A method, system, equipment, and medium for downscaling surface temperature remote sensing products.

Abstract

This invention discloses a method, system, device, and medium for downscaling surface temperature remote sensing products. The method includes the following steps: obtaining T at the target time T. 前 T 后 Raw low spatial resolution land surface temperature data and raw high spatial resolution land surface temperature data; obtain T 前 T 后 The invention obtains low spatial resolution land surface temperature data after sampling and the ID value of each sampled pixel corresponding to the original pixel; it then obtains the pixel-by-pixel rate of change of the low spatial resolution land surface temperature data; uses a pre-trained dynamic multilayer sensing network DyNet to predict the category-by-category rate of change of the high spatial resolution land surface temperature data; and uses a pre-trained weighted sensing network WNet to predict the land surface temperature at the target time T. The technical solution provided by this invention can solve the technical problems that existing linear unmixing models cannot effectively estimate land surface temperature and that existing weighting functions fail near the rate of change of -1.

Description

Technical Field

[0001] This invention belongs to the field of satellite remote sensing Earth observation technology, and specifically relates to a method, system, equipment and medium for downscaling Earth surface temperature remote sensing products. Background Technology

[0002] Land surface temperature (LST) is a crucial component of the Earth's surface energy and water balance, providing information on the spatiotemporal variations of surface equilibrium from local to global levels. Currently, LST is widely used for evapotranspiration estimation, urban heat island monitoring, local climate change monitoring, vegetation monitoring, agricultural monitoring, and forest fire monitoring. However, due to the relatively low energy of surface thermal radiation and the cost and hardware limitations of satellite sensors, current remote sensing data acquired by single satellite sensors suffers from a "spatiotemporal contradiction"—that is, the acquired LST data cannot simultaneously achieve high spatial and temporal resolution, significantly limiting the application needs of LST.

[0003] Regarding the question of how to obtain high spatiotemporal LST, many scholars have proposed different LST downscaling methods, including:

[0004] 1) According to algorithm theory, it is divided into downscaling methods based on scale factor and downscaling methods based on spatiotemporal fusion; among them, downscaling methods based on spatiotemporal fusion can more effectively reduce the model's dependence on auxiliary data, and can achieve downscaling of surface temperature in a low-cost, convenient and efficient manner.

[0005] 2) Based on their implementation principles, methods can be categorized into weight function-based methods, learning-based methods, and unmixing-based methods. Among these, for LST data with strong temporal volatility and limited spectral features, the parameters of weight function-based methods become more sensitive, which also affects the effectiveness of training data in learning-based methods, resulting in a generally lower fusion effect for these two methods. Unmixing-based methods, for LST data that changes over time, can improve the fusion effect of LST data from different time phases by combining the assumption that the ratio of changes in different sensors is equal in regions where the LST change trends are the same.

[0006] Currently, models based on unmixing strategies are the dominant approach. These models can effectively reduce the impact of changes in the land surface environment on prediction accuracy. Compared to learning-based models, these methods have lower learning costs and higher time efficiency. However, existing models still have the following two drawbacks:

[0007] 1) Linear unmixing models cannot effectively balance the number of homogeneous units and unmixing accuracy. Specifically, interpretable, high spatial resolution surface problem information requires a large number of homogeneous surface units. However, due to the insufficient solution capability of current linear unmixing models, when the number of homogeneous units exceeds a certain range, the linear unmixing models cannot solve the problem.

[0008] 2) Due to inherent mathematical limitations, the current weighting function is unpredictable in the range where the rate of change of surface temperature is -1, leading to significant errors in the calculation of surface temperature.

[0009] In summary, existing LST downscaling methods cannot effectively downscale land surface temperature in scenarios with rapid changes in land surface temperature. They suffer from technical defects such as insufficient accuracy of daily high spatial resolution land surface temperature products and inability to provide accurate high spatiotemporal resolution land surface temperature products. Summary of the Invention

[0010] The purpose of this invention is to provide a method, system, device and medium for downscaling land surface temperature remote sensing products, so as to solve the technical problems that existing linear unmixed models cannot effectively estimate land surface temperature and that existing weighting functions fail near the rate of change of -1.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] The first aspect of this invention provides a method for downscaling surface temperature remote sensing products, comprising the following steps:

[0013] Obtain T at target time T 前 T 后 The original low spatial resolution land surface temperature data and the original high spatial resolution land surface temperature data; among which, T 前 Let T be any time before the target time T where data is available. 后 For any time after the target time T that has data;

[0014] T 前 T 后 The original low spatial resolution land surface temperature data was resampled to the same spatial resolution as the original high spatial resolution land surface temperature data to obtain T. 前 T 后 The sampled low spatial resolution surface temperature data and the ID value of each sampled pixel corresponding to the original pixel;

[0015] Based on T 前 T 后 The sampled low spatial resolution surface temperature data, according to T 前 T, T后 The pixel-by-pixel rate of change was calculated at three time points to obtain the pixel-by-pixel rate of change of low spatial resolution land surface temperature data.

[0016] Using the pixel-by-pixel change rate of low spatial resolution land surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the pre-acquired ID of homogeneous land surface units as input, a pre-trained dynamic multilayer sensing network DyNet is used to predict the category-by-category change rate of high spatial resolution land surface temperature data; T 前 T 后 The absolute value of the difference between the original high spatial resolution land surface temperature data and the category-wise change rate of the obtained high spatial resolution land surface temperature data are used as inputs. The land surface temperature at the target time T is predicted by a pre-trained weighted perception network WNet.

