Rice early recognition method and device based on remote sensing data and time sequence model

Abstract

The invention discloses a rice early recognition method and device based on remote sensing data and a time sequence model. The method comprises the following steps: S1, acquiring multi-source remote sensing data and auxiliary data; s2, preprocessing the multi-source remote sensing data and the auxiliary data; s3, constructing a sample set based on the preprocessed multi-source remote sensing data and auxiliary data; s4, constructing a DualMama model, and training the DualMama model on the basis of the sample set; and S5, inputting the multi-source remote sensing data and the auxiliary data obtained in the step S1 into the trained DualMama model to obtain a rice early recognition distribution diagram. According to the method, through the proposed DualMama model, the utilization efficiency of multi-source remote sensing data is remarkably improved, the problem of inconsistent feature dimensions is solved, the early recognition precision of the rice is greatly improved under the complex hill terrain, the F1 score of the rice category can reach 90% within two months before the rice growth period, and accurate early recognition of the rice is achieved.

Description

Technical Field

[0001] The invention belongs to the technical field of agricultural remote sensing classification and time series analysis integration, in particular to an early rice recognition method and device based on remote sensing data and a time series model. Background Art

[0002] Rice is one of the most important food crops in the world, and accurate monitoring of its planting and growth status is of great significance for agricultural production. However, due to the complex terrain and diverse vegetation in hilly areas, traditional recognition methods based on a single remote sensing data source usually cannot provide sufficient spatiotemporal information support, resulting in low recognition accuracy. With the rapid development of remote sensing technology, the combination of multi-source remote sensing data, especially the integrated application of optical and radar data, has provided new possibilities for early rice monitoring. The complementary advantages of different types of data can effectively make up for the limitations of a single data source in spatiotemporal coverage. In addition, the development of deep learning technology has provided a new path for high-precision crop classification of remote sensing data. Combining multi-source remote sensing data with a time series deep learning model has become an effective solution for achieving high-precision classification of early rice. Summary of the invention

[0003] In order to solve the above technical problems, the present invention provides a method and device for early rice identification based on remote sensing data and time series model. The method uses multi-source remote sensing data, including Sentinel-1, Sentinel-2 and GF-3 data, combined with a time series deep learning model, so that the F1 score of the rice category reaches 90% as early as possible, effectively solving the problem of inconsistent remote sensing data features, significantly shortening the time for early rice identification and improving classification accuracy.

[0004] To achieve the above purpose, the technical solution adopted by the present invention is as follows:

[0005] A rice early identification method based on remote sensing data and time series model includes the following steps:

[0006] S1. Acquire multi-source remote sensing data and auxiliary data;

[0007] The present invention utilizes multi-source remote sensing data for early rice identification, and the data include Sentinel-1, Sentinel-2 and GF-3 time series remote sensing data, and a variety of auxiliary data.

[0008] S1-1: Sentinel-1 data is a Level-1 GRD product in wide-swath interferometric mode (IW), with a resolution of 10 meters and including VH and VV polarization features;

[0009] S1-2: GF-3 data is a Level-1 product of the Fine Strip II mode (FSII), providing SAR data of HH and HV polarization characteristics with a resolution of 10 meters;

[0010] S1-3: Sentinel-2 data is a Level-2A surface reflectance product, covering the visible light band, near infrared band, red edge band and short-wave infrared band, a total of 10 bands. All bands are resampled and the final resolution is unified to 10 meters.

[0011] S1-4: Ancillary data include field survey data, land cover / land use data, and digital elevation models (DEMs).

[0012] S2. Preprocess multi-source remote sensing data and auxiliary data;

[0013] S2-1: Radiometric correction, terrain correction, speckle filtering and DB conversion are performed on Sentinel-1 data to generate VH and VV polarization features of the time series;

[0014] S2-2: GF-3 data were converted to complex data, multi-viewed, speckle filtered and geocoded using ENVI SARscape software to generate time series HH and HV polarization features;

[0015] S2-3: Declouding, interpolation and 10-day median synthesis of Sentinel-2 data are performed to generate a time series spectral feature dataset including visible light, near infrared, red edge and short-wave infrared bands. These operations can be completed through Google EarthEngine (GEE) or ESA SNAP software. All these data are pre-processed and uniformly use the EPSG:4326 (WGS84) geographic coordinate system.

