Water and soil conservation monitoring method based on remote sensing image

Abstract

The invention relates to the technical field of water and soil conservation monitoring, and discloses a remote sensing image-based water and soil conservation monitoring method, which comprises the following steps of: 1, integrating optical, infrared and microwave remote sensing data, and constructing a soil information atlas; step 2, capturing soil humidity change in each season by using long-time-sequence data according to the soil information, and determining the influence of rainfall, snow melting and evaporation on the reflectivity; step 3, training data of each wave band through a deep neural network according to the influence information of the reflectivity; through an adaptive learning algorithm and a model parameter regular updating mechanism, the problem that in the prior art, characteristic spectrums of wet soil, vegetation and bare soil are overlapped due to seasonal changes is effectively solved, after the seasonal modulation factors are introduced, the system can accurately capture influences of natural factors such as rainfall, snow melting and evaporation on the soil reflectivity, and the soil reflectivity is improved. And model parameters are adaptively adjusted, so that the interference of reflectivity fluctuation on a monitoring result in wet and drought states is eliminated, and the stability of long-term monitoring is ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of soil and water conservation monitoring, and specifically to a soil and water conservation monitoring method based on remote sensing images. Background Technique

[0002] The soil and water conservation monitoring method based on remote sensing images is a technical system that uses satellite or aerial remote sensing technology to periodically obtain electromagnetic wave information reflected or radiated by the earth's surface, and combines geographic information system (GIS) and computer algorithms to dynamically monitor and analyze key soil and water conservation indicators such as soil erosion, vegetation cover, and surface morphology. Its core advantages lie in the ability to obtain data on a large scale, with high frequency and low cost, as well as the comprehensive analysis ability at multiple spatio-temporal scales.

[0003] For example, the publication number is CN116934753A. This invention relates to the technical field of image processing and proposes a soil and water conservation monitoring method based on remote sensing images, including: performing edge detection on remote sensing images of soil and water conservation areas in two adjacent months respectively to obtain erosion gully edge lines, calculating the midlines of erosion gullies, and then obtaining the gully spans and calculating the gully span ratios; obtaining the highest point of the local soil surface area and the deepest point of the suspected soil erosion area based on the gray values of the pixel points on both sides of the pixel points on the erosion gully edge lines, and then calculating the unilateral gully subsidence degree to obtain the soil erosion significance; obtaining the soil erosion degree weight image, extracting the soil erosion area, and then evaluating the soil erosion degree. The purpose of this invention is to solve the problem of low accuracy when traditional image segmentation is used to segment remote sensing images of soil and water conservation areas.

[0004] However, in the prior art, factors such as precipitation, snowmelt, and seasonal evaporation will cause the soil moisture to fluctuate in different seasons. The reflectance characteristics of the soil in the wet state are significantly different from those in the dry state, which will cover up and mislead the soil stability and erosion conditions. In addition, the change of soil moisture will affect the reflectance of each band in the remote sensing image, especially in the infrared and short-wave infrared bands. The wet soil will have spectral overlap with the characteristics of vegetation and bare soil, making it difficult for the automatic classification algorithm to accurately distinguish different surface features, resulting in inaccurate monitoring results.

[0005] Therefore, those skilled in the art provide a soil and water conservation monitoring method based on remote sensing images to solve the above-mentioned problems. Summary of the Invention

[0006] Aiming at the deficiencies of the prior art, the present invention provides a soil and water conservation monitoring method based on remote sensing images to solve the problems raised in the above background technique.

[0007] To achieve the above objectives, the present invention is realized through the following technical solutions: A soil and water conservation monitoring method based on remote sensing images, including:

[0008] Step 1: Integrate optical, infrared, and microwave remote sensing data to construct a soil information atlas;

[0009] Step 2: Use long-term data to capture soil moisture changes in each season based on the soil information, and clarify the impacts of precipitation, snowmelt, and evaporation on reflectivity;

[0010] Step 3: Train the data of each band through a deep neural network based on the information on the impact of reflectivity to automatically distinguish wet soil, vegetation, and bare soil;

[0011] Step 4: According to the data of each band, combine the real-time humidity data of ground IoT sensors with a physical model to correct the remote sensing data and timely correct the spectral deviation caused by humidity changes;

[0012] Step 5: Update the model parameters regularly according to the corrected information to ensure that the monitoring system can be continuously optimized according to different climate and season conditions.