[0017] A further improvement of the present invention is that the step of obtaining the pre-trained dynamic multilayer perceptron DyNet includes:

[0018] Obtain a first training sample set; wherein each training sample in the first training sample set includes, as input, the pixel-wise change rate of low spatial resolution surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the ID of the surface homogeneous unit, and as labels, the category-wise change rate of high spatial resolution surface temperature data; wherein the category-wise change rate of high spatial resolution surface temperature is based on T 前 T, T 后 The original high spatial resolution surface temperature data, according to T 前 T, T 后 The calculation is performed pixel-by-pixel at three time points, and the mean value of each category is obtained by using the surface homogeneous unit ID; the surface homogeneous unit ID is obtained by clustering time series surface temperature data or land cover type data.

[0019] The pre-built dynamic multilayer sensing network is trained using the MSE Loss function based on the first training sample set. After reaching the preset convergence condition, the pre-trained dynamic multilayer sensing network is obtained.

[0020] A further improvement of the present invention is that the pre-constructed dynamic multilayer perceptron DyNet includes: a temporary de-layer for shielding some input neurons, multiple fully connected layers and a ReLU activation function layer connected in series after the temporary de-layer, and a temporary de-layer for shielding some output neurons.

[0021] A further improvement of the present invention is that the step of obtaining the pre-trained weight-aware network WNet includes:

[0022] Obtain a second training sample set; wherein each training sample in the second training sample set includes high spatial resolution land surface temperature data T as input. 前 and T 后 The absolute value of the difference and the category-wise change rate of the high spatial resolution surface temperature data in the first training sample set, and the high spatial resolution surface temperature of the target date as the label;

[0023] The pre-built weight-aware network is trained using the MSE Loss function based on the second training sample set. After reaching the preset convergence condition, the pre-trained weight-aware network is obtained.

[0024] A further improvement of the present invention is that the pre-built weight-aware network includes: a data mapping layer for input data, and multiple fully connected layers and a ReLU activation function layer connected in series after the data mapping layer.

[0025] A further improvement of the present invention is that, in the pre-built weight-aware network WNet, the data mapping layer performs the following steps for data mapping:

[0026] The label truth data is transformed using the expression ΔLST=LST-LST. 前 In the formula, LST is the actual surface temperature LST at the target time T. 前 For T 前 The surface temperature LST at time t, ΔLST is the temperature difference;

[0027] When T 后 The surface temperature at time T is less than T 前 When the surface temperature value is given at a given time, the transformation expression for the category-by-category rate of change of the high spatial resolution image is: In the formula, For high spatial resolution images, the category-specific rate of change. The transformed rate of change by category;

[0028] When T 后 The surface temperature at time T is greater than T 前 When the surface temperature value is given at a given time, the transformation expression for the category-by-category rate of change of the high spatial resolution image is:

[0029] The expression for the predicted land surface temperature output by the weighted sensing network is as follows: The LST is the land surface temperature predicted by the model.

[0030] A further improvement of the present invention is that the steps of obtaining the first training sample set and obtaining the second training sample set specifically include:

[0031] Obtain the original low-spatial-resolution and high-spatial-resolution land surface temperature data of the sample;

[0032] The low spatial resolution surface temperature data is resampled to the same spatial resolution as the high spatial resolution surface temperature data according to the nearest neighbor sampling principle, forming the sampled low spatial resolution surface temperature data and the ID value of each sampled pixel corresponding to the original pixel.

[0033] Using time-series high spatial resolution land surface temperature data as multiple features, clustering is used to generate homogeneous land surface units;

[0034] Following the temporal order, the pixel-by-pixel rate of change of the sampled low-spatial-resolution and high-spatial-resolution land temperature data was calculated at three time points: before the target time, at the target time, and after the target time. This yielded the pixel-by-pixel rate of change for both low-spatial-resolution and high-spatial-resolution land temperature data. The expression for calculating the rate of change is as follows:

[0035] In the formula, ΔL ke (i,j) and ΔL ok (i,j) represent two periods [t] k ,t e ] and [t0,t k The difference in Landsat remote sensing reflectance between [L] and [L]. e (i,j) is the Landsat remote sensing reflectance at the end date te; L k (i,j) is the Landsat remote sensing reflectance of the target date tk; L0(i,j) is the Landsat remote sensing reflectance of the start date t0.; subscript 0 is the start date, k is the target date; e is the end date; (i,j) represents the pixel in the i-th row and j-th column;

[0036] The average change rate of the generated high spatial resolution surface temperature data per pixel is calculated using the generated homogeneous surface units, thus obtaining the category-specific change rate of the high spatial resolution surface temperature data.

[0037] Construct a first training set; wherein the input features of the first training set are the pixel-wise change rate of low spatial resolution land surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the homogeneous unit of the land surface, and the output label is the category-wise change rate of high spatial resolution land surface temperature data.

[0038] Construct a second QR code training set; wherein, the input features of the second training set are the absolute value of the difference between the high spatial resolution surface temperature data before and after the target time and the category-by-category change rate of the high spatial resolution surface temperature data, and the output label is the surface temperature at the target time.

