Crop classification method combining random mask strategy and plot scale features

Abstract

The invention discloses a crop classification method combining a random mask strategy and plot scale features, and aims to improve the crop classification precision and optimize the final crop drawing effect. The diversity of training data is enhanced based on a time sequence random mask strategy, and spatial information in a cultivated land parcel is fully mined through a gray level co-occurrence matrix and the like. Experimental results show that compared with a traditional method, the method has the advantages that the crop type classification performance is better, and the F1 score of each crop is remarkably increased.

Description

Technical Field

[0001] The invention relates to the technical field of remote sensing ground object information extraction, and in particular to a crop classification method combining a random mask strategy with plot scale characteristics. Background Art

[0002] Accurate and timely spatial distribution information of crop types is an important basis for food crop yield prediction and sustainable agricultural management, and is a basic prerequisite for solving explicit and implicit hunger. However, the traditional use of field surveys and statistical data to obtain crop distribution information is costly, especially for large-scale and long-term monitoring of the distribution of multiple crops. Crop mapping using rapidly developing satellite remote sensing images is an economical and effective alternative to obtain crop planting area and spatial distribution information.

[0003] The current crop type identification methods based on remote sensing images can be divided into two categories: threshold methods based on specific phenological characteristics and machine learning methods. The former identifies specific phenological characteristics of crops during their growth period through pre-set classification rules and thresholds. Since different crops have specific physiological characteristics during their growth stages, crop classification can be achieved by formulating specific classification rules and dividing thresholds. However, the effectiveness of such methods is highly dependent on the stability of the physiological characteristics and has poor generalization. In contrast, machine learning methods have been proven to be strong generalization methods for crop mapping. They are able to establish the relationship between satellite spectral features and crop types, and have been successfully applied to classify different crop types over large areas. However, traditional machine learning methods can only use shallow features, which limits the performance of the model and leads to poor transferability across space and time. In addition, due to the difficulty in collecting a large number of samples, emerging deep learning methods often face the problem of insufficient training data, which may cause the model to focus on training only a small part of the features or noise in the data, resulting in poor classification performance. At the same time, crop classification based on pixel scale is also susceptible to the "salt and pepper noise" of the image, resulting in insufficient classification accuracy. There are many public farmland plot products and open-source farmland plot extraction methods, but there is a lack of crop mapping methods targeting the spatiotemporal spectral characteristics of farmland plots. Therefore, there is an urgent need for a crop classification method that can fully utilize the spatiotemporal spectral information at the plot scale in small sample scenarios to improve the crop recognition accuracy in complex terrain areas. Summary of the invention

[0004] To solve the above problems, the purpose of the present invention is to provide a crop classification method that combines a random masking strategy with plot scale characteristics, which can effectively improve the accuracy of crop mapping based on remote sensing.

[0005] In order to achieve the above object, the following is the technical solution of the present invention:

[0006] A crop classification method combining random masking strategy and plot scale characteristics includes the following steps:

[0007] S1, obtain satellite remote sensing image data of the study area and preprocess it, and generate a time series spectral reflectance feature set through the median synthesis method;

[0008] S2. Collect or generate a data set of cultivated land plots in the study area, and generate a spatiotemporal spectral feature data set based on the cultivated land plots divided by the data set of cultivated land plots, in coordination with the time series spectral reflectance feature set and the gray level co-occurrence matrix;

[0009] S3, using a temporal random mask to process the spatiotemporal spectral feature data set to obtain an amplified feature data set;

[0010] S4, constructing a deep learning model, and performing deep learning model training based on the expanded feature data set and the cultivated land type label of each cultivated land plot in the corresponding cultivated land plot data set;

[0011] S5. Input the spatiotemporal spectral feature dataset to be classified into the trained deep learning model to obtain the corresponding crop type.

[0012] As mentioned above, S1 specifically includes the following steps:

[0013] S11, obtaining satellite remote sensing image data of each study area within the corresponding collection time period;

[0014] S12, preprocessing the satellite remote sensing image data: identifying clouds, cloud shadows and snow for each satellite remote sensing image in the satellite remote sensing image data, and removing the affected low-quality satellite remote sensing images; linearly interpolating two good satellite remote sensing images at adjacent times to fill in the missing values ​​caused by low-quality satellite remote sensing images;

[0015] The preprocessed satellite remote sensing image data are subjected to median synthesis for the selected spectral bands to generate a time series spectral reflectance feature set.

