Method for identifying rice planting area based on multi-temporal optical and polarimetric radar remote sensing images

Abstract

This invention discloses a method for identifying rice-growing areas based on multi-temporal optical and polarimetric radar remote sensing images, belonging to the field of agricultural remote sensing technology. The invention includes: S1. Dataset construction, S2. Data preprocessing, S3. Data partitioning, S4. Model construction, S5. Identification step, S6. Re-acquisition step, S7. Model transfer, and S8. Re-identification step. By fusing multi-temporal and multi-source remote sensing image information, the model can effectively identify rice-growing areas. Compared with traditional supervised classification methods and deep convolutional neural network semantic segmentation algorithms, this model shows significant advantages in classification performance. Furthermore, the model also exhibits good cross-year generalization performance, achieving high recognition accuracy not only in the training year but also obtaining good interpretation results in other years. Moreover, the rice-growing fields in the classification map show a continuous distribution, which is more consistent with the actual situation.

Description

Technical Field

[0001] This invention relates to the field of agricultural remote sensing technology, and in particular to a method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images. Background Technology

[0002] Rice is one of the most important food crops for human survival. Remote sensing technology can provide long-term, continuous Earth observation data to monitor rice-growing areas, understand their distribution patterns, and predict overall yields, enabling precision and smart agriculture. Optical imagery provides good differentiation of crops in the near-infrared and short-wave infrared bands. However, high-quality remote sensing images are often difficult to obtain due to cloud cover, while radar imagery can penetrate clouds and fog, is available in all weather conditions and at all times, and is less affected by weather conditions and sunlight. Furthermore, given the significant phase variation in rice's backscattering coefficient, combining radar and optical imagery allows for better identification and monitoring of rice.

[0003] Compared to traditional machine learning methods, deep neural networks can learn and reflect the essential features of input data, enhancing the ability to identify and utilize feature information and improving image interpretation performance. However, existing methods mainly utilize single-source data and fail to fully leverage the complementary advantages of multi-source remote sensing imagery, limiting the reliability of interpretation algorithms. For example, when using multi-temporal data to identify rice, the differences in rice growth cycles between different years are not considered, and the trained model is directly applied, severely impacting the model's generalization performance and interpretation results.

[0004] In view of this, the present invention uses the band information and polarization information of optical images and PolSAR images from different phenological periods to train a deep convolutional neural network model for rice monitoring. Considering the changes in the rice growth cycle caused by meteorological differences, an adaptive transfer mechanism is adopted before identifying rice in image sets from different years to ensure that it has strong cross-year generalization performance.

[0005] Therefore, proposing a method for identifying rice-growing areas based on multi-temporal optical and polarimetric radar remote sensing images to address the difficulties in existing technologies is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention provides a method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images. It not only achieves high recognition accuracy in the training year, but also obtains good interpretation results in other years. Moreover, the rice planting fields in the classification map show a continuous distribution pattern, which is more in line with the actual situation.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A method for identifying rice-growing areas based on multi-temporal optical and polarimetric radar remote sensing imagery includes the following steps:

[0009] S1. Constructing the dataset: Obtain Landsat-8 and Sentinel-1 imagery data from different years and construct a dataset;

[0010] S2. Data preprocessing: The Landsat-8 and Sentinel-1 image data of a certain year are preprocessed to obtain multi-temporal data;

[0011] S3. Data partitioning: Divide the multi-temporal data into training and validation sets;

[0012] S4. Model Building: Build a model based on multi-temporal data, train the model using the training set, and obtain a deep learning model;

[0013] S5. Identification Steps: Use a deep learning model to obtain the identification results of rice planting areas for a specific year;

[0014] S6. Further acquisition step: Landsat-8 image data and Sentinel-1 image data from another year obtained from the dataset are processed through step S2 to obtain multi-temporal data for another year;

[0015] S7. Model Transfer: The deep learning model obtained in S4 is transferred using an adaptive cross-year transfer mechanism to obtain a transferred deep learning model.

[0016] S8. Re-identification step: Input the multi-temporal data of another year obtained in S6 into the transfer deep learning model to obtain the rice planting area identification results of another year.

[0017] Optionally, Landsat-8 data in S1 is obtained from the geospatial data cloud, and Sentinel-1 data is obtained from the ESA data distribution website.

