3, and spatial structure features for 0.2, resulting in a lower-dimensional comprehensive feature through linear weighting. The weights can be set empirically or determined through previous training processes.

[0047] 3. Extract key feature components from the fused feature matrix using dimensionality reduction algorithms or feature selection methods to obtain the final multidimensional feature vector. Dimensionality reduction algorithms may include principal component analysis, linear discriminant analysis, or nonlinear manifold learning, while feature selection methods may include feature importance based on information gain, LASSO regularization, or tree models. These methods effectively remove redundant features, retaining the feature components that best distinguish different crop types and land management statuses, resulting in a final multidimensional feature vector with appropriate dimensionality and strong discriminative power.

[0048] The multi-dimensional feature vectors obtained through the above rules encompass not only the changing shape of the spectral curve, but also the trend and key node information over time, as well as the texture and boundary features of spatial distribution, thus forming a relatively comprehensive descriptive system at the data input level. Relying solely on a single feature, such as spectral indices, can easily lead to difficulties in distinguishing between fallow and abandoned land; similarly, relying solely on temporal trends may result in misjudging single-season disaster damage as long-term abandonment. Multi-dimensional fusion can minimize such confusion and improve the discriminative power of the model's input data.

[0049] Step S3: Based on the temporal segmentation network, identify the crop type labels of each time phase in the temporal samples of the plot, construct the crop type change sequence, and input the crop type change sequence and the multi-dimensional feature vector of the plot into the plot discrimination model to obtain the preliminary judgment result of the operation status.

[0050] Specifically, the method for identifying crop type labels for each time phase in the temporal samples of the land parcels based on a temporal segmentation network and constructing a crop type change sequence includes: inputting the temporal samples of the land parcels into a trained temporal segmentation network, performing pixel-by-pixel classification on the pixels of each time phase image, and outputting the corresponding crop type labels; aggregating the crop type labels according to the cadastral vector boundaries to generate crop type results corresponding to each land parcel; and encoding the crop type results of each time phase in chronological order to form a crop type change sequence for characterizing the crop evolution process.

[0051] Furthermore, the crop type change sequence and the plot multidimensional feature vector are jointly input into the plot discrimination model to obtain a preliminary judgment result of the management status. This includes: performing time-series encoding on the crop type change sequence to generate a time-series embedding vector representing the crop evolution process; performing feature mapping on the plot multidimensional feature vector to generate a comprehensive feature embedding representing the plot's spectral information, time-series information, and spatial structure information; fusing the time-series embedding vector and the comprehensive feature embedding in a unified feature space and passing it as input to the plot discrimination model; and the plot discrimination model outputs a preliminary judgment result of the management status based on the fused input.

[0052] Specifically, the training rules for the land parcel discrimination model are as follows: historical land parcel time-series samples with land parcel management status labels are collected, and the historical land parcel time-series samples simultaneously contain the corresponding crop type change sequence and the land parcel multi-dimensional feature vector; the crop type change sequence and the land parcel multi-dimensional feature vector are used as inputs, and the land parcel management status labels are used as supervision signals to perform supervised learning training on the land parcel discrimination model; boundary constraint loss and temporal consistency loss are introduced during the training process, and the land parcel discrimination model is obtained after training convergence.

[0053] In the specific implementation process, it is necessary to identify crop type labels for each temporal phase in the temporal samples of the plots based on a temporal segmentation network, and then construct a crop type change sequence based on this. This crop type change sequence, along with the multi-dimensional feature vector of the plot, is then input into the plot discrimination model to obtain a preliminary judgment result on the operational status. This process is adopted because multi-temporal remote sensing image data inherently possesses both temporal and spatial dimensions, while crop growth and land use patterns naturally exhibit an evolutionary process over time. Classification results from a single temporal phase often only reflect the crop cover situation at a specific instant, lacking temporal continuity and overall consistency, and therefore insufficient to distinguish between multiple cropping, fallow, and abandoned states. However, by constructing a crop type change sequence and jointly modeling it with the multi-dimensional feature vector of the plot, both temporal evolution patterns and spatial spectral characteristics can be considered simultaneously, significantly improving the discrimination accuracy.

