Spatial-temporal data fusion method and system combining STARFM algorithm and AI neural network technology

Abstract

The invention discloses a spatio-temporal data fusion method combining an STARFM algorithm and an AI neural network technology, and is applied to the field of remote sensing image processing. According to the method, based on the similarity between low-spatial-resolution remote sensing images and high-spatial-resolution remote sensing images and the concept of continuity of geographic space phenomena at different time points, an STARFM algorithm is utilized to perform preliminary data fusion, and then an AI neural network technology is introduced to optimize a preliminary fusion result, so that the optimal fusion result is obtained. The problems of image quality reduction, blurring or distortion and the like possibly occurring in the fusion process of a traditional algorithm are solved. Fusion processing is carried out by taking MODIS data with high time resolution and low spatial resolution and Landsat satellite data with low time resolution and high spatial resolution as examples, and the result shows that the result quality of a traditional spatio-temporal data fusion method based on the pixel level can be improved, and the fusion result has good stability.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image processing, and specifically relates to a method of integrating the traditional spatio-temporal data fusion algorithm STARFM with artificial intelligence technology to generate satellite images with higher quality and both high temporal resolution and high spatial resolution. Background Art

[0002] With the development of satellite remote sensing technology, an unprecedented amount of high-resolution and multi-spectral image data is emerging continuously. These data are of crucial significance for many application fields such as environmental monitoring, resource management, and disaster assessment. However, different sensors have differences in temporal resolution and spatial resolution, which poses challenges to the application of the data. For example, some sensors can provide high-spatial-resolution images but are limited by the revisit cycle; while other sensors may have higher temporal resolution but lack spatial details. Spatio-temporal data fusion is an important technology for overcoming these problems.

[0003] STARFM is a classic model for fusing remote sensing images with different resolutions, aiming to generate synthetic images with high spatio-temporal resolution by combining low-frequency images with high spatial resolution and high-frequency images with low spatial resolution. The basic principle of STARFM is to utilize the characteristic that the reflectance change of ground objects at the same location within adjacent time periods is relatively small, namely the so-called "similarity hypothesis". Moreover, the algorithm is relatively simple to implement, does not require a large number of parameter adjustments, can adapt to different ground object types, and shows good generality when processing data from multiple sensors. However, the actual surface reflectance change is often non-linear, and the STARFM algorithm based on linear regression to establish the conversion function may lead to insufficient accuracy in some cases, especially when dealing with complex environments and lacking sufficient historical data.

[0004] In recent years, with the rapid development of machine learning, especially in the field of deep learning, AI technology, especially neural network technology, has gradually become an effective tool for dealing with complex image tasks. Compared with traditional statistical or machine learning models, neural networks have stronger learning ability and generalization performance, and are particularly suitable for dealing with non-linear and high-dimensional data. In remote sensing image analysis, convolutional neural networks, recurrent neural networks and their variants such as long short-term memory networks have been widely used, and they have shown excellent effects in feature extraction, classification, object detection, etc. Summary of the Invention

[0005] In view of the limitations of the STARFM algorithm in dealing with complex real-world scenarios and its poor performance in the case of insufficient historical data, the present invention provides a spatio-temporal data fusion method that combines the STARFM algorithm with AI neural network technology. With the powerful learning ability of the AI neural network, an accurate prediction model can be established through a large number of training samples, providing new ideas and effective technical means to solve the above problems. Moreover, the fusion of the two technologies can create a spatio-temporal data fusion framework with higher intelligence, better efficiency, and significantly enhanced adaptability.

[0006] According to one aspect of the specification of the present invention, there is provided a spatio-temporal data fusion method that combines the STARFM algorithm with AI neural network technology, including: Obtain preprocessed satellite images with low spatial resolution and high temporal resolution and satellite image data with high spatial resolution and low temporal resolution within the study area; Input the preprocessed satellite image data into the spatio-temporal data fusion algorithm STARFM to obtain a preliminary data fusion result; Input the preliminary data fusion result into the trained neural network to output the final data fusion result.

