Machine learning method for reconstructing cloudless land surface temperature by combining spatial data information

Abstract

This invention discloses a machine learning method for reconstructing cloudless land surface temperature by combining spatial information. First, raw land surface temperature data is acquired using satellite remote sensing. This data is then combined with auxiliary data such as topography, land use, population density, reflectance, and vegetation index to initially fill in missing values. Next, spatial information of the land surface temperature is calculated based on this data. Then, combined with the aforementioned auxiliary data, a random forest regression is used to obtain the reconstructed land surface temperature. Following this, residuals are obtained from the original data, and Kriging regression is used to predict the missing residuals. Finally, these residuals are used to correct the reconstructed land surface temperature, yielding cloudless land surface temperature data. This invention can accurately reconstruct land surface temperature data, providing an algorithmic foundation for obtaining high spatiotemporal resolution data and solving the key problem of obtaining continuous imagery for studying urban thermal environment changes.

Description

TECHNICAL FIELD

[0001] The present application relates to a method for reconstructing missing data by remote sensing, and in particular to a method for reconstructing cloud-free land surface temperature by machine learning combined with data space information. BACKGROUND

[0002] In order to understand the impact of human activities on the natural ecological environment and the response to global environmental change, land surface temperature (LST) has been widely used in various fields, including urban thermal environment research, global climate change, etc.

[0003] The development of remote sensing technology has greatly facilitated the acquisition of LST data. At present, there are various algorithms for retrieving LST from satellite images according to different sensors (Li Z-L, Tang B-H, Wu H, et al., 2013, Satellite-derived land surface temperature: Current status and perspectives. Remote Sensing of Environment, 131, pp. 14-37.). The retrieved LST has the characteristics of wide spatial coverage, long observation period, and various spatial resolutions. However, due to technical and atmospheric conditions, the spatial resolution and temporal resolution of a single sensor are often inconsistent, which is not conducive to the observation and research of fine and long-time series of LST. For example, the Landsat series of satellites provides high spatial resolution data of 30m, with a return period of 16 days, while the MODIS series of products provides low spatial resolution products of 250m to 1km per day. In addition, due to the fact that the thermal infrared band cannot penetrate the cloud layer, the LST retrieved in the cloud-covered area often has some differences from the true value. Therefore, it is difficult to obtain high-resolution cloud-free satellite images to obtain LST within a certain time and in a certain area.

