Crop identification model construction method and system compatible with any time sequence remote sensing image

Abstract

The invention provides a crop identification model construction method and system compatible with any time sequence remote sensing image, and the method comprises the steps: obtaining a multi-source crop remote sensing image, obtaining the time sequence length of model input data based on the crop remote sensing image, and carrying out the preprocessing of the model input data according to the time sequence length; inputting the processed model input data into an improved Transform model, performing data enhancement on the processed model input data by adopting a random time mask method, and training the enhanced input data to obtain a final crop identification model; the method further comprises the step of identifying a plurality of crop types in the remote sensing image in any time sequence and any wave band through a crop identification model. On the basis of a Transformer model, optical Landsat-8, Sentinel-2 and Sentinel-1 SAR images are used as data sources, a deep learning framework compatible with any remote sensing time sequence and wave band information is designed, remote sensing time sequence characteristics of different crops are learned, and high-precision recognition of the crops in the growing period is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of crop identification, and in particular to a method and system for constructing a crop identification model compatible with any time-series remote sensing images. Background Art

[0002] Accurately grasping the temporal and spatial distribution information of crop types with strong currentness is of great significance for ensuring the balance of food supply and demand, the utilization of cultivated land resources and the reform of crop planting patterns. Traditional early crop identification usually requires field surveys and censuses, which are time-consuming and labor-intensive and cannot complete regional crop type mapping. Remote sensing data can be used for large-scale and rapid ground observations to monitor crop growth conditions. The main idea of ​​crop identification research based on remote sensing is to take crops as the research target, fully explore the differences in spectrum, phenology (time) and texture (space) between crops and other land cover types, and use expert knowledge, machine learning and deep learning to achieve crop classification.

[0003] There are many semi-automatic / automatic classification methods, and the commonly used ones can be classified into three categories: multi-temporal index method, machine learning method and deep learning method. The multi-temporal index curve method, that is, the crop mapping method based on phenology, uses the regularity and uniqueness of the spectral index curve of multi-temporal remote sensing images in the key growth period of crops to distinguish different crops in the field. It usually analyzes the peaks and valleys or mutation characteristics of the spectral index within the year to identify different types of crops. The multi-temporal index method has high requirements on the time series density and quality of remote sensing data. Due to different scenarios such as natural conditions and farming management differences in different regions or different years, the same crop and growth may also change, resulting in intra-class differences in the time series curve. There is a certain uncertainty in the selection of thresholds when analyzing the time series index curves of different crops on a large scale with strong spatial heterogeneity. Most machine learning methods are based on time series image data, and use methods such as k-means clustering, random forests and decision trees for crop recognition research. However, machine learning algorithms can only add pre-defined crop recognition spatial features and cannot adaptively learn their spatial features. At the same time, they cannot effectively learn time series features and require time series regularity data. Denoising methods such as remote sensing data time series interpolation or synthesis are computationally intensive and time-consuming in large-area crop identification. They may also introduce errors or cause time series information loss, which in turn affects the next step of crop identification. Deep learning methods include convolutional neural networks and recurrent neural networks. By abstracting image features in layers, learning the inherent laws of sample data, and characterizing complex textures, spatial structures, and time series changes, they can adaptively learn to summarize and extract image features, highlighting their important role in crop identification. However, existing deep learning methods cannot fully utilize the band information of different remote sensing data sources, which limits the recognition accuracy of crop types. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for constructing a crop recognition model compatible with any time-series remote sensing images, aiming to solve the above-mentioned problems in the prior art.

[0005] The embodiment of the present invention provides a method for constructing a crop recognition model compatible with any time-series remote sensing images, including:

[0006] Acquire multi-source crop remote sensing images, obtain the time series length of model input data based on the crop remote sensing images, and pre-process the model input data according to the time series length; and

[0007] The processed model input data is input into the improved Transformer model, a random time mask method is used to perform data enhancement on the processed model input data, and the enhanced input data is trained to obtain a final crop recognition model.