[0007] As a further technical solution, the preprocessing includes: For the two types of satellite image data, perform geometric correction, radiometric correction, cloud removal, noise reduction, and data screening in sequence. Then, taking the spatial coordinates of the data points of the satellite image with high spatial resolution and low temporal resolution as a reference, upsample the satellite image data with low spatial resolution and high temporal resolution by means of linear interpolation.

[0008] As a further technical solution, the two types of satellite image data have the same observation date and bands.

[0009] As a further technical solution, the training of the neural network includes: Construct a data set, where the data set includes image pairs of the two types of satellite images and the prediction results of the STARFM algorithm; Construct a convolutional neural network, using the grids in the prediction results of the STARFM algorithm as the input of the convolutional neural network, and using the corresponding grids in the satellite image with high spatial resolution and low temporal resolution as the labels of the convolutional neural network; Use the constructed data set for training and verification, and output the trained convolutional neural network.

[0010] As a further technical solution, constructing the data set further includes: Select a set of image pairs, input the satellite images with low spatial resolution and high temporal resolution and other image pairs into the STARFM algorithm to obtain the prediction results for the corresponding dates; Divide the prediction results and the satellite images with high spatial resolution and low temporal resolution in this set of image pairs into multiple small grids in the same way. Each grid in the prediction results is used as the input of the convolutional neural network, and the corresponding grids in the satellite images with high spatial resolution and low temporal resolution are used as the labels during the training of the convolutional neural network; Repeat the above method to obtain multiple sets of data, and construct the training set and validation set required for the training of the convolutional neural network according to the obtained data.

[0011] As a further technical solution, input the preliminary data fusion result into the trained neural network to output the final data fusion result, including: Perform grid processing on the preliminary results of the target date predicted by the spatio-temporal data fusion algorithm STARFM, input each grid into the trained convolutional neural network for optimization, and splice the prediction results of the convolutional neural network to obtain the final data fusion result.

[0012] According to one aspect of the specification of the present invention, there is provided a spatio-temporal data fusion system combining the STARFM algorithm and AI neural network technology, including: A data acquisition module for acquiring preprocessed satellite image data with low spatial resolution and high temporal resolution and satellite image data with high spatial resolution and low temporal resolution within the research area; A preliminary fusion module for inputting the preprocessed satellite image data into the spatio-temporal data fusion algorithm STARFM to obtain a preliminary data fusion result; A fusion optimization module for inputting the preliminary data fusion result into the trained neural network to output the final data fusion result.

[0013] According to one aspect of the specification of the present invention, there is provided a spatio-temporal data fusion device combining the STARFM algorithm and AI neural network technology, including a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology.

[0014] According to one aspect of the specification of the present invention, there is provided a non-transitory computer-readable storage medium storing computer instructions, and the computer instructions cause the computer to execute the steps of the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention provides a spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology. With the powerful learning ability of the AI neural network, an accurate prediction model is established through a large number of training samples, and the prediction model is used for fusion optimization. By fusing satellite images with low temporal resolution and high spatial resolution (tens of meters) and satellite images with high temporal resolution and low spatial resolution (hundreds of meters), a fused image with high spatio-temporal resolution can be obtained. Compared with the existing fusion methods, it has the advantages of higher intelligence, better efficiency and significantly enhanced adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following briefly introduces the drawings used in the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0017] Figure 1 It is a flowchart of the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology provided by the embodiment of the present invention.

[0018] Figure 2 It is a schematic diagram of building a convolutional neural network provided by the embodiment of the present invention.

[0019] Figure 3 It is a schematic diagram of the comparison of the prediction effects of the method provided by the embodiment of the present invention and the traditional method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The present invention provides a spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology, which is used to fuse satellite images with low temporal resolution and high spatial resolution (tens of meters) and satellite images with high temporal resolution and low spatial resolution (hundreds of meters) to obtain a fused image with high spatio-temporal resolution. The premise of this method is that the input image data has been geometrically corrected, radiometrically corrected and cloud-removed.

[0021] The spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology described in the present invention includes the following specific steps: Step 1, obtain MODIS and Landsat image data of the study area. Here, the obtained satellite image data with low spatial resolution and high temporal resolution and satellite image data with high spatial resolution and low temporal resolution in the study area are geometrically corrected, radiometrically corrected and cloud-removed. In addition to having overlapping observation dates, the two types of image data should also have the same bands.