[0054] Specifically, the first step is to use a temporal segmentation network to identify the temporal samples of the land parcels and extract crop type labels for each time phase. The rule for this step is as follows: the temporal samples of the land parcels are used as input, and each time phase is fed into the already trained temporal segmentation network. Pixel-by-pixel classification is performed on each pixel of the image, and the corresponding crop type label is output. Pixel-by-pixel classification means individually determining the crop type corresponding to each pixel in the image at that time phase, such as wheat, corn, rice, or bare land. This process differs from traditional areal classification methods; it can capture subtle differences within the land parcel at the pixel level, thus ensuring more refined and accurate labeling results. With the development of deep learning technology, temporal segmentation networks often adopt a structure combining convolutional neural networks and recurrent neural networks to simultaneously capture spatial and temporal features. During the training phase, they are based on a large amount of labeled historical data, thus enabling reliable pixel classification results to be output during the inference phase.

[0055] After obtaining the crop type labels for each pixel, these pixel-level labels need to be further aggregated according to the cadastral vector boundaries. The rule is as follows: first, identify the cadastral plot number corresponding to each pixel; then, statistically or weightedly integrate all pixel labels under the same plot number to obtain a unified crop type result corresponding to that plot. This is because the plot is the basic unit of agricultural management, while the pixel is merely a technical division at the observation scale. If pixel labels are used directly as the basis for judgment, noise or boundary interference may cause local anomalies, thus affecting the overall judgment. Therefore, it is necessary to aggregate the pixel results based on the cadastral vector boundaries so that the final crop type result corresponds one-to-one with the actual plot unit.

[0056] After extracting the crop type results for each time phase, the results from each time phase need to be encoded chronologically to construct a crop type change sequence characterizing the crop evolution process. The rule is as follows: the crop type labels corresponding to the same plot on different dates are arranged sequentially according to time, forming a discretized sequence. For example, the labels for a plot throughout the four seasons might be "wheat—bare land—corn—bare land," directly reflecting the crop succession and land use changes of that plot during the phenological cycle. In this way, the originally independent single-time phase labels are linked into a continuous evolutionary process, significantly improving the explanatory power of subsequent analyses on operational status.

[0057] Furthermore, the crop type change sequence and the plot multidimensional feature vector need to be input into the plot discrimination model to obtain a preliminary judgment result of the operating status. In this step, feature encoding and mapping need to be performed on the crop type change sequence and the plot multidimensional feature vector respectively. The rules are as follows: First, the crop type change sequence is temporally encoded to generate a temporal embedding vector representing the crop evolution process. Temporal encoding refers to the process of converting discrete crop label sequences into continuous numerical vectors through methods such as recurrent neural networks, attention mechanisms, or positional encoding, thereby maintaining the time order and evolution characteristics in the vector space. For example, independent embedding vectors can be set for different crop types, and then the sequence can be processed through a long short-term memory network to capture the sequential relationship and periodic pattern in the sequence, and output a fixed-length temporal embedding vector.

[0058] Feature mapping is performed on the multidimensional feature vectors of the land parcels to generate comprehensive feature embeddings representing the spectral, temporal, and spatial structural information of the land parcels. The feature mapping rules include fully connected mapping or convolutional mapping, which compress or expand the original multidimensional features to the same dimensional space as the temporal embedding vector through a series of nonlinear transformations, thus facilitating subsequent feature fusion. These multidimensional features include spectral exponential curves, trend information and key nodes, texture statistics and boundary parameters, etc., which, after mapping, can be better aligned with the temporal information to form a unified embedding representation.

[0059] Furthermore, the temporal embedding vector and the comprehensive features are embedded in a unified feature space and fused together, then passed as input to the land parcel discrimination model. The fusion rule can be vector concatenation, weighted summation, or attention weighting; the specific method chosen can be adjusted based on experimental results. Through fusion, the model input can simultaneously contain the temporal patterns of crop evolution and the spatial information of land parcel characteristics, providing a comprehensive basis for discrimination. Finally, based on the fused input, the land parcel discrimination model outputs a preliminary judgment result on the management status. This preliminary judgment result may include categories such as abandoned, fallow, replanting, and normal planting, which are used for subsequent correction rules to further eliminate confusion.