[0008] The embodiment of the present invention provides a crop recognition model construction system compatible with any time-series remote sensing images, including:

[0009] a processing module, used for acquiring multi-source crop remote sensing images, obtaining a time series length of model input data based on the crop remote sensing images, and preprocessing the model input data according to the time series length; and

[0010] A construction module is used to input the processed model input data into the improved Transformer model, perform data enhancement on the processed model input data using a random time mask method, and train the enhanced input data to obtain a final crop recognition model.

[0011] An embodiment of the present invention also provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the steps of the above-mentioned method for constructing a crop recognition model compatible with any time-series remote sensing image are implemented.

[0012] An embodiment of the present invention further provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps of the above-mentioned method for constructing a crop recognition model compatible with any time-series remote sensing image are implemented.

[0013] The use of the embodiments of the present invention can include the following beneficial effects: The embodiments of the present invention are based on the Transformer model, take optical Landsat-8, Sentinel-2 and Sentinel-1SAR images as data sources, design a deep learning framework compatible with any remote sensing time series and band information, learn the remote sensing time series characteristics of different crops, and realize early identification and high-precision identification of crops during the growth period. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate one or more embodiments of this specification or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0015] Figure 1 This is a flow chart of a method for constructing a crop recognition model compatible with any time-series remote sensing images according to an embodiment of the present invention;

[0016] Figure 2 Schematic diagram of the organization of multi-source remote sensing data in an embodiment of the present invention (wherein the dark grey without characters represents the mask value);

[0017] Figure 3 is a schematic diagram of a crop recognition deep learning model framework according to an embodiment of the present invention;

[0018] Figure 4 Schematic diagram of the accuracy change of the model training test samples in an embodiment of the present invention;

[0019] Figure 5 It is a schematic diagram of a crop recognition model construction system compatible with any time-series remote sensing images according to an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the following will be combined with the drawings in one or more embodiments of this specification to clearly and completely describe the technical solutions in one or more embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this document.

[0021] Method Embodiment

[0022] According to an embodiment of the present invention, a method for constructing a crop recognition model compatible with any time-series remote sensing images is provided. Figure 1 is a flow chart of a method for constructing a crop recognition model compatible with arbitrary time-series remote sensing images according to an embodiment of the present invention. Figure 1 As shown, the method for constructing a crop recognition model compatible with any time-series remote sensing image according to an embodiment of the present invention specifically includes:

[0023] Step S101, acquiring multi-source crop remote sensing images, obtaining the time series length of model input data based on the crop remote sensing images, and preprocessing the model input data according to the time series length, specifically includes:

[0024] The crop remote sensing images of Landsat-8, Sentinel-2 and Sentinel-1 SAR sensors are obtained respectively, and the pixel points of known crop types sampled in advance are used as sample points. Based on the corresponding positions of the sample points in the crop remote sensing images, the effective observation time points of each crop remote sensing image in the year are obtained and a union operation is performed to obtain the effective observation number of the sample point in the year. The maximum value of the effective observation number in the year is taken as the time series length of the model input data. The effective observation time points of the sample pixels in each crop remote sensing image are calculated and a union operation is performed to obtain the number of elements in the effective observation time point set of the sample pixels. The number of elements is compared with the time series length. If the number of elements is less than the time series length, the missing part of the data is filled with a mask value to obtain the model input data with the same time series length.

[0025] The model input data includes a time dimension and a feature dimension;

[0026] The characteristic dimensions include image band, time point and year;

[0027] Step S102, inputting the processed model input data into the improved Transformer model, performing data enhancement on the processed model input data using a random time mask method, and training the enhanced input data to obtain a final crop recognition model, specifically including:

[0028] Input the processed model input data into the improved Transformer model, adopt a data enhancement strategy of random time mask for each batch of data in the model training process to generate multimodal training data, construct position encoding for the training data using the regularization method shown in Formula 1, establish a mask matrix if the input at a certain time point in the training data is a mask value, multiply the mask matrix by the feature of the training data to obtain a mask input, add the mask input and the position encoding through a fully connected layer of the same dimension to obtain new input data, output a time series feature vector representation of each sensor data for the new input data through a four-layer transformer encoding module, fuse the time series feature vector representation of each sensor data in the time dimension to generate a time series feature vector, use the time series feature vector as an intermediate input and pass through two layers of transformer encoding modules to obtain a final crop recognition time series feature vector, average the non-mask feature vectors in the crop recognition time series feature vector, and obtain the discrete probability distribution of each crop type through a fully connected layer with a Softmax function to obtain a final crop recognition model;

[0029]

[0030] Among them, PE (i) Indicates time position coding, doy i represents the Julian day (1, 2, 3, ..., 365), i is the index of each time point in the sample time series, mean is the mean, and std is the standard deviation;

[0031] The sub-layer dimension of the four-layer transformer encoding module is the same as the input layer dimension;

[0032] The multi-head self-attention mechanism layer of each encoding module establishes an attention mask matrix according to the mask matrix;

[0033] The method further comprises:

[0034] The crop recognition model is used to recognize several crop types in remote sensing images of any time sequence and any band.

[0035] The above technical solution of the embodiment of the present invention is described in detail below in conjunction with the specific situation of the method for constructing a crop recognition model compatible with any time series remote sensing image in the embodiment of the present invention.

[0036] The embodiment of the present invention is based on the Transformer model, and uses the optical Landsat-8, Sentinel-2 and Sentinel-1 SAR images as data sources to design a deep learning framework that is compatible with any remote sensing time series and band information, learns remote sensing time series features of different crops, and achieves high-precision recognition of crop types. First, based on multi-source remote sensing data, a fixed time series of remote sensing images containing invalid observations but with fewer invalid observations is generated to minimize the amount of model input data; then, based on the encoder module of the Transformer model, the time series features of different remote sensing images are deeply mined by fusing mask mechanisms, optimizing position encoding, and combining intermediate features of different remote sensing sources; finally, the time series data is enhanced by using the time mask method to be compatible with time series input of any length, so as to achieve early recognition of crops during the growth period.

[0037] 1. Fixed time series generation of multi-source remote sensing images

[0038] The time resolution of the Landsat-8 Operational Land Imager (OLI) sensor for earth observation is 16 days, and the time resolution of the Sentinel-2A&2B Multi-spectral instrument (MSI) sensor for earth observation is about 5 days. In addition, due to the influence of fog, cloud shadows, etc., there are differences in the observation time series of Landsat and Sentinel-2 optical images. In contrast, Sentinel-1 SAR observations are not affected by weather, and the data time resolution is 12 days. However, the Transformer-based deep learning model requires that the samples have equal length time series. For this reason, the embodiment of the present invention first obtains the valid observation time points of Landsat-8, Sentinel-2A&2B and Sentinel-1 images in the year (such as Julian day 2020302) based on the positions of sample points (sample points are pixels of known crop types sampled in advance), and performs a union operation on the valid observation time point sets of the three to obtain the valid observation number of the sample point in the year, and finally obtains the valid observation number of all samples, and obtains the maximum value as the time series length N of the model input data. max When organizing the input data, first calculate the effective observation time point set of the target pixel (i.e., pixel-by-pixel prediction during model inference) in the three remote sensing data sources, and find the union of the two sets as the effective observation time point set of the pixel. The number of set elements is N. union In this time point set, at least one remote sensing data source has valid observations, and the data sources with invalid observations are filled with mask values ​​(-9999). union Less than N max , then the other three remote sensing data sources (N max-N union ) The observations and time points at the time point are filled with mask values. This input storage method ensures samples with equal time series and minimizes the sample time series length. Figure 2 The data organization diagram using 6 bands of Landsat-8, 10 bands of Sentinel-2A&2B and 3 bands of Sentinel-1 (Table 1) is shown, where the band is represented by ρ and the data input time dimension is N max , the feature dimension is m, which includes the image band, time point (Julian day number DOY) and year. For Sentinel-2 data, m is 12, for Landsat 8 data, m is 8, and for Sentinel-1 data, m is 5.