[0016] As mentioned above, S2 includes the following steps:

[0017] S21, collecting or generating a data set of cultivated land plots of the same year according to the acquisition time period corresponding to the satellite remote sensing image data of the study area in step S11;

[0018] S22, based on the cultivated land plots divided in the cultivated land plot data set and the time series spectral reflectance characteristics corresponding to the cultivated land plots, the maximum value, average value, minimum value and standard deviation of the reflectance corresponding to each selected spectral band in each cultivated land plot are calculated to obtain a time series spectral reflectance characteristic data set at the cultivated land plot scale;

[0019] S23. From the cultivated land parcels divided in the cultivated land parcel dataset, select the near-infrared, red, and green bands from the corresponding preprocessed satellite remote sensing image data to synthesize a pseudo-color image, convert the color space of the pseudo-color image from RGB to the YUV space, derive the gray-level co-occurrence matrix GLCM based on the grayscale image Y in the YUV space, and extract the spatial texture dataset at the cultivated land parcel scale based on the gray-level co-occurrence matrix GLCM;

[0020] S24. Merge the spatial texture dataset at the cultivated land parcel scale and the temporal spectral reflectance feature dataset at the cultivated land parcel scale to generate a spatio-temporal spectral feature dataset.

[0021] As described above, step S3 includes the following steps:

[0022] S31. Construct multiple temporal random masks;

[0023] S32. Multiply each temporal random mask by the time series data in the spatio-temporal spectral feature dataset respectively to obtain a subset of the augmented feature dataset correspondingly, and the set of subsets of the augmented feature dataset is the augmented feature dataset.

[0024] Constructing the temporal random mask as described above includes the following steps:

[0025] Set the mask threshold to m, where m < t, and t is equal to the length of the time series in the spatio-temporal spectral feature dataset;

[0026] Define the temporal random mask as [M1, M2, …, M t , and the values of m elements in the elements M1, M2, …, M t are 0, and the values of (t - m) elements are 1. In different temporal random masks, the temporal positions of the elements with a value equal to 0 are random.

[0027] As described above, the deep learning model sequentially includes a double-layer bidirectional long short-term memory network module, an attention mechanism module, and an output layer.

[0028] As described above, extracting the spatial texture dataset at the cultivated land parcel scale based on the gray-level co-occurrence matrix GLCM specifically includes the following steps:

[0029] Select the types of co-occurrence offsets and the types of texture features;

[0030] For each cultivated land parcel divided in the cultivated land parcel dataset, for each pixel in the cultivated land parcel, calculate the texture feature values of various texture features corresponding to each co-occurrence offset;

[0031] For each texture feature of each co-occurrence offset corresponding to each cultivated land parcel, take the average value of the texture feature values based on all the pixels in the cultivated land parcel to obtain the corresponding plot co-occurrence offset texture feature value;

[0032] For each texture feature corresponding to each cultivated land parcel, the symbiotic offset texture feature values ​​of the parcels are averaged based on the symbiotic offset, and the texture feature mean value of each texture feature corresponding to each cultivated land parcel is obtained.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. This method uses random masking technology to fully augment time series remote sensing image data to overcome the negative impact of insufficient training data on deep learning methods, which is more conducive to crop mapping and recognition tasks in small sample scenarios.

[0035] 2. This method works with the plot dataset to overcome the influence of the "salt and pepper phenomenon" at the pixel scale, while rationally integrating the temporal image texture and optical features at the plot scale. By combining with the BiLSTM model, it can significantly improve the accuracy of crop classification and optimize the mapping effect.