[0004] To address this problem, several algorithms have been proposed to obtain seamless LSTs. These methods can be broadly categorized into three types. The first type is based on spatial information. Since LSTs possess spatial continuity and autocorrelation, missing values ​​can be reconstructed by utilizing the spatial information of the effective values ​​surrounding the target reconstructed pixel. For example, spline interpolation (Zhang Jun, Qin Zhihao, Liu Mei, Tu Lili, Zhou Yi, Yang Qiang., 2011, Feasibility study on estimating cloud-covered pixel surface temperature using spatial interpolation. Geography and Geographic Information Science, 27(06), pp.45-49+115.), geostatistical interpolation method (Bhattacharjee S, Mitra P, Ghosh S K., 2014, Spatial Interpolation to Predict Missing Attributes in GIS Using Semantic Kriging. IEEE Transactions on Geoscience and Remote Sensing, 52(8), pp.4771-4780.), inverse distance weighting, etc. (Liu Z, Wu P, Duan S, et al., 2017, Spatiotemporal Reconstruction of Land Surface Temperature Derived From FengYun Geostationary Satellite Data. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 10(10), pp. 4531-4543.). However, when the missing data is large, the limited effective spatial information of spatial information-based methods, which only refer to their own spatial information, will lead to poor reconstruction results (Artusi A, Banterle F, Chetverikov D., 2011, A Survey of Specularity Removal Methods. Computer Graphics Forum, 30(8), pp. 2208-2230.). The second type is based on multi-temporal information methods, which use data from the same area at different times to reconstruct cloudless data.Xu Y et al. (Xu Y, Shen Y., 2013, Reconstruction of the land surface temperature timeseries using harmonic analysis. Computers & Geosciences, 61, pp. 126-132.) developed a time-series harmonic analysis algorithm (HANTS) to reconstruct 8-day MODISLST data. Peng Fu et al. (Fu P, Weng Q H., 2016, Consistent land surface temperature data generation from irregularly spaced Landsat imagery. Remote Sensing of Environment, 184, pp. 175-187.) used LST data retrieved from continuous Landsat satellite imagery to reconstruct seamless LST. While multi-temporal information-based methods perform well in areas with large amounts of missing data, their performance weakens when land cover types change (Wu P, Yin Z, Zeng C, et al., 2021, Spatially Continuous and High-Resolution Land Surface Temperature Product Generation: A review of reconstruction and spatiotemporal fusion techniques. IEEE Geoscience and Remote Sensing Magazine, 9(3), pp. 112-137.). The last category is hybrid methods based on spatiotemporal information, which contain rich information and are considered the most promising methods.Xia et al. (Xia HP, Chen YH, Li Y, et al., 2019, Combining kernel-driven and fusion-based methods to generate daily high-spatial-resolution land surface temperatures. Remote Sensing of Environment, 224, pp. 259-274.) combined the basic theories of downscaling and data fusion to develop a combination of kernel-driven and fusion-based methods to generate daily high spatial resolution land surface temperatures (LSTs). Feng G et al. (Feng G, Masek J, Schwaller M, et al., 2006, On the blending of the Landsat and MODIS surface reflectance: predicting daily Landsat surface reflectance. IEEE Transactions on Geoscience and Remote Sensing, 44(8), pp. 2207-2218.) proposed the Spatiotemporal Adaptive Reflectance Fusion Model (STARFM) algorithm, which generates high spatiotemporal LSTs by combining Landsat and MODIS LST data. Due to the convenient acquisition of the required data and good prediction performance, this method is widely used. However, STARFM is sensitive to heterogeneous regions. Although researchers have proposed various methods to obtain seamless LST data, their performance is affected by factors such as the limited availability of raw data, land cover changes, and accuracy verification. Summary of the Invention

[0005] The purpose of this invention is to provide a machine learning method for reconstructing cloudless land surface temperature by combining spatial data information, so as to improve the accuracy and applicability of land surface temperature observation.

[0006] To address the aforementioned technical problems, this invention provides a technical solution: a machine learning method for reconstructing cloudless land surface temperature by combining spatial data information, comprising the following steps:

[0007] S1. Acquire Landsat satellite remote sensing image data and auxiliary data including topographic data, land use data, population density, soil type, emissivity and spectral index;

[0008] S2. The Landsat satellite remote sensing image data acquired in S1 is inverted to retrieve the land surface temperature, followed by cloud removal and resampling to obtain preliminary raw land surface temperature data.

[0009] S3. Combining the auxiliary data described in S1 and the preliminary raw surface temperature data obtained in S2, construct a classical regression random forest model;

[0010] S4. Adjust the parameters of the classical regression random forest model based on the feedback from the out-of-bag validation data until the best results are obtained. Then use the adjusted parameters to obtain the initial fill of the original surface temperature data.

[0011] S5. Based on the preliminary fill-in original surface temperature data obtained in S4, according to the principle that the closer to the pixel, the greater the weight, the more complete the fill-in original surface temperature data and its spatial information are obtained.

[0012] S6. Based on the original surface temperature data and spatial information obtained in S5, and combined with the auxiliary data described in S1, construct a random forest regression model that considers spatial information.

[0013] S7. Adjust the parameters of the random forest regression model considering spatial information based on the out-of-bag validation data feedback until the best results are obtained. Then use the adjusted parameters to obtain the reconstructed land surface temperature considering spatial information.