[0039] In order to reduce the difference between the six optical bands corresponding to Sentinel-2 and Landsat 8, the embodiment of the present invention adopts NASA Harmonized Landsat Sentinel-2 (HLS), Sentinel-2 Multispectral Instrument (MSI) (HLSS30) and Landsat-8 / 9 Operational Land Imager (OLI) (HLSL30) surface reflectance (V2.0). The data spatial resolution is 30m and the temporal resolution is 2-3 days. HLS data processing includes Landsat and Sentinel sensor radiation adjustment, atmospheric correction, cloud, cloud shadow detection, spatial alignment, bidirectional reflectance distribution function normalization and band adjustment, which reduces the difference between the two data sources themselves. Sentinel-1SAR data uses the ground-range multi-view product (GRD) VH and VV polarization data of the interferometric wide-band mode (IW), with a spatial resolution of 10 meters and a temporal resolution of 12 days.

[0040] Table 1 Sample feature dimension information based on multi-source remote sensing data

[0041]

[0042] 2. Construction of deep learning model for crop identification

[0043] Transformer is a deep learning model that uses a self-attention mechanism to perform differential weighting on each part of the input data to extract time series features. At the same time, feature learning at different time points can be independent and parallel, which is more efficient than the LSTM model in training and prediction. The classic Transformer model is mainly composed of an encoding module and a decoding module. The embodiment of the present invention improves the classic Transformer model when modeling, and represents the output of the encoding module as a time series feature vector, such as Figure 3 shown.

[0044] ①Model input

[0045] Will Figure 2 The sample time series of the organization shown in the figure is used as input. Different from the classic encoding based on sin and cos functions of the Transformer model, the embodiment of the present invention adopts a regularization method to construct positional encodings (PE), as shown in Formula 1. If the input at a certain time point is an invalid value (i.e., masked), the Masked mechanism is introduced to establish N max ×m mask matrix, multiplied with the input data features to form mask input, and the mask input and position encoding are added through a fully connected Dense layer of the same dimension (d) to achieve model input data preprocessing and dynamic extraction of relative time position relationship. max -N union The observations and DOY at the time point are masked, resulting in the model input being completely masked, and some remote sensing observations are masked, but the DOY position encoding is retained. The three sensor data are independently masked and the relative time position relationship is dynamically extracted.

[0046]

[0047] Among them, PE (i) Indicates time position coding, doy i represents the Julian day (1, 2, 3, ..., 365), i is the index of each time point in the sample time series, mean is the mean, and std is the standard deviation.

[0048] ②Encoding module

[0049] Taking ① as input, the three data sources are respectively passed through the four-layer transformer encoding module to extract the data time series features. Similar to the classic Transformer model, the encoding module contains two sub-layers: a multi-head self-attention mechanism (Multihead) and a fully connected feedforward neural network (FFN). A residual block is introduced after each sub-layer, and the output result of the sub-layer is added to the residual matrix and normalized. To achieve the connection between each sub-layer, the dimensions of all sub-layers in the encoder are the same as the dimensions of the input layer. The multi-head self-attention mechanism layer can establish an attention mask based on the mask matrix, automatically shielding the time points containing invalid values ​​so that they do not participate in differential weighting. Multiple heads can be added to the model to make the model pay attention to more time series features as much as possible. The multi-head attention mechanism layer is expressed as follows:

[0050] MultiHead(Q,K,V)=Concat(head1,head2,…,head n )W O (2);

[0051]

[0052] In the formula, the index i of the multi-head attention head is 1, 2, ..., n, W i Q ∈R d×m , W i K ∈R d×m , W O ∈R d×i×m are all parameter matrices, n is the number of heads of the multi-head attention mechanism, d is the dimension of the input data after the dense layer, m is the feature dimension of the input variable, and d k is the dimension of matrix K, and its value is d / T. Q, K, and V respectively represent three matrices obtained by concatenating vector query, key Key, and value Value according to the corresponding position dimensions. In the embodiment of the present invention, vector, key, and value are all set as input data feature vectors.