[0013] Preferably, the optical remote sensing data includes image data from high-resolution satellites and aerial platforms, which is used to provide surface detail information; the infrared remote sensing data includes thermal infrared images, which are used to capture the surface temperature distribution, thereby assisting in reflecting the soil moisture state; the microwave remote sensing data is synthetic aperture radar (SAR) data, and its all-weather cloud-penetrating ability is used to detect soil moisture changes.

[0014] Preferably, the long-term data covers at least one year or a longer time period to fully reflect the impact changes of factors such as precipitation, snowmelt, and evaporation on soil reflectivity. The specific steps include:

[0015] Step 1.1 Obtain multi-source remote sensing data covering at least one year or a longer time period, and simultaneously collect the corresponding meteorological data to ensure that the data can fully reflect the environmental changes in different seasons;

[0016] Step 1.2 Perform radiometric correction, geometric correction, and spatio-temporal alignment on the data of each time phase to eliminate the differences between sensors and in time, and perform normalization processing on the meteorological data to ensure the comparability between each data source. The normalization processing algorithm formula is:

[0017]

[0018] where x represents the original data value, x min and x max are respectively the minimum and maximum values in the dataset, α represents a hyperparameter, and ln represents the natural logarithm function;

[0019] Step 1.3 According to the timestamp of the data, the long-time series data is segmented monthly, and the statistical features of soil reflectance and the corresponding precipitation, snowmelt, and evaporation indexes within each time period are extracted respectively to construct a feature set reflecting the changes in environmental impacts. The feature algorithm formula is;

[0020]

[0021] Among them, R i represents the feature value at time i, N represents the total number of times in the time series data, ε represents a small constant introduced to avoid division by zero, and the first term represents the relative change amplitude between consecutive times, capturing one change. |R i+1 -2R i +R i-1 | is the second-order difference, which is used to measure the acceleration or curvature of the change. σ R represents the standard deviation of R in the entire time series data, and β represents the weight coefficient;

[0022] Step 1.4 Using the extracted time series features, the Seasonal Influence Weighted Reflectance Change Index (SIWRCI) is adopted to quantitatively calculate the fluctuation of soil reflectance caused by natural factor changes between adjacent time periods, so as to reveal the comprehensive influence of precipitation, snowmelt, and evaporation on soil reflectance. The algorithm formula is:

[0023]

[0024] Among them, A i is the average soil reflectance of the i-th time period, A i+1 represents the average soil reflectance of the subsequent time period, P i , S i and E i respectively represent the precipitation, snowmelt, and evaporation indexes within the i-th time period, w P , w S and w E represent the weight parameters of natural factors.

[0025] Preferably, in Step 1.5, a trend analysis model is constructed based on the above reflectance change index, and the model is verified and parameter-corrected using on-site monitoring data to achieve accurate description and prediction of the change of soil reflectance under long-time series environmental impacts. Specifically, it includes:

[0026] Step 1.51 Using the reflectance change index to construct a preliminary trend model, and its formula is:

[0027]

[0028] Among them, represents the predicted soil reflectance at time t;

[0029] Let \(t\) represent the time variable, and \(TCWFI(t)\) represent the time - series change weighted feature index, which is used to quantify the mutation and trend information in time - series data;

[0030] \(\theta_0\), \(\theta_1\) and \(\theta_2\) represent the model parameters to be estimated, representing the baseline, linear trend and index weight respectively;

[0031] Step 1.52 To improve the prediction accuracy of the model, the on - site monitoring data \(R(t)\) is used to obs perform real - time verification on the model, and the prediction error is fed back for parameter update. First, calculate the prediction error:

[0032]

[0033] Then, the self - calibration update formula is used to dynamically adjust the parameters:

[0034] \(\theta\) t+1 =\(\theta\) t -\(\eta\cdot[\Delta(t)\cdot X(t)+\lambda\) t \(\cdot(\theta\) t -\(\theta\) ref )],

[0035] where, represents the parameter vector at the \(t\) - th moment;

[0036] is the feature vector, \(\eta\) represents the basic learning rate, \(\lambda\) t represents the adaptive penalty factor, \(\theta\) ref represents the reference parameter vector;

[0037] To make the penalty factor adaptively adjusted, the update formula is further defined:

[0038] \(\lambda\) t+1 =\(\lambda\) t +\(\gamma\cdot(\Delta(t)\) 2 -\(\delta)\),

[0039] where, \(\gamma\) represents the penalty factor update step size, and \(\delta\) represents the preset target error threshold.

