An algorithm for inversion of forest fuel moisture content based on remote sensing data

Abstract

The present invention discloses an algorithm for inverting the moisture content of forest combustibles based on remote sensing data, and belongs to the field of deep learning. The present invention adopts an MLP model, establishes an algorithm for linking remote sensing spectral reflectance with the moisture content of forest canopy vegetation and surface litter, and pre-processes remote sensing data, uses its spectral reflectance to invert the moisture content of canopy vegetation and surface litter, and the final model fitting degree can reach about 0.8. The algorithm also proposes an optimization scheme for poor penetration of optical remote sensing between the canopy and the surface, and also provides a theoretical basis for large-scale determination of regional canopy and surface litter moisture content by remote sensing estimation method.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and specifically relates to an inversion algorithm for forest fuel moisture content based on remote sensing data. Background Art

[0002] At present, the methods for determining the moisture content of forest fuel mainly include the equilibrium moisture content method, meteorological element method and remote sensing estimation method. The equilibrium moisture content method considers the equilibrium moisture content, the initial moisture content of fuel, time and time lag factors under ideal conditions, and then predicts the moisture content change over a period of time through the model. The meteorological element regression method mainly establishes statistical models of various meteorological factors and the moisture content of fuel, mainly including the fire risk scale model method, the comprehensive index method, the Rothermel model and the BEHAVE model. The remote sensing estimation method has been widely used with the development of remote sensing technology. With the rapid development of computers and the advancement of satellite technology, the application direction of remote sensing technology has been expanded to the detection of soil and vegetation moisture in the 1970s. Hyperspectral technology, which appeared in the 1990s, can use optical sensors to obtain spectral data from various regions, and the spectral information mainly comes from fuel.

[0003] Compared with the other two methods, the advantages of remote sensing estimation method are low cost and large measurement scale. At present, there are many studies on the estimation of moisture content of live combustibles using remote sensing spectroscopy technology. However, in practice, the moisture content of dead combustibles is lower than that of live combustibles, so it has a greater impact on the occurrence of fires. Therefore, the use of remote sensing technology to estimate the moisture content of canopy vegetation and surface litter is of great significance in fire forecasting. The traditional estimation of the moisture content of regional combustibles is based on a large amount of manual measured data. Although this method has high accuracy, it is very inefficient, consumes a lot of manpower and material resources, and causes certain damage to the regional ecology.

[0004] Therefore, a new method with convenient acquisition, high timeliness and long detection distance is needed to provide data for the inversion of regional combustible moisture content. Summary of the invention

[0005] In view of the above problems existing in the prior art, the object of the present invention is to provide an algorithm for inverting the moisture content of forest fuels based on remote sensing data.

[0006] In order to solve the above problems, the technical solution adopted by the present invention is as follows:

[0007] An algorithm for inverting forest fuel moisture content based on remote sensing data includes the following steps:

[0008] Step 1: Extract remote sensing data, preprocess the remote sensing data, and cut and select sample points;

[0009] Step 2: Go to the sample point marked in step 1 for field sampling;

[0010] Step 3: Use the MLP deep learning model to invert the moisture content of canopy fuels;

[0011] Step 4: Use the MLP deep learning model to invert the moisture content of surface combustibles.

[0012] The steps of preprocessing remote sensing data are as follows:

[0013] Step 1: Atmospheric correction was performed on the 10m resolution band and the 20m resolution band in turn, and two groups of L2A-level data of the 10m resolution band and two groups of 20m resolution band were obtained respectively;

[0014] Step 2: Resample the data obtained in step 1 to a 10m resolution band using the nearest neighbor method;

[0015] Step 3: Perform band synthesis on the data obtained in step 2 to generate a true color image.

[0016] The field sampling in step 2 is to collect canopy vegetation and litter on the ground at the sample point, and measure and record the latitude and longitude, tree species, temperature and humidity, and atmospheric pressure of the sample point, and calculate the moisture content of all samples by the following formula:

[0017] Absolute moisture content

[0018]

[0019] Relative moisture content

[0020]

[0021] Among them, W H is the wet weight of combustibles (g), W D is the dry weight of combustibles (g).