[0016] S3. Construct a sample set based on preprocessed multi-source remote sensing data and auxiliary data;

[0017] S3-1: Based on the field survey data and combined with the visual interpretation results, rice and non-rice areas are marked and classification labels are generated;

[0018] S3-2: Extract the time series features corresponding to DOY from the preprocessed multi-source remote sensing data to generate a comprehensive sample set combining optical features and SAR features;

[0019] S3-3: Based on the comprehensive sample set, the time nodes are obtained according to different remote sensing data, and incremental sample sets containing all time nodes are gradually produced to provide sufficient data support for incremental training;

[0020] S3-4: Divide the sample set into training set, validation set and test set in a ratio of 6:2:2.

[0021] S4. Build a DualMamba model and train the DualMamba model based on the sample set;

[0022] S4-1: Construct DualMamba deep learning model. This model consists of two parallel Mamba branches, which process Sentinel-2 optical data and Sentinel-1 and GF-3 SAR data respectively. The first Mamba branch first extracts the temporal features from Sentinel-2 data through a convolutional layer, then applies batch normalization to standardize the temporal features, and the selective state space module extracts the temporal dependency features from the standardized temporal features, and finally captures the temporal features through the SiLU activation function; the second Mamba branch first extracts the temporal features from Sentinel-1 and GF-3 data through a convolutional layer, combines batch normalization and standardizes the temporal features, and the selective state space module extracts the temporal dependency features from the standardized temporal features, and finally captures the temporal features through the SiLU activation function; the temporal features extracted by the two Mamba branches are fused through the adaptive fusion module, and the fused high-dimensional temporal features are mapped to the output space through the fully connected layer, and the probability distribution of rice and non-rice plots is calculated through the Softmax layer to complete the temporal binary classification task;

[0023] S4-2: The model training strategy includes the following steps: The number of model training is set to 1000 times, and the batch size is 64. The optimizer is AdamW, the initial learning rate is 1e-4, and the learning rate is gradually adjusted through cosine annealing according to the training progress to ensure convergence. The Dropout mechanism is used during training to prevent overfitting, and the Dropout rate is 0.3;

[0024] An incremental training strategy is adopted, starting from the starting DOY (start_doy) of rice sowing, and gradually introducing new remote sensing data for training. The model input data of each stage contains all time series data from the starting DOY to the current training moment, ensuring that the model can effectively capture the changes in time series features. Through training on the incremental sample set, the date when the F1 score of the rice category first reaches 0.9 is determined as the earliest recognition date (early_doy);

[0025] During model training, evaluation metrics include overall accuracy (OA) and F1 score:

[0026] ,

[0027] ,

[0028] ,

[0029] ,

[0030] Among them, TP (True Positive) represents the number of pixels correctly classified as rice, FP (False Positive) represents the number of non-rice pixels misclassified as rice, TN (True Negative) represents the number of pixels correctly classified as non-rice, FN (False Negative) represents the number of rice pixels misclassified as non-rice, OA (Overall Accuracy) represents the overall classification accuracy of the model for all pixels, Recall represents the proportion of correctly identified actual rice pixels, Precision represents the proportion of pixels identified as rice that are actually rice, and the F1 score is the harmonic average of precision and recall, which is used to comprehensively evaluate model performance.

[0031] S5. Input the multi-source remote sensing data and auxiliary data obtained in step S1 into the trained DualMamba model to obtain a rice early identification distribution map.

[0032] S5-1: Based on the model evaluation results of S4, determine the earliest identification date of rice (early_doy), that is, the date when the F1 score of the rice category first reaches 0.9;

[0033] S5-2: Generate the initial early rice distribution map using all available Sentinel-1, GF-3 and Sentinel-2 data from start_doy to early_doy in combination with the DualMamba model. Post-process the DEM data to generate the final early rice identification distribution map.

[0034] In the method for early crop identification, in step S1, the ground sample data is the crop type and its geographical location, and is used to verify the accuracy of remote sensing image classification.

[0035] In the method for early crop identification, in step S2-1, the time series length of the Sentinel-1 time series image finally obtained is 15, covering the VH and VV polarization characteristics of the entire rice growth cycle.

[0036] In the method for early crop identification, in step S2-2, the time series length of the GF-3 time series image finally obtained is 12, covering the VH and VV polarization characteristics of the entire rice growth cycle.

