A crop classification method of an adaptive SAM and ConvxLSTM hybrid model

Abstract

The application discloses a crop classification method of a self-adaptive SAM and ConvxLSTM hybrid model, and relates to the technical field of remote sensing image processing. The method comprises the following steps: constructing a time series dataset facing crop phenology, filtering and screening an optimal classification time phase band combination through J-M distance; generating unmarked samples according to the optimal classification time phase band combination, generating crop pseudo-labels with category annotations based on a self-adaptive SAM model; processing the time series dataset based on a ConvxLSTM network to classify crops and output crop classification results. The application can ensure automatic acquisition of high-quality crop pseudo-labels, provide crop category annotations for the pseudo-labels, greatly reduce the excessive dependence of a model on artificially annotated samples, more effectively mine the space-time features of data, and thus improve the crop classification precision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image processing, and particularly relates to a crop classification method based on an adaptive SAM and ConvxLSTM hybrid model. BACKGROUND

[0002] Timely and accurate estimation of crop area and distribution is crucial for food security. Optical remote sensing has been a key technology for obtaining regional and even global crop area and condition. Due to frequent weather influences in some areas, optical remote sensing images are often unavailable, which brings great challenges to crop classification. Therefore, combining optical and radar data, or multi-day data synthesis products, has become the mainstream of current research.

[0003] Classic machine learning methods such as random forest and support vector machine are not designed for processing time series data, and are prone to ignore the time dependence of time series. With the rise of deep learning, it has a good advantage in fusing spatial and temporal features. Current one-dimensional convolutional neural networks and recurrent neural networks have been proven to be effective deep learning methods for solving end-to-end time series classification problems. Long short-term memory network (LSTM) and gated recurrent unit (GRU) as variants of RNN can solve the problem of gradient disappearance or explosion with the increase of time series. Compared with classic machine learning methods, deep learning model training requires a large number of labeled samples, which mainly come from manual annotation and are time-consuming and laborious. Especially for remote sensing images, different types of sensors, different weather conditions, and different research areas will affect the labeled samples, so the weakly supervised classification method becomes particularly important. The present application makes adaptive improvements to the current most popular segmentation all model (SAM) to generate high-quality pseudo-labels of labeled crop class labels to reduce manual annotation work. At the same time, based on the latest xLSTM network, a new ConvxLSTM module is designed, and a crop classification network based on ConvxLSTM is constructed to realize accurate classification of crops.

[0004] However, traditional deep learning methods require a large number of high-quality labeled samples, and the diversity of crops and the fast-changing nature of the growth period make it difficult to obtain enough labeled samples. Due to the large domain difference between the training dataset and the downstream test dataset, SAM is not very robust and generalizable in multiple downstream tasks, and lacks annotation of target samples. Moreover, the traditional LSTM model method cannot always meet the demand for high-precision classification in some scenarios, and the research of new LSTM models needs further exploration. SUMMARY

[0005] Therefore, the present application aims to provide a crop classification method based on a hybrid model of adaptive SAM and ConvxLSTM, by constructing a FY3D MERSI-II NVI decile product time series product set oriented to crop phenology, designing an adaptive SAM model for generating high-quality crop pseudo-labels, and using a ConvxLSTM-based neural network for high-precision crop classification, which can ensure the automatic acquisition of high-quality crop pseudo-labels, provide crop class labels for the pseudo-labels, greatly reduce the over-reliance of the model on manually labeled samples, more effectively mine the spatiotemporal features of the data, and thus improve the crop classification accuracy.

[0006] To achieve the above object, the present application provides the following technical scheme.

[0007] Based on the above object, in a first aspect, the present application provides a crop classification method based on a hybrid model of adaptive SAM and ConvxLSTM, comprising the following steps:

[0008] Constructing a time series data set oriented to crop phenology, filtering and J-M distance screening the best classification time phase band combination;

[0009] Generating unlabeled samples according to the best classification time phase band combination, and generating crop pseudo-labels with class labels based on an adaptive SAM model;

[0010] Performing crop classification based on a ConvxLSTM network processing time series data set, and outputting crop classification results.