[0022] Step 2, preprocess the above satellite image data. This preprocessing reduces the noise signals in the satellite image data with low spatial resolution and high temporal resolution through noise reduction processing. The image data of different bands all need to be subjected to noise reduction processing. This preprocessing filters the corresponding satellite images with low spatial resolution and high temporal resolution and satellite image data with high spatial resolution and low temporal resolution according to the research area and date. This preprocessing also takes the spatial coordinates of the data points of the satellite images with high spatial resolution and low temporal resolution as a reference, and upsamples the satellite image data with low spatial resolution and high temporal resolution by linear interpolation to align the two different satellite image data.

[0023] Step 3, input the preprocessed remote sensing image data into the spatio-temporal data fusion algorithm STARFM to obtain a preliminary data fusion result.

[0024] Step 4, construct an AI neural network to further optimize the image result obtained by the preliminary fusion to obtain the final data fusion result.

[0025] The specific steps of Step 1 include: Step 1.1, obtain the image data in the MCD43A4 NBAR surface reflectance product of the MODIS satellite. This data includes visible light bands (blue, green, red), near-infrared band and 2 short-wave infrared bands, and its spatial resolution is 500 m; Step 1.2, obtain the image data in the Landsat 8 OLI Level-2 Collection 2 surface reflectance dataset. This data includes visible light bands (blue, green, red), near-infrared band and 2 short-wave infrared bands, and its spatial resolution is 30 m.

[0026] The specific steps of Step 2 include: Step 2.1, use the fast non-local means denoising function in the open-source computer vision library OpenCv to perform noise reduction processing on each band data of the MODIS satellite image. This method reduces the noise signal by averaging the globally similar parts of the picture and can achieve a high-quality noise reduction effect; Step 2.2, filter the corresponding MODIS satellite and Landsat satellite image data according to the research area and date; Step 2.3, take the spatial coordinates of the data points of the Landsat satellite image as a reference, and upsample the MODIS satellite image data by linear interpolation to align the two different satellite image data.

[0027] The specific steps of Step 3 include: In each input of the spatio-temporal data fusion algorithm STARFM, at least one pair of MODIS and Landsat image data with the same observation date, and MODIS image data of the prediction date are input into the STARFM algorithm to fuse and obtain the satellite image with preliminary high spatial resolution of the prediction date . For the prediction of any pixel value, it is not only related to the pixel value at that position in the input satellite image pair, but also related to the values of its surrounding pixels: .

[0028] In the formula: F and C respectively represent the satellite images of Landsat with high spatial resolution and low temporal resolution, and MODIS with low spatial resolution and high temporal resolution; is the prediction date; is the time of the existing image pair ; is the pixel position; n is the number of image pairs; w is the size of the search window; the weight determines the contribution of the predicted value of each pixel in the search window to the predicted value of the central pixel, and it is related to the spectral distance , the time distance and the spatial distance . In the calculation of the spatial distance , A is a constant that defines the relative importance of the spatial distance to the spectral and time distances. In addition, not every pixel point in the search window participates in the calculation. Only when a pixel point in the search window is similar enough to the pixel point at the center of the window, will it participate in the calculation. The STARFM algorithm contains 3 different criteria to judge whether pixel points are similar:

[0029] In the formula: is the error of the Landsat satellite image data; is the error of the MODIS satellite image data; is the standard deviation of the pixel values of the Landsat satellite image within the window; is the standard deviation of the pixel values of the satellite image with high spatial resolution and low temporal resolution within the window; m is the number of pixel categories (such as water body or land, etc.).

[0030] Step 4 specifically includes: Step 4.1, Use MODIS and Landsat satellite images other than the predicted date to construct the dataset required for the neural network. First, select a set of image pairs. Input the MODIS satellite image in this set and other image pairs into the STARFM algorithm to obtain the predicted results for the corresponding date. Then, divide the predicted results and the Landsat satellite image in this set of image pairs into multiple smaller grids in the same way. The grids divided can have pixel overlaps along the length or width direction. Each grid in the predicted results serves as the input data for the neural network, while the corresponding grid in the Landsat satellite image acts as the label or target during the training of the neural network. Repeat the above method to obtain multiple sets of data, and then construct the training set and validation set required for neural network training based on these obtained data; Step 4.2, Use pytorch to build a convolutional neural network and train this neural network with the grid data obtained above; Step 4.3, Perform grid processing on the preliminary results of the target date predicted by the STARFM algorithm, then input each grid into the trained neural network for optimization, and then splice the predicted results of all neural networks to obtain the final high-spatial-resolution satellite image.

