Low-altitude unmanned aerial vehicle spectral radiation correction method and system based on optimized BP neural network

Abstract

The application provides a low-altitude unmanned aerial vehicle spectral radiation correction method and system for optimizing a BP neural network, and comprises the following steps: acquiring and preprocessing low-altitude multispectral image data and calibration data; generating an atmospheric transmittance lookup table, wherein the atmospheric transmittance lookup table comprises atmospheric transmittance data under different conditions; generating a low-altitude unmanned aerial vehicle radiation correction training set combined with the atmospheric transmittance; obtaining an optimal BP neural network training model for the optimization of the single input parameter of the BP neural network; and performing radiation correction by using the optimal BP neural network training model. The application solves the problem of unmanned aerial vehicle image radiation correction, and provides an efficient, reliable and feasible scheme for real-time unmanned aerial vehicle ground object identification and ground object surveying.

Description

Technical Field

[0001] This invention belongs to the field of radiation imaging, specifically relating to a radiometric correction scheme for UAV multispectral remote sensing images that combines an atmospheric transmittance lookup table method with a BP neural network optimized by the BP neural network. Background Technology

[0002] Since the 1970s, satellite remote sensing technology has primarily acquired two-dimensional information about landmarks and features directly through vertical observation of the Earth, or classified target features using spectral characteristics acquired by sensors. With economic development and the continuous improvement of remote sensing technology, simple qualitative remote sensing methods such as feature classification can no longer meet the requirements of researchers. Researchers began to explore using sensors to acquire electromagnetic information about target features, thereby accurately retrieving detailed information about features in remote sensing images—this is quantitative remote sensing. Supported by rapidly developing computer technology, quantitative remote sensing can establish the relationship between remote sensing images and target parameters through complex mathematical models, and perform inversion and calculation of target information or atmospheric parameters using corresponding models. At the same time, the methods of acquiring remote sensing images are becoming increasingly diversified. The rapid development of UAV low-altitude remote sensing has made quantitative remote sensing a method with a wide observation range, simple data acquisition methods, repeatable observations, and high timeliness, and it is widely used in water body monitoring and identification, smart agriculture, ecological environment monitoring, and emergency disaster relief.

[0003] In the application of low-altitude remote sensing by UAVs, a crucial issue is how to deduce the true reflectance of ground features from the DN (Digital Number) values ​​of remotely sensed images acquired by UAVs. Ideally, after solar radiation reaches the Earth's surface, a portion of the energy is absorbed by ground features, and another portion is reflected and received by the UAV. The ratio of this latter portion to the total solar radiation energy is the surface reflectance of the ground feature. However, in actual data acquisition, even at relatively low UAV altitudes, environmental factors and transmission media can still affect the energy reflected by ground features. This leads to absorption and scattering by atmospheric particles, weakening the radiation signal of the target ground feature. Furthermore, background interference prevents the UAV from accurately reflecting the true spectral reflectance, spectral radiance, and other physical characteristics of the target ground feature. Therefore, the radiation signal received by the UAV sensor, before atmospheric correction, cannot accurately reflect the true radiation characteristics of the target ground feature. Although significant progress has been made in atmospheric correction research, most current atmospheric correction methods based on radiative transfer models are designed for satellite sensors, such as the MODTRAN model. These methods can only complete the entire process of radiometric correction for satellite imagery and cannot effectively perform radiometric correction for UAV low-altitude imagery. Furthermore, atmospheric radiometric correction methods for UAV low-altitude remote sensing imagery data suffer from complex model parameters, stringent input conditions, and long computation times, failing to fully utilize the low cost and simple acquisition methods of UAV sensors.

[0004] Meanwhile, due to the low image acquisition altitude of low-altitude UAVs, unlike the radiometric correction process for satellite imagery, the influencing factors of low-altitude remote sensing imagery are more complex and variable, and are significantly affected by terrain. Commonly used linear empirical radiometric correction methods yield a linear relationship between DN values ​​and reflectivity, which is not applicable to situations with complex and variable terrain and uneven environmental conditions. Therefore, a highly robust and accurate radiometric correction method for UAV imagery is needed. Summary of the Invention

[0005] This invention provides a radiometric correction method for UAV multispectral remote sensing images that combines an atmospheric transmittance lookup table method with a BP neural network optimized for this purpose. The true reflectance of ground objects is directly obtained by training and inverting from the DN values ​​of the UAV images, solving the problems of complex parameters and cumbersome calculations required by traditional UAV radiometric correction. The method utilizes a multispectral sensor mounted on a UAV to acquire multispectral remote sensing images of target ground objects. Radiometric correction is performed using an optimal BP neural network model trained with different parameter combinations. An acceleration scheme is proposed to modify the network training process to address specific calculation issues. Furthermore, GPU acceleration is applied to leverage the high repeatability of image processing, enabling rapid acquisition of reflectance information for target ground objects and achieving real-time surveying and characteristic analysis.

