Production method of high-spatial-resolution seamless leaf area index product

Abstract

The invention discloses a production method of a high-spatial-resolution seamless leaf area index product. Comprising the following steps: S1, producing a high-spatial-resolution seamless reflectivity image: reconstructing missing data of a high-spatial-resolution and low-spatial-resolution HLS image by using a high-temporal-resolution and low-spatial-resolution MODIS image so as to produce 30m spatial-resolution seamless reflectivity data per 12 days; and step S2, constructing a leaf area index inversion model based on a Transform model. The invention provides a Transform-based leaf area index inversion model, overcomes the defect that the existing leaf area index inversion model based on a long-short term memory neural network and a bidirectional long-short term memory neural network is difficult to fully utilize time sequence image information, and improves the inversion precision of the leaf area index. The technical process for producing the high-spatial-resolution seamless leaf area index product is provided, and the defect that a large amount of data is lost in an existing high-spatial-resolution leaf area index product is overcome.

Description

Technical Field

[0001] The present invention relates to a product production method, and particularly to a method for producing a high-spatial-resolution seamless leaf area index product. Background Art

[0002] The leaf area index is defined as half of the green leaf area of vegetation per unit land area. It is a key variable in processes such as photosynthesis, respiration, and precipitation interception. Achieving continuous monitoring of the leaf area index at a fine scale helps to understand the growth dynamics of vegetation and is crucial for ecosystem health assessment, precision management, and global change research. The monitoring methods of the leaf area index mainly include traditional methods and remote sensing inversion. Traditional methods use destructive sampling to obtain leaf area index information. Although it is accurate, it has a large workload, poor timeliness, and can only be used for single-point measurements. In contrast, remote sensing technology can achieve large-area synchronous monitoring, facilitating the production of spatially seamless leaf area index products. Therefore, it is of great significance to achieve high-resolution mapping of the leaf area index based on remote sensing.

[0003] Regarding the production of remote-sensing-based leaf area index products, there are mainly two methods: the physical model method and the empirical model method. Among them, the physical model method uses a leaf optical model with clear physical meaning and a canopy structure simulation model to simulate the canopy reflectance spectrum with parameters such as leaf physiological and biochemical component information, leaf structure information, canopy structure information, solar zenith angle, solar azimuth angle, observation zenith angle, and even a series of scene construction information (3D model) as inputs, and realizes the inversion of the leaf area index based on its inverse process. To improve the inversion efficiency, the look-up table method is often used in the physical model method. The look-up table method forms a table by pre-defining possible parameter combination forms and using the physical model to simulate the corresponding spectral reflectance, and then finds the most matching entry in the table according to the observed data, thereby realizing the inversion of the leaf area index. The advantage of the physical model method is that its mechanism meaning is clear, and the disadvantages are low computational efficiency and easy occurrence of "ill-conditioned" inversion problems. The empirical model method is a method that directly establishes a quantitative relationship model between image features and the leaf area index based on statistical methods and then uses the model to predict the leaf area index. Among them, training samples are required to establish the quantitative relationship between image features and the leaf area index. The leaf area index values of these samples usually come from the leaf area index values inverted by using the selected image pure pixels combined with the physical model method or ground-measured leaf area index data; the models for establishing the quantitative relationship usually include machine learning methods such as spectral index method, random forest, and bidirectional long short-term memory neural network. The advantages of the empirical model method are high computational efficiency and no "ill-conditioned" inversion problems, and the disadvantages are that a large number of training samples are required, and a suitable model is needed to accurately learn the relationship between image features and the leaf area index through the samples, so as to realize the inversion of the leaf area index.

[0004] Based on the physical model method and the empirical model method, many leaf area index products have been produced, including the MODIS leaf area index product of the National Aeronautics and Space Administration of the United States, the CYCLOPES leaf area index product of the European Space Agency, and the GLASS leaf area index product of China. However, the uncertainty of these products is greater than 15%, and the accuracy is difficult to meet the monitoring requirements of the vegetation leaf area index. Moreover, most of them have a relatively low spatial resolution, about 500m. Although there are a few leaf area index products made based on high-resolution satellite images with a spatial resolution of 16-30m, due to the long revisit period and narrow swath of high-spatial-resolution satellites, they are easily affected by cloudy and rainy weather, resulting in a large number of missing data in the products. Therefore, there is an urgent need to design a leaf area index inversion method to propose a model with strong learning ability to effectively learn the relationship between image features and the leaf area index based on samples, improve the inversion accuracy of the leaf area index, and consider relevant technologies to eliminate the influence of cloudy and rainy weather to achieve the production of high-resolution seamless leaf area index products.