[0018] Optionally, in S2, ENVI software is used to perform radiometric correction, atmospheric correction, resampling, and cropping preprocessing on Landsat-8 image data; and polarimetric speckle filtering and geocoding preprocessing on Sentinel-1 image data.

[0019] Optionally, the multi-temporal data obtained in S2 includes, but is not limited to, multi-band characteristics of optical images of rice at different growth stages, and multi-polarization amplitude, phase and polarization decomposition parameters of PolSAR images.

[0020] Optionally, the ratio of training set to validation set in the S3 data partition is 7:3.

[0021] Optionally, the specific content of model construction in S4 is as follows:

[0022] Build the U-Net network architecture model framework;

[0023] Input the training set and train the model to learn;

[0024] Adjust the model's hyperparameters in real time based on the model's training and learning progress;

[0025] Save the trained model.

[0026] Optionally, the specific details of the adaptive cross-year migration mechanism in S7 are as follows:

[0027] By employing a similar image pair matching strategy, image pairs similar to those used in the training year of the rice recognition model are found in other years, enabling adaptive adjustment of the input image date and allowing the model to be applied in other years.

[0028] Optionally, the specific details of the similar image pair matching strategy are as follows:

[0029] The similarity between pixel pairs at the same location within the rice-growing area on the SAR image is calculated using the Wishart likelihood ratio, and the average similarity of all pixel pairs is taken as the scattering similarity between image pairs. In other years, the image with the closest similarity to the image used in the year the rice recognition model was trained is taken as the similar image pair.

[0030] Optionally, the Wishart likelihood ratio is calculated using the following formula:

[0031]

[0032] Taking the logarithm of equation (1) and discarding the constant term, we obtain the final likelihood ratio test value:

[0033] Q2=ln|X|+ln|Y|-2|X+Y| (2)

[0034] Where X and Y are polarization covariance matrices, L is the number of image views, and |·| represents the matrix determinant.

[0035] As can be seen from the above technical solution, compared with the prior art, the present invention provides a method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images, which has the following beneficial effects:

[0036] (1) This invention takes the Sanjiang Plain as the research area and proposes a deep learning model for rice planting areas based on multi-temporal and multi-source images. The model integrates rich band information in optical images and polarization information in PolSAR data for training. Considering the differences in rice growth period in different years caused by factors such as climate conditions, a transfer mechanism is introduced to ensure that the model can be applied to rice identification in different years.

[0037] (2) By fusing remote sensing image information from multiple time periods and sources, the model of this invention can effectively identify rice-growing areas. Compared with traditional supervised classification methods and deep convolutional neural network semantic segmentation algorithms, this model shows significant advantages in classification performance; moreover, the model also demonstrates good cross-year generalization performance, achieving high recognition accuracy not only in the training year but also obtaining good interpretation results in other years. Furthermore, the rice-growing fields in the classification map show a continuous distribution pattern, which is more in line with the actual situation. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0039] Figure 1 Flowchart of the method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images provided by the present invention;

[0040] Figure 2 The deep learning recognition model framework provided by this invention;

[0041] Figure 3 This invention provides a multi-temporal, multi-source remote sensing image dataset.

[0042] Figure 4 A 2020 image scatter plot provided for an embodiment of the present invention;

[0043] Figure 5 A scatter plot of images from 2021 that are close to the date of 2020, provided for an embodiment of the present invention;

[0044] Figure 6 A scatter plot of 2021 images obtained according to a similar image pair matching strategy provided in an embodiment of the present invention;

[0045] Figure 7 The 2020 single-phase Landsat-8 dataset provided for embodiments of the present invention;

[0046] Figure 8 The 2020 multi-temporal Landsat-8 dataset provided for embodiments of the present invention;

[0047] Figure 9 The 2020 single-phase Sentinel-1 dataset provided for embodiments of the present invention;

[0048] Figure 10The 2020 multi-temporal Sentinel-1 dataset provided for embodiments of the present invention;

[0049] Figure 11 The original Landsat-8 image from 2020 provided for embodiments of the present invention;

[0050] Figure 12 The original polarimetric SAR imagery of 2020 provided for embodiments of the present invention;

[0051] Figure 13 The 2020 tagged image provided for an embodiment of the present invention;

[0052] Figure 14 The classification results of this model on the 2020 single-phase Landsat-8 dataset provided in this embodiment of the invention;