[0014] S8. Based on the preliminary raw surface temperature data obtained in S2 and S7 respectively and the reconstructed surface temperature considering spatial information, the residuals are obtained. Kriging interpolation regression is used to predict the missing parts of the residuals to obtain the complete residual data.

[0015] S9, the complete residual data obtained by coupling S8, and the reconstructed surface temperature considering spatial information obtained by S7 are used to obtain the final reconstructed surface temperature.

[0016] According to the above scheme, in S1, the spatial resolution of the Landsat satellite remote sensing image data is 30m×30m, and the temporal resolution is 16 days; the terrain data includes latitude and longitude, digital elevation model, and slope data, the slope data is calculated by the digital elevation model, and its spatial resolution is 30m×30m; the spatial resolution of the land use data is 30m×30m; the soil type includes three soil textures, namely sandy soil, clay, and clay, with a spatial resolution of 250m×250m; the reflectance is the reflectance in the mid-infrared band, with a spatial resolution of 250m×250m and a temporal resolution of 1 day; the spectral indices include enhanced vegetation index and solar zenith angle, with a spatial resolution of 250m×250m and a temporal resolution of 1 day.

[0017] According to the above scheme, the temperature inversion algorithm used in S2 is a single-channel algorithm; the cloud removal process is completed using the QA band built into Landsat; the resampling process specifically involves upsampling the 30m×30m spatial resolution surface temperature data to a 100m×100m spatial resolution using bilinear interpolation.

[0018] According to the above scheme, the specific formula of the classic regression random forest model described in S3 is as follows.

[0019] LST ori =f(x) terrain x pop x LUCC x mir x index x soil )

[0020] Among them, LST ori For raw surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0021] According to the above scheme, in S4, the parameters of the classical regression random forest model subject to feedback adjustment include the number of decision trees, depth, and number of features involved in the modeling; the specific formula for the preliminary filling of the original surface temperature data is as follows.

[0022] LST gapfill =f(x) terrain x pop x LUCC x mir x index x soil )

[0023] Among them, LST gapfill For preliminary filling of the original surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0024] According to the above scheme, the spatial information acquisition of the original surface temperature data to be filled in, as described in S5, specifically involves...

[0025]

[0026] Among them, LST sf This represents the spatial characteristics of the Earth's surface temperature at the center pixel m, where N represents the total number of pixels and n represents the neighboring pixels near the center pixel m. LST represents the spatial distance between pixels m and n. gapfill_n It is the initial fill of the raw surface temperature data at pixel n.

[0027] According to the above scheme, the random forest regression model in S6 that considers spatial information is specifically as follows:

[0028] LST x =f(x) terrain x pop x LUCC x mir x index x soil x sf )

[0029] Among them, LST x To fill in the original surface temperature data, x sf To fill in the spatial information of the original surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0030] According to the above scheme, the method for obtaining the reconstructed surface temperature considering spatial information in S7 is as follows.

[0031] LST sfrfrm =f(x) terrain x pop x LUCC x mir x index x soil x sf )

[0032] Among them, LST sfrfrm To reconstruct surface temperature taking into account spatial information, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil For soil type, x sf Spatial information to fill in the original surface temperature data.

[0033] According to the above scheme, the specific method for obtaining complete residual data using kriging interpolation in S8 is as follows:

[0034]

[0035] Where ε is the complete residual data after prediction, and λ i is the weight at pixel i, and n is the total number of pixels.

[0036] According to the above scheme, the specific method for obtaining the final reconstructed landmark temperature in S9 is as follows:

[0037] LST sfrfkrm =LST sfrfrm +ε

[0038] Among them, LST sfrfkrm For the final reconstruction of surface temperature, LST sfrfrm To reconstruct the surface temperature taking into account spatial information, ε represents the complete residual data.