[0011] As a further scheme of the present application, the time series data set oriented to crop phenology comprises the following steps:

[0012] Obtaining the phenology of crops in the study area, and determining the key growth nodes of the crops based on the phenology;

[0013] Collecting multi-temporal remote sensing image data corresponding to the key growth nodes, and constructing a time series data set;

[0014] Filtering the time series data set using a Savitzky-Golay filter;

[0015] Screening the best classification time phase band combination using a J-M distance method, wherein the best classification time phase band combination is composed of three time phase bands and has the maximum J-M distance.

[0016] As a further scheme of the present application, when constructing the time series data set, the data of the FY3D MERSI-II NVI decile product of the key growth nodes of crop sowing, emergence, jointing, heading, transplanting, tillering, flowering, milk maturity, and harvesting are obtained based on the phenology.

[0017] As a further aspect of the present invention, the multi-temporal remote sensing image data is FY3D MERSI-II NVI ten-day product or MODIS surface reflectance data.

[0018] As a further aspect of the present invention, generating pseudo-labels for crops with category annotations based on an adaptive SAM model includes the following steps:

[0019] Unlabeled samples are generated using the obtained optimal combination of time phase bands;

[0020] An adaptive SAM network is constructed by inserting an adaptive module into a pre-trained SAM model; where:

[0021] Two adaptive modules are inserted into the encoder: the first module is located after the multi-head attention layer and before the residual connection, and the second module is located on the residual path of the multilayer perceptron (MLP) layer.

[0022] Three adaptive modules are inserted into the decoder: the first module is located after the multi-head cross-attention layer, the second module is located on the residual path of the MLP layer, and the third module is located after the residual connection of the image embedding.

[0023] The adaptive SAM encoder knowledge is distilled into the contrastive language-image pre-trained (CLIP) encoder using a teacher-student model, and the CLIP decoder knowledge is transferred to the SAM decoder to generate pseudo-labels with crop category tags.

[0024] As a further embodiment of the present invention, the adaptive module is a bottleneck structure, comprising a lower projection layer, a ReLU activation function, and an upper projection layer in sequence, with the insertion positions including:

[0025] The last two points of the multi-head attention layer in the SAM encoder;

[0026] The SAM decoder has three locations: after the multi-head cross-attention layer, after the MLP residual path, and after the image embedding residual connection.

[0027] As a further aspect of the present invention, knowledge transfer employs the Open-Vocabulary learning framework to achieve semantic alignment.

[0028] As a further aspect of the present invention, crop classification is performed based on time-series datasets processed by a ConvxLSTM network, including the following steps:

[0029] Constructing a ConvxLSTM module: Add a convolutional module to the input of the xLSTM module;

[0030] Construct a seven-layer classification network:

[0031] The first four layers are ConvxLSTM layers.

[0032] The fifth layer is a 3D convolutional (Conv3D) layer.

[0033] The sixth layer is a two-dimensional convolutional (Conv2D) layer.

[0034] The seventh layer is the logistic regression loss layer;

[0035] The filtered time-series data is input into the network for training, and the crop classification results are output.

[0036] As a further aspect of the present invention, the ConvxLSTM module is composed of convolutional layers and xLSTM connected in series, and the network architecture includes:

[0037] 4 layers of ConvxLSTM → 1 layer of Conv3D → 1 layer of Conv2D → 1 layer of logistic regression loss layer.

[0038] As a further embodiment of the present invention, the Conv3D layer is used to extract spatiotemporal joint features, and the Conv2D layer is used for spatial feature dimensionality reduction and compression.

[0039] As a further embodiment of the present invention, the ConvxLSTM module can be replaced with a ConvLSTM module.

[0040] Compared with existing technologies, the crop classification method based on an adaptive SAM and ConvxLSTM hybrid model proposed in this invention has the following advantages:

[0041] 1. This invention significantly reduces the reliance on manual annotation. By designing an adaptive SAM model, it can automatically generate high-quality pseudo-labels with crop category labels. It utilizes the Open-Vocabulary concept to perform knowledge distillation, transferring the knowledge of the SAM encoder to the contrastive language-image pre-trained encoder. This solves the problem of traditional deep learning methods requiring a large number of manually annotated samples and greatly reduces annotation costs.

