Convolutional neural network-based land surface temperature downscaling method, product, medium and equipment
By building a multi-source high-score database and convolutional neural network model, the problem of spatial and temporal resolution contradiction between satellite remote sensing data is solved, and the land temperature data reduction scale from low resolution to high resolution is achieved, which improves the spatial and temporal resolution of the data is improved. , support urban heat island research.
Patent Information
- Application Number
- CN202411003698.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-25
- Publication Date
- 2024-11-12
AI Technical Summary
There are contradictions in existing satellite remote sensing data in space-time resolution, and it is difficult to meet the needs of high spatial resolution and high temporal resolution at the same time. Especially in urban thermal environment research, resulting in insufficient data support.
Using the land table temperature reduction method based on convolutional neural network, the land table temperature reduction model is constructed by constructing a multi-source high-score database and convolutional neural network downscale model, and low-resolution land table temperature data and high-resolution auxiliary data are trained to achieve low resolution Rate to high resolution data reduction scale, improve spatiotemporal resolution.
It effectively improves the spatial and temporal resolution of land surface temperature data, provides data support with high spatial and temporal resolution, and is suitable for urban heat island research in small-scale areas, solving the problem of spatial and temporal resolution contradiction in the existing technology.
Smart Images

Figure CN118940036A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of land temperature observation technology, and in particular to a land surface temperature downscaling method, product, medium and equipment based on a convolutional neural network. Background Art
[0002] Land Surface Temperature (LST) comprehensively reflects the result of material and energy exchange during the interaction between the earth's surface and the atmosphere. It is an important physical parameter for studying the water and heat balance of the earth system. It is of great significance for understanding surface processes, monitoring environmental changes, and responding to climate change. It is now widely used in soil moisture inversion, evapotranspiration inversion, environmental inversion, urban heat island research, and drought monitoring.
[0003] Satellite remote sensing is an efficient method to obtain LST. In recent years, with the development of remote sensing technology, land surface temperature products obtained by satellite remote sensing have played an irreplaceable role in estimating surface thermal inertia, surface humidity and surface evapotranspiration, measuring the degree of land desertification and urban heat island. However, due to the physical performance limitations of satellite sensors, it is often difficult to obtain both spatial resolution and temporal resolution in a single remote sensing data, which may limit some research or applications that require high spatial resolution and high temporal resolution data. For example, high temporal resolution data is difficult to meet the needs of urban-scale refined land surface temperature research due to its low spatial resolution, while high spatial resolution data cannot study the continuous change of land surface temperature in a short time due to its low temporal resolution. Compared with natural land surfaces, urban underlying surfaces are more complex and their land surface temperatures have higher spatial differences. Therefore, high spatiotemporal resolution land surface temperature is the key data for urban thermal environment research. However, current satellite remote sensing data generally have the problem of spatial and temporal resolution contradiction. Summary of the invention
[0004] The purpose of the present invention is to provide a land surface temperature downscaling method, product, medium and equipment based on convolutional neural network, which can achieve low-resolution to high-resolution land surface temperature data downscaling and improve the spatiotemporal resolution of land surface temperature data.
[0005] To achieve the above object, the present invention provides the following solution.
[0006] In one aspect, the present invention provides a land surface temperature downscaling method based on a convolutional neural network, comprising:
[0007] Acquire high-resolution land surface temperature data and high-resolution auxiliary data of the same phase or adjacent phases and construct a multi-source high-resolution database; the high-resolution auxiliary data includes multi-spectral RGB data, vegetation index, building index and surface elevation;
[0008] Constructing a convolutional neural network downscaling model; the input of the convolutional neural network downscaling model is low-resolution land surface temperature data and high-resolution auxiliary data, and the output is high-resolution land surface temperature data;
[0009] Construct a deep learning training sample set based on a multi-source high-score database;
[0010] The convolutional neural network downscaling model is trained and tested using a deep learning training sample set. After the training is completed, a high-resolution land surface temperature downscaling model is obtained.
[0011] For the low-resolution land surface temperature data to be downscaled, a high-resolution land surface temperature downscaling model is used to downscale them to obtain high-resolution land surface temperature data.