[0039] The beneficial effects of this invention are as follows: This method utilizes multi-source remote sensing data to construct the model, improving its stability and applicability. Simultaneously, it considers the spatial information of the data, further enhancing the model's predictive performance. This invention not only offers better prediction accuracy but also preserves finer textures in the images. This greatly alleviates the problem of remote sensing data contamination caused by sensor factors, cloud cover, and other influences, and has significant practical implications for observing detailed long-term urban thermal environment changes. It solves a crucial problem for natural resources departments in observing and studying urban thermal environments, possessing significant theoretical, practical, and application value. Attached Figure Description

[0040] Figure 1 This is a raw surface temperature data map according to an embodiment of the present invention;

[0041] Figure 2 This is a residual distribution diagram according to an embodiment of the present invention;

[0042] Figure 3 This is a final reconstructed seamless surface temperature data map according to an embodiment of the present invention;

[0043] Figure 4 This is a scatter density map of the final reconstructed seamless surface temperature data and the original surface temperature data according to an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0045] See Figures 1-3 A machine learning method for reconstructing cloudless land surface temperature by combining spatial information of data includes the following steps.

[0046] S1. Acquire Landsat satellite remote sensing image data and auxiliary data including topographic data, land use data, population density, soil type, emissivity and spectral index;

[0047] S2. The Landsat satellite remote sensing image data acquired in S1 is inverted to retrieve the land surface temperature, followed by cloud removal and resampling to obtain preliminary raw land surface temperature data.

[0048] S3. Combining the auxiliary data described in S1 and the preliminary raw surface temperature data obtained in S2, construct a classical regression random forest model;

[0049] S4. Adjust the parameters of the classical regression random forest model based on the feedback from the out-of-bag validation data until the best results are obtained. Then use the adjusted parameters to obtain the initial fill of the original surface temperature data.

[0050] S5. Based on the preliminary fill-in original surface temperature data obtained in S4, according to the principle that the closer to the pixel, the greater the weight, the more complete the fill-in original surface temperature data and its spatial information are obtained.

[0051] S6. Based on the original surface temperature data and spatial information obtained in S5, and combined with the auxiliary data described in S1, construct a random forest regression model that considers spatial information.

[0052] S7. Adjust the parameters of the random forest regression model considering spatial information based on the out-of-bag validation data feedback until the best results are obtained. Then use the adjusted parameters to obtain the reconstructed land surface temperature considering spatial information.

[0053] S8. Based on the preliminary raw surface temperature data obtained in S2 and S7 respectively and the reconstructed surface temperature considering spatial information, the residuals are obtained. Kriging interpolation regression is used to predict the missing parts of the residuals to obtain the complete residual data.

[0054] S9, the complete residual data obtained by coupling S8, and the reconstructed surface temperature considering spatial information obtained by S7 are used to obtain the final reconstructed surface temperature.

[0055] According to the above scheme, in S1, the spatial resolution of the Landsat satellite remote sensing image data is 30m×30m, and the temporal resolution is 16 days; the terrain data includes latitude and longitude, digital elevation model, and slope data, the slope data is calculated by the digital elevation model, and its spatial resolution is 30m×30m; the spatial resolution of the land use data is 30m×30m; the soil type includes three soil textures, namely sandy soil, clay, and clay, with a spatial resolution of 250m×250m; the reflectance is the reflectance in the mid-infrared band, with a spatial resolution of 250m×250m and a temporal resolution of 1 day; the spectral indices include enhanced vegetation index and solar zenith angle, with a spatial resolution of 250m×250m and a temporal resolution of 1 day.

[0056] According to the above scheme, the temperature inversion algorithm used in S2 is a single-channel algorithm; the cloud removal process is completed using the QA band built into Landsat; the resampling process specifically involves upsampling the 30m×30m spatial resolution surface temperature data to a 100m×100m spatial resolution using bilinear interpolation.

[0057] According to the above scheme, the specific formula of the classic regression random forest model described in S3 is as follows.