[0022] The canopy moisture content inversion is to use the MLP deep learning model to select red light, green light, near infrared and two short-wave infrared in the original data as input ends, directly invert the moisture content of the canopy combustibles, and perform machine learning with the data obtained from actual sampling multiple times to optimize the model.

[0023] The inversion of the moisture content of the surface combustibles is to process the original data using the bidirectional reflectance distribution function to obtain multi-angle remote sensing data, which is shown as follows:

[0024]

[0025] Where λ is the wavelength, θ is i is the angle between the incident direction of sunlight and the zenith angle, θr is the angle between the observation direction and the zenith angle, and They refer to the angles of the incident direction and the observation direction in azimuth respectively;

[0026] Substitute the obtained data into the 4-scale model in the radiation transfer model again. The reflectivity relationship of the 4-scale model is:

[0027] R=R T K T +R G K G +R ZT K ZT +R ZG K ZG

[0028] Where: R T represents the reflectance of the canopy illuminated surface;

[0029] K T Represents the probability that the sensor observes the illuminated surface on the ground

[0030] R G Indicates the reflectivity of the illuminated surface of the ground;

[0031] K G Represents the probability that the sensor observes the illuminated surface on the ground

[0032] R ZT represents the canopy background surface reflectance;

[0033] K ZT Represents the probability that the sensor observes the canopy background surface

[0034] R ZG Represents the reflectivity of the ground background surface;

[0035] K ZG Represents the probability that the sensor observes the ground background surface;

[0036] Then use the following formula to obtain the surface remote sensing data;

[0037]

[0038] Where: M is the multiple scattering factor, A, B, C are the relationship between M and K components;

[0039] Finally, the MLP deep learning model was used to select red light, green light, near infrared and two short-wave infrared light from the surface remote sensing data as input to invert the moisture content of surface combustibles. The machine learning was then performed multiple times with the data obtained from actual sampling to optimize the model.

[0040] Compared with the prior art, the present invention uses remote sensing data to invert the moisture content based on the MLP deep learning model, and only uses a small amount of measured data as a test of the accuracy of the model. In addition, the remote sensing data can be used for long-distance detection over a large range, which is convenient to obtain and has high timeliness. The present invention also provides a new method for inverting the moisture content of regional combustibles. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the overall flow of the algorithm of the present invention;

[0042] Figure 2 It is a true color remote sensing image;

[0043] Figure 3 is the distribution of sample points in the sampling area;

[0044] Figure 4 This is the structure diagram of the MLP deep learning model;

[0045] Figure 5 This is a comparison chart of the model training and actual errors of canopy fuel moisture content;

[0046] Figure 6 Comparison chart between the predicted value and the true value of the canopy MLP model:

[0047] Figure 7 This is a comparison chart of the model training and actual errors of surface combustible moisture content;

[0048] Figure 8 This is a comparison chart between the predicted values ​​and the true values ​​of the surface MLP model;

[0049] Fig. 9 This is the inversion effect diagram of the moisture content of canopy fuel;

[0050] Fig.10 This is the inversion effect diagram of the moisture content of surface combustibles. DETAILED DESCRIPTION

[0051] The present invention is further described below in conjunction with specific embodiments.

[0052] Example 1

[0053] The inversion structure diagram of this algorithm is shown in Figure 1 shown.