[0040] Specifically, the present invention makes the model able to adaptively adjust parameters according to seasonal factors such as precipitation, snowmelt and evaporation by introducing a seasonal modulation factor,

[0041] avoiding the influence of soil reflectance fluctuations in wet and dry states on the monitoring results and ensuring the long - term stability of the monitoring system.

[0042] Preferably, the deep neural network is a convolutional neural network (CNN), which is used to automatically extract features from multi - band data and achieve accurate classification of surface types. Its specific steps include:

[0042] Step 2.1: For data of different bands, fusion is performed through the weighted entropy fusion algorithm to form a high-dimensional feature space, enabling the CNN to make full use of the information of each band. The fusion algorithm includes:

[0043]

[0044] Among them, F represents the fused feature representation, w i represents the weight of the i-th band, and Entropy(X i ) represents the entropy value of the i-th band;

[0045] Step 2.2: Use the CNN to extract features from the multi-band data. The CNN automatically learns local features through multiple convolutional layers and reduces the spatial dimension of the features through the pooling layer. To enhance the model's ability, an adaptive convolutional kernel size is adopted, and the size of the convolutional kernel is dynamically adjusted according to the different characteristics of the input data, so as to better capture local and global features:

[0046]

[0047] Among them, k adapt represents the adaptive convolutional kernel size, k0 represents the initial convolutional kernel size,

[0048] Var(X) represents the variance of the input data, and μ X represents the mean of the input data;

[0049] Through this adaptive adjustment, the convolutional neural network can more flexibly adapt to the changes of different data and extract more effective features.

[0050] Preferably, in Step 2.3, by training the convolutional neural network model, through the multi-task learning framework, combining the classification task and the regression task, the generalization ability and accuracy of the model are improved. The algorithm formula of the multi-task learning framework is:

[0051] L total =B·L classification +β·L regression ,

[0052] Among them, L total represents the total loss function, L classification represents the classification loss, and the cross-entropy or softmax loss function is adopted. L regression represents the regression loss, and B and β represent the weight coefficients that control the importance of the two tasks;

[0053] The CNN can not only perform land surface type classification but also provide additional regression information for each pixel point;

[0054] Step 2.4 adopts a dynamic error adjustment mechanism to dynamically adjust the learning rate and optimization strategy of the model according to the loss difference between the training set and the validation set. The formula of the dynamic error adjustment algorithm is as follows:

[0055] η t+1 = η t ·(1 + λ·|L train (t) - L val (t)|),

[0056] where η t+1 represents the updated learning rate, L train (t) and L val (t) are the losses of the training set and the validation set respectively. After the model training in Step 2.5 is completed, cross-validation and adaptive parameter optimization methods are used to verify the model to ensure its generalization ability:

[0057]

[0058] where represents the gradient of the loss function, and δ(t) represents the correction term related to the validation error.

[0059] Specifically, by combining ground sensing data with adaptive parameter updates, the present invention can dynamically adjust model parameters, reduce the spectral overlap problem between wet soil and vegetation, bare soil, significantly improve the accuracy of automatic classification, and thus improve the accuracy of monitoring results.

[0060] Preferably, the ground IoT sensor includes a soil moisture sensor deployed in key areas, which is used to collect soil water content data in real time and compare and correct it with remote sensing data. The specific steps are as follows:

[0061] Step 3.1 Update the correction factor C(t) using the dynamic correction algorithm according to the difference between the IoT data M LoT (t) and the remote sensing estimate S rs (t);

[0062]

[0063] where M LoT (t) represents the soil moisture directly measured by the IoT sensor at time t, S rs (t) represents the soil moisture preliminarily estimated by the remote sensing data at time t, and ζ represents the momentum factor;

[0064] Step 3.2 The formula for calculating the corrected soil moisture:

[0065] S corr (t) = S rs (t)·(1 + C(t)),

[0066] The remotely sensed soil moisture is dynamically corrected by multiplying a correction factor of 1 + C(t), so that S corr (t) is closer to the on-site IoT data.