[0005] Leaf area index is an important parameter reflecting the growth status of vegetation, and it is of great significance to achieve its accurate monitoring. Aiming at the defects existing in the existing methods, the present invention intends to solve the following technical problems: 1) Considering the satellite transit period and image resolution, most of the existing international leaf area index inversion products are produced based on MODIS reflectance images and AVHRR reflectance images, and their resolutions are mostly around 500m. Because at this resolution, the availability of optical images is relatively high. However, there will be many mixed pixels at a 500m spatial resolution, resulting in low accuracy of these leaf area index products and difficulty in meeting user needs. Although there are currently a small number of leaf area index products with a resolution of 16 - 30m inverted based on Chinese high-resolution satellite reflectance images and US Landsat satellite reflectance images, due to the influence of satellite transit period and cloud and rain weather, these products have a large amount of missing data and are difficult to utilize. 2) The leaf area index of vegetation changes regularly and dynamically. Images at different times can support each other to capture this change, thus achieving a better leaf area index inversion effect. Most of the existing leaf area index inversion models are established using single-temporal images based on spectral index methods, machine learning methods such as random forests, etc., and cannot utilize temporal information, resulting in relatively large errors; a few studies have established leaf area index inversion models using temporal images based on methods such as long short-term memory neural networks and bidirectional long short-term memory neural networks, which can utilize temporal information. Although the above methods improve the inversion accuracy of the leaf area index, since long short-term memory neural networks and bidirectional long short-term memory neural networks process temporal data step by step in the order of information input time, they cannot well retain the information transmitted from earlier time steps in the time series, making it difficult to utilize all the information of temporal images and limiting the inversion accuracy of the model. Therefore, the existing methods cannot effectively learn the relationship between image features and leaf area index using sample data. In addition, the leaf area index products produced by existing studies either have a low spatial resolution or have a large amount of missing data, and it is difficult to form a high-spatial-resolution seamless leaf area index product to support vegetation health monitoring. Summary of the Invention

[0006] In order to solve the above-mentioned deficiencies in the technology, the present invention provides a method for producing a high-spatial-resolution seamless leaf area index product.

[0007] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for producing a high-spatial-resolution seamless leaf area index product, the method comprising the following steps:

[0008] Step S1, production of high-spatial-resolution seamless reflectance images: By using MODIS images with high temporal resolution and low spatial resolution to reconstruct the missing data of HLS images with high spatial resolution and low temporal resolution to produce seamless reflectance data with a 30m spatial resolution every 12 days;

[0009] Step S2: Construction of Leaf Area Index Inversion Model Based on Transformer Model: Based on the constructed leaf area index inversion model, using the reflectance data of each band produced in Step S1 as input, run the model to obtain a high-spatial-resolution seamless leaf area index product.

[0010] Preferably, Step S1 includes the following steps:

[0011] Step S11: Cloud and cloud shadow removal;

[0012] Step S12: 12-day reflectance synthesis based on the maximum coverage strategy;

[0013] Step S13: Production of spatially seamless MODIS images;

[0014] Step S14: Automatic spatial registration of MODIS images and HLS images;

[0015] Step S15: Reconstruction of missing data in HLS images;

[0016] Step S16: Optimization of the reconstructed HLS images.

[0017] Preferably, the specific process of Step S11 is as follows:

[0018] For MODIS images, use its quality layer to identify cloud and cloud shadow areas;

[0019] For HLS images, first use its Fmask layer to remove cloud and cloud shadow areas, and then, use the HOT index method to further remove cloud areas, and use the automatic threshold method to determine the threshold of the HOT index.

[0020] Preferably, the specific process of Step S12 is as follows: For the non-overlapping areas of spatially adjacent images, first determine the cloud coverage of each image within 12 days; secondly, in the order of increasing cloud coverage range, select images to fill the target area in turn to minimize the number of missing values in the area, thereby forming a composite image.