[0037] In the method for early crop identification, in step S2-3, the Sentinel-2 time-series images in the entire growing season are synthesized by median value for 10 days, and the length of the time-series images finally obtained is 18.

[0038] In the method for early crop identification, in step S5-2, the specific operation of post-processing the initial rice distribution map using DEM data is to first generate a slope map using DEM, and then set the rice area with a slope greater than 10° to 0, so that the post-processing operation can be completed.

[0039] In another aspect, the present invention provides a device for early identification of rice based on remote sensing data and a time series model, comprising:

[0040] A data acquisition unit, used for acquiring multi-source remote sensing data and auxiliary data;

[0041] A data preprocessing unit, used for preprocessing multi-source remote sensing data and auxiliary data;

[0042] A sample set construction unit, used to construct a sample set based on preprocessed multi-source remote sensing data and auxiliary data;

[0043] A model training unit, used to build a DualMamba model and train the DualMamba model based on a sample set;

[0044] The distribution map acquisition unit is used to input the multi-source remote sensing data and auxiliary data acquired by the data acquisition unit into the trained DualMamba model to obtain the early rice identification distribution map.

[0045] In a third aspect, the present invention provides an electronic device, characterized in that it comprises: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned early rice identification method based on remote sensing data and time series model.

[0046] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned early rice identification method based on remote sensing data and a time series model.

[0047] The beneficial effects of the present invention are:

[0048] 1. The present invention significantly improves the frequency of remote sensing data acquisition during the rice growing season by combining optical remote sensing data (Sentinel-2) and synthetic aperture radar data (Sentinel-1, GF-3), shortens the time for early rice identification, and ensures that the F1 score of the rice category can reach 90% as early as possible within the first two months of the rice growing season. Especially in complex hilly areas, the application of the present invention effectively improves the accuracy and reliability of rice identification.

[0049] 2. This paper proposes an innovative DualMamba deep learning model to solve the problem of inconsistency in temporal characteristics of multi-source remote sensing time series data. Through a dual-branch structure that processes optical data and SAR data separately, this paper can effectively fuse different types of remote sensing data and ultimately achieve high-precision early rice identification in complex hilly areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is a flow chart of a method for early rice identification based on remote sensing data and a time series model of the present invention;

[0051] Figure 2 This is the DualMamba structure diagram proposed by the present invention;

[0052] Figure 3 It is the time coverage of multi-source time series remote sensing data;

[0053] Figure 4 It is the comparison result of integrating multi-source time series remote sensing data and single time series remote sensing data as input;

[0054] Figure 5 The following is a comparison chart of the recognition accuracy of DualMamba, Random Forest (RF) and DualLSTM models. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0056] like Figure 1 As shown, it is a flow chart of a method for early rice identification based on remote sensing data and time series model of the present invention, which specifically includes:

[0057] Step S1, obtain multi-source remote sensing data and auxiliary data. The study area is Lujiang County, Anhui Province. The screening time range of remote sensing data is from April 20, 2024 to October 10, 2024, and the DOY is 111 to 284. This time period covers the entire rice growth phenological period. Figure 3 As shown, the multi-source remote sensing data collected by the present invention includes time-series images of Sentinel-1, Sentinel-2 and GF-3, as follows:

[0058] Sentinel-1 data: Sentinel-1 uses GRD products in wide-swath interferometric mode (IW), including VH and VV polarization features. A total of 16 Sentinel-1 images were collected between April 20, 2024 and October 10, 2024 (DOY 111 to 284).

[0059] Sentinel-2 includes B2, B3, B4 (visible light band), B8 (near infrared band), B5, B6, B7, B8A (red edge band) and B11, B12 (shortwave infrared band). Each band is uniformly resampled to a resolution of 10 meters. During the observation period, a total of 18 Sentinel-2 images were generated due to the 10-day median synthesis of Sentinel-2.

[0060] GF-3 data: Fine Strip II mode (FSII), including HH and HV polarization features, with a resolution of 10 meters. During the observation period from April 20 to October 10, 2024 (DOY 111 to 284), GF-3 includes 12 images.

[0061] Each time point of the image is represented by DOY (Day of Year). For example, the DOY on October 10, 2024 is 284.