[0060] During training, the land parcel discrimination model requires supervised learning training using labeled historical land parcel time-series samples to ensure accurate output of operational status. The training rules are as follows: collect historical land parcel time-series samples labeled with operational status. These samples simultaneously contain corresponding crop type change sequences and multi-dimensional feature vectors of the land parcels. Use these as inputs, and use operational status labels provided by manual surveys or authoritative databases as supervisory signals to train the land parcel discrimination model. To enhance the model's performance in complex scenarios, boundary constraint loss and temporal consistency loss are introduced during training. The boundary constraint loss rule is: when calculating the classification loss, pixels or regions close to the cadastral vector boundary are assigned higher weights to reduce errors caused by contamination between land parcels. The temporal consistency loss rule is: constrain the temporal continuity of crop type change sequences within the same land parcel, requiring that the model's prediction results do not exhibit unreasonable jumps between adjacent time phases, such as jumping directly from rice to wheat without any bare land or transitional stages in between. By introducing these constraints, the robustness of the model in scenarios with complex boundaries and temporal breaks can be significantly improved. After multiple rounds of training iterations, when the loss function converges, the final land parcel discrimination model that can be used for inference is obtained.

[0061] The above rules form a complete data link: starting from the time series samples of the plots, the crop type labels are obtained pixel by pixel through the time series segmentation network, aggregated and encoded into crop type change sequences, and then fused with multi-dimensional feature vectors and input into the discrimination model, finally outputting the preliminary judgment result of the management status.

[0062] Step S4: Based on the preset confusion resolution rule set, the preliminary judgment result of the operating status is corrected, and the result of the abandoned land identification is output.

[0063] Specifically, the confusion resolution rule set predefines multiple cropping discrimination rules, fallow discrimination rules, and disaster damage detection rules, and stores them in a rule base. The initial judgment result of the management status is compared with the crop type change sequence and the multi-dimensional feature vector of the plot, and the confusion resolution rule set is called to verify them one by one. If the multiple cropping discrimination rule is satisfied, the initial judgment result of the management status is corrected to multiple cropping status; if the fallow discrimination rule is satisfied, the initial judgment result of the management status is corrected to fallow status; if the disaster damage detection rule is satisfied, the initial judgment result of the management status is corrected to disaster damage status; if none of the multiple cropping discrimination rule, fallow discrimination rule, and disaster damage detection rule are satisfied, the initial judgment result of the management status is confirmed as abandoned status, and the abandoned plot identification result is output.

[0064] In this embodiment of the invention, after completing the inference of the plot discrimination model based on crop type change sequences and multi-dimensional feature vectors of plots, the preliminary judgment results of the management status often still have certain ambiguity and uncertainty. This is because actual agricultural production scenarios are complex, and plots in different states may have high similarity in spectral performance and temporal evolution. For example, multiple-cropping plots, due to multiple sowing and harvesting throughout the year, may appear as empty land or low-coverage states at certain times, very similar to abandoned plots in surface features; fallow plots, although temporarily unplanted, often show some traces of artificial covering or irrigation, and their spectral performance partially overlaps with abandoned plots; disaster-damaged plots may suddenly lose their normal crop growth curves when subjected to natural disasters, but may resume planting in subsequent seasons. If identification is based solely on images from a single season, it is highly likely to be misjudged as abandoned. Therefore, based on the preliminary judgment results of the model's management status, an additional set of confusion resolution rules needs to be introduced to verify and correct the preliminary judgment results item by item, in order to eliminate confusion between multiple similar states.

[0065] Specifically, the confusion resolution rule set is a set of discrimination rules that are predefined and stored in the rule base, including three categories: multiple cropping discrimination rules, fallow discrimination rules, and disaster damage detection rules.

[0066] The criteria for determining multiple cropping are based on the multi-peak structure of crop type change sequences. The rule is as follows: if two or more distinct vegetation peaks appear in the crop type change sequence within the same phenological cycle, and the temporal characteristics of the corresponding multidimensional feature vectors show a recurring growth trend, then the plot can be determined to be in a multiple cropping state. If the initial assessment of the management status is abandonment, but the above criteria for multiple cropping are met, then it is revised to multiple cropping.

[0067] The fallow classification rules primarily rely on cover indexes and irrigation / drainage characteristics within the spatial structural features. The rules are as follows: if a typical crop growth curve does not appear in the crop type change sequence, but the plot's multidimensional feature vector shows spectral characteristics of surface cover crop residues or straw returned to the field, and the boundary density and irrigation / drainage traces maintain high integrity, then it indicates that although the plot has not been planted with a major crop, it is still under artificial management and can be corrected to a fallow state. If the initial assessment of the management status is marked as abandoned, but it conforms to the above fallow rules, it is corrected.