[0053] Figure 15 The classification results of this model on the 2020 multi-temporal Landsat-8 dataset provided in this embodiment of the invention;

[0054] Figure 16 The classification results of this model on the 2020 single-temporal SAR dataset provided in this embodiment of the invention;

[0055] Figure 17 The classification results of this model on the 2020 multi-temporal SAR dataset provided in this embodiment of the invention;

[0056] Figure 18 The classification results of this model on the 2020 single-temporal Landsat-8+SAR dataset provided in this embodiment of the invention;

[0057] Figure 19 The classification results of this model on the 2020 multi-temporal Landsat-8+SAR dataset provided in this embodiment of the invention;

[0058] Figure 20 The images show the identification results of rice planting areas in 2020 using different methods provided in the embodiments of the present invention, where a is the original polarimetric SAR image, b is the labeled image, c is the RF supervised classification result, d is the Wishart supervised classification result, e is the SVM supervised classification result, f is the Deeplab model classification result, g is the FCN model classification result, and h is the classification result of the model of the present invention.

[0059] Figure 21 The identification of rice planting areas in 2021 provided in this embodiment of the invention includes, where a is the original polarimetric SAR image, b is the labeled image, c is the deep learning identification result without using the transfer mechanism, and d is the deep learning identification result using the transfer mechanism. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Reference Figure 1 As shown, this invention discloses a method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images, comprising the following steps:

[0062] S1. Constructing the dataset: Obtain Landsat-8 and Sentinel-1 imagery data from different years and construct a dataset;

[0063] S2. Data preprocessing: The Landsat-8 and Sentinel-1 image data of a certain year are preprocessed to obtain multi-temporal data;

[0064] S3. Data partitioning: Divide the multi-temporal data into training and validation sets;

[0065] S4. Model Building: Build a model based on multi-temporal data, train the model using the training set, and obtain a deep learning model;

[0066] S5. Identification Steps: Use a deep learning model to obtain the identification results of rice planting areas for a specific year;

[0067] S6. Further acquisition step: Landsat-8 image data and Sentinel-1 image data from another year obtained from the dataset are processed through step S2 to obtain multi-temporal data for another year;

[0068] S7. Model Transfer: The deep learning model obtained in S4 is transferred using an adaptive cross-year transfer mechanism to obtain a transferred deep learning model.

[0069] S8. Re-identification step: Input the multi-temporal data of another year obtained in S6 into the transfer deep learning model to obtain the rice planting area identification results of another year.

[0070] Furthermore, Landsat-8 data in S1 is obtained from the geospatial data cloud, and Sentinel-1 data is obtained from the ESA data distribution website.

[0071] Furthermore, in S2, ENVI software is used to perform radiometric correction, atmospheric correction, resampling, and cropping preprocessing on Landsat-8 image data; and polarimetric speckle filtering and geocoding preprocessing on Sentinel-1 image data.

[0072] Furthermore, the multi-temporal data obtained in S2 includes, but is not limited to, multi-band characteristics of optical images of rice at different growth stages, and multi-polarization amplitude, phase, and polarization decomposition parameters of PolSAR images.

[0073] Furthermore, the ratio of training set to validation set in the S3 data partition is 7:3.

[0074] Furthermore, the specific details of model construction in S4 are as follows:

[0075] Build the U-Net network architecture model framework;

[0076] Input the training set and train the model to learn;

[0077] Adjust the model's hyperparameters in real time based on the model's training and learning progress;

[0078] Save the trained model.

[0079] Furthermore, the specific details of the adaptive cross-year migration mechanism in S7 are as follows:

[0080] By employing a similar image pair matching strategy, image pairs similar to those used in the training year of the rice recognition model are found in other years, enabling adaptive adjustment of the input image date and allowing the model to be applied in other years.

[0081] Furthermore, the specific details of the similar image pair matching strategy are as follows:

[0082] The similarity between pixel pairs at the same location within the rice-growing area on the SAR image is calculated using the Wishart likelihood ratio, and the average similarity of all pixel pairs is taken as the scattering similarity between image pairs. In other years, the image with the closest similarity to the image used in the year the rice recognition model was trained is taken as the similar image pair.