[0042] 2. This invention inserts a bottleneck-type adaptive module at a key position in the SAM network to enhance the model's adaptability to remote sensing crop classification tasks. It overcomes the problems of insufficient generalization and lack of target sample annotation caused by domain differences in the original SAM in downstream tasks. The ConvxLSTM module enables the model to directly process the spatiotemporal features of image samples. The seven-layer hybrid network architecture deeply integrates temporal and spatial information, significantly improving the accuracy of crop classification. It solves the defect that traditional LSTM cannot meet the high-precision requirements and optimizes the accuracy of spatiotemporal feature extraction.

[0043] 3. This invention also constructs a FY3D MERSI-II NVI ten-day product time-series dataset based on crop phenological calendars, and uses Savitzky-Golay filtering to reduce noise, and uses JM distance to select the best temporal band combination to reduce data redundancy, improve model training efficiency, support data source replacement (such as MODIS surface reflectance data replacing FY3D data) and module replacement (such as ConvLSTM replacing ConvxLSTM), and adapt to the needs of different application scenarios.

[0044] These or other aspects of this application will become more apparent from the following description of embodiments. It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the application. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or related technologies, the accompanying drawings used in the description of the exemplary embodiments or related technologies will be briefly introduced below. The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation thereof. In the drawings:

[0046] Figure 1 This is a technical execution block diagram of a crop classification method using an adaptive SAM and ConvxLSTM hybrid model, according to an embodiment of the present invention.

[0047] Figure 2 This is a flowchart of a crop classification method using a hybrid model of adaptive SAM and ConvxLSTM, according to an embodiment of the present invention.

[0048] Figure 3 This is a flowchart illustrating the construction of a time-series dataset oriented towards crop phenological calendars in a crop classification method based on an adaptive SAM and ConvxLSTM hybrid model, as described in an embodiment of the present invention.

[0049] Figure 4 This is a flowchart illustrating the generation of pseudo-labels for crops with category annotations based on an adaptive SAM model in a crop classification method using a hybrid adaptive SAM and ConvxLSTM model, as described in an embodiment of the present invention.

[0050] Figure 5 This is a flowchart illustrating the crop classification method based on a ConvxLSTM network for processing time-series datasets, as described in an embodiment of the present invention, which employs a hybrid adaptive SAM and ConvxLSTM model. Detailed Implementation

[0051] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings, but it should be understood that the scope of protection of the present invention is not limited to the specific embodiments.

[0052] Unless otherwise expressly stated, throughout the specification and claims, the term "comprising" or its variations such as "including" or "comprises" shall be understood to include the stated elements or components without excluding other elements or other components.

[0053] Traditional deep learning methods require a large number of high-quality labeled samples, but the diversity of crops and the rapid variability of their growth periods make it difficult to obtain sufficient labeled samples. Due to the significant domain differences between the training dataset and the downstream test dataset, SAM (Synthetic Amplifier) ​​does not exhibit strong robustness and generalization in various downstream tasks and lacks labeled target samples. Furthermore, traditional LSTM models cannot consistently meet the high-precision classification requirements in certain scenarios, necessitating further research into novel LSTM models. This invention proposes a crop classification method using an adaptive SAM and ConvxLSTM hybrid model. By constructing a FY3D MERSI-II NVI ten-day product time-series set based on crop phenological calendars, an adaptive SAM model is designed to generate high-quality crop pseudo-labels, and a ConvxLSTM-based neural network is used for high-precision crop classification. This method ensures the automatic acquisition of high-quality crop pseudo-labels while also providing crop category annotations for the pseudo-labels, significantly reducing the model's over-reliance on manually labeled samples and more effectively mining the spatiotemporal features of the data, thereby improving crop classification accuracy.

[0054] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a crop classification method using a hybrid model of adaptive SAM and ConvxLSTM, the method comprising the following steps:

[0055] Step S10: Construct a time-series dataset for crop phenological calendars, and filter and select the best combination of time-phase bands for classification using filtering and JM distance.

[0056] In this step, see Figure 3 As shown, constructing a time-series dataset for crop phenological calendars includes the following steps:

[0057] Step S101: Obtain the phenological calendar of crops in the study area and determine the key growth nodes of crops based on the phenological calendar;

[0058] Step S102: Collect multi-temporal remote sensing image data corresponding to the key growth nodes and construct a time-series dataset;

[0059] Step S103: Use the Savitzky-Golay filter to filter the time series dataset;

[0060] Step S104: Use the JM distance method to select the optimal classification time phase band combination, wherein the optimal classification time phase band combination consists of three time phase bands with the largest JM distance.