[0068] The disaster damage assessment rules primarily rely on cross-seasonal recoverability and proximity consistency. The rules state: if an abnormal decline or interruption occurs in the crop type variation sequence of a given season, resulting in an incomplete growth curve for that season, but a normal crop peak is observed in subsequent seasons, and similar anomalies are observed in surrounding plots during the same period, then the plot can be determined to be disaster-damaged rather than abandoned. If the initial assessment of the operational status is abandonment, but the disaster damage assessment rules are met, the status is revised to disaster-damaged.

[0069] In practical applications, the preliminary judgment result of the management status is compared with the crop type change sequence and the multi-dimensional feature vector of the plot, and the confusion resolution rule set is called to verify each one. The execution order can be as follows: multiple cropping discrimination, fallow discrimination, and disaster damage inspection. When a certain rule condition is detected, the preliminary judgment result of the management status is corrected; if none of the above three rules are met, the preliminary judgment result of the management status is confirmed as a fallow state, and the fallow plot identification result is finally output.

[0070] By conducting secondary screening and correction of the initial assessment results of the operational status, the potential for misjudgment in the direct output of the model was effectively resolved. The multiple cropping discrimination rule ensures that areas with intermittent bare land are not mistakenly labeled as abandoned; the fallow discrimination rule ensures that policy-mandated fallow land is not incorrectly counted as abandoned land; and the disaster damage verification rule prevents short-term anomalies caused by natural disasters from being misinterpreted as long-term abandonment. The introduction of these rules makes the final abandonment identification results more consistent with actual agricultural production conditions, providing land management departments with more reliable data support and facilitating subsequent policy formulation and resource allocation.

[0071] In one possible implementation, during a farmland monitoring exercise in a plain area, multi-temporal remote sensing images of the area for a given year are acquired. After preprocessing and masking / cropping, a temporal sample of 1200 plots is formed. Taking plot A101 as an example, this plot has an area of ​​6.2 hectares. Based on the identification results of the temporal segmentation network, this plot is labeled as "wheat" in spring (March to May) images, appears as "bare land" in summer (June to July), is identified as "corn" again in autumn (August to October), and is vacant land in winter (November to December) images. If relying solely on the initial model judgment, this plot would be classified as "abandoned" due to the continuous bare land appearance in summer.

[0072] The initial assessment of the land's operational status was input into the confusion resolution rule set for verification. First, the multiple cropping discrimination rule was invoked: comparing the crop type change sequence of plot A101, it was observed that two complete crop growth peaks (wheat and corn) occurred within the same year. The spectral index curves reached peak values ​​of 0.72 and 0.69 in April and September, respectively, both exceeding the preset multiple cropping threshold of 0.6. Simultaneously, the temporal feature extraction of this plot showed two distinct growth trend segments, corresponding to the wheat greening stage and the corn jointing stage. After meeting the multiple cropping discrimination criteria, the system revised the initial operational status assessment from "abandoned" to "multiple cropping."

[0073] Plot B305, with an area of ​​4.8 hectares, showed no significant crop cover in the year-round imagery, with a spectral index below 0.25 throughout the year. However, further examination of the plot's multidimensional feature vector revealed that the boundary compactness in the spatial structure remained at 0.91, and there was persistent straw cover reflection in the red band. Based on the fallow criteria, although no major crop was grown, there were clear signs of human management, therefore the initial assessment of the management status was revised to "fallow."

[0074] Plot C512, covering an area of ​​5.5 hectares, showed a spectral index peak of only 0.32 in June during summer imagery, while similar anomalies were observed in neighboring plots (C510, C511, and C513). By September, the spectral index of C512 had recovered to 0.71. According to the disaster damage assessment rules, this anomaly was interpreted as a temporary reduction in yield due to summer flooding, with subsequent replanting. Therefore, the initial assessment of the operational status was revised from "abandoned" to "disaster damaged."

[0075] If the initial operational status assessment from the discriminant model were used directly, A101, B305, and C512 would all be classified as abandoned land. However, after correction using the confusion resolution rule, they were accurately identified as replanted, fallow, and disaster-damaged land, respectively. This rule effectively avoids misjudgments under different circumstances, improving the accuracy and reliability of the abandoned land identification results.