[0036] 3. Compared with the use of time series masks in the prior art, the random mask of the present invention only fixes the number of mask time series, and the position of the mask time series is random. While improving the efficiency of the model, it reduces the redundancy of the training data, and randomly masks the different time series positions of different samples in each training, thereby improving the robustness and generalization of the model. Compared with the invention using cultivated land plot data in the prior art, the present invention is not only limited to object-oriented classification based on cultivated land plot data, but also further uses the spectral texture features of cultivated land plots as units in the classification process. Including but not limited to the maximum value, average value, minimum value and standard deviation of different band features within the plot range, as well as homogeneity, entropy value, angular second-order moment and correlation texture features calculated based on the gray-level co-occurrence matrix. This improvement makes full use of the information of the cultivated land plot unit and further improves the final classification accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the present invention;

[0038] Figure 2a This is an enlarged view of the cultivated land plot corresponding to the study area A, and the corresponding cultivated land plot data sampling points; the study area A is France;

[0039] Figure 2b This is an enlarged view of the cultivated land plot corresponding to the study area B, and the corresponding cultivated land plot data sampling points; the study area B is the Inner Mongolia Autonomous Region;

[0040] Figure 2c This is an enlarged view of the cultivated land plot corresponding to the study area C, and the corresponding cultivated land plot data sampling points; the study area C is Hubei Province;

[0041] Figure 3 It is a structural diagram of the deep learning model of the present invention; wherein, Mask-OLSTM represents a long short-term memory network Mask Object-based LongShort-Term Memory based on random mask and plot-scale spatiotemporal spectral features, which is the deep learning model adopted by the present invention; BiLSTM represents a two-layer bidirectional long short-term memory network, Attention represents an attention mechanism module, Output layer represents an output layer, and LSTM represents a long short-term memory network;

[0042] Figure 4 It is the crop classification accuracy evaluation result of the present invention; wherein, RF represents the random forest method, SVM represents the support vector machine method, Mask-OLSTM (MaskObject-based Long Short-Term Memory, Mask-OLSTM) represents a crop classification method combining random mask strategy with plot scale characteristics of the present invention, and F1, PA, and UA in the figure refer to F1 score (F1-score), producer accuracy, and user accuracy, respectively. DETAILED DESCRIPTION

[0043] In order to facilitate ordinary technicians in the field to understand and implement the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and examples. It should be understood that the embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0044] A crop classification method combining random masking strategy with plot scale characteristics is used to solve the problems in the existing technology that the number of samples is insufficient, the mapping results are seriously affected by salt and pepper noise, resulting in low crop recognition and verification accuracy, and the final crop mapping results are unreliable.

[0045] like Figures 1 to 3 As shown in FIG. 1 , a crop classification method combining a random masking strategy with plot scale features includes the following steps:

[0046] S1. Obtain satellite remote sensing image data of the study area and preprocess it, and generate a time series spectral reflectance feature set through the median synthesis method.

[0047] In this embodiment, Sentinel-2 remote sensing image data is acquired and preprocessed based on the Google Earth Engine platform, and a time series spectral reflectance feature set is generated by a 20-day median synthesis method; in addition, satellite remote sensing image data such as MODIS satellite remote sensing image data and Landsat series satellite remote sensing image data are also applicable to the method of the present invention.

[0048] The specific process of step S1 is:

[0049] S11. Download satellite remote sensing image data of each study area in the corresponding collection time period through Google Earth Engine (GEE) platform:

[0050] The present invention downloads satellite remote sensing image data for crop classification through the Google Earth Engine platform. Specifically, the satellite remote sensing image data of study area A, study area B and study area C were downloaded to obtain the surface reflectance data set (the distribution of study areas and field sampling points is shown in Figure 2a-2c ). According to the crop phenology and field sampling date of each study area, this embodiment selects Sentinel-2 remote sensing image data within the following time periods: the collection time period corresponding to study area A is from November 1, 2019 to September 16, 2020, the collection time period corresponding to study area B is from May 1, 2019 to December 27, 2019, and the collection time period corresponding to study area C is from May 1, 2021 to October 27, 2021.

[0051] S12. Preprocess the downloaded satellite remote sensing image data and generate a time series spectral reflectance feature set by median synthesis:

[0052] Selection of spectral bands: In order to capture the complex spatiotemporal characteristics of areas with strong heterogeneity, the present invention selects four spectral bands with a spatial resolution of 10 meters (i.e., blue light, green light, red light, and near-infrared bands).