[0062] The field survey data were obtained through field surveys to make sample labels. The specific methods are as follows: planning field experiments: Lujiang County, Anhui Province was selected as the research area to investigate the rice and non-rice plots in the area; location records: using tools such as Aowei Map, the geographical locations of rice and non-rice features were recorded and the types of features were marked. Label production: based on the field survey data, shapefile files were generated as labels for model samples.

[0063] Step S2, preprocessing the multi-source remote sensing data and auxiliary data;

[0064] Step S2-1 preprocesses Sentinel-1 time series data based on the SNAP platform, including: radiation calibration, terrain correction, speckle filtering, and decibel conversion. The final acquired time series Sentinel-1 time series image covers 16 time nodes during the rice growth period;

[0065] Step S2-2 preprocesses the GF-3 time series data through the ENVI SARscape platform, including: multi-view processing, speckle filtering, geocoding, radiometric calibration, and the final GF-3 time series image contains 12 time nodes;

[0066] Step S2-3 preprocesses the Sentinel-2 time series data based on the Google Earth Engine (GEE) platform. The steps include: cloud removal, interpolation, and 10-day median synthesis. The preprocessed Sentinel-2 image has 18 time nodes.

[0067] Step S3: Construct a sample set based on the preprocessed multi-source remote sensing data and auxiliary data.

[0068] Step S3-1: Create labels for rice and non-rice crops based on field survey data and visual interpretation of high-resolution images;

[0069] Step S3-2: Extract multi-temporal features corresponding to the labels from the pre-processed Sentinel-1, Sentinel-2 and GF-3 time series images, including optical and SAR features, to form a sample set. Each sample data is organized by DOY to reflect the time nodes when different features are obtained;

[0070] Step S3-3: Incrementally combine the image features according to the time nodes to create incremental sample sets, so that each incremental sample set corresponds to a separate DOY;

[0071] Step S3-4: Divide each type of samples in the incremental sample set into a training set, a validation set, and a test set in a ratio of 6:2:2.

[0072] Step S4. Constructing a DualMamba model and training the DualMamba model based on the sample set;

[0073] Step S4-1: DualMamba architecture design and details: Figure 2As shown in the figure, the present invention adopts the innovatively designed DualMamba architecture to specifically process the time series characteristics of multi-source remote sensing data. The core of the model consists of two parallel Mamba branches, which process different types of time series remote sensing features respectively. The optical branch first extracts the time series features in the Sentinel-2 data through a convolutional layer, and then applies batch normalization to standardize these time series features to eliminate the time differences. Subsequently, the Selective State Space Model (SSM) further extracts the time series dependency features, and finally captures the complex time features through the SiLU activation function. The SAR branch also processes the SAR time series data from Sentinel-1 and GF-3, extracts the time series features through the convolutional layer, combines batch normalization and SSM modules, fully mines the dynamic information in the time series, and further enhances the time series dependency through the SiLU activation function. The time series features extracted by the two Mamba branches are fused through an adaptive fusion module. Finally, the model maps the fused high-dimensional time series features to the output space through a fully connected layer, and calculates the probability distribution of rice and non-rice plots through a Softmax layer to complete the time series binary classification task;

[0074] Step S4-2: When training the DualMamba model, the number of training times is set to 1000, the batch size is 64, and the AdamW optimizer is used for parameter update. The initial learning rate is set to 1e-4, and the cosine annealing strategy is used to dynamically adjust the learning rate so that the learning rate can be reduced in the later stage of training to ensure the convergence stability of the model. In addition, in order to prevent overfitting, the model introduces a Dropout mechanism with a Dropout rate of 0.3;

[0075] Step S4-3: Incremental training starts from DOY=113, and DualMamba is used to train the incremental data sets corresponding to different DOYs. The performance of the model on different DOYs is evaluated to determine the earliest DOY at which the F1 score of the rice category reaches 0.9 for the first time;

[0076] Step S4-4: During the model training process, the evaluation indicators include the overall accuracy (OA) and the F1 score of the rice category. The calculation formulas for the overall accuracy (OA) and the F1 score are:

[0077] ,

[0078] ,

[0079] ,

[0080] ,

[0081] Among them, TP (True Positive) represents the number of pixels correctly classified as rice, FP (False Positive) represents the number of non-rice pixels misclassified as rice, TN (True Negative) represents the number of pixels correctly classified as non-rice, FN (False Negative) represents the number of rice pixels misclassified as non-rice, OA (Overall Accuracy) represents the overall classification accuracy of the model for all pixels, Recall represents the proportion of correctly identified actual rice pixels, Precision represents the proportion of pixels identified as rice that are actually rice, and the F1 score is the harmonic average of precision and recall, which is used to comprehensively evaluate model performance.