[0076] Figure 3 This is a system structure diagram of an abandoned land identification system based on crop type changes, provided by one embodiment of the present invention. Figure 3 As shown, this invention provides a system for identifying abandoned land parcels based on crop type changes. The system includes: an acquisition unit for acquiring multi-temporal remote sensing images and cadastral vector boundaries of a target area, performing preprocessing on the multi-temporal remote sensing images to obtain preprocessed images, and performing spatial registration and vector mask cropping on the preprocessed images based on the cadastral vector boundaries to generate land parcel time-series samples corresponding to each parcel; a feature extraction unit for extracting multi-dimensional feature vectors of land parcels, including spectral features, temporal features, and spatial structure features, from the land parcel time-series samples within a preset phenological window; a judgment unit for identifying crop type labels for each temporal phase in the land parcel time-series samples based on a temporal segmentation network, constructing a crop type change sequence, and inputting the crop type change sequence and the land parcel multi-dimensional feature vectors into a land parcel discrimination model to obtain a preliminary judgment result of the management status; and an output unit for correcting the preliminary judgment result of the management status based on a preset confusion resolution rule set and outputting the abandoned land parcel identification result.

[0077] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0078] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0079] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

Claims

1. A method for identifying abandoned land plots based on crop type changes, characterized by, The method comprises: acquiring a multi-temporal remote sensing image of a target area and a cadastral vector boundary, performing preprocessing on the multi-temporal remote sensing image to obtain a preprocessed image, and performing spatial registration and vector mask cropping on the preprocessed image based on the cadastral vector boundary to generate a land parcel time sequence sample corresponding to each land parcel; extracting a land parcel multi-dimensional feature vector including spectral features, time sequence features and spatial structure features from the land parcel time sequence sample within a preset phenology window; identifying a crop type label of each time phase in the land parcel time sequence sample based on a time sequence segmentation network, constructing a crop type change sequence, and inputting the crop type change sequence and the land parcel multi-dimensional feature vector into a land parcel discrimination model to obtain an initial judgment result of the management state; performing correction on the initial judgment result of the management state based on a preset confusion resolution rule set to output a recognition result of abandoned land parcels; wherein performing correction on the initial judgment result of the management state based on a preset confusion resolution rule set to output a recognition result of abandoned land parcels, including that the confusion resolution rule set is pre-defined with a multiple cropping discrimination rule, a fallow discrimination rule and a disaster damage inspection rule, and is stored in a rule library; the initial judgment result of the management state is compared with the crop type change sequence and the land parcel multi-dimensional feature vector, and the confusion resolution rule set is called one by one for checking; if the multiple cropping discrimination rule is satisfied, the initial judgment result of the management state is corrected to a multiple cropping state; if the fallow discrimination rule is satisfied, the initial judgment result of the management state is corrected to a fallow state; if the disaster damage inspection rule is satisfied, the initial judgment result of the management state is corrected to a disaster damage state; and if none of the multiple cropping discrimination rule, the fallow discrimination rule and the disaster damage inspection rule is satisfied, the initial judgment result of the management state is confirmed as an abandoned state, and a recognition result of abandoned land parcels is output.