[0012] Optionally, in the multi-source high-resolution database, high-resolution land surface temperature data and high-resolution auxiliary data are stored step by step in the order of data source, spatial resolution, time, data type and data value, and the data are backed up regularly.
[0013] Optionally, the constructing of a convolutional neural network downscaling model specifically includes:
[0014] The high-resolution land surface temperature data is convolved to obtain the first convolution layer Conv1;
[0015] Conv1 is downsampled and subjected to residual convolution to obtain the second convolution layer Conv2;
[0016] Conv2 is downsampled and subjected to residual convolution to obtain the third convolution layer Conv3;
[0017] The low-resolution land surface temperature data with the corresponding spatial range to the high-resolution land surface temperature data is upsampled using the nearest neighbor sampling to the upsampled land surface temperature data, and connected with Conv3 on the channel to obtain the connection layer Cat;
[0018] Cat is convolved to obtain the fourth convolutional layer Conv4;
[0019] Conv4 is upsampled and convolved, and then cross-layered with Conv2 to obtain the fifth convolutional layer Conv5.
[0020] Conv5 is upsampled and convolved, and then cross-layered with Conv1 to obtain the sixth convolutional layer Conv6.
[0021] Conv6 is convolved to obtain high-resolution land surface temperature data and output it.
[0022] Optionally, constructing a deep learning training sample set based on a multi-source high-score database specifically includes:
[0023] According to the temporal and spatial range of low-resolution land surface temperature data, high-resolution auxiliary data and high-resolution land surface temperature data of the same phase or adjacent phase are selected from the multi-source high-resolution database and mask processing is performed;
[0024] The low-resolution land surface temperature data, high-resolution auxiliary data and high-resolution land surface temperature data after mask processing are cropped into multiple groups of training samples of fixed size, which together constitute a deep learning training sample set; each group of training samples includes low-resolution land surface temperature image blocks and corresponding high-resolution auxiliary image blocks and high-resolution land surface temperature image blocks.
[0025] Optionally, the deep learning training sample set is used to train and test the convolutional neural network downscaling model, and after the training is completed, a high-resolution land surface temperature downscaling model is obtained, which specifically includes:
[0026] Set the hyperparameters required for model training and use the deep learning training sample set to train the convolutional neural network downscaling model;
[0027] During the training process, the model parameters are continuously updated to minimize the loss function of the training samples;
[0028] After the model training is completed, an independent validation set is used to test and evaluate the model, and the model is optimized based on the test and evaluation results, ultimately obtaining a high-resolution land surface temperature downscaling model with excellent performance.
[0029] Optionally, downscaling the low-resolution land surface temperature data to be downscaled using a high-resolution land surface temperature downscaling model to obtain high-resolution land surface temperature data specifically includes:
[0030] According to the time information of the low-resolution land surface temperature data to be downscaled, the high-resolution auxiliary data of the nearest phase is searched from the multi-source high-resolution database;
[0031] The high-resolution auxiliary data of the nearest phase are clipped in spatial range to be consistent with the low-resolution land surface temperature data to be downscaled, and the invalid values in the low-resolution land surface temperature data to be downscaled are used to mask the high-resolution auxiliary data of the nearest phase;
[0032] The processed low-resolution land surface temperature data and the high-resolution auxiliary data of the nearest phase are input into the high-resolution land surface temperature downscaling model, and the corresponding high-resolution land surface temperature data are output.
[0033] On the other hand, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the land surface temperature downscaling method based on a convolutional neural network.
[0034] On the other hand, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the land surface temperature downscaling method based on a convolutional neural network.
[0035] On the other hand, the present invention also provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the land surface temperature downscaling method based on a convolutional neural network.