[0053] Then, the output of the multi-head self-attention mechanism layer is used as the input of the feedforward neural network sublayer (FFN). The transformation method is linear change, and the activation function selects the ReLU function, which is expressed as follows:

[0054] FFN(x)=max(0,xW1+b1)W2+b2 (4);

[0055] Among them, FFN(x) means x passes through a feed-forward neural network sublayer FFN, x represents the output vector of the multi-head self-attention mechanism layer, and the matrix W1∈R d×d , W2∈R d×d , the row vector b1∈R 1×d , b2∈R1×d The elements in are all parameters that need to be trained. Finally, a sample encoder outputs a time series feature vector representing h′∈R Nmax×d .

[0056] Finally, the time series feature vectors h′ output by the three sensor data are concatenated in the time dimension, and the three data time series features are fused to generate the time series feature vector h * ∈R Nmax×3d , then h * As input, it passes through two layers of transformer encoding modules to generate the final crop recognition time series feature vector h∈R Nmax×3d , all non-mask feature vectors are averaged in the time dimension and passed through a fully connected layer with a Softmax function to generate a discrete probability distribution for each crop type, thus achieving high-precision crop recognition.

[0057] 3. Data enhancement based on temporal mask

[0058] The embodiment of the present invention proposes a random mask data enhancement strategy to enable the model to adapt to various multimodal data inputs and increase the number of model samples. First, a DOY is randomly generated in t∈{1,2,3,…,366} r (DOY r Indicates the image acquisition time, expressed in Julian days, i.e. 1, 2, 3…), and then the time series of the training samples with DOY greater than DOY r The observations are reassigned to mask values, which are less than DOY r The observations remain unchanged and a new time series is generated. This operation means that the input data of the model can be compatible with any time series during the crop growth period and can be used for crop identification during the growth period. In particular, this random mask data augmentation strategy is not to randomly generate once, pre-store the data, and then read the input model, but to randomly mask and train in each batch of model training. This method not only saves data storage, but also enables the same batch of training data at different Epochs to generate different masks, greatly increasing the multimodality of the data and improving the recognition accuracy of the model.

[0059] In the example, the embodiment of the present invention uses Sentinel-2tile (11SKA) as the research scope, uses CDL (Cropland Data Layer) data, collects crop type samples, and obtains 16,436 samples of 14 types of crops such as soybeans, corn, winter wheat, and almonds for model training. The accuracy of the model test samples changes over time as shown in the following figure: Figure 4 shown.

[0060] System Example

[0061] According to an embodiment of the present invention, a crop recognition model construction system compatible with any time-series remote sensing images is provided. Figure 5 is a schematic diagram of a crop recognition model construction system compatible with any time series remote sensing images according to an embodiment of the present invention. Figure 5 As shown, the crop recognition model construction system compatible with any time-series remote sensing images according to an embodiment of the present invention specifically includes:

[0062] A processing module 50 is used to obtain multi-source crop remote sensing images, obtain the time series length of model input data based on the crop remote sensing images, and pre-process the model input data according to the time series length;

[0063] A construction module 52 is used to input the processed model input data into the improved Transformer model, perform data enhancement on the processed model input data using a random time mask method, and train the enhanced input data to obtain a final crop recognition model;

[0064] The system further comprises:

[0065] The recognition module is used to recognize several crop types in remote sensing images of any time sequence and any band through the crop recognition model.

[0066] The embodiment of the present invention is a system embodiment corresponding to the above-mentioned method embodiment. The specific operations of each module can be understood by referring to the description of the method embodiment, which will not be repeated here.

[0067] Device Example 1

[0068] An embodiment of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program implements the steps described in the method embodiment when executed by the processor.

[0069] Device Example 2

[0070] An embodiment of the present invention provides a computer-readable storage medium, on which a program for implementing information transmission is stored. When the program is executed by a processor, the steps described in the method embodiment are implemented.