[0006] To achieve the above objectives, this invention provides a method for spectral radiometric correction of low-altitude unmanned aerial vehicles (UAVs) using an optimized BP neural network, comprising the following steps:

[0007] Step 1: Acquire and preprocess low-altitude multispectral image data and calibration data;

[0008] Step 2: Generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions;

[0009] Step 3: Generate a low-altitude UAV radiation correction training set that incorporates atmospheric transmittance.

[0010] Step 4: Based on the results obtained in Step 3, optimize the single input parameter of the BP neural network to obtain the optimal BP neural network training model.

[0011] Step 5: Use the optimal BP neural network to train the model for radiation correction.

[0012] Moreover, step 1 is implemented by arranging a reflector, using a drone equipped with a sensor to capture images, and performing geometric correction and image processing.

[0013] Furthermore, in step 2, the atmospheric transmittance data under different conditions includes atmospheric transmittance under different atmospheric profiles and different aerosol model conditions.

[0014] Furthermore, step 3 is implemented as follows:

[0015] Obtain the DN value of the reflector, including using the mean of the three neighborhoods of the center of the reflective cloth or reflector as its DN value on the corresponding image;

[0016] Calculate the solar altitude angle;

[0017] Based on the different input conditions of different images, the atmospheric transmittance lookup table is substituted, and the output transmittance data is used as the necessary input condition. The DN value, altitude value, solar altitude angle and wavelength are used as alternative input parameters, and the reflectivity of the reflector is used as the output parameter. The two are combined to form a training set.

[0018] Furthermore, step 4 is implemented as follows:

[0019] Step 4.1: Combine the input parameters to form a series of input parameter combinations;

[0020] Step 4.2: Calculate the number of hidden layer nodes using Komokov's theorem, and combine it with the input layer to form different combinations;

[0021] Step 4.3: For the training process of the BP neural network, set an adaptive learning rate and optimize the initial weights;

[0022] Step 4.4: Substitute different combinations into the BP neural network for training, and calculate the R value and MSE value of each training result. The optimal result is taken as the best parameter combination.

[0023] Furthermore, step 5 is implemented as follows:

[0024] The input image is divided into bands and pixels to form a DN value array, which is then combined with atmospheric parameters, image parameters, and environmental parameters to form the input parameters to be corrected.

[0025] After inputting into the optimal BP neural network training model, the output results are combined to obtain the reflectance results for each band.

[0026] On the other hand, the present invention provides a low-altitude UAV spectral radiometric correction system with optimized BP neural network, for implementing the low-altitude UAV spectral radiometric correction method with optimized BP neural network as described above.

[0027] Moreover, it includes the following modules,

[0028] The first module is used for acquiring and preprocessing low-altitude multispectral image data and calibration data;

[0029] The second module is used to generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions.

[0030] The third module is used to generate a training set for low-altitude UAV radiation correction that incorporates atmospheric transmittance.

[0031] The fourth module is used to optimize the BP neural network training model based on the results obtained from the third module, taking into account the singleness of the input parameters.

[0032] The fifth module is used to train a model using the optimal BP neural network for radiation correction.

[0033] Alternatively, it may include a processor and a memory, with the memory used to store program instructions and the processor used to call the stored instructions in the memory to execute a low-altitude UAV spectral radiometric correction method based on an optimized BP neural network as described above.

[0034] Alternatively, it may include a readable storage medium storing a computer program that, when executed, implements a low-altitude UAV spectral radiometric correction method for optimizing a BP neural network as described above.

[0035] The present invention has the following positive effects:

[0036] 1) This invention utilizes the atmospheric transmittance lookup table method to pre-calculate atmospheric transmittance under different conditions, avoiding redundant calculations that reduce efficiency. At the same time, it uses MODTRAN to calculate atmospheric transmittance, eliminating the need to measure complex atmospheric parameters. Combined with easily measurable data, it can obtain true and reliable results.