[0021] For the overlapping areas between spatially adjacent images, adopt the image synthesis strategy of NASA's MODIS data products.

[0022] Preferably, the specific process of Step S13 is as follows: Perform SG filtering on the 12-day MODIS composite image to generate a spatio-temporally seamless MODIS reflectance product, and during the SG filtering process, set the polynomial fitting order to 4 and the sliding window size to 8.

[0023] Preferably, the specific process of step S14 is as follows: First, resample and clip the MODIS image to make its spatial resolution and range the same as that of the HLS image; then, use the automatic registration and orthorectification algorithm to perform an affine transformation on the MODIS image to make its spatial position consistent with the corresponding HLS image.

[0024] Preferably, the specific process of step S15 is as follows: First, use the partial least squares model to determine the relationship between the reflectance of the target date and the reflectance of adjacent dates in the low-spatial-resolution image; then, based on the assumption of local scale invariance of the above relationship, apply the model to the high-spatial-resolution image to reconstruct the missing data of the target date through the data of adjacent dates; during this process, when selecting the HLS images of the adjacent dates before and after the target date, the image with the least data missing has the highest priority, and the original image has a higher priority than the reconstructed image.

[0025] Preferably, the specific process of step S15 is: Perform SG filtering on the reconstructed HLS data to further fill in the data missing and correct the outliers.

[0026] Preferably, in step S2, the leaf area index inversion model includes:

[0027] An input layer: The input layer processes the spectral reflectance vectors of each band at the current time step and the previous two time steps.

[0028] An output layer: The output layer obtains the leaf area index value of the current time step.

[0029] A hidden layer: The hidden layer contains a basic Transformer encoding module, which specifically includes a multi-head self-attention mechanism module, two layer normalization modules, and a feed-forward network module.

[0030] Preferably, the multi-head self-attention mechanism module contains 8 heads, and each head uses 16 nodes to obtain the correlation between different time step information; the feed-forward network module consists of 20 nodes;

[0031] Feature linear embedding and position encoding are performed before the information is transmitted to the hidden layer; when training the model based on samples, the batch scale is set to 64, the maximum number of training epochs is 150, the initial learning rate is 0.0001, and the optimization function is Adam.

[0032] The object of the present invention is to propose a technical method for producing a spatio-temporally seamless leaf area index product with high spatio-temporal resolution. In this process, firstly, a technical process for reconstructing the missing data of HLS images based on MODIS images is proposed to produce a reflectance image with a 30m resolution that is spatio-temporally seamless every 12 days; secondly, based on the produced spatio-temporally seamless 30m reflectance image, the Transformer model has a multi-head self-attention mechanism that can calculate the spectral correlation matrix between any two time points in the input time series, and the results of each head are weighted and fused, which can fully learn the dependencies between time series and will not suffer from information attenuation due to changes in the length of the time series. Combining with the time series reflectance image, a leaf area index inversion model based on Transformer is constructed to overcome the defect that the existing leaf area index inversion models based on long short-term memory neural networks and bidirectional long short-term memory neural networks are difficult to make full use of the information in the time series images and improve the inversion accuracy of the leaf area index.

[0033] The present invention proposes a method for producing a high-spatial-resolution seamless leaf area index product. It includes two parts: 1) generation of a high-spatial-resolution seamless reflectance image; 2) construction of a leaf area index inversion model based on Transformer. Among them, in terms of generating a high-spatial-resolution seamless reflectance image, the present invention proposes a technical process for generating a high-spatial seamless reflectance image through steps such as cloud and cloud shadow removal, reflectance synthesis based on the maximum coverage model, production of a spatial seamless MODIS image, automatic spatial registration of the image, reconstruction of missing data of the HLS image, and optimization of the reconstructed image; in terms of constructing a leaf area index inversion model based on Transformer, the network architecture of the Transformer model for inverting the leaf area index is proposed, thus overcoming the problems existing in the prior art. Description of the Drawings

[0034] Figure 1 It is the technical flow chart of the present invention.

[0035] Figure 2 It is the architecture diagram of the leaf area index inversion model based on the Transformer model of the present invention.

[0036] Figure 3 It is the leaf area index data diagram of the embodiment of the present invention.