[0083] Furthermore, the Wishart likelihood ratio is calculated as follows:

[0084]

[0085] Taking the logarithm of equation (1) and discarding the constant term, we obtain the final likelihood ratio test value:

[0086] Q2=ln|X|+ln|Y|-2|X+Y| (2)

[0087] Where X and Y are polarization covariance matrices, L is the number of image views, and |·| represents the matrix determinant.

[0088] In a specific embodiment, the details are as follows:

[0089] The Sanjiang Plain was selected as the study area. This region cultivates single-season rice, typically sown from late April to early May and harvested between September and October, with a growth cycle of approximately 200-240 days. The study area is flat and vast, with highly mechanized agricultural production, ensuring synchronized rice sowing and harvesting. This provides a highly feasible approach for identifying large-scale rice-growing areas using multi-source, multi-temporal imagery.

[0090] Landsat-8 and Sentinel-1 were used as data sources. Landsat-8 data can be downloaded from the geospatial data cloud. Considering cloud cover and quality in the study area, only three months of imagery data were selected. Sentinel-1 data can be downloaded from the ESA data distribution website. ENVI software was used to preprocess the optical images, including radiometric correction, atmospheric correction, resampling, and cropping. The preprocessing stage of Sentinel-1 data mainly included basic processing such as polarimetric speckle filtering and geocoding.

[0091] To differentiate rice from other vegetation with similar spectral reflectance and scattering characteristics, a deep learning model for the rice-growing area was trained using multi-temporal data based on rice's growth cycle and phenological characteristics. Three Landsat-8 images and five Sentinel-1 images from the study area in 2020, representing the rice seedling, tillering, heading, and maturity stages, were selected. Table 1 lists the specific imaging times of the images used for model training.

[0092] Table 1. Information on the Landsat-8 and Sentinel-1 datasets used for model training.

[0093]

[0094]

[0095] The deep learning-based rice-growing area recognition model proposed in this invention is built upon the U-Net network architecture. The model structure is as follows: Figure 2 As shown, the encoder consists of two parts: a contraction path and a symmetric expansion path. The contraction path is the encoder part, which uses two 3×3 convolutional layers for feature extraction and downsampling through ReLU and 2×2 max pooling operations. In each downsampling step, the number of feature channels is doubled, and feature information is extracted. The symmetric expansion path is the decoder part, which reduces the number of feature channels by half through upsampling operations and 3×3 convolutional layers. Simultaneously, symmetric skip connections connect the corresponding layers of the contraction and expansion paths.

[0096] The model fully utilizes the rich feature information from multi-source remote sensing imagery to achieve accurate rice identification. To this end, it integrates multi-band features (B1–B7) from optical images of rice at different growth stages, and multi-polarization amplitude, phase, and polarization decomposition parameters (polarization entropy H / inverse entropy A / dihedral angle Alpha) from PolSAR images as input feature data for the model (e.g., ...). Figure 3 (As shown).

[0097] The training set for the model consists of images from 2020. To apply the model to rice-growing area identification in other years, an adaptive transfer mechanism is proposed to ensure strong cross-year generalization performance and accurate identification of rice-growing areas in different years. The core idea of ​​this transfer mechanism is to use a similar image pair matching strategy to find similar image pairs between the network training dataset and the dataset of the year to be interpreted. This enables adaptive adjustment of the input image dates during the interpretation process, allowing the model to have better fitting performance in new years.

[0098] To measure the scattering similarity of rice paddies in two images, the Wishart likelihood ratio statistical test is first used to calculate the similarity between pixel pairs at the same location within the rice-growing area of ​​the SAR image. The average similarity of all pixel pairs is then taken as the scattering similarity between the image pairs. For the polarization covariance matrices X and Y, the Wishart likelihood ratio is calculated as follows:

[0099]

[0100] Taking the logarithm of the above equation and discarding the constant term, we obtain the final likelihood ratio test value:

[0101] Q2=ln|X|+ln|Y|-2|X+Y| (2)

[0102] In the formula, L is the number of image views, and |·| represents the matrix determinant.

[0103] Figure 4-6 The images shown are scatter plots of 2020 images obtained by matching the Wishart likelihood ratio calculation method, scatter plots of 2021 images with dates close to 2020, and scatter plots of 2021 images obtained by matching similar image pairs; where the x-axis and y-axis represent the amplitude values ​​of different polarizations, and the z-axis represents the phase difference. Figure 4 and Figure 5 The green scatter plots showed a highly similar distribution pattern, but the red and blue scatter plots showed significant differences, indicating that there were certain differences in the phenological stages of rice between the two years. Figure 6 The scatter distribution state and Figure 4 More consistent, that is, the proposed similar images are effective for the matching strategy.