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention. In addition, the technical features in each embodiment or individual embodiment provided by the present invention can be combined with each other arbitrarily to form a new technical solution. This combination is not restricted by the order of steps and / or the structural composition mode, but must be based on the ability of those of ordinary skill in the art to implement. When the combination of technical solutions results in contradictions or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the protection scope required by the present invention.

[0032] Figure 1Schematically shows the main implementation idea of the present invention. The research and demonstration area of the present invention is located in the high-altitude area in the northwest of Nepal. The satellite image data used includes research areas with three different snow cover change situations, namely research areas with three change types of snow melting, snowfall, and snow with little change. First, perform geometric correction and radiometric correction on the Landsat images of the obtained research areas, and perform geometric correction, radiometric correction, and cloud removal on the MODIS images. Then, use the fast non-local means denoising function in the open-source computer vision library OpenCv to denoise each band data of the MODIS satellite images. After that, pair the MODIS and Landsat images with the same observation date in each research area, and then upsample the MODIS satellite image data by linear interpolation so that the two satellite images focus on the same regional location. Subsequently, input the MODIS satellite image data of the target date and the image pairs of other dates into the STARFM algorithm to obtain preliminary prediction results. Then, construct a data set based on the MODIS and Landsat image pairs except for the prediction date. The method of constructing neural network training data is to first select a set of image pairs, input the MODIS satellite image in it and the image pairs of other observation dates into the STARFM algorithm to obtain the prediction results corresponding to the dates, and then divide the prediction results and the Landsat satellite images in this set of image pairs into multiple smaller grids in the same way. The size of the divided grids is 32×32, and adjacent grids will overlap by 20 pixel units along the length or width direction of the image. The grids in the prediction results are used as the input of the neural network, while the corresponding grids in the Landsat satellite images are used as the labels or targets of the neural network. Repeat the above method to obtain multiple sets of data. Then use pytorch to build a convolutional neural network, use 80% of the data obtained above as training data, and the remaining 20% as validation data to train the neural network. Finally, use this neural network to optimize the image results of the target date predicted by the STARFM algorithm. Before inputting into the neural network, perform grid segmentation on the preliminary prediction results generated by the STARFM, and then input these gridded data into the trained neural network model in sequence for optimization. After the optimization is completed, splice all the output results of the neural network into a complete image to obtain a satellite image with significantly improved spatial resolution and higher prediction accuracy. During the splicing process, an overlapping area is designed between adjacent grids to ensure smooth transition. For the pixel values in the overlapping part, the final value is calculated by taking the average according to the number of overlaps.

[0033] Figure 2Shows the convolutional neural network architecture constructed in the embodiments of the present invention. The input of the network is a matrix of size 32×32×6, where 32 is the length and width of the matrix, and the depth of the matrix is 6, corresponding to 6 different band data of satellite images. During the training process, each time an image sample matrix with a batch size of 128 is used as the input for training, and the number of training epochs is 400 times. Then there are 3 convolutional layers to process the input data, and different convolutional layers are connected by the non-linear activation function Relu. The initial learning rate of the first 2 convolutional layers is , and the initial learning rate of the third convolutional layer is . In addition, the kernel size of the first convolutional layer is 9×9, the number of input channels is 6, the number of output channels is 64, and the size of the bias vector is 64; the kernel size of the second convolutional layer is 5×5, the number of input channels is 64, the number of output channels is 32, and the size of the bias vector is 32; the kernel size of the third convolutional layer is 5×5, the number of input channels is 32, the number of output channels is 6, and the size of the bias vector is 6. The data matrix finally output by the neural network has the same shape as the input data matrix. In addition, the optimizer of this neural network is Adam, and the loss function is MSELoss. During the training process, PSNR is used to evaluate the quality of the image calculated by the neural network. PSNR is a common metric in the field of image synthesis. Generally, the larger its value, the closer the calculated image is to the target image.