[0039] A second aspect of the present invention provides a surface temperature remote sensing product downscaling system, comprising:

[0040] The data acquisition module is used to obtain the T of the target time T. 前 T 后 The original low spatial resolution land surface temperature data and the original high spatial resolution land surface temperature data; among which, T 前 Let T be any time before the target time T where data is available. 后 For any time after the target time T that has data;

[0041] The data processing module is used to process T 前 T 后 The original low spatial resolution land surface temperature data was resampled to the same spatial resolution as the original high spatial resolution land surface temperature data to obtain T. 前 T 后 The sampled low spatial resolution surface temperature data and the ID value of each sampled pixel corresponding to the original pixel;

[0042] The rate of change acquisition module is based on T. 前 T 后 The sampled low spatial resolution surface temperature data, according to T 前 T, T 后 The pixel-by-pixel rate of change was calculated at three time points to obtain the pixel-by-pixel rate of change of low spatial resolution land surface temperature data.

[0043] The prediction module takes the pixel-by-pixel rate of change of low spatial resolution land surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the pre-acquired ID of homogeneous land surface units as input, and uses a pre-trained dynamic multilayer sensing network DyNet to predict the category-by-category rate of change of high spatial resolution land surface temperature data; and uses T... 前 T 后 The absolute value of the difference between the original high spatial resolution land surface temperature data and the category-wise change rate of the obtained high spatial resolution land surface temperature data are used as inputs. The land surface temperature at the target time T is predicted by a pre-trained weighted perception network WNet.

[0044] A third aspect of the present invention provides an electronic device comprising:

[0045] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor to enable the at least one processor to perform a method for downscaling surface temperature remote sensing products as described in any of the present invention.

[0046] The fourth aspect of the present invention provides a computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the surface temperature remote sensing product downscaling method described in any one of the present invention.

[0047] Compared with the prior art, the present invention has the following beneficial effects:

[0048] This invention leverages the strong nonlinear fitting capability and high prediction stability of multilayer sensing networks (MLNs) in the -1 rate of change range to provide a novel downscaling method for land surface temperature remote sensing products. The technical solution employs a dynamic MLN to construct a nonlinear unmixing model, enabling dynamic unmixing from low-resolution to high-resolution satellite pixels. Furthermore, the weighted sensing network effectively avoids the failure of existing weighting functions near the -1 rate of change. In summary, the land surface temperature remote sensing product downscaling method provided by this invention can effectively estimate land surface temperature, thereby improving the accuracy of land surface temperature products. Attached Figure Description

[0049] 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 are briefly introduced below; obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0050] Figure 1 This is a flowchart illustrating a method for downscaling surface temperature remote sensing products provided in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the process for obtaining the training sample dataset in an embodiment of the present invention;

[0052] Figure 3 This is a schematic diagram of the training process of the dynamic multilayer perceptron (DyNet) in an embodiment of the present invention;

[0053] Figure 4 This is a schematic diagram of the training process of the weight-aware network (WNet) in an embodiment of the present invention;

[0054] Figure 5This is a flowchart illustrating the process of predicting target date data in an embodiment of the present invention;

[0055] Figure 6 This is a schematic diagram of the data post-processing process in an embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0057] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0058] The present invention will now be described in further detail with reference to the accompanying drawings:

[0059] Please see Figure 1 The present invention provides a method for downscaling surface temperature remote sensing products, comprising the following steps:

[0060] Step S1, obtain T at the target time T. 前 T 后 The original low spatial resolution land surface temperature data and the original high spatial resolution land surface temperature data; among which, T 前 For the moment preceding the target time T, T 后 For the time immediately following the target time T;

[0061] Step S2, T 前 T 后 The original low spatial resolution land surface temperature data was resampled to the same spatial resolution as the original high spatial resolution land surface temperature data to obtain T. 前 T 后The sampled low spatial resolution surface temperature data and the ID value of each sampled pixel corresponding to the original pixel;

[0062] Step S3, based on T 前 T 后 Low spatial resolution surface temperature data after sampling, according to T 前 T, T 后 The pixel-by-pixel rate of change was calculated at three time points to obtain the pixel-by-pixel rate of change of low spatial resolution land surface temperature data.

[0063] Step S4: Using the pixel-by-pixel change rate of low spatial resolution land surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the pre-acquired homogeneous land surface units as input, the pre-trained dynamic multilayer sensing network is used to predict the category-by-category change rate of high spatial resolution land surface temperature data; T 前 T 后 The absolute value of the difference between the original high spatial resolution land surface temperature data and the category-wise change rate of the predicted high spatial resolution land surface temperature data are used as inputs, and the land surface temperature at the target time T is predicted using a pre-trained weighted perception network.

[0064] The method provided in this invention, based on a pre-trained dynamic multilayer sensing network and a weighted sensing network, can effectively downscale the land surface temperature, downscaling the original daily low spatial resolution data into daily high spatial resolution data, thereby improving the accuracy of land surface temperature products and serving high-precision, high-frequency satellite remote sensing monitoring of land surface temperature.

[0065] The steps for obtaining the pre-trained dynamic multilayer perceptron and weighted perceptron in the technical solution provided by the embodiments of the present invention can be summarized as follows:

[0066] Step 1, data preprocessing, includes: performing necessary spatial registration, homogeneous unit division, rate of change calculation, and pixel index calculation on the original low spatial resolution land surface temperature data and high spatial resolution land surface temperature data, so as to provide sufficient training data for subsequent model training;

[0067] Step 2, model training, includes: generating training samples using the data obtained in Step 1, and training the multilayer perceptron (DyNet) and weighted perceptron (WNet).

[0068] As exemplary in this embodiment of the invention, surface temperature prediction is carried out for a target date while retaining the optimal model from step 2.