[0104] To train the deep neural network for rice-growing area identification, 400 sample blocks of 256×256 pixels were selected from the 2020 imagery of the study area, with the area ratio of non-rice-growing areas to rice-growing areas approximately 3:1. To prevent overfitting, data augmentation was employed, increasing the sample size to 1600 through vertical and horizontal flipping. The samples were randomly divided into training and validation parts in a 7:3 ratio. The model input data included band information from three preprocessed optical images at different time phases and polarization information from five PolSAR images at different time phases, totaling 56 feature channels.

[0105] To objectively evaluate the classification performance of different rice monitoring methods, four commonly used evaluation metrics are introduced: overall accuracy (OA), precision, recall, and F1 score. Among them, OA and F1 are more comprehensive metrics that can more objectively and comprehensively evaluate the performance of classifiers.

[0106] In another specific embodiment, to verify the advantages of multi-temporal and multi-source information, the experimental part also constructed four other different dataset combinations for model training (such as...). Figure 7-10 (As shown in Table 2). Comparison Table 2 and Figure 11-19 The identification results clearly show that using a multi-temporal dataset significantly improves model performance. Furthermore, it was found that using multi-source imagery also effectively improves the model's classification accuracy to some extent.

[0107] Table 2 Quantitative evaluation values ​​of recognition results for different datasets

[0108] Dataset OA (%) Precision (%) Recall (%) F1(％) Single-phase Landsat-8 75.61 81.57 16.77 27.82 Multi-phase Landsat-8 82.01 88.39 41.21 56.21 Single-phase SAR 86.27 82.36 64.31 72.60 Multi-temporal SAR 93.11 85.68 90.57 88.06 Single-phase Landsat-8+SAR 90.18 80.21 86.23 83.11 Multi-temporal Landsat-8+SAR 95.71 91.47 89.90 90.68

[0109] To fully verify the performance of the rice planting area identification model proposed in this invention, the interpretation results of this model are compared with those of traditional supervised classification methods (Random Forest (RF) supervised classification, Wishart supervised classification, and Support Vector Machine (SVM) supervised classification) and classic semantic segmentation algorithms (DeepLab model and FCN model). Figure 20 To demonstrate the identification results of different methods for rice planting areas in 2020, where a is the original polarimetric SAR image, b is the labeled image, c is the RF supervised classification result, d is the Wishart supervised classification result, e is the SVM supervised classification result, f is the Deeplab model classification result, g is the FCN model classification result, and h is the classification result of the model of this invention; Table 3 lists the quantitative evaluation values ​​of the identification accuracy of different methods.

[0110] Table 3 Quantitative evaluation results of different methods

[0111]

[0112]

[0113] As can be seen, the rice planting area identification model proposed in this invention performs well overall and has significant advantages. Traditional supervised classification methods all achieve F1 scores below 80%. Among semantic segmentation algorithms, the DeepLab model has the lowest recognition accuracy.

[0114] In traditional supervised classification, the classification maps of RF, Wishart, and SVM contain many points with blurred boundaries, severely affecting classification accuracy. This indicates that traditional interpretation methods only consider low-level features of the target while ignoring contextual semantic information and high-level structural information of the image. Furthermore, Wishart and SVM methods exhibit severe misclassification of ground features with similar scattering characteristics to rice. The three deep learning semantic segmentation algorithms used in the experiment take into account the contextual information of the image, the reflection information of ground features in optical images, and the polarization scattering information of ground features in SAR images, effectively reducing the impact of noise in the image and significantly improving the recognition effect of rice-growing areas. Visually, the targets in the classification map of the model proposed in this invention are more continuous, and the boundaries between rice-growing areas and other areas are clearer, better reflecting reality.