[0037] Figure 4 It is the corresponding relationship between the measured leaf area index and the inverted leaf area index based on the verification sample data set in the embodiment of the present invention. Detailed Embodiment

[0038] The present invention will be further described in detail below in conjunction with the drawings and specific embodiments.

[0039] AsFigure 1 A method for producing a high-spatial-resolution seamless leaf area index product, which mainly consists of two parts: the production of high-spatial-resolution seamless reflectance images and the construction of a leaf area index inversion model based on the Transformer model.

[0040] (1) High-spatial-resolution seamless reflectance images:

[0041] In the present invention, the missing data of high-spatial-resolution and low-temporal-resolution HLS images are reconstructed by using MODIS images with high temporal resolution and low spatial resolution to produce 30m spatial seamless reflectance data every 12 days. The whole process includes six steps: cloud and cloud shadow removal, 12-day reflectance synthesis based on the maximum coverage strategy, production of spatial seamless MODIS images, automatic spatial registration of MODIS images and HLS images, reconstruction of missing data of HLS images, and optimization of the reconstructed images. The technical route is as Figure 1 shown and is described in detail as follows:

[0042] 1) Cloud and cloud shadow removal

[0043] In this process, for MODIS images, its quality layer is used to identify cloud and cloud shadow areas; for HLS images, first its Fmask layer is used to remove cloud and cloud shadow areas, and then, considering that the Fmask layer cannot fully display cloud areas, the HOT index method is used to further remove cloud areas. To determine the threshold of the HOT index, the automatic threshold method designed by Liu et al. (Remote Sensing of Envrionment, 2013, 133: 21-37) is adopted.

[0044] 2) 12-day reflectance synthesis based on the maximum coverage strategy

[0045] To suppress the problem that the pixels of the synthesized image may have abrupt changes in values due to different dates, the present invention proposes maximum coverage synthesis to synthesize 12-day images. In this process, for the non-overlapping areas of spatially adjacent images, the cloud coverage of each image within 12 days is first determined; secondly, in the order of increasing cloud coverage range, images are sequentially selected to fill the target area to minimize the number of missing values within the area. For the overlapping areas between spatially adjacent images, the image synthesis strategy of MODIS is adopted by the present invention.

[0046] 3) Production of spatial seamless MODIS images

[0047] The 12-day MODIS synthesized image is filtered by Savitzky-Golay (SG) to generate a spatio-temporally seamless MODIS reflectance product. In the SG filtering process, the polynomial fitting order is set to 4 and the sliding window size is set to 8.

[0048] 4) Spatial Automatic Registration of MODIS Images and HLS Images

[0049] In this process, first, the MODIS images are resampled and clipped to have the same spatial resolution and extent as the HLS images. Then, the algorithm proposed by Gao et al. (Journal of Applied Remote Sensing, 2009, 3(1):033515) is used to perform an affine transformation on the MODIS images to make their spatial positions consistent with the corresponding HLS images.

[0050] 5) Reconstruction of Missing Data in HLS Images

[0051] The missing data in HLS are reconstructed by the spatiotemporal fusion method incorporating spectral autocorrelation (FIRST). First, the partial least squares model is used to determine the relationship between the reflectance of the target date and the reflectance of adjacent dates in the low-resolution image. Then, based on the assumption of local scale invariance of the above relationship, the model is applied to the high-resolution image to reconstruct the missing data of the target date using the data of adjacent dates. In this process, when selecting the HLS images of the adjacent dates before and after the target date, the image with the least data missing has the highest priority, and the original image has a higher priority than the reconstructed image.

[0052] 6) Optimization of the Reconstructed Images

[0053] The reconstructed HLS data are filtered by Savitzky-Golay (SG) filter to further fill in the missing data and correct the outliers. For the SG filter, the polynomial fitting order and the sliding window size are set to 4 and 8, respectively.

[0054] (2) Construction of the Leaf Area Index Inversion Model Based on the Transformer Model: Based on the constructed leaf area index inversion model, using the reflectance data of each band produced in step S1 as the input, the model is run to obtain the high-spatial-resolution seamless leaf area index product.