[0069] In a preferred embodiment of the present invention, after obtaining the predicted surface temperature value, the data post-processing is further included, including: fixing the target date, selecting multiple sets of observation data pairs in the time series, repeatedly predicting the surface temperature on the target date, integrating the multiple prediction results according to certain rules, and synthesizing the final data product.

[0070] Please see Figure 2 Specifically, as exemplified in this embodiment of the invention, the step of obtaining training sample data for training a dynamic multilayer perceptron and a weighted perceptron specifically includes:

[0071] Step 1) Resample historical low spatial resolution images (for example, MODIS 1km scale land surface temperature products) to the same spatial resolution as high spatial resolution images (for example, Landsat 30m scale land surface temperature products) according to the nearest neighbor sampling principle to form a sampled low spatial resolution image (which can be denoted as M30); and the ID value of each sampled pixel corresponding to the original MODIS pixel (which can be denoted as MID);

[0072] Step 2) Using historical high spatial resolution images (for example, Landsat 30m scale land surface temperature products) as multiple features, the unsupervised clustering model ISOData algorithm is used to cluster them into multiple homogeneous land surface units, denoted as ClassID;

[0073] Step 3): In chronological order, the pixel-by-pixel change rate of the sampled low-spatial-resolution image and high-spatial-resolution image is calculated at every three time points (before T, during T, after T) to finally obtain the pixel-by-pixel change rate of the low-spatial-resolution image (which can be denoted as M_ratio) and the pixel-by-pixel change rate of the high-spatial-resolution image (which can be denoted as L_ratio).

[0074] The expression for calculating the rate of change is:

[0075]

[0076] In the formula, ΔL ke (i,j) and ΔL ok (i,j) represent two periods [t] k ,t e ] and [t0,t k The difference in Landsat remote sensing reflectance between [L] and [L]. e (i,j) is the Landsat remote sensing reflectance at the end date te; L k(i,j) is the Landsat remote sensing reflectance of the target date tk; L0(i,j) is the Landsat remote sensing reflectance of the start date t0.; subscript 0 is the start date, k is the target date; e is the end date; (i,j) represents the pixel in the i-th row and j-th column;

[0077] Step 4): Using the surface homogeneous unit (ClassID) generated in Step 2), the pixel-by-pixel change rate of the high spatial resolution image generated in Step 3) is averaged to obtain the high spatial resolution class-by-class change rate (denoted as L_ratio_C).

[0078] Step 5) Construct a training dataset for a dynamic multilayer perceptron (DyNet). The input to the dataset is a three-channel feature image, with the three channels representing the pixel-wise change rate (M_ratio) of the low spatial resolution image, the ID value (MID) of each sampled pixel corresponding to the original MODIS pixel, and the ClassID of the surface homogeneous unit. The output of the dataset is a single-channel feature image, which is the class-wise change rate (L_ratio_C) of the high spatial resolution image.

[0079] Step 6) Construct the WNet training dataset for the weighted network; the input of the dataset is a two-channel feature image, the two channels being the absolute value of the difference between the high-resolution images before time T and after time T, and the class-by-class change rate of the high spatial resolution image, respectively; the output of the data is a single-channel feature image, which is the true surface temperature (LST in T) at the target time T.

[0080] Please see Figure 3 Specifically, as exemplified in this embodiment of the invention, the steps for training a dynamic multilayer perceptron include:

[0081] In the dynamic multilayer sensing network, a dropout layer is set up for input data. The dropout layer is used to activate only the neural nodes that exist in the current input data and block the neural nodes that do not exist in the current input data by utilizing the MID in the input data. After the dropout layer, five fully connected layers (FC) and ReLU activation function layers are passed in succession. Specifically, each fully connected layer can have 128 neurons. After the fifth fully connected layer (FC) and ReLU activation function layer, another dropout layer is set up. Whether the neurons in the dropout layer are activated or not is determined according to the ClassID of the homogeneous surface unit in the input data.

[0082] During training, the root mean square loss (MSE Loss) is calculated based on the predicted class-specific rate of change (P_ratio_C) and the ground truth (L_ratio_C) of the training samples. After backpropagation, the trained dynamic multilayer perceptron is obtained when the preset convergence condition is met.

[0083] Please see Figure 4 Specifically, as exemplified in this embodiment of the invention, the steps for training a weight-aware network include:

[0084] The weighted sensing network includes a data mapping layer; after the data mapping layer, five fully connected layers (FC) and a ReLU activation function layer are set consecutively. Specifically, for example, each fully connected layer can have 40 neurons.

[0085] During training, the input parameters pass through a data mapping layer, where the data mapping is performed according to the following formula:

[0086] The transformation of the true data is: ΔLST = LST - LST per ;

[0087] When the difference between the surface temperature at time T and the surface temperature at time T is less than 0, the class-specific rate of change (L_ratio_C) of the high spatial resolution image is transformed as follows:

[0088] When the difference between the surface temperature at time T and the surface temperature at time T is greater than 0, the class-specific rate of change (L_ratio_C) of the high spatial resolution image is transformed as follows:

[0089] The expression for converting the model output into predicted surface temperature values ​​is as follows:

[0090] Among them, for the predicted surface temperature The root mean square loss (MSE Loss) is calculated based on the surface temperature (true value) on the target date. After backpropagation, the trained weighted sensing network is obtained once the preset convergence condition is met.

[0091] The dynamic multilayer sensing network in this embodiment of the invention can be replaced by other deep networks or deep convolutional networks that include multiple fully connected layer structures and temporary de-convolutional layer structures. The weight sensing network in this invention can be replaced by other deep networks or deep convolutional networks that include multiple fully connected layer structures.