[0036] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:
[0037] In order to solve the problem of temporal and spatial resolution contradiction that is common in current satellite remote sensing data, the present invention proposes a land surface temperature downscaling method, product, medium and equipment based on convolutional neural network. By constructing a multi-source high-resolution database and a convolutional neural network downscaling model, the model is trained using low-resolution land surface temperature data and remote sensing high-resolution information of adjacent time phases (including high-resolution land surface temperature data and high-resolution auxiliary data), so that the model can effectively extract features and achieve downscaling of low-resolution to high-resolution land surface temperature data, thereby improving the temporal and spatial resolution of land surface temperature data, and providing reliable data support for research that requires high temporal and spatial resolution land surface temperature data (such as small-scale regional urban heat island research). BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0039] Figure 1 A flow chart of the land surface temperature downscaling method based on convolutional neural network provided by the present invention;
[0040] Figure 2 A schematic diagram of the principle of the land surface temperature downscaling method based on convolutional neural network provided by the present invention;
[0041] Figure 3 A schematic tree diagram of the hierarchical storage of data in the multi-source high-resolution database of the present invention;
[0042] Figure 4 It is a schematic table diagram of the data storage method in the multi-source high-resolution database of the present invention;
[0043] Figure 5It is a schematic diagram of the convolutional neural network downscaling model structure and processing process of the present invention. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] The present invention aims to provide a land surface temperature downscaling method, product, medium and equipment based on convolutional neural network, which utilizes the spatiotemporal correlation between high-resolution remote sensing information of adjacent time phases and low-resolution land surface temperature data, combined with a convolutional neural network downscaling model, to downscale the low-resolution land surface temperature data to high resolution, thereby improving the spatiotemporal resolution of the land surface temperature data.
[0046] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0047] Land surface temperature (abbreviated as LST) is the direct driving force of the exchange of longwave radiation and ground-atmosphere turbulent heat flux. It is an important input parameter for various land surface process models and environmental simulations. It is now widely used in the fields of soil moisture inversion, evapotranspiration inversion, environmental inversion, urban heat island research, and drought monitoring. Due to the complex underlying surface, the surface temperature is spatially heterogeneous, so high spatiotemporal resolution data is needed to study the spatial distribution and temporal changes of the surface temperature. However, it is difficult for LST data inverted by a single sensor to have both high spatial resolution and high temporal resolution.
[0048] The current land surface temperature downscaling methods are not effective in refining surface temperature studies at small regional scales (such as cities). For example, the land surface temperature downscaling fusion method based on all-weather observation passive microwave data and high spatial resolution optical data can downscale the resolution of 25km to 1km, but still cannot provide land surface temperature data with a spatial resolution of hundreds of meters or meters. The land surface temperature downscaling method based on statistical regression only considers the linear relationship between land surface temperature and other driving factors, and cannot be applied to areas with complex underlying surfaces. The land surface temperature downscaling methods based on machine learning models such as random forests and support vector machines have weak generalization capabilities and perform poorly on data other than training data.
[0049] In view of the above problems, the present invention proposes a land surface temperature downscaling method based on convolutional neural network, which is used to generate high-resolution (high-resolution) land surface temperature (LST) data from low-resolution (low-resolution) LST data. The key to the method of the present invention is to use the same phase or adjacent phase data to build a multi-source high-resolution database, build and train a convolutional neural network downscaling model, and use the model to downscale the low-resolution LST data to be downscaled. The process of the method of the present invention is as follows Figure 1 As shown, the principle is as Figure 2 shown.
[0050] See also Figure 1 and Figure 2 The present invention provides a land surface temperature downscaling method based on a convolutional neural network, comprising:
[0051] Step 1: Obtain high-resolution land surface temperature data and high-resolution auxiliary data of the same or adjacent phases and construct a multi-source high-resolution database.
[0052] Search and download high-resolution / low-resolution LST data, multispectral RGB data, vegetation index (NDVI), building index (NDBI), surface elevation (DEM), slope, longitude and latitude data from relevant data websites (official websites, platforms, scientific research institutions, etc.), whose data sources are MODIS, Landsat, Sentinel, AVHRR, etc. RGB, NDVI, NDBI, DEM, slope, longitude and latitude data are collectively referred to as high-resolution auxiliary data. The collected and acquired high-resolution / low-resolution LST data and related high-resolution auxiliary data need to be data of the same phase or adjacent phases, and all data should be stored step by step in the order of data source, spatial resolution, time, data type and data value, and the data should be backed up regularly to prevent data loss. The data can be stored step by step in the following way: Figure 3 and Figure 4 As shown, data can be classified and stored layer by layer through a tree structure or a table structure, and each level represents a different classification standard. This classification method helps to organize and manage large amounts of data, making it easier to find, access, and analyze data. Building a multi-source high-score database and storing data layer by layer is the basis. The subsequent steps require obtaining data from the built multi-source high-score database.