[0071] The computer-readable storage medium in this embodiment includes, but is not limited to, ROM, RAM, magnetic disk or optical disk, etc.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a crop recognition model compatible with any time series remote sensing images, characterized in that include: Acquire multi-source crop remote sensing images, obtain the time series length of model input data based on the crop remote sensing images, and pre-process the model input data according to the time series length; as well as The processed model input data is input into the improved Transformer model, a random time mask method is used to perform data enhancement on the processed model input data, and the enhanced input data is trained to obtain a final crop recognition model.

2. The method according to claim 1, characterized in that The method further comprises: The crop recognition model is used to recognize several crop types in remote sensing images of any time sequence and any band.

3. The method according to claim 1, characterized in that Acquiring multi-source crop remote sensing images, obtaining the time series length of model input data based on the crop remote sensing images, and preprocessing the model input data according to the time series length specifically includes: The crop remote sensing images of Landsat-8, Sentinel-2 and Sentinel-1 SAR sensors are obtained respectively, and the pixel points of known crop types sampled in advance are used as sample points. Based on the corresponding positions of the sample points in the crop remote sensing images, the effective observation time points of each crop remote sensing image in the year are obtained and a union operation is performed to obtain the effective observation number of the sample point in the year. The maximum value of the effective observation number in the year is taken as the time series length of the model input data. The effective observation time points of the sample pixels in each crop remote sensing image are calculated and a union operation is performed to obtain the number of elements in the effective observation time point set of the sample pixels. The number of elements is compared with the time series length. If the number of elements is less than the time series length, the mask value is used to fill the missing part of the data to obtain the model input data with the same time series length.

4. The method according to claim 3, characterized in that The model input data includes a time dimension and a feature dimension; The characteristic dimensions include image band, time point and year.

5. The method according to claim 4, characterized in that The processed model input data is input into the improved Transformer model, the processed model input data is enhanced by using a random time mask method, and the enhanced input data is trained to obtain the final crop recognition model, which specifically includes: Input the processed model input data into the improved Transformer model, adopt a data enhancement strategy of random time mask for each batch of data in the model training process to generate multimodal training data, construct position encoding for the training data using the regularization method shown in Formula 1, establish a mask matrix if the input at a certain time point in the training data is a mask value, multiply the mask matrix by the feature of the training data to obtain a mask input, add the mask input and the position encoding through a fully connected layer of the same dimension to obtain new input data, output a time series feature vector representation of each sensor data for the new input data through a four-layer transformer encoding module, fuse the time series feature vector representation of each sensor data in the time dimension to generate a time series feature vector, use the time series feature vector as an intermediate input and pass through two layers of transformer encoding modules to obtain a final crop recognition time series feature vector, average the non-mask feature vectors in the crop recognition time series feature vector, and obtain the discrete probability distribution of each crop type through a fully connected layer with a Softmax function to obtain a final crop recognition model; Among them, PE (i) Indicates time position coding, doy i represents the Julian day (1, 2, 3, ..., 365), i is the index of each time point in the sample time series, mean is the mean, and std is the standard deviation.

6. The method according to claim 5, characterized in that The sub-layer dimension of the four-layer transformer encoding module is the same as the input layer dimension; The multi-head self-attention mechanism layer of each encoding module establishes an attention mask matrix according to the mask matrix.

7. A crop recognition model construction system compatible with any time series remote sensing image, characterized in that include: A processing module, used for acquiring multi-source crop remote sensing images, obtaining a time series length of model input data based on the crop remote sensing images, and preprocessing the model input data according to the time series length; as well as A construction module is used to input the processed model input data into the improved Transformer model, perform data enhancement on the processed model input data using a random time mask method, and train the enhanced input data to obtain a final crop recognition model.

8. The system according to claim 7, characterized in that The system further comprises: The recognition module is used to recognize several crop types in remote sensing images of any time sequence and any band through the crop recognition model.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for constructing a crop recognition model compatible with arbitrary time-series remote sensing images as described in any one of claims 1 to 6.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores an implementation program for information transmission, and when the program is executed by the processor, the steps of the method for constructing a crop recognition model compatible with any time-series remote sensing image as described in any one of claims 1 to 6 are implemented.

Images