[0037] 2) This invention trains by combining different parameters to explore the optimal parameter combination that affects radiation correction, eliminating the image of different atmospheric conditions in different regions, and has good applicability.

[0038] 3) This invention specifically optimizes the training process of the BP neural network and adjusts the training iteration process, which greatly improves the efficiency of radiometric correction.

[0039] 4) This invention directly inverts the reflectivity of ground objects based on the image DN value and the reflectivity calibrated by the reflector. It is low in cost, high in accuracy, fast in speed and simple to operate. It can effectively improve the efficiency of radiometric correction processing of UAV images and improve the quality of radiometric correction, thus providing a guarantee for UAV real-time ground reconnaissance.

[0040] This invention optimizes and applies MODTRAN and BP neural network technologies to the field of radiometric correction of UAV multispectral images, solving the problem of radiometric correction of UAV images. It provides an efficient, reliable and feasible method for real-time UAV ground feature identification and surveying, with advantages of low cost, high efficiency and strong operability, and has broad application prospects in the fields of ground feature surveying and agricultural remote sensing. Attached Figure Description

[0041] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0042] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0043] This invention utilizes MODTRAN simulation to calculate atmospheric transmittance under different aerosol modes and creates corresponding lookup tables. It studies the inversion of reflectance measured by a reflector based on a BP neural network with optimal parameter combination, eliminating errors caused by different conditions, and proposes acceleration schemes for the network training process and image processing process.

[0044] This embodiment proposes a radiometric correction method for UAV multispectral remote sensing images using a BP neural network optimized with an atmospheric transmittance lookup table. A reflector and reflective cloth are deployed as targets. Multispectral data is collected using a multispectral sensor mounted on a UAV, and the data is preprocessed using Pix4D software. The average DN value of the 3*3 neighborhood of the reflector's center point is used as the reflector's DN value. A training set is formed by combining atmospheric transmittance calculated using MODTRAN with various image and environmental parameters. Different combinations of input parameters are set and fed into the BP neural network for training. The training results of different combinations are statistically analyzed to find the optimal combination. Radiometric correction is performed using the model with the optimal combination of training results. Substituting the required parameters for the optimal input parameters and the original image yields the corrected image. Finally, an acceleration scheme is designed for the model training and image processing processes.

[0045] See Figure 1 This invention proposes a radiometric correction method for UAV multispectral remote sensing images that combines an atmospheric transmittance lookup table method with a BP neural network optimized by the BP neural network. The method includes the following steps:

[0046] Step 1), Low-altitude multispectral image data and calibration data. In this embodiment, it is preferable to use a UAV to collect low-altitude multispectral image data, that is, to acquire and preprocess UAV multispectral images. This step mainly involves taking multispectral images of target objects and targets at different altitudes and preprocessing them using software.

[0047] The implementation method of step 1 in the embodiment is as follows:

[0048] First, reflectors need to be set up at the experimental site. A series of reflectors with a reflectivity of 5%-70% should be placed sequentially, and images should be taken at a height difference of 20m between 20m and 1000m. In practice, reflective cloth can also be used instead.

[0049] Secondly, image geometric correction and image registration are performed. Image geometric distortions are generally divided into two categories: systematic and unsystematic. Systematic errors are usually caused by the sensor itself, are regular and predictable, and can be corrected using sensor models; unsystematic distortions are irregular, and can be caused by instability in the height and attitude of the sensor platform itself, changes in the curvature of the earth and atmospheric refraction, and changes in terrain, etc.

[0050] Generally, geometric calibration involves three steps: 1) selecting control points; 2) establishing a geometric calibration model; and 3) image resampling.

[0051] In this implementation plan, control points will be acquired using GPS instruments, and all control points have been photographed in the field and labeled on RGB imagery. Geometric correction models and image resampling will be performed using the latest commercial drone software, Pix4D, to obtain optimal results.

[0052] Finally, orthophotos are generated. Most drone imagery is acquired in batches, requiring stitching to generate orthophotos for accurate results. Currently, drone orthophoto generation technology is relatively mature, and most commercial drone image post-processing software offers orthophoto generation functionality.

[0053] In this embodiment, all orthophoto generation will be performed using Pix4D software. The implementation process includes the following steps:

[0054] Step 1.1: Arrange reflectors or reflective cloths in sequence at the test site.

[0055] Step 1.2: Use a drone equipped with sensors to capture images.

[0056] Step 1.3: Use Pix4D software to complete geometric correction and image registration, including: 1) selecting control points; 2) establishing a geometric correction model; 3) image resampling.