[0053] The present invention uses high-resolution data of the same phase or adjacent phase (including high-resolution LST data and related high-resolution auxiliary data) to build a multi-source high-resolution database to provide comprehensive surface information. The establishment of a multi-source high-resolution database provides a basis for obtaining high-resolution information and provides reliable support for the processing and analysis of land surface temperature data.
[0054] Step 2: Build a convolutional neural network downscaling model.
[0055] A convolutional neural network (CNN) is used to build a downscaling model. The model structure is as follows: Figure 5 As shown. The training data of the model consists of low-score LST data and corresponding high-score auxiliary data and high-score LST data from a multi-source high-score database. The input of the model is low-score LST data and high-score auxiliary data, and the output is high-score LST data.
[0056] See also Figure 5 , the specific construction steps of the feedforward part of the convolutional neural network downscaling model are as follows:
[0057] Step 2.1: Perform a 3×3 convolution operation on the high-resolution LST data with a pixel size of (h, w, c) to obtain the first convolution layer Conv1 with a size of (h, w, 256); where h and w are the height and width of the data, and c is the number of channels of the data;
[0058] Step 2.2: Conv1 is subjected to a 2x downsampling and 3×3 residual convolution operation to obtain the second convolution layer Conv2 with a size of (h / 2, w / 2, 256);
[0059] Step 2.3: Conv2 is subjected to a 2x downsampling and 3×3 residual convolution operation to obtain the third convolution layer Conv3 with a size of (h / 4, w / 4, 256);
[0060] Step 2.4: Use the nearest neighbor to sample the low-resolution LST data in the spatial range corresponding to the high-resolution LST data to upsample the LST data to a size of (h / 4, w / 4, 1), and connect it with Conv3 on the channel to obtain a connection layer Cat with a size of (h / 4, w / 4, 257);
[0061] Step 2.5: Cat is subjected to a 3×3 convolution operation to obtain the fourth convolution layer Conv4 with a size of (h / 4, w / 4, 256);
[0062] Step 2.6: Conv4 is subjected to a 2x upsampling and 3×3 convolution operation, and is cross-connected with Conv2 to obtain the fifth convolutional layer Conv5 with a size of (h / 2, w / 2, 256);
[0063] Step 2.7: Conv5 is subjected to a 2x upsampling and 3×3 convolution operation again, and is cross-connected with Conv1 to obtain the sixth convolutional layer Conv6 with a size of (h, w, 256);
[0064] Step 2.8: Perform a 3×3 convolution operation on Conv6 to obtain high-resolution LST data of size (h, w, 1) and output it.
[0065] The model structure of the present invention is introduced below by taking 100m low-resolution LST data and 20m high-resolution data as examples. The model structure is divided into two parts. One part uses the nearest neighbor method to upsample the 100m low-resolution LST data to 60m upsampled LST data; the other part has two convolution sampling units. First, the 20m high-resolution auxiliary data is convolved to extract the image features of the low-resolution LST data to obtain the convolution layer Conv1. Then, residual convolution is performed on Conv1 and downsampled to 40m to obtain the convolution layer Conv2. Repeat the previous operation on Conv2 to obtain the convolution layer Conv3 with a resolution of 60m. At this time, Conv3 and the upsampled LST data of 60m are channel-connected to obtain the connection layer Cat. Then convolve Cat to learn the image features and obtain the convolution layer Conv4. Then, convolve Conv4 and upsample it to 40m, and add it to the intermediate layer output Conv2 to obtain the convolution layer Conv5. Repeat the previous operation on Conv5 and add it to the intermediate layer output Conv1 to obtain the 20m resolution convolution layer Conv6. Then convolve Conv6 to obtain 20m high-resolution LST data.