[0054] First, the remote sensing data of the target area is obtained through the Sentinel-2 satellite. The target area selected in this embodiment is Chongli District, Zhangjiakou City, Hebei Province. The remote sensing data of the study area was generated by the satellite Sentinel-2B on orbit R075 at 3:05:49 on August 25, 2020. Figure 1As shown, the data level is L1C data that has completed geometric correction, radiation calibration and atmospheric reflectance calculation. Data preprocessing operations include atmospheric correction, resampling and cropping. After atmospheric correction of the 10m resolution band and the 20m resolution band by the Sen2Cor plug-in of SNAP (Sentinel Application Platform), four sets of L2A level data (two sets of 10m and 20m resolution) can be obtained. The data is then resampled to 10m resolution using the nearest neighbor method, and then a true color image is generated by band synthesis (R:G:B=Band4:Band3:Band2) ( Figure 1 ) and then cropped to finally obtain the remote sensing image within the study area, and randomly select several sample points (Region Of Interest, ROI) in the vegetation coverage area for field sampling.

[0055] It is necessary to obtain and study the moisture content of the vegetation canopy fuel and the surface litter. 200 sample points were evenly and widely selected from the cropped remote sensing images of the study area and marked on the Chongli District vector boundary map, such as Figure 2 As shown in the figure, the size of each sample point is set to 45m×45m. After the sample points were located on the spot, the canopy vegetation and the litter on the surface of the sample points were collected in the target area by direct acquisition method from July 1, 2021 to September 1, 2021, and the latitude and longitude, tree species, temperature, humidity and atmospheric pressure of the sample points were measured and recorded.

[0056] After retrieving all samples, first weigh the samples and record them as the wet weight of the combustibles. Then put them into a baking oven for continuous constant temperature drying. After the weight becomes constant, measure the weight after drying and record it as the wet weight of the combustibles. Finally, calculate the moisture content of all samples according to Formula 1.

[0057] Absolute Moisture Content (AMC)

[0058]

[0059] Relative Moisture Content (RMC)

[0060]

[0061] Where W_H is the wet weight of the combustible (g), and W_D is the dry weight of the combustible (g).

[0062] The near-infrared band is located in the high reflectivity area of ​​plants and also in the strong absorption area of ​​water bodies. The short-wave infrared band is located between the absorption bands of water bodies, and the moisture content of combustibles has a significant correlation with the spectral reflectivity of these two bands. The spectral moisture index method mainly calculates the spectral index based on the spectral reflectivity and compares it with the measured data to calculate the moisture content of canopy combustibles. For the moisture content of surface litter, the remote sensing data needs to be processed according to the radiation transmission model. First, the remote sensing of the ground is initially obtained through satellite remote sensing data, and the moisture content of the combustibles is inverted through the spectral reflectivity and spectral moisture index.

[0063] At the same time, due to the canopy shading problem, the radiation transfer model is used to solve the canopy shading problem, and the correlation between the moisture content of the combustible material and the spectral reflectance is analyzed based on the MLP model.

[0064] After collecting the remote sensing data and the data collected on the spot, they are input into the computer for deep learning of the inversion model. The present invention adopts the MLP deep learning model.

[0065] The network structure of MLP includes input layer, hidden layer and output layer. It is a commonly used model in deep learning. It learns the characteristics of input data by building a multi-layer neural network. It has strong adaptability and is currently widely used in regression prediction research.

[0066] Based on the above model, the acquired L1C-level data is first preprocessed by atmospheric correction, resampling, etc., and processed into L2A-level remote sensing data, and then the target study area is cropped as the sample point for model inversion. Then, after correlation analysis based on all the bands provided by the remote sensing data, a total of 5 characteristic variables, including red light (B3), green light (B4), near infrared (B8) and two short-wave infrared (B11, B12) bands, are selected as the multi-independent variable input of the MLP model. According to the spectral characteristics of water content, the most suitable MLP deep learning model is selected. The model structure is as follows: Figure 4 As shown in the figure, the two fully connected layers each contain 64 nodes and use ReLU (Rectified Linear Unit) as the activation function, and the output layer uses a linear function as the activation function.