[0053] First, the invention preprocesses the downloaded satellite remote sensing image data, identifies clouds, cloud shadows and snow according to the QA10 quality band in the satellite remote sensing image data, and removes the affected low-quality satellite remote sensing images in the satellite remote sensing image data. Linear interpolation is performed by two good satellite remote sensing images of adjacent time to fill the missing values ​​caused by low-quality satellite remote sensing images, and obtain preprocessed satellite remote sensing image data.

[0054] For the preprocessed satellite remote sensing image data, based on a 20-day synthesis period, median synthesis is performed on the selected spectral bands to generate a time series spectral reflectance feature set.

[0055] S2. Collect or generate a data set of cultivated land plots in the study area, and generate a spatiotemporal spectral feature data set based on the cultivated land plots divided by the data set of cultivated land plots, in coordination with the time series spectral reflectance feature set and the gray-level co-occurrence matrix.

[0056] The specific process of step S2 is:

[0057] S21. Collect or generate a dataset of cultivated land plots in the study area;

[0058] According to the acquisition time period corresponding to the satellite remote sensing image data of the study area in step S11, the cultivated land plot data set of the same year is collected or generated:

[0059] This example uses the 2020 cultivated land plot dataset from the Registre Parcellaire Graphique (RPG) database (https: / / geoservices.ign.fr / rpg) in study area A. The RPG database is an open-access georeferenced vector dataset generated through field surveys, produced and maintained by the Agency for Services and Payments (ASP). Each cultivated land plot was manually digitized by farmers to centimeter-level resolution and the corresponding crop type was recorded. The RPG database has been widely used in crop type mapping in various scientific research communities.

[0060] For study area B and study area C, due to the lack of open access to cultivated land plot datasets in China, this embodiment obtains the cultivated land plot datasets of these two study areas based on the DSTFNet model: First, the DSTFNet model is used to generate the initial cultivated land plot datasets of the two study areas in 2019 and 2021, respectively. Then, the cultivated land plot datasets are refined, checked and corrected by manual means. Finally, cultivated land plots with an area of ​​less than 100 square meters are eliminated to correspond to the spatial resolution of satellite remote sensing image data (i.e., 10-meter spatial resolution), thereby generating a cultivated land plot dataset that can be used for subsequent crop classification.

[0061] S22, based on the cultivated land parcels divided in the cultivated land parcel data set obtained in step S21 and the time series spectral reflectance feature sets corresponding to the cultivated land parcels, generate a time series spectral reflectance feature data set at the cultivated land parcel scale;

[0062] In previous classification studies based on the scale of cultivated land plots, satellite remote sensing data in each cultivated land plot are usually averaged to reduce the salt and pepper noise phenomenon. However, this method often ignores the spectral changes of crops in the spatial domain within the cultivated land plots, and thus cannot fully utilize the spatial information within the cultivated land plots. In order to address this limitation, the present invention calculates the maximum value, average value, minimum value and standard deviation of the reflectivity corresponding to each selected spectral band in each cultivated land plot based on the cultivated land plots divided in the cultivated land plot data set and the time series spectral reflectance characteristics corresponding to the cultivated land plots to obtain a time series spectral reflectance feature data set at the scale of cultivated land plots, thereby more comprehensively reflecting the spectral dynamics of crops in space, especially in the early stages of crop growth.

[0063] S23, generating a spatial texture dataset at the scale of cultivated land plots based on the cultivated land plots divided in the cultivated land plot dataset, the pre-processed satellite remote sensing image data and the gray-level co-occurrence matrix;

[0064] The present invention adopts the Gray-Level Co-occurrence Matrix (GLCM) method to fit the spatial correlation between adjacent pixels. In order to highlight the characteristics of vegetation, the present invention selects near-infrared, red and green bands from the pre-processed satellite remote sensing image data obtained in the corresponding step S21 to synthesize a pseudo-color image according to the cultivated land plots divided in the cultivated land plot data set, and converts the color space of the pseudo-color image from RGB to YUV space, where the Y layer represents the corresponding grayscale image; then, based on the grayscale co-occurrence matrix GLCM derived from the grayscale image Y, the spatial texture data set at the cultivated land plot scale is extracted based on the grayscale co-occurrence matrix GLCM;