[0034] Figure 3 Is a comparison of the prediction results between the algorithm provided by the present invention and the traditional spatio-temporal fusion algorithm STARFM. In areas less affected by snowfall and during snowmelt, both the traditional algorithm and the algorithm proposed by the present invention exhibit relatively ideal prediction performance. However, in terms of detail presentation, the results predicted by the algorithm provided by the present invention are better than those of the traditional algorithm. Compared with the algorithm provided by the present invention, there are more outliers or noises in the prediction results of the STARFM algorithm. In the research area of more complex snowfall scenarios, the differences in the prediction results of the two algorithms are more significant. The prediction results of the STARFM algorithm have the problem of significantly insufficient spatial resolution, resulting in a large loss of image detail information; while the prediction results of the algorithm provided by the present invention can highly restore the real image information. From the final effect, the algorithm provided by the present invention is not only better than the traditional algorithm, but also shows better applicability in dealing with various complex scenarios.

[0035] The spatio-temporal fusion algorithm proposed by the present invention will be tested below. The statistical results in Table 1 are the average values of the statistical results of the study areas of three change types. As can be seen from Table 1, the spatio-temporal fusion method proposed by the present invention exhibits excellent performance in all three statistical indicators. Regardless of which band, the error between the fused image and the real image is less than 0.1, showing high accuracy. Except that the correlation coefficient of the short-wave infrared band 2 is 0.7749, the correlation coefficients of the prediction results of most bands reach above 0.8, and the correlation coefficients of all bands are higher than 0.75, indicating a strong correlation between the fused image and the real image.

[0036] Table 1 Statistical evaluation results of the average difference, root mean square error, and correlation coefficient between the prediction results of the spatio-temporal fusion algorithm provided by the present invention and the real image

[0037] The implementation basis of each embodiment of the present invention is achieved through programmed processing by a device with processor functions. Therefore, in engineering practice, the technical solutions and functions of each embodiment of the present invention are encapsulated into various modules. Based on this actual situation, on the basis of the above embodiments, an embodiment of the present invention provides a spatio-temporal data fusion system combining the STARFM algorithm and AI neural network technology, which is used to execute the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology in the above method embodiments.

[0038] The system includes: a data acquisition module for acquiring preprocessed satellite images with low spatial resolution and high temporal resolution and satellite image data with high spatial resolution and low temporal resolution in the study area; a preliminary fusion module for inputting the preprocessed satellite image data into the spatio-temporal data fusion algorithm STARFM to obtain a preliminary data fusion result; and a fusion optimization module for inputting the preliminary data fusion result into the trained neural network to output the final data fusion result.

[0039] The spatio-temporal data fusion system combining the STARFM algorithm and AI neural network technology provided by the embodiment of the present invention aims at the limitations of the STARFM algorithm in dealing with complex real-world scenarios and its poor performance in the case of insufficient historical data. By using the aforementioned several modules and relying on the powerful learning ability of the AI neural network, an accurate prediction model can be established through a large number of training samples, providing new ideas and effective technical means to solve the above problems. Moreover, the fusion of the two technologies can create a spatio-temporal data fusion framework with higher intelligence, better efficiency, and significantly enhanced adaptability.

[0040] It should be noted that the system embodiments provided by the present invention are used not only to implement the methods in the above method embodiments, but also to implement the methods in other method embodiments provided by the present invention. The difference lies only in setting corresponding functional modules, and the principle is basically the same as that of the above system embodiments provided by the present invention. As long as those skilled in the art, on the basis of the above system embodiments, refer to the specific technical solutions in other method embodiments, obtain corresponding technical means by combining technical features, and the technical solutions composed of these technical means, and on the premise of ensuring the practicability of the technical solutions, improve the modules in the above system embodiments to obtain corresponding system-like embodiments for implementing the methods in other method-like embodiments.

[0041] Based on the same inventive concept as the foregoing embodiments, the embodiments of the present invention further provide a spatio-temporal data fusion device combining the STARFM algorithm and AI neural network technology, including a memory and a processor. The memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology.