[0092] Please see Figure 5 The specific application process of the method provided in this embodiment of the invention includes: based on the full training of the dynamic multilayer sensing network DyNet and the weighted sensing network WNet, using high-resolution and low-resolution digital image data from the previous time step (before T) and the next time step (after T), prediction of data for the target date (time in T) can be carried out; the specific process includes the following steps:

[0093] The low spatial resolution image in the current image pair (T front-T middle-T back) is resampled to the same spatial resolution as the high spatial resolution image according to the nearest neighbor sampling principle, forming the sampled low spatial resolution image and the ID value of each sampled pixel corresponding to the original MODIS pixel.

[0094] Calculate the pixel-by-pixel rate of change of the current image pair (Tbefore-Tmiddle-Tafter) of the low spatial resolution image;

[0095] A dynamic multilayer sensing network (DyNet) dataset for the current image pair is constructed. The dataset consists of three-channel feature images, with the three channels representing the pixel-wise change rate (M_ratio) of the low spatial resolution image, the ID value (MID) of each sampled pixel corresponding to the original MODIS pixel, and the ClassID of the surface homogeneous unit. The trained dynamic multilayer sensing network (DyNet) is used to predict the change rate of the homogeneous unit to obtain the prediction results.

[0096] Construct a weighted sensing network (WNet) dataset for the current image pair. The input to the dataset consists of two-channel feature images. The two channels are the absolute value of the difference between the high-resolution images (L30) before time T and after time T, and the predicted value of the rate of change of homogeneous units, respectively. Use the trained weighted sensing network (WNet) to predict the surface temperature on the target date (time T).

[0097] Please see Figure 6 In a preferred embodiment of the present invention, data post-processing is also included, which involves using multiple sets of dates to predict the same day from the image, merging the multiple prediction values ​​to reduce prediction errors; the specific process includes:

[0098] Select multiple date pairs under the same target date and generate multiple sets of prediction results;

[0099] Based on the actual reasonable range of surface temperature, outliers in multiple sets of prediction results are masked.

[0100] By using the difference in surface temperature between different dates, a threshold is set to mask out pixels with small differences and predict pixels with low confidence.

[0101] For multiple sets of prediction results, the median value in the time series is taken as the final high-resolution land surface temperature prediction result for the target date.

[0102] In summary, the technical solution provided by this invention constructs a nonlinear unmixing model by designing a novel dynamic multilayer sensing network, achieving dynamic unmixing from low-resolution satellite pixels (using MODIS satellite as an example in this invention) to high-resolution satellite pixels (using Landsat satellite as an example in this invention). The established dynamic multilayer sensing network effectively solves the problem that existing linear unmixing models cannot effectively estimate surface temperature, fundamentally avoiding the unsolvable problem caused by the contradiction between the number of homogeneous surface units and the number of available solution equations. This invention also effectively avoids the failure problem of existing weight functions near the rate of change of -1 by designing a novel multilayer weighted sensing network. The network model established by this invention is based on mathematical mechanisms and can achieve model generalization across regions and time periods under the premise that the solution conditions remain unchanged.

[0103] In this embodiment of the invention, simulations and experiments were conducted in the Shenzhen pilot zone, where internationally accepted evaluation indicators were used for measurement. The existing STARFM model and ESTARFM model were compared, and the comparative experimental results are shown in Table 1.

[0104] Table 1. Comparative Experimental Results

[0105]

[0106] As shown in Table 1, in the experimental groups with multiple prediction dates, the system constructed in this invention outperforms the two internationally used models (STARFM and ESTARFM) in terms of quantitative indicators such as peak signal-to-noise ratio (PNSR), structural similarity (SSIM), correlation coefficient (CC), root mean square error (RMSE), and mean error (MAE).

[0107] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not omitted in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0108] An embodiment of the present invention provides a surface temperature remote sensing product downscaling system, comprising:

[0109] The data acquisition module is used to obtain the T of the target time T. 前 T 后 The original low spatial resolution land surface temperature data and the original high spatial resolution land surface temperature data; among which, T 前 For the moment preceding the target time T, T 后 For the time immediately following the target time T;

[0110] The data processing module is used to process T 前 T 后The original low spatial resolution land surface temperature data was resampled to the same spatial resolution as the original high spatial resolution land surface temperature data to obtain T. 前 T 后 The sampled low spatial resolution surface temperature data and the ID value of each sampled pixel corresponding to the original pixel;

[0111] The rate of change acquisition module is used to obtain the rate of change based on T. 前 T 后 Low spatial resolution surface temperature data after sampling, according to T 前 T, T 后 The pixel-by-pixel rate of change was calculated at three time points to obtain the pixel-by-pixel rate of change of low spatial resolution land surface temperature data.

[0112] The prediction module takes the pixel-by-pixel rate of change of low spatial resolution land surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and pre-acquired homogeneous land surface units as input, and uses a pre-trained dynamic multilayer sensing network to predict the category-by-category rate of change of high spatial resolution land surface temperature data; T 前 T 后 The absolute value of the difference between the original high spatial resolution land surface temperature data and the predicted rate of change of the high spatial resolution land surface temperature data by category are used as inputs, and the land surface temperature at the target time T is predicted by a pre-trained weighted perception network.