[0082] Step S5. Input the multi-source remote sensing data and auxiliary data obtained in step S1 into the trained DualMamba model to obtain a rice early identification distribution map;

[0083] Step S5-1: According to the result of S4, Figure 4 As shown in the figure, the earliest recognition date of rice (early_doy) is determined to be DOY=155, which is the time point when the F1 score of the rice category first reaches 0.9. At this time, the model can perform early recognition of rice with high accuracy;

[0084] Step S5-2: Generate an early rice distribution map using multi-source remote sensing data from start_doy (DOY=113) to the earliest identification date (early_doy, DOY=155) in combination with the DualMamba deep learning model. First, generate an initial rice distribution map using the classification results of the DualMamba model. To further improve the accuracy of the classification results, the initial distribution map is post-processed using DEM data to remove areas with slopes greater than 10°, ensuring that the identified rice areas meet the actual planting conditions, and finally generate an optimized early rice identification map.

[0085] In order to further verify the superiority of the DualMamba model in early rice identification, the classic deep learning model DualLSTM and the traditional random forest (RF) model were selected for comparative testing. The recognition accuracy of the model was calculated at different DOY time nodes, and the performance of each model was compared when the model input features were the same. The comparison results are shown in Figure 2. Figure 5 As shown, the results show that the accuracy of the DualMamba model in early rice recognition is significantly better than that of the RF and DualLSTM models, proving the reliability and superiority of the DualMamba model in early rice recognition.

[0086] On the other hand, the present invention provides a rice early identification device based on remote sensing data and time series model, and the various units included in the device can implement the aforementioned steps, specifically including:

[0087] A data acquisition unit, used for acquiring multi-source remote sensing data and auxiliary data;

[0088] A data preprocessing unit, used for preprocessing multi-source remote sensing data and auxiliary data;

[0089] A sample set construction unit, used to construct a sample set based on preprocessed multi-source remote sensing data and auxiliary data;

[0090] A model training unit, used to build a DualMamba model and train the DualMamba model based on a sample set;

[0091] The distribution map acquisition unit is used to input the multi-source remote sensing data and auxiliary data acquired by the data acquisition unit into the trained DualMamba model to obtain the early rice identification distribution map.

[0092] In a third aspect, the present invention provides an electronic device, characterized in that it comprises: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned early rice identification method based on remote sensing data and time series model.

[0093] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned early rice identification method based on remote sensing data and a time series model.

[0094] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for early rice identification based on remote sensing data and time series model, characterized in that: The following steps are involved: Step S1, acquiring multi-source remote sensing data and auxiliary data; Step S2, preprocessing multi-source remote sensing data and auxiliary data; Step S3, constructing a sample set based on the preprocessed multi-source remote sensing data and auxiliary data; Step S4, constructing a DualMamba model and training the DualMamba model based on the sample set; Step S5: input the multi-source remote sensing data and auxiliary data obtained in step S1 into the trained DualMamba model to obtain an early rice identification distribution map.

2. The method for early rice identification based on remote sensing data and time series model according to claim 1, characterized in that: The step S1 includes that the multi-source remote sensing data includes Level-1 GRD data in the wide-band interferometric mode (IW) of Sentinel-1, Level-1 dual-polarization characteristics in the fine strip 2 mode of GF-3, and Level-2A surface reflectance products of Sentinel-2; the auxiliary data includes field survey data, land cover / land use data and digital elevation model data, and the field survey data includes information on the type of land features and their geographical locations.

3. The method for early rice identification based on remote sensing data and time series model according to claim 1, characterized in that: The step S2 comprises: Step S2-1, performing radiation correction, terrain correction, speckle filtering and DB conversion on the Sentinel-1 data to generate a time series feature set of VH and VV polarizations; Step S2-2, performing multi-view processing, speckle filtering, geocoding and DB conversion on the GF-3 data to generate a time series feature set of HH and HV polarizations; Step S2-3, perform cloud removal, linear interpolation and median synthesis on the Sentinel-2 data, extract multispectral features, and generate a time series feature set including visible light, near infrared, red edge band and shortwave infrared band.