[0058] LST ori =f1(x terrain x pop x LUCC， x mir x index x soil )

[0059] Among them, LST ori For raw surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0060] According to the above scheme, in S4, the parameters of the classical regression random forest model subject to feedback adjustment include the number of decision trees, depth, and number of features involved in the modeling; the specific formula for the preliminary filling of the original surface temperature data is as follows.

[0061] LST gapfill =f1(x terrain， x pop xLUCC x mir x index x soil )

[0062] Among them, LST gapfill For preliminary filling of the original surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0063] According to the above scheme, the spatial information acquisition of the original surface temperature data to be filled in, as described in S5, specifically involves...

[0064]

[0065] Among them, LST sf This represents the spatial characteristics of the Earth's surface temperature at the center pixel m, where N represents the total number of pixels and n represents the neighboring pixels near the center pixel m. LST represents the spatial distance between pixels m and n. gapfill_n It is the initial fill of the raw surface temperature data at pixel n.

[0066] According to the above scheme, the random forest regression model in S6 that considers spatial information is specifically as follows:

[0067] LST x =f2(x terrain x pop x LUCC x mir x index x soil x sf )

[0068] Among them, LST x To fill in the original surface temperature data, x sf To fill in the spatial information of the original surface temperature data, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil Soil type.

[0069] According to the above scheme, the method for obtaining the reconstructed surface temperature considering spatial information in S7 is as follows.

[0070] LSTsfrfrm =f2(x terrain x pop x LUCC x mir x index x soil x sf )

[0071] Among them, LST sfrfrm To reconstruct surface temperature taking into account spatial information, x terrain For terrain data, x pop For population density, x LUCC For land use data, x mir x is the reflectivity. index x is the spectral index. soil For soil type, x sf Spatial information to fill in the original surface temperature data.

[0072] According to the above scheme, the specific method for obtaining complete residual data using kriging interpolation in S8 is as follows:

[0073]

[0074] Where ε is the complete residual data after prediction, and λ i is the weight at pixel i, and n is the total number of pixels.

[0075] According to the above scheme, the specific method for obtaining the final reconstructed landmark temperature in S9 is as follows:

[0076] LST sfrfkrm =LST sfrfrm +ε

[0077] Among them, LST sfrfkrm For the final reconstruction of surface temperature, LST sfrfrm To reconstruct the surface temperature taking into account spatial information, ε represents the complete residual data.

[0078] Final Reconstruction of Land Surface Temperature Prediction Accuracy Evaluation: The accuracy of the final reconstructed land surface temperature prediction in S9 was verified by cross-validation, comparing it with the original land surface temperature, and a scatter plot was plotted. The coefficient of determination, root mean square error, and mean absolute error were used as evaluation parameters for the cross-validation. The accuracy evaluation results are as follows: Figure 4 As shown in Table 1, the accuracy evaluation metrics of ordinary random forest and the accuracy evaluation metrics of this method are compared.

[0079]

[0080]

[0081]

[0082] Table 1 Comparison of prediction results of this method with the accuracy of simple random forest estimation.

[0083]

[0084] As shown in Table 1, the traditional random forest regression method, which combines spatial data information, demonstrates high prediction accuracy for reconstructing cloudless surface temperature data. All three test indicators show varying degrees of improvement.