[0113] In existing technologies, weighted function-based methods assume the consistency of temporal variations in surface reflectance across spatial scales. They combine information from all input images using weighted functions to predict the high spatial resolution pixel values ​​of the target image. Gao et al. first proposed the Spatial and Temporal Adaptive Reflectance Fusion Model (STARFM). This model assumes pixels in coarse-resolution images are clean pixels, introduces the concepts of moving windows and similar pixels, and calculates weights based on spectral differences, temporal differences, and spatial distance. However, for areas with complex topography where land types change, the STARFM method cannot accurately predict the true land cover types in unknown temporal phases, leading to significant errors in the fusion results. To address the shortcomings of the STARFM method, Hilker et al. proposed a novel Spatial Temporal Adaptive Algorithm for Mapping Reflectance Change (STAARCH), which uses tasseled cap transformation to detect reflectance changes, improving the ability to detect spatiotemporal changes in land cover from low-resolution images. However, the change detection in STAARCH is only applicable to vegetated surfaces. Therefore, Zhu et al. proposed the Enhanced Spatial and Temporal Adaptive Reference Fusion Model (ESTARFM), which introduces conversion coefficients and linear spectral unmixing theory, using spectral correlation coefficients to replace spatial distances, preserving more spatial details while improving prediction accuracy in complex terrains. To improve the accuracy of the ESTARFM method in predicting land cover changes with shape variations over time, Zhao et al. proposed the Robust Adaptive Spatial and Temporal Fusion Model (RASTFM). RASTFM consists of a weighted average module based on nonlocal linear regression and an image super-resolution module based on nonlocal linear regression, used to predict non-shape changes (including phenological changes and land cover changes without shape changes) and shape changes (land cover changes with shape changes), respectively. While the STARFM series of models has simple parameters and high computational efficiency, it still has limitations in predicting complex land cover changes and changes in land cover types across different time phases. A learning-based approach uses machine learning algorithms to model high- and low-resolution image pairs in a non-linear manner, predicting high-spatial-temporal resolution images.Huang et al. proposed a sparse representation-based spatiotemporal reflection fusion model (SPSTFM), which uses dictionary pairs to learn and establish the corresponding reflectance variation relationship between low-spatial-resolution and high-spatial-resolution images, and predicts the target high-resolution image through time weighting. To address the dictionary alignment perturbation problem, Wu et al. proposed error-bound-regularized semi-coupled dictionary learning (EBSCDL), which uses error-bound regularization to solve dictionary perturbations and constructs an optimized semi-coupled dictionary to address the differences between low-spatial-resolution and high-resolution images, improving fusion accuracy. However, the above sparse representation-based fusion methods require manual design of dictionary primitives. In the algorithm implementation, dictionary learning, sparse coding, and image reconstruction steps are separated, increasing the algorithm's instability and complexity. To address this, Song et al. proposed a Spatiotemporal Fusion Method Based on Deep Convolutional Neural Networks (STFDCNN), which combines convolutional neural networks and nonlinear mapping models to design a dual convolutional neural network. This achieves automatic extraction of image features and improves prediction accuracy, effectively overcoming the shortcomings of sparse representation methods. Learning-based methods can capture more surface spatial details and are suitable for heterogeneous regions, but they rely heavily on training samples and model parameters, resulting in high time complexity. Unmixing-based methods effectively compensate for the shortcomings of the above methods. These methods are based on linear spectral mixing theory, extracting land cover categories and abundances from high spatial resolution images to decompose low spatial resolution pixels and obtain the spectral values ​​of the categories. The implementation process generally includes endmember selection, abundance calculation, and unmixing. This type of algorithm was first proposed by Zhukov et al., which assumes that land cover types have equal reflectance. It unmixes the reflectance of low spatial resolution images within a moving window and then assigns the unmixed results to high spatial resolution images of unknown temporal phases. However, this method has serious errors in the unmixing process, and the assumptions in the algorithm do not hold true for areas where land cover types change over time in reality.In response, Zurita-Milla et al. introduced constraints during the linear unmixing process to handle negative and outlier values ​​in the unmixing results. Wu et al. proposed the Spatial Temporal Data Fusion Approach (STDFA), which, based on the assumption that the temporal changes of various land cover types are always the same, incorporates temporal variation information to improve the accuracy of prediction results. Addressing the issue that the STDFA algorithm does not consider the spatial heterogeneity of image endmember reflectance, Zhang et al., based on the STDFA algorithm, combined it with multi-scale segmentation and ISODATA algorithms to generate classification maps, and used the moving window method to unmix low spatial resolution images. Finally, they introduced the concept of temporal weights to predict images with unknown temporal phases. To improve the model's prediction accuracy in areas with changing land cover types, Huang et al. proposed the Unmixing-based Spatio-Temporal Reflectance Fusion Model (U-STFM), which assumes that the reflectance of land cover pixels changes in homogeneous units. The rate of change is equal across regions (HCRs). Superpixels obtained using a multi-scale segmentation algorithm are used for unmixing, thereby improving the fusion effect of remote sensing images from different time phases on the spatial changes of land cover types. This type of algorithm has low computational cost and high operability, improving the fusion accuracy of remote sensing images showing changes in land cover types over time, and is often used in reflectance data. In summary, for LST data with strong temporal fluctuations and limited spectral features, the parameters of weight function-based methods become more sensitive, affecting the effectiveness of training data for learning-based methods, resulting in a lower overall fusion effect for both methods. However, unmixing-based methods, for LST data that changes over time, combine the assumption that the ratio of changes in different sensors is equal in regions with the same LST change trend, improving the fusion effect of LST from different time phases. Based on this, the technical solution provided in this embodiment of the invention constructs a nonlinear unmixing model through a dynamic multilayer sensing network, enabling dynamic unmixing from low-resolution satellite pixels to high-resolution satellite pixels. The technical solution provided in this embodiment of the invention, through a weighted sensing network, can effectively avoid the failure problem of existing weight functions near the rate of change of -1.