[0061] In steps S101 and S102, when constructing the time-series dataset, data from FY3D MERSI-II NVI ten-day products for key growth nodes of crops—sowing, emergence, jointing, heading, transplanting, tillering, flowering, milk stage, and harvest—are obtained based on the phenological calendar. The multi-temporal remote sensing image data are either FY3D MERSI-II NVI ten-day products or MODIS surface reflectance data.

[0062] In this embodiment, phenological calendars of crops within the study area are acquired. Based on these calendars, FY3D MERSI-II NVI ten-day product data for key growth stages such as sowing, emergence, jointing, heading, transplanting, tillering, flowering, milk stage, and harvest are obtained, and a time-series dataset is constructed. This dataset is then filtered using a Savitzky-Golay filter. The JM distance method is used to explore the optimal combination of classification phase bands based on the time-series dataset; specifically, selecting a data combination with the largest JM distance and consisting of three phase bands reduces the amount of data required for sample construction.

[0063] Step S20: Generate unlabeled samples based on the optimal classification phase band combination, and generate crop pseudo-labels with category annotations based on the adaptive SAM model.

[0064] In this step, see Figure 4 As shown, generating pseudo-labels for crops with category annotations based on an adaptive SAM model includes the following steps:

[0065] Step S201: Generate unlabeled samples using the optimal combination of time phase bands obtained in step S10, and select vit_b as the pre-trained model of SAM.

[0066] Step S202: Insert adaptive modules into the pre-trained SAM model to construct an adaptive SAM network; wherein:

[0067] Two adaptive modules are inserted into the encoder: the first module is located after the multi-head attention layer and before the residual connection, and the second module is located on the residual path of the multilayer perceptron (MLP) layer.

[0068] Three adaptive modules are inserted into the decoder: the first module is located after the multi-head cross-attention layer, the second module is located on the residual path of the MLP layer, and the third module is located after the residual connection of the image embedding.

[0069] Step S203: Distill the knowledge of the adaptive SAM encoder into the contrastive language-image pre-trained (CLIP) encoder through the teacher-student model, and transfer the knowledge of the CLIP decoder to the SAM decoder to generate pseudo-labels with crop category tags.

[0070] In this embodiment, to improve the adaptability of SAM in crop classification tasks, an adaptive SAM network is designed, that is, an adaptive module is inserted into a specific position in the SAM network. Specifically, the adaptive module is a bottleneck structure, which includes a lower projection layer, a ReLU activation function, and an upper projection layer in sequence. The insertion positions include:

[0071] The last two points of the multi-head attention layer in the SAM encoder;

[0072] The SAM decoder has three locations: after the multi-head cross-attention layer, after the MLP residual path, and after the image embedding residual connection.

[0073] This invention designs an adaptive SAM model that, while ensuring the automatic acquisition of high-quality crop pseudo-labels, can also provide crop category annotations for the pseudo-labels, significantly reducing the model's over-reliance on manually labeled samples. Two adaptive modules are inserted into the SAM encoder: the first is inserted after the standard vision transformer's multi-head attention and before the residual connection; the second is inserted into the residual path of the MLP layer after the standard vision transformer's multi-head attention. Three adaptive modules are inserted into the decoder: the first is inserted after multi-head cross-attention; the second is inserted into the residual path of the MLP layer after multi-head attention; and the third is inserted after the residual connection of the image embedding.

[0074] In this process, the knowledge transfer adopts the Open-Vocabulary learning framework to achieve semantic alignment.

[0075] Step S30: Process the time-series dataset using the ConvxLSTM network to classify crops and output the crop classification results.

[0076] In this step, see Figure 5 As shown, crop classification based on time-series datasets using a ConvxLSTM network includes the following steps:

[0077] Step S301: Construct the ConvxLSTM module: Add a convolution module to the input of the xLSTM module;

[0078] Step S302: Construct a seven-layer classification network:

[0079] The first four layers are ConvxLSTM layers.

[0080] The fifth layer is a 3D convolutional (Conv3D) layer.

[0081] The sixth layer is a two-dimensional convolutional (Conv2D) layer.

[0082] The seventh layer is the logistic regression loss layer;

[0083] Step S303: Input the filtered time series data into the network for training and output the crop classification results.