[0057] Step 1.4: Generate orthophotos using Pix4D software.

[0058] Step 2) Generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions.

[0059] The present invention further proposes the following specific implementation process:

[0060] Step 2.1: Process atmospheric parameters and integrate the required control parameters, sensor parameters, surface parameters, etc.

[0061] Step 2.2: Draft a series of input conditions for the target situation, and make the input condition parameters into control cards according to the requirements of the MODTRAN official documentation.

[0062] Step 2.3: Input the control card into the MODTRAN software to obtain the output card and compile the transmittance data.

[0063] Step 2.4: Integrate atmospheric transmittance data under different atmospheric contour lines, different aerosol patterns, etc., to form an atmospheric transmittance lookup table.

[0064] For ease of implementation and reference, the following implementation suggestions are provided:

[0065] First, you need to create control cards for the input parameters according to the requirements in the MODTRAN software manual. MODTRAN input parameters are divided into five categories:

[0066] 1) Control operating parameters. This includes choosing the appropriate radiative transfer program and whether to perform multiple scattering calculations. These are primarily handled in CARD1, while CARD5 offers the option for multiple repetition calculations.

[0067] 2) Sensor parameters. These include sensor channel parameters and the observed beam (wavelength range). CARD1A has an option to input the sensor channel response function. In CARD1A3, enter the filename of the channel response function. In CARD4, enter the wavelength range for simulation calculation.

[0068] 3) Atmospheric parameters. The atmospheric profile type is determined through options in CARD1. Other specific parameters (horizontal meteorological visibility (VIS), various atmospheric molecular profiles), including aerosols, are mainly set through CARD2.

[0069] 4) Observation of geometric conditions. CARD1 offers options for geometric conditions, while CARD39 primarily provides options for inputting geometric parameters. MODTRAN uses various combinations to input geometric parameters. Choose the option that best suits your calculation needs.

[0070] 5) Surface parameters. These include spectral reflectance, geographic information, etc. CARD1 provides preliminary options for setting surface parameters, and then CARD4 allows for specific settings of the surface parameters based on the parameters set in CARD1.

[0071] The system simulates a series of parameters required by the input control card and calculates atmospheric transmittance under different atmospheric profiles (tropical model, mid-latitude summer model, mid-latitude winter model, near-Arctic summer model, near-Arctic winter model, 1976 US standard model) and different aerosol models (no aerosol model, rural aerosol model - visibility 23km, rural aerosol model - visibility 5km, naval marine aerosol model, marine aerosol model - visibility 23km, urban aerosol model - visibility 5km, tropospheric aerosol model - visibility 5km, convective fog aerosol model - visibility 0.2km, radiation fog aerosol model - visibility 0.5km, desert aerosol model) using MODTRAN, and establishes corresponding lookup tables.

[0072] Step 3) Generate a low-altitude UAV radiation correction training set that incorporates atmospheric transmittance. In this example, a BP neural network training dataset is generated.

[0073] The present invention further proposes the following specific implementation process:

[0074] Step 3.1, obtain the DN value of the reflector / reflective cloth: take the mean of the three neighborhoods of the center of the reflector / reflective cloth as its DN value on the corresponding image.

[0075] Step 3.2, Calculate the solar altitude angle: The formula for calculating the solar altitude angle at any given time is:

[0076]

[0077] Where H is the solar altitude angle, φ is the local latitude, δ is the solar declination for the day, and t is the solar hour angle at that time. The formula for calculating the solar declination angle is:

[0078]

[0079] N represents accumulated days.

[0080] Step 3.3: Use transmittance, DN value, height value, solar altitude angle, wavelength, etc. as input parameters, and use the reflectivity corresponding to the reflector as output parameter to form a training set.

[0081] Specifically, based on the different input conditions of different images, the atmospheric transmittance lookup table is substituted, the output transmittance data is used as the necessary input condition, and the DN value, altitude value, solar altitude angle, and wavelength are used as alternative input parameters. The reflectivity of the reflector / reflective cloth is used as the output parameter, and the two are combined to form a training set.

[0082] In this embodiment, firstly, the solar altitude angle is calculated using latitude, solar declination, solar hour angle, and accumulated day data. Then, since a series of standard ground features with known reflectivity and DN values ​​are required, the DN values ​​of the reflective fabric / plate and sample points for each multispectral image are obtained, and the mean of the three neighborhoods of the center of the reflective fabric / plate is used as its DN value on the corresponding image. Simultaneously, the reflectivity of the corresponding reflective fabric / plate is recorded and organized into the training set data for a BP neural network.