[0114] In another embodiment of the present invention, a computer device is provided, comprising a processor and a memory. The memory stores a computer program, which includes program instructions. The processor executes the program instructions stored in the computer storage medium. The processor may be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. It is the computing and control core of the terminal, suitable for implementing one or more instructions, specifically suitable for loading and executing one or more instructions from the computer storage medium to achieve a corresponding method flow or function. The processor described in this embodiment of the present invention can be used for the operation of a downscaling method for surface temperature remote sensing products.

[0115] In another embodiment of the present invention, a storage medium is provided, specifically a computer-readable storage medium (Memory), which is a memory device in a computer device used to store programs and data. It is understood that the computer-readable storage medium here can include both the built-in storage medium in the computer device and extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, the storage space also stores one or more instructions suitable for loading and execution by a processor. These instructions can be one or more computer programs (including program code). It should be noted that the computer-readable storage medium here can be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor can load and execute one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the surface temperature remote sensing product downscaling method in the above embodiments.

[0116] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0117] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0118] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0119] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A land surface temperature remote sensing product downscaling method, characterized in that, The method comprises the following steps: Obtain T at target time T 前 T 后 The original low spatial resolution land surface temperature data and the original high spatial resolution land surface temperature data; among which, T 前 Let T be any time before the target time T where data is available. 后 For any time after the target time T that has data; T 前 , T 后 , the original low spatial resolution land surface temperature data is resampled to the spatial resolution consistent with the original high spatial resolution land surface temperature data, obtaining T 前 , T 后 , the sampled low spatial resolution land surface temperature data and the ID value corresponding to the original pixel of each pixel after sampling. Based on T 前 , T 后 , the low spatial resolution land surface temperature data after sampling, the pixel-by-pixel change rate is calculated according to T 前 , T, T 后 at three time points, and the low spatial resolution land surface temperature data pixel-by-pixel change rate is obtained. The low spatial resolution land surface temperature data, the pixel-by-pixel change rate, the ID value corresponding to the original pixel after sampling of each pixel and the pre-acquired land homogeneous unit ID are taken as inputs, and a high spatial resolution land surface temperature data class-by-class change rate is obtained by using a pre-trained dynamic multi-layer perception network DyNet for prediction. 前 , T 后 The absolute value of the difference of the original high spatial resolution land surface temperature data and the obtained high spatial resolution land surface temperature data class-by-class change rate is taken as input, and the land surface temperature at the target moment T is obtained by using a pre-trained weight perception network WNet for prediction. The obtaining of the pre-trained dynamic multi-layer perception network DyNet comprises: obtaining a first training sample set; wherein each training sample in the first training sample set comprises, as input, a low spatial resolution land surface temperature data pixel-wise change rate, an ID value corresponding to an original pixel and a land surface homogeneous unit ID after sampling, and, as a label, a high spatial resolution land surface temperature data category-wise change rate; wherein the high spatial resolution land surface temperature category-wise change rate is calculated based on original high spatial resolution land surface temperature data of T 前 , T 后 , T 前 , T 后 three time points, and the mean value of each category is obtained by using the land surface homogeneous unit ID; the land surface homogeneous unit ID is obtained by clustering time series land surface temperature data or land cover type data; The pre-built dynamic multi-layer perception network DyNet is trained based on the first training sample set by using an MSE Loss loss function, and the pre-trained dynamic multi-layer perception network is obtained after a preset convergence condition is reached. The obtaining step of the pre-trained weight-aware network WNet includes: obtaining a second training sample set; each training sample in the second training sample set includes high spatial resolution land surface temperature data T 前 and the absolute value of the difference between T 后 and the category-by-category change rate of the high spatial resolution land surface temperature data in the first training sample set, and high spatial resolution land surface temperature of a target date as a label; based on the second training sample set, the pre-constructed weight-aware network is trained by using an MSE Loss loss function, and after a preset convergence condition is reached, the pre-trained weight-aware network is obtained.

2. The method according to claim 1, wherein, The pre-built dynamic multi-layer perception network DyNet comprises a temporary return layer for shielding part of input neurons, a plurality of fully connected layers and a Relu activation function layer connected in series after the temporary return layer, and a temporary return layer for shielding part of output neurons.

3. The method of claim 1, wherein, The pre-built weight perception network comprises a data mapping layer for input data, and a plurality of fully connected layers and a Relu activation function layer connected in series after the data mapping layer.

4. The method according to claim 3, wherein, In the pre-built weight perception network WNet, the data mapping performed by the data mapping layer comprises: Transforming the label truth data, the transformation expression is, ; In the formula, LST is the real ground surface temperature LST at the target time T, is T 前 The ground surface temperature LST at the time T, Temperature difference; When T 后 The surface temperature at time T is less than T 前 When the surface temperature value is given at a given time, the transformation expression for the category-by-category rate of change of the high spatial resolution image is: In the formula, For high spatial resolution images, the category-specific rate of change. The transformed rate of change by category; When the surface temperature value at time T 后 is greater than the surface temperature value at time T 前 , the transformation expression of the high spatial resolution image per category change rate is, ; The expression of the ground surface temperature prediction value output by the weight-aware network is , is the model-predicted ground surface temperature LST.