[0084] In this embodiment, the concept of Open-Vocabulary is incorporated to obtain pseudo-labels with crop type markings. Specifically, the knowledge in the adaptive SAM model encoder is first distilled into the contrastive language-image pre-trained encoder using a teacher-student model. For decoding, the knowledge in the contrastive language-image pre-trained encoder is transferred to the SAM decoder. By training the adaptive SAM model incorporating Open-Vocabulary, pseudo-labels with crop markings are obtained. The ConvxLSTM module consists of concatenated convolutional layers and xLSTM, and the network architecture includes:

[0085] 4 layers of ConvxLSTM → 1 layer of Conv3D → 1 layer of Conv2D → 1 layer of logistic regression loss layer.

[0086] The Conv3D layer is used to extract spatiotemporal joint features, and the Conv2D layer is used for spatial feature dimensionality reduction and compression.

[0087] In this embodiment, the ConvxLSTM module can be replaced with a ConvLSTM module.

[0088] This invention, based on a novel xLSTM network, designs a new ConvxLSTM module and a seven-layer network for crop classification, which can more effectively mine the spatiotemporal features of the data, thereby improving the accuracy of crop classification. By adding a convolutional module to the input of the xLSTM module, a ConvxLSTM module is generated, enabling xLSTM to process crop image samples. In terms of network architecture, a seven-layer network is used: the first four layers are ConvxLSTM layers, the fifth layer is a Conv3D layer, the sixth layer is a Conv2D layer, and the seventh layer is a logistic regression loss layer. Temporal data after optical Savitzky-Golay filtering is fed into the ConvxLSTM neural network for training to obtain crop classification results.

[0089] This invention significantly reduces the reliance on manual annotation. By designing an adaptive SAM model, it can automatically generate high-quality pseudo-labels with crop category labels. It utilizes the Open-Vocabulary concept to perform knowledge distillation, transferring the knowledge of the SAM encoder to a contrastive language-image pre-trained encoder. This solves the problem of traditional deep learning methods requiring a large number of manually annotated samples and greatly reduces annotation costs.

[0090] This invention inserts a bottleneck-type adaptive module at a key position in the SAM network to enhance the model's adaptability to remote sensing crop classification tasks. It overcomes the problems of insufficient generalization and lack of target sample annotation caused by domain differences in the original SAM in downstream tasks. The ConvxLSTM module enables the model to directly process the spatiotemporal features of image samples. The seven-layer hybrid network architecture deeply integrates temporal and spatial information, significantly improving the accuracy of crop classification. It solves the shortcomings of traditional LSTM in meeting high-precision requirements and optimizes the accuracy of spatiotemporal feature extraction.

[0091] This invention also constructs a FY3D MERSI-II NVI ten-day product time-series dataset based on crop phenological calendars, and uses Savitzky-Golay filtering for noise reduction, and utilizes JM distance to select the optimal combination of temporal bands, reducing data redundancy, improving model training efficiency, supporting data source replacement (such as MODIS surface reflectance data replacing FY3D data) and module replacement (such as ConvLSTM replacing ConvxLSTM), adapting to the needs of different application scenarios.

[0092] It should be noted that the above figures are merely illustrative of the processes included in the method according to exemplary embodiments of the present invention, and are not intended to be limiting. It is readily understood that the processes shown in the above figures do not indicate or limit the temporal order of these processes. Furthermore, it is readily understood that these processes may, for example, be executed synchronously or asynchronously in multiple modules.

[0093] It should be understood that although the above description follows a certain order, these steps are not necessarily executed in that order. Unless otherwise expressly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, some steps in this embodiment may include multiple steps or multiple stages, which are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least a portion of steps or stages in other steps.

[0094] The above are exemplary embodiments disclosed in this invention. However, it should be noted that various changes and modifications can be made without departing from the scope of the embodiments of this invention as defined by the claims. The functions, steps, and / or actions of the methods according to the disclosed embodiments described herein do not need to be performed in any particular order. Furthermore, although the elements disclosed in the embodiments of this invention may be described or claimed individually, they may be understood as multiple unless explicitly limited to a singular number.