[0085] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method of machine learning reconstruction of cloud-free land surface temperature incorporating data space information, characterized by: The method comprises the following steps, S1, acquiring Landsat satellite remote sensing image data and auxiliary data including terrain data, land use data, population density, soil type, emissivity and spectral index; S2, inverting the surface temperature of the Landsat satellite remote sensing image data acquired in S1, and then performing cloud removal and resampling processing to obtain preliminary raw surface temperature data; S3, combining the auxiliary data in S1 and the preliminary raw surface temperature data obtained in S2 to construct a classical regression random forest model; S4, adjusting the parameters of the classical regression random forest model according to the out-of-bag validation data of the classical regression random forest model until the best effect is obtained, and then using the adjusted parameters to obtain preliminary filled raw surface temperature data; S5, according to the preliminary filled raw surface temperature data obtained in S4, obtaining filled raw surface temperature data and spatial information according to the principle that the closer to the pixel point, the greater the weight; S6, according to the filled raw surface temperature data and spatial information obtained in S5, combining the auxiliary data in S1 to construct a random forest regression model considering spatial information; S7, adjusting the parameters of the random forest regression model considering spatial information according to the out-of-bag validation data of the random forest regression model considering spatial information until the best effect is obtained, and then using the adjusted parameters to obtain reconstructed surface temperature considering spatial information; S8, obtaining residual error according to the preliminary raw surface temperature data and the reconstructed surface temperature considering spatial information obtained in S2 and S7 respectively, predicting the missing part of the residual error using Kriging interpolation regression to obtain complete residual data; S9, coupling the complete residual data obtained in S8 and the reconstructed surface temperature considering spatial information obtained in S7 to obtain the final reconstructed surface temperature; In S4, the parameters of the classical regression random forest model adjusted by feedback include the number of decision trees, the depth, and the number of features participating in modeling; the specific formula of the preliminary filled raw surface temperature data is as follows, wherein, is the preliminary filled raw land surface temperature data, is the terrain data, is the population density, is the land use data, is the reflectance, is the spectral index, is the soil type; The spatial information of the filled raw surface temperature data in S5 is obtained as follows, wherein, wherein represents the spatial feature of the surface temperature at the center pixel m, N represents the total pixels, n represents the adjacent pixel near the center pixel m, represents the spatial distance between the pixel m and the pixel n, is the preliminary filled original surface temperature data at the pixel n; The method for obtaining the reconstructed surface temperature considering spatial information in S7 is as follows, wherein, is a reconstructed land surface temperature taking into account spatial information, is topographic data, is population density, is land use data, is albedo, is a spectral index, is soil type, is spatial information of the filled raw land surface temperature data.

2. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: In S1, the spatial resolution of the Landsat satellite remote sensing image data is 30m×30m, and the temporal resolution is 16 days; the terrain data includes longitude and latitude, digital elevation model, and slope data, the slope data is calculated from the digital elevation model, and the spatial resolution is 30m×30m; the spatial resolution of the land use data is 30m×30m; the soil type includes three soil textures, namely sandy soil, clay and clay, and the spatial resolution is 250m×250m; the reflectivity is the reflectivity of the middle infrared band, and the spatial resolution is 250m×250m, and the temporal resolution is 1 day; the spectral index includes enhanced vegetation index and solar zenith angle, and the spatial resolution is 250m×250m, and the temporal resolution is 1 day.

3. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: The temperature inversion algorithm used by S2 is a single-channel algorithm; the cloud removal process uses the QA band provided by Landsat; the resampling process specifically upsamples the 30m×30m spatial resolution land surface temperature data to 100m×100m spatial resolution using bilinear interpolation.

4. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: The specific formula of the classical regression random forest model described in S3 is as follows, wherein, is raw land surface temperature data, is topography data, is population density, is land use data, is albedo, is spectral index, is soil type.

5. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: The random forest regression model considering spatial information in S6 is specifically, wherein, is the filled original land surface temperature data, is spatial information of the filled original land surface temperature data, is terrain data, is population density, is land use data, is albedo, is a spectral index, is soil type.

6. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: The method for obtaining complete residual data by Kriging interpolation in S8 is specifically, in, It is the complete residual data after prediction. It is a pixel. The weight of the position, That is the total number of pixels.

7. The machine learning reconstruction of cloud-free land surface temperature method incorporating data space information of claim 1, wherein: The specific acquisition method of the final reconstructed land surface temperature in S9 is, wherein, is the final reconstructed surface temperature, is the reconstructed surface temperature taking into account spatial information, is the complete residual data.

Images