[0115] Table 4. Quantitative evaluation results of rice planting area identification in 2021

[0116] OA (%) Precision (%) Recall() F1(％) Do not utilize migration mechanisms 87.43 85.82 73.59 77.24 Utilizing migration mechanisms 89.02 82.26 81.44 80.63

[0117] To improve the cross-year generalization performance of the model in this invention, a transfer mechanism is introduced. Before identifying rice-growing areas in other years, a similar image matching strategy is used to find similar image pairs between the training dataset and the dataset to be interpreted. To verify the effectiveness of this transfer mechanism, Figure 21 To identify rice-growing areas in 2021, the identification results are shown with and without the transfer mechanism. In the table, a represents the original polarimetric SAR image, b represents the labeled image, c represents the deep learning identification result without the transfer mechanism, and d represents the deep learning identification result with the transfer mechanism. Table 4 lists the corresponding quantitative evaluation results. Clearly, the increase in the two most important metrics, OA and F1, indicates that the transfer mechanism proposed in this invention can effectively improve the generalization ability of the deep learning rice identification model.

[0118] Table 5. Similar image pairs between 2020 and 2021

[0119]

[0120] Furthermore, it can be observed that the image dates found using the image similarity principle are generally shifted later than those found based on the date proximity principle. According to relevant information provided on the Heilongjiang Provincial Meteorological Bureau website, affected by the cold wave in mid-April, the rice sowing time in the Sanjiang Plain in 2021 was generally delayed compared to 2020. This is logically consistent with the findings in Table 5, further verifying the reliability of the similar image pair matching strategy proposed in this invention.

[0121] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images, characterized in that, Includes the following steps: S1. Constructing the dataset: Obtain Landsat-8 and Sentinel-1 imagery data from different years and construct a dataset; S2. Data preprocessing: The Landsat-8 and Sentinel-1 image data of a certain year are preprocessed to obtain multi-temporal data; S3. Data partitioning: Divide the multi-temporal data into training and validation sets; S4. Model Building: Build a model based on multi-temporal data, train the model using the training set, and obtain a deep learning model; S5. Identification Steps: Use a deep learning model to obtain the identification results of rice planting areas for a specific year; S6. Further acquisition step: Landsat-8 image data and Sentinel-1 image data from another year obtained from the dataset are processed through step S2 to obtain multi-temporal data for another year; S7. Model Transfer: The deep learning model obtained in S4 is transferred using an adaptive cross-year transfer mechanism to obtain a transferred deep learning model. S8. Re-identification step: Input the multi-temporal data of another year obtained in S6 into the transfer deep learning model to obtain the rice planting area identification results of another year; The multi-temporal data obtained in S2 includes, but is not limited to, multi-band characteristics of optical images of rice at different growth stages, and multi-polarization amplitude, phase and polarization decomposition parameters of PolSAR images; The specific details of the adaptive cross-year migration mechanism in S7 are as follows: By employing a similar image pair matching strategy, image pairs similar to those used in the training year of the rice recognition model are found in other years, enabling adaptive adjustment of the input image date and allowing the model to be applied in other years. The specific details of the similar image pair matching strategy are as follows: The similarity between pixel pairs at the same location in the rice-growing area of ​​the SAR image is calculated using the Wishart likelihood ratio, and the average of the similarity of all pixel pairs is taken as the scattering similarity between the image pairs. In other years, the image with the closest similarity to the image used in the year in which the rice recognition model was trained is taken as the similar image pair. The formula for calculating the Wishart likelihood ratio is: （1） Taking the logarithm of equation (1) and discarding the constant term, we obtain the final likelihood ratio test value: （2） in, and The polarization covariance matrix is... L For image views, Represents the determinant of a matrix.

2. The method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images according to claim 1, characterized in that, Landsat-8 data in S1 is obtained from the geospatial data cloud, and Sentinel-1 data is obtained from the ESA data distribution website.

3. The method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images according to claim 1, characterized in that, In S2, ENVI software was used to perform radiometric correction, atmospheric correction, resampling, and cropping preprocessing on Landsat-8 image data; and polarimetric speckle filtering and geocoding preprocessing were performed on Sentinel-1 image data.

4. The method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images according to claim 1, characterized in that, In the S3 data partition, the ratio of training set to validation set is 7:

3.

5. The method for identifying rice planting areas based on multi-temporal optical and polarimetric radar remote sensing images according to claim 1, characterized in that, The specific content of model construction in S4 is as follows: Build the U-Net network architecture model framework; Input the training set and train the model to learn; Adjust the model's hyperparameters in real time based on the model's training and learning progress; Save the trained model.

Images