[0066] The connection and addition in the convolution sampling unit can pass the learned surface temperature information to the deeper layers of the network. For surface temperature data with strong surface heterogeneity and high dynamic changes, this can more comprehensively extract and map surface temperature features at different levels to prevent the loss of shallow features. Compared with ordinary convolutional neural networks, it can better utilize effective information and is conducive to enhancing the surface temperature downscaling effect.
[0067] Step 3: Construct a deep learning training sample set based on a multi-source high-score database.
[0068] The convolutional neural network downscaling model of the present invention adopts an end-to-end supervised learning method. Supervised learning cannot be performed without annotation, and the annotation needs to be consistent with the final output in data format and data form. Since it is difficult to directly obtain high-resolution land surface temperature data suitable for refined surface temperature research through satellite remote sensing, the method of the present invention is proposed. In the method of the present invention, only a weak form of self-similarity is required, that is, the constructed convolutional neural network downscaling model only needs to learn how to transmit high-frequency details from the existing high-resolution bands to achieve downscaling of land surface temperature data. The high-resolution land surface temperature data stored in the multi-source high-resolution database is used as annotation information to prepare samples required for model training. The annotation acquisition methods include manual annotation, manual revision, and acquisition from existing historical data. Manual participation in annotation revision can only meet the sample production requirements. Assuming that the high-resolution land surface temperature data ultimately required is daily in a certain area, manual participation in annotation revision will not be able to meet the requirements in terms of time efficiency. Historical data may not necessarily meet the final demand. Assuming that there is historical data of high-resolution land surface temperature in region A, and what is ultimately needed is high-resolution land surface temperature data in region B, then the high-resolution land surface temperature data in region A can be used to prepare samples, and the model can be trained using the method of the present invention, and the model can be used for downscaling in region B. Or assuming that what is ultimately needed is land surface temperature data of region A with a resolution of 100m for multiple consecutive days, when Landsat satellite 100m resolution land surface temperature data is selected as a sample, high-resolution land surface temperature data that meets the time requirements cannot be directly obtained due to the problem of time resolution. In this case, we can only find some high-resolution data when passing through region A to prepare samples, train the model using the method of the present invention, and then use the model for downscaling of low-resolution daily land surface temperature in region A.
[0069] The low-resolution LST and high-resolution LST in the deep learning training sample set are obtained from relevant data websites. When the low-resolution LST of a given time resolution and spatial range can be directly obtained, the more mature and reliable LST data on the relevant data website are directly used; when the low-resolution LST data of a given time resolution and spatial range cannot be directly obtained, the high-resolution LST can be obtained by downsampling according to the spatial resolution of the existing low-resolution LST. The high-resolution LST data in the deep learning training sample set is obtained from a multi-source high-resolution database that stores data step by step according to the spatial resolution to be super-resolved.
[0070] Specifically, when constructing the deep learning training sample set, according to the spatial range and time range of the low-resolution LST data, high-resolution auxiliary data and high-resolution LST data of the same phase or adjacent phase are selected from the multi-source high-resolution database and mask processing is performed. When selecting high-resolution LST of the same phase or adjacent phase as the low-resolution LST from the multi-source high-resolution database, the data closest to the low-resolution LST date and with good quality are preferred. For example, the high-resolution LST of MODIS is generally considered to be adjacent to the low-resolution LST data 3 days before and after the time. When the data quality is good and the ground objects have not changed, the high-resolution LST data and high-resolution auxiliary data of the previous day or the next day are preferred as training data. The LST in some areas may be affected by seasonal changes, climate fluctuations, natural disasters and other factors, which can be determined according to research needs and data availability. Selecting an appropriate adjacent time range can improve the spatiotemporal continuity and accuracy of the model and help better capture the changes and trends of LST.