[0083] The number of nodes n in the hidden layer is calculated using Kolmogorov's theorem, as follows:

[0084]

[0085] Where m is the number of input layer nodes, p is the number of output layer nodes, and a is an integer between 1 and 10.

[0086] Step 4) Optimization of the single input parameter of the BP neural network: Different parameter combinations are generated to find the optimal parameter model. Since it is uncertain whether the existing parameters achieve the best prediction effect, the input parameters and the number of hidden layers of the BP neural network are to be changed to find the optimal combination of "input parameters + number of hidden layers" and construct a BP neural network radiometric correction model with the best effect. The output parameters of each combination are the true ground reflectance. It is preferred that the number of input layers be an integer between 2 and 6, and the number of hidden layers be an integer between 3 and 14.

[0087] The present invention further proposes the following specific implementation process:

[0088] Step 4.1: Combine the input parameters to form a series of input parameter combinations.

[0089] Step 4.2: Calculate the number of hidden layer nodes using Komokov's theorem and combine them with the input layer to form different combinations.

[0090] Step 4.3 proposes an acceleration scheme for the network training process, which involves adaptive learning rate and optimization of initial weights.

[0091] Step 4.4: Substitute different combinations into the BP neural network for training, and calculate the R value and MSE value of each training result. The optimal result is taken as the best parameter combination.

[0092] The calculation method using Komokov's theorem is existing technology and will not be described in detail in this invention.

[0093] This invention addresses the slow training speed identified through visualization analysis during training. The main issues are difficulty in converging to a suitable range after parameter initialization and the tendency for oscillations to occur during convergence. The invention proposes an improved method for initializing model weights and designs an adaptive learning rate. It statistically analyzes the error changes after each training iteration and compares these results with previous error statistics. If the error change is small, the learning rate can be increased to find the minimum point more quickly. If the error jumps too large and fails to escape the oscillation region, the learning rate should be decreased to avoid the "minimum trap" and achieve acceleration.

[0094] In this embodiment, the method for initializing the weights is as follows:

[0095] a) Calculate the scaling factor , where n is the number of input layer components and N is the number of hidden layer components.

[0096] b) Initialize the weights of any layer for Random values ​​between.

[0097] c) Update the weights using the following formula:

[0098]

[0099] Adaptive learning rates need to be adjusted according to different training scenarios. After each training session, the training error is calculated. If the error remains within an acceptable range, the learning rate remains unchanged. However, if the error changes significantly, the learning rate should be decreased to avoid local minima. If the error decreases, it indicates that the weight changes are correct, and the learning rate should be increased appropriately. The specific formula is as follows:

[0100]

[0101] in Mean square error, This represents the maximum error rate.

[0102] For each set of parameters, DN value, altitude, solar altitude angle, temperature, and atmospheric transmittance are used as inputs, and reflectance is used as the output. Each set of inputs and outputs constitutes one sample. Training, validation, and test sample points are randomly selected at ratios of 70%, 15%, and 15%, respectively. The optimal model combination is explored using R-value and MSE value as indicators.

[0103] Step 5), use the optimal BP neural network to train the model for radiation correction.

[0104] In practice, the original image and the required environmental parameters are input into the optimal BP neural network training model. Based on the operating environment, an appropriate acceleration scheme is selected to accelerate the image correction process. For example, the user's hardware configuration is detected, and a GPU acceleration scheme or a CPU acceleration scheme is formulated. Finally, the corrected reflectivity image is output through band fusion.

[0105] The present invention further proposes the following specific implementation process:

[0106] Step 5.1: Divide the input image into bands and pixels to form a DN value array, and combine it with atmospheric parameters, image parameters and environmental parameters to form the input parameters to be corrected.

[0107] Step 5.2: After inputting into the optimal BP neural network training model, the output results are combined to obtain the reflectivity results for each band.

[0108] In specific implementation, the method proposed in the technical solution of this invention can be automatically executed by those skilled in the art using computer software technology. System devices for implementing the method, such as computer-readable storage media storing the corresponding computer program of the technical solution of this invention and computer equipment including the computer program running the corresponding computer program, should also be within the protection scope of this invention.

[0109] In some possible embodiments, a low-altitude UAV spectral radiometric correction system with optimized BP neural network is provided, comprising the following modules:

[0110] The first module is used for acquiring and preprocessing low-altitude multispectral image data and calibration data;

[0111] The second module is used to generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions.