[0042] Based on the same inventive concept as the foregoing embodiments, the embodiments of the present invention further provide a non-transitory computer-readable storage medium. The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the steps of the spatio-temporal data fusion method combining the STARFM algorithm and AI neural network technology.

[0043] In summary of the above embodiments, the present invention first performs preliminary data fusion using the STARFM algorithm based on the similarity between low-spatial-resolution and high-spatial-resolution remote sensing images and the concept of the continuity of geospatial phenomena at different time points, and then introduces AI neural network technology to optimize the initial fusion result, solving problems such as image quality degradation, blurring, or distortion that may occur in the fusion process of traditional algorithms. Taking high-temporal-resolution, low-spatial-resolution MODIS data and low-temporal-resolution, high-spatial-resolution Landsat satellite data as examples for fusion processing, the results show that the present invention can improve the result quality of traditional pixel-level spatio-temporal data fusion methods, and the fusion results have good stability.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A spatiotemporal data fusion method combining STARFM algorithm and AI neural network technology, characterized in that: include: Obtain pre-processed low spatial resolution, high temporal resolution satellite images and high spatial resolution, low temporal resolution satellite image data in the study area; The preprocessed satellite image data is input into the spatiotemporal data fusion algorithm STARFM to obtain preliminary data fusion results; The preliminary data fusion results are input into the trained neural network and the final data fusion results are output.

2. The spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology according to claim 1, characterized in that: The preprocessing comprises: For the two types of satellite image data, geometric correction, radiation correction, cloud removal, noise reduction and data screening are performed in sequence. Then, the satellite image data with low spatial resolution and high temporal resolution are upsampled by linear interpolation with the spatial coordinates of the data points of the satellite image with high spatial resolution and low temporal resolution as a reference.

3. The spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology according to claim 1, characterized in that: The two satellite image data have the same observation date and band.

4. The spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology according to claim 1, characterized in that: The training of the neural network comprises: Constructing a data set, wherein the data set includes image pairs of two satellite images and prediction results of a STARFM algorithm; Construct a convolutional neural network, use the grids in the STARFM algorithm prediction results as the input of the convolutional neural network, and use the corresponding grids in the high spatial resolution and low temporal resolution satellite images as the labels of the convolutional neural network; Use the constructed dataset for training and verification, and output the trained convolutional neural network.

5. The spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology according to claim 4, characterized in that: Building a dataset also includes: Select a set of image pairs, and input the satellite images with low spatial resolution and high temporal resolution into the STARFM algorithm along with other image pairs to obtain the prediction results for the corresponding date; The prediction result and the high spatial resolution and low temporal resolution satellite images in the image pair are divided into multiple small grids in the same way, where each grid in the prediction result is used as the input of the convolutional neural network, and the corresponding grid in the high spatial resolution and low temporal resolution satellite images is used as the label for convolutional neural network training; Repeat the above method to obtain multiple sets of data, and construct the training set and validation set required for convolutional neural network training based on the obtained data.

6. The spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology according to claim 1, characterized in that: Input the preliminary data fusion results into the trained neural network and output the final data fusion results, including: The preliminary results of the target date predicted by the spatiotemporal data fusion algorithm STARFM are gridded, and each grid is input into the trained convolutional neural network for optimization. The prediction results of the convolutional neural network are spliced ​​to obtain the final data fusion result.

7. The spatiotemporal data fusion system combining STARFM algorithm and AI neural network technology is characterized by: include: The data acquisition module is used to obtain the pre-processed low spatial resolution, high temporal resolution satellite images and high spatial resolution, low temporal resolution satellite image data in the study area; The preliminary fusion module is used to input the preprocessed satellite image data into the spatiotemporal data fusion algorithm STARFM to obtain preliminary data fusion results; The fusion optimization module is used to input the preliminary data fusion results into the trained neural network and output the final data fusion results.

8. A spatiotemporal data fusion device combining STARFM algorithm and AI neural network technology, characterized in that: It includes a memory and a processor, the memory stores program instructions executed by the processor, and the processor calls the program instructions to execute the steps of the spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which enable the computer to execute the steps of the spatiotemporal data fusion method combining the STARFM algorithm and the AI ​​neural network technology as described in any one of claims 1 to 6.

Images