2. The crop type change-based abandoned field plot identification method of claim 1, wherein, The rules for performing preprocessing on the multi-temporal remote sensing image are: radiometric calibration, atmospheric correction and geometric correction are sequentially performed on the multi-temporal remote sensing image to obtain preprocessed images with spectral comparability and spatial consistency between different time images; wherein radiometric calibration is used to convert original pixel values into physical reflectivity; atmospheric correction is used to eliminate the interference of atmospheric scattering and absorption on the spectral curve; geometric correction is used to spatially align each time phase image with the cadastral vector boundary. 3.The crop type change-based abandoned field plot identification method of claim 1, wherein, Based on the cadastral vector boundary, the preprocessed image is spatially registered and vector mask cropped to generate a land parcel time sequence sample corresponding to each land parcel, including: coordinate system and geometric registration are performed on the cadastral vector boundary and the preprocessed image to ensure one-to-one correspondence between land parcel boundaries and image pixels; vector mask operation is performed on the registered image using the cadastral vector boundary to remove non-land parcel area pixels and only keep pixel sets consistent with the land parcel boundary range; the pixel sets are grouped according to the block identification of the cadastral vector boundary to form land parcel time sequence samples corresponding to each land parcel. 4.The crop type change-based abandoned field plot identification method of claim 1, wherein, The land parcel multi-dimensional feature vector including spectral features, time sequence features and spatial structure features is extracted from the land parcel time sequence sample within a preset phenology window, including: calculating a spectral index curve based on the sequence of pixel reflectance of the plot time series sample within the phenology window, and taking the spectral index curve as a spectral feature; extracting trend information and key node information of crop growth stages based on the time series variation of the plot time series sample within the phenology window, and taking the trend information and the node information as a time series feature; extracting boundary tightness and internal spatial texture parameters based on the image texture distribution and cadastral vector boundary morphology of the plot time series sample, and taking the boundary tightness and the internal spatial texture parameters as a spatial structure feature; fusing the spectral feature, the time series feature, and the spatial structure feature to obtain a multi-dimensional feature vector. 5.The crop type change-based abandoned field plot identification method of claim 4, wherein, fusing the spectral feature, the time series feature, and the spatial structure feature to obtain a multi-dimensional feature vector, comprising: performing normalization processing on the spectral feature, the time series feature, and the spatial structure feature respectively; splicing or weighted summing the normalized features of each type according to a preset weight in the feature dimension to form a fusion feature matrix; extracting key feature components on the fusion feature matrix through a dimension reduction algorithm or a feature selection method to obtain a final multi-dimensional feature vector. 6.The crop type change-based abandoned field plot identification method of claim 1, wherein, identifying crop type labels of each phase in the plot time series sample based on a time series segmentation network, and constructing a crop type change sequence, comprising: inputting the plot time series sample into a trained time series segmentation network, performing pixel-by-pixel classification on the pixels of each phase image, and outputting the corresponding crop type label; aggregating the crop type labels according to the cadastral vector boundary to generate a crop type result corresponding to each plot; encoding the crop type results of each phase in chronological order to form a crop type change sequence for representing the crop evolution process. 7.The crop type change-based abandoned field plot identification method of claim 1, wherein, inputting the crop type change sequence and the plot multi-dimensional feature vector into a plot discrimination model to obtain an initial management state judgment result, comprising: performing time series encoding on the crop type change sequence to generate a time series embedding vector representing the crop evolution process; performing feature mapping on the plot multi-dimensional feature vector to generate a comprehensive feature embedding representing the spectral information, time series information, and spatial structure information of the plot; fusing the time series embedding vector and the comprehensive feature embedding in a unified feature space, and inputting them into the plot discrimination model; the plot discrimination model outputs an initial management state judgment result based on the fused input. 8.The crop type change-based abandoned field plot identification method of claim 7, wherein, The training rule of the plot discrimination model is: collecting historical plot time series samples with plot management state labels, which contain corresponding crop type change sequences and plot multi-dimensional feature vectors; taking the crop type change sequence and the plot multi-dimensional feature vector as input, and taking the plot management state label as a supervision signal, performing supervised learning training on the plot discrimination model; introducing boundary constraint loss and time series consistency loss during training, and obtaining the plot discrimination model after training convergence. 9.A system for identifying abandoned land plots based on crop type changes, the system comprising: The system is used to perform the abandoned plot recognition method based on crop type change according to any one of claims 1-8, and the system comprises: The collection unit is used for acquiring multi-temporal remote sensing images of a target area and cadastral vector boundaries, performing preprocessing on the multi-temporal remote sensing images to obtain preprocessed images, and performing spatial registration and vector mask cropping on the preprocessed images based on the cadastral vector boundaries to generate land parcel time series samples corresponding to each land parcel; The feature extraction unit is used for extracting a land parcel multi-dimensional feature vector including a spectral feature, a time series feature and a spatial structure feature from the land parcel time series samples within a preset phenology window; The judgment unit is used for identifying crop type labels of each time phase in the land parcel time series samples based on a time series segmentation network, constructing a crop type change sequence, and inputting the crop type change sequence and the land parcel multi-dimensional feature vector into a land parcel discrimination model to obtain an initial judgment result of the management state; The output unit is used for performing correction on the initial judgment result of the management state based on a preset confusion resolution rule set and outputting a land parcel abandonment recognition result.

Images