[0112] The third module is used to generate a training set for low-altitude UAV radiation correction that incorporates atmospheric transmittance.

[0113] The fourth module is used to optimize the BP neural network training model based on the results obtained from the third module, taking into account the singleness of the input parameters.

[0114] The fifth module is used to train a model using the optimal BP neural network for radiation correction.

[0115] In some possible embodiments, a low-altitude UAV spectral radiometric correction system with an optimized BP neural network is provided, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the stored instructions in the memory to execute a low-altitude UAV spectral radiometric correction method with an optimized BP neural network as described above.

[0116] In some possible embodiments, a low-altitude UAV spectral radiometric correction system with optimized BP neural network is provided, including a readable storage medium storing a computer program, which, when executed, implements a low-altitude UAV spectral radiometric correction method with optimized BP neural network as described above.

[0117] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A method for spectral radiometric correction of low-altitude unmanned aerial vehicles using an optimized BP neural network, characterized in that: Includes the following steps, Step 1: Acquire and preprocess low-altitude multispectral image data and calibration data; Step 2: Generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions; Step 3: Generate a low-altitude UAV radiation correction training set incorporating atmospheric transmittance; the implementation method is as follows: Obtain the DN value of the reflector, including using the mean of the three neighborhoods of the center of the reflective cloth or reflector as its DN value on the corresponding image. DN stands for Digital Number. Calculate the solar altitude angle; Based on the different input conditions of different images, the atmospheric transmittance lookup table is substituted, the output transmittance data is used as the necessary input condition, DN value, height value, solar altitude angle and wavelength are used as alternative input parameters, and the reflectivity of the reflector is used as the output parameter, and the combination is used to form a training set. Step 4: Based on the results obtained in Step 3, optimize the BP neural network to address the single input parameter limitation, thus obtaining the optimal BP neural network training model; the implementation method is as follows. Step 4.1: Combine the input parameters to form a series of input parameter combinations; Step 4.2: Calculate the number of hidden layer nodes using Komokov's theorem, and combine it with the input layer to form different combinations; Step 4.3: For the training process of the BP neural network, set an adaptive learning rate and optimize the initial weights; Step 4.4: Substitute different combinations into the BP neural network for training, and calculate the R value and MSE value of each training result. The optimal result is taken as the best parameter combination. Step 5: Train the model using the optimal BP neural network for radiometric correction, as follows. The input image is divided into bands and pixels to form a DN value array, which is then combined with atmospheric parameters, image parameters, and environmental parameters to form the input parameters to be corrected. After inputting into the optimal BP neural network training model, the output results are combined to obtain the reflectance results for each band.

2. The method for spectral radiometric correction of low-altitude UAVs using an optimized BP neural network according to claim 1, characterized in that: Step 1 is implemented by setting up a reflector, using a drone equipped with a sensor to capture images, and then performing geometric correction and image processing.

3. The method for spectral radiometric correction of low-altitude UAVs using an optimized BP neural network according to claim 1, characterized in that: In step 2, the atmospheric transmittance data under different conditions include atmospheric transmittance under different atmospheric profiles and different aerosol model conditions.

4. A low-altitude unmanned aerial vehicle (UAV) spectral radiometric correction system with optimized BP neural network, characterized in that: This method is used to implement the low-altitude UAV spectral radiometric correction method with an optimized BP neural network as described in any one of claims 1-3.

5. The low-altitude UAV spectral radiometric correction system with optimized BP neural network according to claim 4, characterized in that: Includes the following modules, The first module is used for acquiring and preprocessing low-altitude multispectral image data and calibration data; The second module is used to generate an atmospheric transmittance lookup table, which contains atmospheric transmittance data under different conditions. The third module is used to generate a training set for low-altitude UAV radiation correction that incorporates atmospheric transmittance. The fourth module is used to optimize the BP neural network training model based on the results obtained from the third module, taking into account the singleness of the input parameters. The fifth module is used to train a model using the optimal BP neural network for radiation correction.

6. A low-altitude unmanned aerial vehicle (UAV) spectral radiometric correction system with optimized BP neural network, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the stored instructions in the memory to execute the low-altitude UAV spectral radiometric correction method according to any one of claims 1-3.

7. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which, when executed, implements a method for spectral radiometric correction of low-altitude unmanned aerial vehicles using an optimized BP neural network as described in any one of claims 1-3.

Images