[0065] The spatial texture dataset of cultivated land plot scale is extracted based on the gray level co-occurrence matrix GLCM, which includes the following steps:

[0066] (1) Select the type of symbiotic offset and the type of texture feature; In this embodiment, four commonly used texture features are extracted: homogeneity (Hom), entropy (Ent), angular second moment (ASM) and correlation (Cor). Different symbiotic offsets will produce different texture feature values, so the present invention uses four symbiotic offsets (0, 1), (1, 0), (1, 1) and (-1, -1) of the gray-level symbiotic matrix GLCM for calculation; Table 1 shows the calculation formulas for selecting four commonly used texture features in this embodiment.

[0067] (2) For each cultivated land parcel divided in the cultivated land parcel data set, for each pixel in the cultivated land parcel, the texture feature values ​​of various texture features corresponding to each symbiotic offset are calculated. In this embodiment, each pixel corresponds to 16 texture feature values;

[0068] (3) For each texture feature of each symbiotic offset corresponding to each cultivated land parcel, an average value of the texture feature values ​​is taken based on all pixels in the cultivated land parcel to obtain the corresponding symbiotic offset texture feature value of the parcel. In this embodiment, each cultivated land parcel corresponds to 16 symbiotic offset texture feature values ​​of the parcel;

[0069] (4) For each texture feature corresponding to each cultivated land plot, the symbiotic offset texture feature values ​​of the plots are averaged based on the symbiotic offset, so that each cultivated land plot obtains the texture feature mean of each texture feature. In this embodiment, each cultivated land plot obtains four texture feature means (i.e., the mean of homogeneity, the mean of entropy, the mean of angular second-order moment, and the mean of correlation).

[0070] Table 1 Commonly used calculation formulas for texture features

[0071]

[0072] Wherein, N is the total number of gray levels in the gray level co-occurrence matrix, and in the present invention, N is set to 64. The element P in the gray level co-occurrence matrix i,j represents the co-occurrence probability of gray level i and gray level j. μ and σ 2 They are the mean and variance of the gray-level co-occurrence matrix, and the specific μ i Represents the element P in the gray-level co-occurrence matrix i,0 , P i,1 …、P i,N The mean of Represents the element P in the gray-level co-occurrence matrix i,0 , P i,1 …、P i,N Variance of j Represents the element P in the gray-level co-occurrence matrix 0,j , P 1,j …、P N,j The mean of Represents the element P in the gray-level co-occurrence matrix 0,j , P 1,j …、P N,j The variance of .

[0073] S24, merging the spatial texture dataset at the scale of cultivated land plots and the temporal spectral reflectance feature dataset at the scale of cultivated land plots to generate a spatiotemporal spectral feature dataset. In the data structure of the spatiotemporal spectral feature dataset, for each cultivated land plot and each temporal position, the maximum value, average value, minimum value and standard deviation of the reflectance corresponding to each spectral band calculated in step S22, and the texture feature mean value of each texture feature calculated in step S23 (4 texture feature means in this embodiment) are included.

[0074] S3, construct a random mask strategy, and use the temporal random mask to process the spatiotemporal spectral feature data set to obtain an amplified feature data set;

[0075] S31, construct multiple timing random masks;

[0076] In view of the complex crop planting patterns and the data missing problem caused by adverse observation conditions such as clouds and rainfall, the traditional deep learning model may face the problem of overfitting due to insufficient training data, which leads to the decline of the prediction ability of the deep learning model. In order to solve this problem, the present invention adopts a temporal random mask strategy to perform training data amplification processing to enhance the generalization ability and robustness of the deep learning model.

[0077] Specifically, suppose a time series data [X1, X2, …, Xt , X1, X2, …, X t represent elements in time - series data. First, set a masking threshold m. m represents the number of time - series positions to be masked randomly selected from the time series, and m < t. In the present invention, t is equal to the length of the time series of the spatio - spectral feature dataset. Then, define the time - series random mask as [M1, M2, …, M t , and among the elements M1, M2, …, M t , the values of m elements are 0 and the values of t - m elements are 1. In different time - series random masks, the time - series positions of the elements with value 0 are random.