5. The method of claim 1, wherein, The steps of obtaining the first training sample set and the second training sample set specifically comprise: Obtaining original low-spatial-resolution land surface temperature data and high-spatial-resolution land surface temperature data of samples; Resampling the low-spatial-resolution land surface temperature data to a spatial resolution consistent with the high-spatial-resolution land surface temperature data according to a nearest neighbor sampling principle to form resampled low-spatial-resolution land surface temperature data and ID values of each pixel corresponding to original pixels; Clustering to generate land homogenous units by taking the time-series high-spatial-resolution land surface temperature data as a plurality of features; In time sequence, the low spatial resolution land surface temperature data and the high spatial resolution land surface temperature data after sampling are calculated for pixel-by-pixel change rate according to three time points of a time point before a target time point, the target time point and a time point after the target time point, to obtain low spatial resolution land surface temperature data pixel-by-pixel change rate and high spatial resolution land surface temperature data pixel-by-pixel change rate; wherein, the change rate calculation expression is, ; wherein, ; ; ; ; ; ; and ) ; Performing mean calculation on pixel-by-pixel change rates of the generated high-spatial-resolution land surface temperature data by using the generated land homogenous units to obtain high-spatial-resolution land surface temperature data change rates of each category; Constructing a first training sample set; wherein input features of the first training sample set are pixel-by-pixel change rates of low-spatial-resolution land surface temperature data, ID values of each pixel corresponding to original pixels after resampling, and land homogenous units, and an output label is high-spatial-resolution land surface temperature data change rates of each category; Constructing a second training sample set; wherein input features of the second training sample set are absolute values of differences between high-spatial-resolution land surface temperature data of a time point before a target time point and high-spatial-resolution land surface temperature data of a time point after the target time point, and high-spatial-resolution land surface temperature data change rates of each category, and an output label is land surface temperature of the target time point.

6. A land surface temperature remote sensing product downscaling system, characterized in that, The method comprises the following steps: a data acquisition module, configured to acquire original low spatial resolution land surface temperature data and original high spatial resolution land surface temperature data of a target time T 前 , T 后 ; wherein T 前 is any data time before the target time T, and T 后 is any data time after the target time T; a data processing module, configured to resample the original low spatial resolution land surface temperature data of T 前 , T 后 to a spatial resolution consistent with the original high spatial resolution land surface temperature data, to obtain the resampled low spatial resolution land surface temperature data of T 前 , T 后 and the ID value corresponding to each pixel of the resampled low spatial resolution land surface temperature data to the original pixel. The rate of change acquisition module is based on T. 前 T 后 The sampled low spatial resolution surface temperature data, according to T 前 T, T 后 The pixel-by-pixel rate of change was calculated at three time points to obtain the pixel-by-pixel rate of change of low spatial resolution land surface temperature data. The prediction module is configured to take the low spatial resolution land surface temperature data, the pixel-by-pixel change rate, the ID value corresponding to the original pixel of each pixel after sampling and the pre-acquired land homogenous unit ID as inputs, and utilize a pre-trained dynamic multi-layer perception network DyNet to predict the high spatial resolution land surface temperature data to obtain a category-by-category change rate. 前 , T 后 The prediction module is configured to take the low spatial resolution land surface temperature data, the pixel-by-pixel change rate, the ID value corresponding to the original pixel of each pixel after sampling and the pre-acquired land homogenous unit ID as inputs, and utilize a pre-trained dynamic multi-layer perception network DyNet to predict the high spatial resolution land surface temperature data to obtain a category-by-category change rate. 前 , T 后 The prediction module is configured to take the low spatial resolution land surface temperature data, the pixel-by-pixel change rate, the ID value corresponding to the original pixel of each pixel after sampling and the pre-acquired land homogenous unit ID as inputs, and utilize a pre-trained dynamic multi-layer perception network DyNet to predict the high spatial resolution land surface temperature data to obtain a category-by-category change rate. The obtaining of the pre-trained dynamic multi-layer perception network DyNet comprises: Obtain a first training sample set; wherein each training sample in the first training sample set includes, as input, the pixel-wise change rate of low spatial resolution surface temperature data, the ID value of each sampled pixel corresponding to the original pixel, and the ID of the surface homogeneous unit, and as labels, the category-wise change rate of high spatial resolution surface temperature data; wherein the category-wise change rate of high spatial resolution surface temperature is based on T 前 T, T 后 The original high spatial resolution surface temperature data, according to T 前 T, T 后 The calculation is performed pixel-by-pixel at three time points, and the mean value of each category is obtained by using the surface homogeneous unit ID; the surface homogeneous unit ID is obtained by clustering time series surface temperature data or land cover type data. The pre-built dynamic multi-layer perception network DyNet is trained based on the first training sample set by using an MSE Loss loss function, and the pre-trained dynamic multi-layer perception network is obtained after a preset convergence condition is reached. The steps for obtaining the pre-trained weighted sensing network WNet include: obtaining a second training sample set; wherein each training sample in the second training sample set includes high spatial resolution land surface temperature data T as input. 前 and T 后 The absolute value of the difference and the category-by-category change rate of the high spatial resolution surface temperature data in the first training sample set, as well as the high spatial resolution surface temperature of the target date as the label; based on the second training sample set, the pre-constructed weighted sensing network is trained using the MSE Loss loss function, and after reaching the preset convergence condition, the pre-trained weighted sensing network is obtained.

7. An electronic device, comprising: The method comprises the following steps: At least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the land surface temperature remote sensing product downscaling method according to any one of claims 1 to 5.

8. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to implement the land surface temperature remote sensing product downscaling method according to any one of claims 1 to 5.

Images