[0071] When selecting high-resolution auxiliary data of the same phase or adjacent phase, it is necessary to select one or more high-resolution auxiliary data according to the actual situation. For example, RGB, NDVI, DEM, slope and longitude and latitude have different effects on the spatial distribution pattern of LST. Generally, high-resolution RGB data contains the most high-resolution information. NDVI and slope will have global effects, and DEM and longitude and latitude will have local effects. Increasing high-resolution auxiliary data can improve the effect of LST downscaling, but it does not mean that the more high-resolution auxiliary data, the better the downscaling effect. Appropriate and representative high-resolution auxiliary data can provide more comprehensive and accurate surface information, thereby improving the downscaling effect of LST. For example, when doing urban heat island, using high-resolution RGB as high-resolution auxiliary data is the best choice; when doing high-resolution LST in non-urban areas, if the land type in the area is mainly vegetation and there is relatively good high-resolution NDVI data, NDVI can be used as high-resolution auxiliary data; if the altitude in the area has a great impact on LST and there is good DEM data, DEM can be considered as high-resolution auxiliary data.
[0072] It should be noted that masking should also be performed when making samples, which is not necessary during model training. Masking invalid values or outliers in training data (such as clouds, shadows, water bodies, etc.) can prevent them from interfering with model training and prediction results, thereby improving the stability and accuracy of the model.
[0073] The low-resolution LST data, high-resolution auxiliary data and high-resolution LST data after mask processing are cropped into multiple sets of training samples of fixed size, which together constitute a deep learning training sample set. Each set of training samples includes a low-resolution LST image block and the corresponding high-resolution auxiliary image block and high-resolution LST image block. The low-resolution LST image block can be obtained by downsampling the high-resolution LST image block, or by obtaining the real low-resolution LST.
[0074] Step 4: Use the deep learning training sample set to train and test the convolutional neural network downscaling model. After training, a high-resolution land surface temperature downscaling model is obtained.
[0075] Set the hyperparameters such as the cycle, batch size, and learning rate required for model training, and use the cropped samples to train the convolutional neural network downscaling model.
[0076] After preprocessing, the data is input into the convolutional neural network downscaling model for training. During the training process, the model parameters are continuously updated to minimize the loss function of the training data; its weights and parameters are optimized to improve the performance and accuracy of the model. After the model training is completed, it is tested and evaluated using independent validation set data, and the model is optimized based on the test and evaluation results. Finally, a high-resolution land surface temperature downscaling model with excellent performance is obtained.
[0077] Step 5: For the low-resolution land surface temperature data to be downscaled, use the high-resolution land surface temperature downscaling model to downscale and obtain high-resolution land surface temperature data.
[0078] Input high-resolution auxiliary data and low-resolution LST data to be downscaled, use the trained model to downscale the low-resolution land surface temperature, and finally obtain high-resolution land surface temperature data. The specific steps are as follows:
[0079] Step 5.1: Based on the time information of the low-resolution LST data to be downscaled, search and obtain the high-resolution auxiliary data of the nearest phase from the multi-source high-resolution database, where the data source, spatial resolution, and data type of the high-resolution auxiliary data are consistent with those of the data when the sample is cropped;
[0080] Step 5.2: Clip the high-resolution auxiliary data of the nearest phase in spatial range to be consistent with the low-resolution LST data to be downscaled, and use the invalid values in the low-resolution LST data to be downscaled to mask the high-resolution auxiliary data of the nearest phase, so that the invalid value area does not participate in the downscaling operation;
[0081] Step 5.3: Input the processed low-resolution LST data and the high-resolution auxiliary data of the nearest phase into the trained high-resolution LST downscaling model to output the corresponding high-resolution LST data.
[0082] The method of the present invention applies deep learning algorithms to satellite observation systems, which can solve the problems of land surface temperature at the hundred-meter level and meter level in existing downscaling methods. By constructing a multi-source high-resolution database and applying a downscaling model, the method of the present invention can significantly improve the processing accuracy and efficiency of land surface temperature data, providing important support for the real-time application of satellite observation systems.
[0083] In some embodiments, the present invention also provides a computer program product, including a computer program, which, when executed by a processor, implements the land surface temperature downscaling method based on a convolutional neural network.
[0084] The land surface temperature downscaling method based on convolutional neural network proposed in the present invention constructs a multi-source high-resolution database, selects high-resolution land surface temperature and high-resolution data of adjacent phases with corresponding spatial and temporal resolutions from it to generate deep learning training samples, then constructs a convolutional neural network downscaling model structure and uses the samples to train the model, finally obtains high-resolution auxiliary data of adjacent phases from the multi-source high-resolution database according to the low-resolution land surface temperature data to be downscaled, inputs the low-resolution LST data and the high-resolution auxiliary data into the trained model, and outputs high-resolution land surface temperature data.