[0078] S32: Multiply each time - series random mask with the time - series data in the spatio - spectral feature dataset respectively to obtain subsets of the augmented feature dataset. The set of subsets of the augmented feature dataset is the augmented feature dataset;

[0079] The masking threshold m is selected according to the time series of the spatio - spectral feature dataset. In the present invention, the value of the masking threshold m is set to one - quarter of the length of the time series of the spatio - spectral feature dataset. That is, before the deep - learning model, in the spatio - spectral feature dataset, one - quarter of the time - series data will be randomly masked to simulate the possible observation loss in actual data and improve the generalization of the deep - learning model. In the present invention, the time - series data in the spatio - spectral feature dataset is multiplied with each time - series random mask to obtain subsets of the augmented feature dataset. The set of subsets of the augmented feature dataset is the augmented feature dataset.

[0080] S4: Construct a deep - learning model and train the deep - learning model based on the augmented feature dataset and the cultivated land type labels of each cultivated land plot in the corresponding cultivated land plot dataset;

[0081] To effectively mine the dependencies between different time features in the augmented feature dataset, the deep - learning model constructed in the present invention sequentially includes a double - layer bidirectional long - short - term memory network (BiLSTM) module, an attention mechanism module (Attention), and an output layer. Among them, the long - short - term memory network (LSTM) in the double - layer bidirectional long - short - term memory network module introduces memory cells and gating mechanisms, which can retain the information stored at the previous time and the current time. The bidirectional mechanism processes the information in the input sequence simultaneously forward and backward to help the deep - learning model explore high - level time features in the input time - series data. In addition, the attention mechanism module is a fully - connected layer with a softmax function, which dynamically assigns weights to the hidden features from different times to aggregate their information. Through these aggregated features, the ability of the model to identify crops can be significantly enhanced. The final model is as Figure 3 shown.

[0082] In this embodiment, the cultivated land type label of each cultivated land plot in the cultivated land plot data set and the amplified feature data set subset in the corresponding amplified feature data set are used as sample pairs. Based on the different corresponding temporal random masks, the sample pairs are divided into training sets and test sets. In each single cycle of model training, the sample pairs in the training set are first input into the double-layer bidirectional long short-term memory network BiLSTM, and the features output by the double-layer bidirectional long short-term memory network BiLSTM are input into the attention mechanism module and connected to the output layer, and finally the corresponding predicted crop type is obtained. The Focal Loss loss function is used to coordinate the iterative optimization during the training of the deep learning model. In this embodiment, the deep learning model is independently trained and verified in areas A, B and C to verify the generalization of the model in different areas.

[0083] S5. Input the spatiotemporal spectral feature dataset to be classified into the trained deep learning model to obtain the corresponding crop type.

[0084] In this embodiment, the subset of the amplified feature dataset in the test set is used as the spatiotemporal feature dataset to be classified, input into the trained deep learning model, and the F1 score is used to evaluate the accuracy of the predicted crop type. Figure 4 Compared with the random forest method (RF) and the support vector machine method (SVM), the method of the present invention has higher accuracy and better recognition effect.

[0085] The above contents are further explanations of the present invention in combination with specific implementation methods, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions or substitutions can be made without departing from the concept of the present invention, which should be regarded as falling within the protection scope of the present invention.

Claims

1. A crop classification method combining random masking strategy and plot scale characteristics, characterized in that: It includes the following steps: S1. Obtain the satellite remote sensing image data of the study area and perform preprocessing, and generate a time-series spectral reflectance feature set through the median synthesis method; S2. Collect or generate a cultivated land plot data set within the study area. Based on the cultivated land plots divided by the cultivated land plot data set, synergistically generate a spatio-temporal spectral feature data set with the time-series spectral reflectance feature set and the gray-level co-occurrence matrix; S3. Use time-series random masking to process the spatio-temporal spectral feature data set to obtain an augmented feature data set; S4. Construct a deep learning model, and perform deep learning model training based on the augmented feature data set and the cultivated land type labels of each cultivated land plot in the corresponding cultivated land plot data set; S5. Input the spatio-temporal spectral feature data set to be classified into the trained deep learning model to obtain the corresponding crop types.