[0067] Based on the MLP deep learning model, the canopy fuel moisture content inversion was performed. The reflectance of 5 bands, B3, B4, B8, B11, and B12, was selected as input. 70% of the data (a total of 140) were selected as training samples. The mean square error (MSE) was used as the loss function, and the training was iterated 1000 times. The comparison between the training error and the actual error during the training process is shown in the figure below. Figure 5As shown in the figure, both can control the error within 1, and the model training effect is good. After the training is completed, 30% of the data in the sample (a total of 60) are selected as test samples, and the model is used to make predictions and compare them with the actual values, and a line graph is drawn. Figure 6 , and calculated the actual fit (R^2), the result was 0.843.

[0068] In addition to the combustibles in the canopy, it is also necessary to invert the moisture content of the fallen combustibles on the ground. The moisture content of the fallen combustibles on the ground is predicted using the bidirectional reflectance distribution function, which is defined as the radiant illumination reflected along the reflection direction (i.e., the observation direction). The radiation intensity of the observed target surface The ratio between them is as follows:

[0069]

[0070] Where λ is the wavelength (nm), θ i is the angle between the incident direction of sunlight and the zenith angle, θ r is the angle between the observation direction and the zenith angle, and They refer to the angles in azimuth between the incident direction and the observation direction respectively.

[0071] The reflectivity relationship of the 4-scale model of the radiation transfer model is:

[0072] R=R T K T +R G K G +R ZT K ZT +R ZG K ZG (4)

[0073] Where: R T represents the reflectance of the canopy illuminated surface;

[0074] K T Represents the probability that the sensor observes the illuminated surface on the ground

[0075] R G Indicates the reflectivity of the illuminated surface of the ground;

[0076] K G Represents the probability that the sensor observes the illuminated surface on the ground

[0077] R ZT represents the canopy background surface reflectance;

[0078] K ZT Represents the probability that the sensor observes the canopy background surface

[0079] R ZG Represents the reflectivity of the ground background surface;

[0080] K ZG Represents the probability that the sensor observes the ground background surface.

[0081] Canopy spectral reflectance R at observation angles α and β α and R β The relationship is as follows:

[0082]

[0083] Where: M is the multiple scattering factor, A, B, C are the relationship between M and K components.

[0084] There is a problem of canopy occlusion when using remote sensing data to invert the moisture content of surface combustibles. The spectral reflectance of the vegetation area is determined by factors such as leaves and soil. It is not a planar rigid body. Radiation can pass through the vegetation canopy and then undergo multiple scattering effects. Finally, it escapes from the upper layer of the vegetation and is received by remote sensing. Remote sensing data obtains a two-dimensional plane model, while the vegetation area is a three-dimensional model. Therefore, in order to obtain the surface reflectance, this paper first uses BRDF to process the original remote sensing data through ENVI to obtain multi-angle remote sensing data, and then substitutes it into the 4-scale model in the radiation transfer model. Based on the measured data and formula 4, the multiple scattering factor M and the observation probability K are obtained, and then the surface remote sensing data is obtained according to formula 5.

[0085] Based on remote sensing data and MLP deep learning model, the inversion of surface litter moisture content selected 5 band reflectances, B3, B4, B8, B11, and B12, as input, 70% of the data (a total of 140) were selected as training samples, and the mean square error was used as the loss function. The training was iterated 1000 times. The comparison between the training error and the actual error during the training process is shown in the figure below: Figure 7 As shown in the figure, the training error is continuously reduced during the training process. After the actual error becomes stable, no iteration is performed to prevent overfitting. The actual mean absolute error is 7.69. After the training is completed, 30% of the data (a total of 60) are selected as test samples, and the model is used to make predictions and compare them with the actual values ​​to draw a line graph. Figure 8 And calculate the actual fit (R 2 ), the calculated result is 0.448.

[0086] The time of the remote sensing data image selected in this embodiment is August 25, 2020, which is consistent with the time period of the field survey from July 1, 2021 to September 1, 2021, and there is no obvious change in the vegetation conditions.

[0087] Fig. 9It is a grayscale map and distribution map of the moisture content in Chongli District inverted based on the MLP model with spectral reflectance as the input variable. The main tree species in the western part of Chongli District are shrubs, and the main tree species in the eastern part are trees. The canopy leaf water absorption process of shrubs and trees is similar. The canopy interception of some shrubs is higher than that of tree canopies. Therefore, the canopy vegetation moisture content in Chongli District will be high in the west and low in the east.