[0085] Among them, the multi-source high-resolution database constructed by the present invention stores data in a step-by-step order according to the data source, spatial resolution, time, data type, and data value. The multi-source high-resolution database is mainly used to generate deep learning training samples for land surface temperature downscaling and to search and obtain high-resolution auxiliary data of adjacent time phases when downscaling low-resolution land surface temperature.
[0086] The convolutional neural network downscaling model constructed by the present invention first learns the underlying texture features of high-resolution auxiliary data through a small number of residual convolutions and downsampling operations, then connects the resampled low-resolution land surface temperature with the deep learning feature layer on the channel, and finally outputs the high-resolution land surface temperature through convolution and upsampling operations and cross-layer connections.
[0087] When constructing the corresponding relationship between low-resolution land surface temperature and high-resolution data, in order to cope with the problem of difficulty in obtaining low-resolution land surface temperature and high-resolution auxiliary data of the same phase due to the low temporal resolution of high spatial resolution data, the present invention proposes to establish a relationship between low-resolution land surface temperature and high-resolution data of adjacent phases based on the small changes and high similarity of land features in the short term, which can significantly improve the processing accuracy and efficiency of land surface temperature data.
[0088] In some embodiments, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the land surface temperature downscaling method based on a convolutional neural network.
[0089] In some embodiments, the present invention further provides a computer device, which includes a processor, a memory, an input / output interface (Input / Output, referred to as I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store pending transactions. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the land surface temperature downscaling method based on a convolutional neural network can be implemented.
[0090] The land surface temperature downscaling method based on convolutional neural network of the present invention firstly constructs a multi-source high-resolution database containing remote sensing image data from high-resolution satellites based on the high-resolution remote sensing information of the adjacent time phase, then constructs a convolutional neural network downscaling model, and uses low-resolution land surface temperature data and high-resolution data of the adjacent time phase selected from the multi-source high-resolution database as auxiliary data to train the model. The model extracts effective features from it and uses these features to downscale the land surface temperature data. Finally, the model is used to downscale the low-resolution land surface temperature data to obtain high-resolution land surface temperature data.
[0091] The multi-source high-resolution database constructed by the present invention can provide high-quality, easy-to-obtain, and fast-updated remote sensing high-resolution image data, which can improve the downscaling efficiency and reduce the downscaling cost. In addition, by constructing a convolutional neural network downscaling model and utilizing the spatiotemporal correlation between low-resolution land surface temperature data and the adjacent phase remote sensing high-resolution information in the multi-source high-resolution database, the accuracy of the downscaled land surface temperature data can be improved. At the same time, the powerful generalization ability of the convolutional neural network enables the method of the present invention to be applied to various types of research areas, providing land surface temperature data with high spatiotemporal resolution for urban-scale refined surface temperature research.
[0092] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A land surface temperature downscaling method based on convolutional neural network, characterized in that: include: Acquire high-resolution land surface temperature data and high-resolution auxiliary data of the same phase or adjacent phases and construct a multi-source high-resolution database; the high-resolution auxiliary data includes multi-spectral RGB data, vegetation index, building index and surface elevation; Constructing a convolutional neural network downscaling model; the input of the convolutional neural network downscaling model is low-resolution land surface temperature data and high-resolution auxiliary data, and the output is high-resolution land surface temperature data; Construct a deep learning training sample set based on a multi-source high-score database; The convolutional neural network downscaling model is trained and tested using a deep learning training sample set. After the training is completed, a high-resolution land surface temperature downscaling model is obtained. For the low-resolution land surface temperature data to be downscaled, a high-resolution land surface temperature downscaling model is used to downscale them to obtain high-resolution land surface temperature data.
2. The land surface temperature downscaling method based on convolutional neural network according to claim 1, characterized in that: In the multi-source high-resolution database, high-resolution land surface temperature data and high-resolution auxiliary data are stored step by step in the order of data source, spatial resolution, time, data type and data value, and the data are backed up regularly.