4. The method for early rice identification based on remote sensing data and time series model according to claim 1, characterized in that: The step S3 comprises: Step S3-1, making rice and non-rice labels based on the acquired field survey data combined with visual interpretation methods; Step S3-2, extracting the temporal characteristics of the sample area from the preprocessed multi-source remote sensing data to generate a comprehensive sample set containing multi-source optical and SAR characteristics; Step S3-3, based on the comprehensive sample set, according to the time node of acquiring the multi-source remote sensing data, construct an incremental sample set containing all features before the time node; Step S3-4, divide the incremental sample set into a training set, a validation set, and a test set in a ratio of 6:2:

2.

5. The method for early rice identification based on remote sensing data and time series model according to claim 1, characterized in that: The step S4 comprises: Step S4-1, constructing a DualMamba model, wherein the DualMamba model learns the optical and SAR data respectively through two parallel branches, and the outputs of the two branches are feature fused through an adaptive fusion module, and finally a binary classification result of rice and non-rice is generated through a fully connected layer; Step S4-2, the DualMamba model is trained using the AdamW optimizer, and the learning rate is adjusted by cosine annealing to ensure the effective convergence of the DualMamba model; starting from the starting DOY of rice sowing, incremental training is performed using the training samples corresponding to each DOY, and the date when the F1 score of the rice category first reaches 0.9 is finally determined as the earliest recognition date; during the model training process, the overall accuracy OA and the F1 score of the rice category are used for performance evaluation: ， ， ， ， Among them, TP (True Positive) represents the number of pixels correctly classified as rice, FP (False Positive) represents the number of non-rice pixels misclassified as rice, TN (True Negative) represents the number of pixels correctly classified as non-rice, FN (False Negative) represents the number of rice pixels misclassified as non-rice, OA (Overall Accuracy) represents the overall classification accuracy of the model for all pixels, Recall represents the proportion of correctly identified actual rice pixels, Precision represents the proportion of pixels identified as rice that are actually rice, and the F1 score is the harmonic average of precision and recall, which is used to comprehensively evaluate model performance.

6. The method for early rice identification based on remote sensing data and time series model according to claim 5, characterized in that: The DualMamba model in step S4-1 includes two parallel Mamba branches, which process optical and SAR data respectively, wherein the first Mamba branch first extracts the time series features in the Sentinel-2 data through a convolution layer, then applies batch normalization to standardize the time series features, and the selective state space module extracts the time series dependency features from the standardized time series features, and finally captures the time features through the SiLU activation function; the second Mamba branch first extracts the time series features of the Sentinel-1 and GF-3 data through a convolution layer, combines batch normalization and standardization of the time series features, and the selective state space module extracts the time series dependency features from the standardized time series features, and finally captures the time features through the SiLU activation function; the time features extracted by the two Mamba branches are fused through the adaptive fusion module, and the fused high-dimensional time series features are mapped to the output space through the fully connected layer, and the probability distribution of rice and non-rice plots is calculated through the Softmax layer to complete the time series binary classification task.

7. The method for early rice identification based on remote sensing data and time series model according to claim 1, characterized in that: The step S5 comprises: using all Sentinel-1, GF-3 and Sentinel-2 data from the initial DOY to the earliest identification date, combined with the trained DualMamba model, to generate a rice early identification distribution map, and post-processing the distribution map in combination with DEM data.

8. A device for early rice identification based on remote sensing data and time series model, characterized in that: include: A data acquisition unit, used for acquiring multi-source remote sensing data and auxiliary data; A data preprocessing unit, used for preprocessing multi-source remote sensing data and auxiliary data; A sample set construction unit, used to construct a sample set based on preprocessed multi-source remote sensing data and auxiliary data; A model training unit, used to build a DualMamba model and train the DualMamba model based on a sample set; The distribution map acquisition unit is used to input the multi-source remote sensing data and auxiliary data acquired by the data acquisition unit into the trained DualMamba model to obtain the early rice identification distribution map.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Wherein, when one or more programs are executed by the one or more processors, the one or more processors implement the early rice identification method based on remote sensing data and time series model as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, and when the instructions are executed by the processor, the processor can implement the early rice identification method based on remote sensing data and a time series model as described in any one of claims 1 to 7.

Images