[0095] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples. Within the framework of the invention, technical features of the above embodiments or different embodiments can be combined, and many other variations of different aspects of the invention exist, which are not provided in the details for the sake of brevity. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the invention should be included within the protection scope of the invention.

[0096] The foregoing description of specific exemplary embodiments of the invention is for illustrative and explanatory purposes. These descriptions are not intended to limit the invention to the precise forms disclosed, and it will be apparent that many changes and variations can be made in accordance with the foregoing teachings. The exemplary embodiments were chosen and described in order to explain the specific principles of the invention and its practical application, thereby enabling those skilled in the art to implement and utilize various different exemplary embodiments of the invention, as well as various different choices and variations. The scope of the invention is intended to be defined by the claims and their equivalents.

Claims

1. A crop classification method of a self-adaptive SAM and ConvxLSTM hybrid model, characterized in that, The method comprises the following steps: The method comprises the following steps: Construct a time series dataset oriented to crop phenology, filter and screen the best classification time phase band combination by J-M distance; Generate unlabeled samples according to the best classification time phase band combination, and generate crop pseudo-labels with category annotations based on an adaptive SAM model; Crop classification based on a ConvxLSTM network processing time series dataset, outputting crop classification results; Wherein, generating crop pseudo-labels with category annotations based on an adaptive SAM model comprises the following steps: Generate unlabeled samples using the obtained best time phase band combination; Insert an adaptive module into the pre-trained SAM model to construct an adaptive SAM network; wherein: Two adaptive modules are inserted into the encoder: the first module is located after the multi-head attention layer and before the residual connection, and the second module is located on the residual path of the multi-layer perception layer; Three adaptive modules are inserted into the decoder: the first module is located after the multi-head cross-attention layer, the second module is located on the residual path of the MLP layer, and the third module is located after the residual connection of the image embedding; Through a teacher-student model, the adaptive SAM encoder knowledge is distilled to the contrast language-image pre-training encoder, and the CLIP decoder knowledge is migrated to the SAM decoder to generate pseudo-labels with crop category labels; Wherein, the adaptive module is a bottleneck structure, which comprises a lower projection layer, a ReLU activation function and an upper projection layer in sequence, and the insertion positions include: 2 places after the multi-head attention layer in the SAM encoder; 3 places after the multi-head cross-attention layer in the SAM decoder, the MLP residual path and the image embedding residual connection; Wherein, crop classification based on a ConvxLSTM network processing time series dataset comprises the following steps: Construct a ConvxLSTM module: add a convolution module to the input end of the xLSTM module; Build a seven-layer classification network: The first four layers are ConvxLSTM layers, The fifth layer is a three-dimensional convolution layer, The sixth layer is a two-dimensional convolution layer, The seventh layer is a logistic regression loss layer; Input the filtered time series data into the network for training, and output the crop classification results; 2.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 1, wherein, The ConvxLSTM module is composed of a convolution layer and an xLSTM in series. Construct a time series dataset oriented to crop phenology, comprising the following steps: Obtain the phenology of crops in the study area, and determine the key growth nodes of crops based on the phenology; Collect multi-temporal remote sensing image data corresponding to the key growth nodes to construct a time series dataset; Filter the time series dataset using a Savitzky-Golay filter; 3.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 2, wherein, Screen the best classification time phase band combination using the J-M distance method, which is composed of three time phase bands and has the maximum J-M distance. 4.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 3, wherein, When constructing the time series dataset, the FY3D MERSI-II NVI decile product data of the key growth nodes of crop sowing, emergence, jointing, heading, transplanting, tillering, flowering, milk maturity and harvesting are obtained based on the phenology. The multi-temporal remote sensing image data is FY3D MERSI-II NVI decile product or MODIS surface reflectance data. 5.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 4, wherein, The knowledge transfer adopts an Open-Vocabulary learning framework to realize semantic alignment. 6.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 5, wherein, The network architecture of the ConvxLSTM module comprises: 4 layers of ConvxLSTM→1 layer of Conv3D→1 layer of Conv2D→1 layer of a logistic regression loss layer. 7.The crop classification method of adaptive SAM and ConvxLSTM hybrid model according to claim 6, wherein, The Conv3D layer is used for extracting spatio-temporal joint features, and the Conv2D layer is used for spatial feature dimension reduction and compression; and the ConvxLSTM module can be replaced by a ConvLSTM module.

Images