[0055] Compared with the complete Transformer architecture, the present invention only uses the encoder layer to construct the leaf area index inversion model. The architecture of the leaf area index inversion model proposed by the present invention is as Figure 2As shown. It includes an input layer, a hidden layer, and an output layer. The input layer processes the spectral reflectance vectors of each band at the current time step and the previous two time steps, and the output layer obtains the leaf area index value at the current time step. The hidden layer contains a basic Transformer encoding module, specifically including a multi-head self-attention mechanism module, two layer normalization modules, and a feed-forward network module. The multi-head self-attention mechanism module contains 8 heads, and each head uses 16 nodes to obtain the correlation between different time step information; the feed-forward network module consists of 20 nodes. In addition, feature linear embedding and position encoding are performed before the information is transmitted to the hidden layer. When training the model based on samples, the batch size can be set to 64, the maximum number of training epochs to 150, the initial learning rate to 0.0001, and the optimization function to Adam.

[0056] The present invention proposes a leaf area index inversion model based on Transformer, which overcomes the defect that the existing leaf area index inversion models based on long short-term memory neural network and bidirectional long short-term memory neural network are difficult to fully utilize the temporal image information, and improves the inversion accuracy of the leaf area index. In addition, the present invention proposes a technical process for producing high-resolution seamless leaf area index products, which overcomes the defects that there are a large number of data missing in the existing high-resolution leaf area index products, and the low-resolution products are difficult to meet the application requirements.

[0057] Embodiment

[0058] The following combines specific embodiments to further elaborate on a method for producing high-spatial-resolution seamless leaf area index products disclosed in the present invention.

[0059] In this embodiment, MODIS reflectance images, HLS images, and ESRI land cover data of Jiangsu Province in 2023 were collected, and ground leaf area sampling work was carried out from August 25th to August 29th, 2023. A spatially seamless leaf area index product with a resolution of 30m every 12 days in Jiangsu Province was produced using the method proposed in the present invention. The specific steps are as follows:

[0060] 1) Produce a reflectance product with a resolution of 30m at 12-day intervals based on the 6 steps of producing the aforementioned high-spatial-resolution seamless reflectance image.

[0061] 2) First, Jiangsu Province was divided into 25km×25km grids; then, based on ESRI land cover data, 200 sampling points were allocated in each grid using stratified random sampling according to the proportion of the area occupied by different vegetation types in each grid. It should be noted that in the boundary areas, they do not cover the entire grid. The number of samples allocated to these areas is calculated based on the ratio of their area to the grid area multiplied by 200. A total of 17,822 sample points were selected. The PROSAIL model was used to invert the leaf area index values ​​of these sample points every 12 days using the lookup table method as the simulation data samples. In addition, the ground measured leaf area index data were randomly divided into a modeling measured sample data set and a verification measured sample data set at a ratio of 1:1.

[0062] 3) According to the content of the leaf area index inversion model based on the Transformer model, the leaf area index Transformer inversion model architecture is constructed; secondly, the model is first trained using simulated data samples; then, the last layer of model parameters is discarded, and the model is retrained with the modeling measured sample data set to generate the last layer of parameters, thereby forming a leaf area index inversion model.

[0063] 4) Based on the generated leaf area index inversion model, the reflectivity product generated previously is used to obtain the leaf area index product with a resolution of 30m every 12 days, and the accuracy of the produced product is evaluated using the verified measured sample data set. Among them, the leaf area index data inverted on August 29, 2023 is as follows Figure 3 As shown in a, Figure 3 b is the corresponding land cover data; Figure 4 This is a diagram showing the correspondence between the measured leaf area index and the inverted leaf area index based on the validation sample data set.

[0064] First, from Figure 3 a It can be seen that the leaf area index product produced is spatially seamless and has no missing values. Figure 3 b It can be seen that Jiangsu Province is an important grain-producing area in China, with a high proportion of cultivated land, mainly distributed in the north and central regions; the southern region mainly includes cities and industrial areas, and the proportion of non-vegetated areas is relatively high. Figure 3 It can be seen that the corresponding Figure 3 b. The leaf area index estimation results are consistent with the spatial distribution of vegetation and non-vegetation areas. The leaf area index in the vegetation area is high, and the leaf area index in the non-vegetation area is 0. In addition, since this date coincides with the grain filling period of corn and the heading-flowering period of rice, the crops are in a vigorous growth stage. Figure 3 a shows that the leaf area index level of cultivated land is higher than that of forest. The leaf area index values ​​of different land cover types are cultivated land > forest land > grassland > wetland vegetation, which is logically consistent with the actual situation. Figure 4It can be seen that, based on the verified measured sample data set, there is a good correspondence between the actually measured leaf area index and the leaf area index inverted based on the method of the present invention. The average relative error between the two is 14.80%, meeting the requirement of less than 15% uncertainty in leaf area index inversion in fields such as global change research, and being higher than the accuracy of existing leaf area index products.