3. The land surface temperature downscaling method based on convolutional neural network according to claim 1, characterized in that: The constructing of the convolutional neural network downscaling model specifically includes: The high-resolution land surface temperature data is convolved to obtain the first convolution layer Conv1; Conv1 is downsampled and subjected to residual convolution to obtain the second convolution layer Conv2; Conv2 is downsampled and subjected to residual convolution to obtain the third convolution layer Conv3; The low-resolution land surface temperature data with the corresponding spatial range to the high-resolution land surface temperature data is upsampled using the nearest neighbor sampling to the upsampled land surface temperature data, and connected with Conv3 on the channel to obtain the connection layer Cat; Cat is convolved to obtain the fourth convolutional layer Conv4; Conv4 is upsampled and convolved, and then cross-layered with Conv2 to obtain the fifth convolutional layer Conv5. Conv5 is upsampled and convolved, and then cross-layered with Conv1 to obtain the sixth convolutional layer Conv6. Conv6 is convolved to obtain high-resolution land surface temperature data and output it.
4. The land surface temperature downscaling method based on convolutional neural network according to claim 1, characterized in that: The deep learning training sample set is constructed based on the multi-source high-score database, specifically including: According to the temporal and spatial range of low-resolution land surface temperature data, high-resolution auxiliary data and high-resolution land surface temperature data of the same phase or adjacent phase are selected from the multi-source high-resolution database and mask processing is performed; The low-resolution land surface temperature data, high-resolution auxiliary data and high-resolution land surface temperature data after mask processing are cropped into multiple groups of training samples of fixed size, which together constitute a deep learning training sample set; each group of training samples includes low-resolution land surface temperature image blocks and corresponding high-resolution auxiliary image blocks and high-resolution land surface temperature image blocks.
5. The land surface temperature downscaling method based on convolutional neural network according to claim 1, characterized in that: The deep learning training sample set is used to train and test the convolutional neural network downscaling model, and a high-resolution land surface temperature downscaling model is obtained after the training is completed, specifically including: Set the hyperparameters required for model training and use the deep learning training sample set to train the convolutional neural network downscaling model; During the training process, the model parameters are continuously updated to minimize the loss function of the training samples; After the model training is completed, an independent validation set is used to test and evaluate the model, and the model is optimized based on the test and evaluation results, ultimately obtaining a high-resolution land surface temperature downscaling model with excellent performance.
6. The land surface temperature downscaling method based on convolutional neural network according to claim 1, characterized in that: The method of downscaling the low-resolution land surface temperature data to be downscaled using a high-resolution land surface temperature downscaling model to obtain high-resolution land surface temperature data specifically includes: According to the time information of the low-resolution land surface temperature data to be downscaled, the high-resolution auxiliary data of the nearest phase is searched from the multi-source high-resolution database; The high-resolution auxiliary data of the nearest phase are clipped in spatial range to be consistent with the low-resolution land surface temperature data to be downscaled, and the invalid values in the low-resolution land surface temperature data to be downscaled are used to mask the high-resolution auxiliary data of the nearest phase; The processed low-resolution land surface temperature data and the high-resolution auxiliary data of the nearest phase are input into the high-resolution land surface temperature downscaling model, and the corresponding high-resolution land surface temperature data are output.
7. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the land surface temperature downscaling method based on a convolutional neural network described in any one of claims 1 to 6 is implemented.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the land surface temperature downscaling method based on a convolutional neural network described in any one of claims 1 to 6 is implemented.
9. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the land surface temperature downscaling method based on a convolutional neural network as described in any one of claims 1 to 6.
Citation Information
Patent Citations
Parallel meteorological element downscaling method and device, electronic equipment and storage medium
CN110619605A
Visual image enhancement generation method, system and device and storage medium
CN113066013A
GWDM-based surface temperature downscaling method
CN114972986A
Method for reconstructing cloud-free surface temperature by machine learning in combination with data space information
CN115907036A
Mode forecast product downscaling method based on deep learning
CN116467946A