2. A crop classification method combining random masking strategy and plot scale characteristics according to claim 1, characterized in that: The specific steps in S1 include the following: S11. Obtain the satellite remote sensing image data of each study area during the corresponding acquisition time period; S12. Perform preprocessing on the satellite remote sensing image data: identify clouds, cloud shadows, and snow in each satellite remote sensing image in the satellite remote sensing image data, and eliminate the affected low-quality satellite remote sensing images; perform linear interpolation on two adjacent good satellite remote sensing images to fill in the missing values caused by the low-quality satellite remote sensing images; For the preprocessed satellite remote sensing image data, perform median synthesis for the selected spectral bands to generate a time-series spectral reflectance feature set.

3. A crop classification method combining random masking strategy and plot scale characteristics according to claim 1, characterized in that: The steps in S2 include the following: S21. According to the acquisition time period corresponding to the satellite remote sensing image data of the study area in step S11, collect or generate a cultivated land plot data set of the same year; S22. Based on the cultivated land plots divided by the cultivated land plot data set and the corresponding time-series spectral reflectance features of the cultivated land plots, calculate the maximum value, average value, minimum value, and standard deviation of the reflectance corresponding to each selected spectral band in each cultivated land plot to obtain a time-series spectral reflectance feature data set at the cultivated land plot scale; S23. According to the cultivated land plots divided by the cultivated land plot data set, select the near-infrared, red, and green bands from the corresponding preprocessed satellite remote sensing image data to synthesize a pseudo-color image, convert the color space of the pseudo-color image from RGB to YUV space, derive the gray-level co-occurrence matrix GLCM based on the gray image Y in the YUV space, and extract a spatial texture data set at the cultivated land plot scale based on the gray-level co-occurrence matrix GLCM; S24. Merge the spatial texture data set at the cultivated land plot scale and the time-series spectral reflectance feature data set at the cultivated land plot scale to generate a spatio-temporal spectral feature data set.

4. The crop classification method combining random masking strategy and plot scale characteristics according to claim 1, characterized in that: The steps in S3 include the following: S31. Construct multiple time-series random masks; S32. Multiply each time-series random mask by the time-series data in the spatio-temporal spectral feature data set respectively to obtain subsets of the augmented feature data set, and the set of subsets of the augmented feature data set is the augmented feature data set.

5. A crop classification method combining random masking strategy and plot scale characteristics according to claim 4, characterized in that: Constructing the time-series random mask includes the following steps: Set the mask threshold to m, where m < t and t is equal to the length of the time series in the spatio-temporal spectral feature data set; Define the temporal random mask as [M1, M2, …, M t ], elements M1, M2, ..., M t There are m elements with a value of 0, (tm) elements with a value of 1, and in different temporal random masks, the temporal positions of the elements with a value equal to 0 are random.

6. The crop classification method combining random masking strategy and plot scale characteristics according to claim 1, characterized in that: The deep learning model sequentially includes a double-layer bidirectional long short-term memory network module, an attention mechanism module, and an output layer.

7. The crop classification method combining random masking strategy and plot scale characteristics according to claim 3, characterized in that: The method of extracting a spatial texture dataset at the scale of cultivated land plots based on a gray level co-occurrence matrix GLCM specifically includes the following steps: Select the type of symbiotic offset and the type of texture feature; For each cultivated land parcel divided in the cultivated land parcel data set, for each pixel in the cultivated land parcel, the texture feature values ​​of various texture features corresponding to each symbiotic offset are calculated; For each texture feature of each symbiotic shift corresponding to each cultivated land parcel, the average value of the texture feature value is taken based on all pixels in the cultivated land parcel to obtain the corresponding symbiotic shift texture feature value of the parcel; For each texture feature corresponding to each cultivated land parcel, the symbiotic offset texture feature values ​​of the parcels are averaged based on the symbiotic offset, and the texture feature mean value of each texture feature corresponding to each cultivated land parcel is obtained.

Images