[0065] The above embodiments are not limitations on the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by those skilled in the art within the scope of the technical solution of the present invention also fall within the protection scope of the present invention.

Claims

1. A method for producing a seamless leaf area index product with high spatial resolution, characterized in that: The method comprises the following steps: Step S1, high spatial resolution seamless reflectance image production: produce 30m spatial resolution seamless reflectance data every 12 days by using high temporal resolution, low spatial resolution MODIS images to reconstruct the missing data of high spatial resolution, low temporal resolution HLS images; Step S2, constructing a leaf area index inversion model based on the Transformer model: Based on the constructed leaf area index inversion model, taking the reflectance data of each band produced in step S1 as input, running the model to obtain a high spatial resolution seamless leaf area index product.

2. The method for producing a seamless leaf area index product with high spatial resolution according to claim 1, characterized in that: The step S1 comprises the following steps: Step S11, removing clouds and cloud shadows; Step S12: 12-day reflectivity synthesis based on the maximum coverage strategy; Step S13, spatial seamless MODIS image production; Step S14, automatic spatial registration of MODIS image and HLS image; Step S15, HLS image missing data reconstruction; Step S16: Optimize the reconstructed HLS image.

3. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S11 is as follows: For MODIS images, use its quality layer to identify cloud and cloud shadow areas; For HLS images, the Fmask layer is first used to remove the cloud and cloud shadow areas. Then, the HOT index method is used to further remove the cloud area, and the automatic threshold method is used to determine the threshold of the HOT index.

4. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S12 is as follows: for the non-overlapping areas of spatially adjacent images, firstly determine the cloud coverage of each image within 12 days; secondly, select images to fill the target area in order from small to large cloud coverage to minimize the number of missing values ​​in the area, thereby forming a synthetic image.

5. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S13 is: SG filtering is performed on the 12-day MODIS synthetic image to generate a spatiotemporally seamless MODIS reflectance product, and in the SG filtering process, the polynomial fitting order is set to 4 and the sliding window size is set to 8.

6. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S14 is: first, the MODIS image is resampled and sheared to make its spatial resolution and range the same as those of the HLS image; then, the MODIS image is affine transformed using an automatic registration and orthorectification algorithm to make its spatial position consistent with that of the corresponding HLS image.

7. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S15 is as follows: first, a partial least squares model is used to determine the relationship between the reflectance of the target date and the reflectance of the adjacent dates in the low spatial resolution image; then, based on the assumption that the above relationship is invariant at a local scale, the model is applied to the high spatial resolution image, and the missing data of the target date is reconstructed through the data of the adjacent dates; in this process, when selecting HLS images of adjacent dates before and after the target date, the image with the least data missing has the highest priority, and the original image has a higher priority than the reconstructed image.

8. The method for producing a seamless leaf area index product with high spatial resolution according to claim 2, characterized in that: The specific process of step S16 is: performing SG filtering on the reconstructed HLS data to further fill in data gaps and correct outliers.

9. The method for producing a seamless leaf area index product with high spatial resolution according to claim 1, characterized in that: In step S2, the leaf area index inversion model includes: An input layer: The input layer processes the spectral reflectance vectors for each band at the current time step and the previous two time steps; An output layer: The output layer obtains the leaf area index value of the current time step; One hidden layer: The hidden layer contains a basic Transformer encoding module, which specifically includes a multi-head self-attention mechanism module, two layer normalization modules and a feedforward network module.

10. The method for producing a seamless leaf area index product with high spatial resolution according to claim 9, characterized in that: The multi-head self-attention mechanism module contains 8 heads, each of which uses 16 nodes to obtain the correlation between information at different time steps; the feedforward network module consists of 20 nodes; Perform feature linear embedding and position encoding before transferring information to hidden layers; When training the model based on samples, the batch size is set to 64, the maximum training cycle is set to 150, the initial learning rate is set to 0.0001, and the optimization function is set to Adam.

Images