[0088] Fig.10 It is a grayscale map and distribution map of the moisture content in Chongli District inverted based on the MLP model with spectral reflectance as the input variable. The main tree species in the southeastern part of Chongli District are trees. The field investigation period is in midsummer. In summer, the litter under the tree forest has a more obvious effect of intercepting surface runoff and inhibiting the evaporation of soil moisture. The moisture content is higher at this time, so the moisture content of the litter in this area is higher than that in the west.

[0089] The average moisture content of canopy vegetation in the entire Chongli area is 35%, and the average moisture content of surface litter is 52%. The inversion accuracy of canopy fuel moisture content is relatively high, with a fitting degree of 0.843. Although there is a canopy shading problem for surface litter, the inversion model with good accuracy is obtained after processing remote sensing data through the radiation transfer model, with a fitting degree of 0.448.

Claims

1. An algorithm for inversion of forest fuel moisture content based on remote sensing data, characterized in that: The following steps are involved: Step 1: Extract remote sensing data, preprocess the remote sensing data, and cut and select sample points; Step 2: Go to the sample point marked in step 1 for field sampling; Step 3: Use the MLP deep learning model to invert the moisture content of canopy fuels; Step 4: Use the MLP deep learning model to invert the moisture content of surface combustibles; The canopy moisture content inversion is to use the MLP deep learning model to select red light, green light, near infrared and two short-wave infrared in the original data as input ends, directly invert the moisture content of the canopy fuel, and perform machine learning with the data obtained by actual sampling multiple times to optimize the model; The inversion of the moisture content of the surface combustible material is to process the original data using the bidirectional reflectance distribution function to obtain multi-angle remote sensing data, and the formula is as follows: In the formula, is the wavelength, is the angle between the incident direction of sunlight and the zenith angle. is the angle between the observation direction and the zenith angle, and They refer to the angles of the incident direction and the observation direction in azimuth respectively; Substitute the obtained data into the 4-scale model in the radiation transfer model again. The reflectivity relationship of the 4-scale model is: in: represents the canopy illuminated surface reflectance; Represents the probability that the sensor observes the illuminated surface on the ground Indicates the reflectivity of the illuminated surface of the ground; Represents the probability that the sensor observes the illuminated surface on the ground represents the canopy background surface reflectance; Represents the probability that the sensor observes the canopy background surface Represents the reflectivity of the ground background surface; Represents the probability that the sensor observes the ground background surface; Then use the following formula to obtain the surface remote sensing data; Where: M is the multiple scattering factor, A, B, C are the relationship between M and K components; Finally, the MLP deep learning model was used to select red light, green light, near infrared and two short-wave infrared light from the surface remote sensing data as input to invert the moisture content of surface combustibles. The machine learning was then performed multiple times with the data obtained from actual sampling to optimize the model.

2. The forest fuel moisture content inversion algorithm based on remote sensing data according to claim 1 is characterized in that: The steps of preprocessing remote sensing data are as follows: Step 1: Atmospheric correction was performed on the 10m resolution band and the 20m resolution band in turn, and two groups of L2A-level data of the 10m resolution band and two groups of 20m resolution band were obtained respectively; Step 2: Resample the data obtained in step 1 to a 10m resolution band using the nearest neighbor method; Step 3: Perform band synthesis on the data obtained in step 2 to generate a true color image.

3. The forest fuel moisture content inversion algorithm based on remote sensing data according to claim 1 is characterized in that: The field sampling in step 2 is to collect canopy vegetation and litter on the ground at the sample point, measure and record the latitude and longitude, tree species, temperature and humidity, and atmospheric pressure information of the sample point, and calculate the moisture content of all samples by the following formula: Absolute moisture content Relative moisture content in, is the wet weight of combustibles (g), is the dry weight of combustibles (g).

Images