[0067] Preferably, the physical model is based on a calibration curve between soil moisture and reflectance in each band, and is used to correct the spectral deviation caused by humidity changes. The specific steps include:

[0068] Step 4.1 Use IoT sensors to collect soil moisture data M LoT (t) in real time in different regions, and collect multi-band remote sensing reflectance data R band (t), covering bands of visible light, infrared, and microwave;

[0069] Step 4.2 Establish a calibration curve model between soil moisture and reflectance in each band through long-term monitoring data:

[0070] Assume that there is a non-linear relationship between soil moisture M IoT (t) and reflectance R band (t), and use a piecewise function to represent the relationship in different humidity ranges:

[0071]

[0072] where a1, b1, c1 and a2, b2, c2 represent parameters fitted according to experimental data, and are used to represent the relationship between reflectance and humidity in different humidity ranges. M threshold represents the critical value of humidity;

[0073] Step 4.3 Combine the real-time collected soil moisture M IoT (t) and reflectance R band (t) data, and dynamically correct the reflectance through the calibration curve to correct the spectral deviation caused by humidity changes:

[0074] Δθ(t) = η·[R band (t) - f(M IoT (t))]·X(t),

[0075] where Δθ(t) represents the parameter update amount at the current moment, and f(M IoT (t)) represents the predicted reflectance calculated through the current soil moisture;

[0076] The updated reflectance is calculated by the following formula:

[0077] R corr (t) = R band (t) + Δθ(t),

[0078] R corr (t) represents the corrected reflectance value, which can more accurately reflect the influence of soil moisture on reflectance after dynamic adjustment.

[0079] Specifically, through continuous learning and optimization of remote sensing and ground sensing data, the present invention can adaptively adjust model parameters under different climate conditions, ensuring that the monitoring system always maintains high precision and long-term stability, and improving the reliability of the system.

[0080] Preferably, based on the change of meteorological data, a meteorological influence correction factor C met (t) is constructed and integrated into the dynamic correction model. The relationship between this correction factor and factors such as rainfall, temperature, and wind speed can be represented by a weighted model:

[0081] C met (t) = α·P(t) + β·T(t) + γ·V(t),

[0082] where P(t) represents rainfall, T(t) represents temperature, and V(t) represents wind speed;

[0083] In the process of adjusting the reflectance based on the basic factor of the traditional dynamic correction formula for soil moisture and reflectance, the new correction formula is:

[0084]

[0085] By combining the dynamic correction factor and the meteorological correction factor, calculate the corrected reflectance R corr (t):

[0086] R corr (t) = R band (t) + Δθ(t) + λ·C met (t),

[0087] where λ·C met (t) represents the correction brought by meteorological data.

[0088] Preferably, the periodic update of the model parameters adopts an adaptive learning algorithm, which can continuously optimize according to the latest remote sensing and ground data, ensuring the long-term stability and accuracy of the monitoring system under different climate and season conditions. The specific steps include:

[0089] Step 5.1 Update the cumulative gradient square using the exponentially weighted moving average method to obtain G(t):

[0090] G(t) = D·G(t - 1) + (1 - D)·[g(t)] 2 ,

[0091] Where D is the attenuation coefficient;

[0092] Step 5.2 is to construct a seasonal modulation factor S(t) to adapt to different climate and seasonal changes, so that the parameter update has environmental sensitivity:

[0093]

[0094] Where E represents the modulation amplitude, T season represents the seasonal cycle, and φ represents the phase shift;

[0095] Step 5.3 is to calculate the adaptive learning rate:

[0096]

[0097] Step 5.4 combines the above items, and the update formula for the model parameters is:

[0098] θ(t + 1) = θ(t) - η(t)·g(t)·S(t) + ζ·[θ(t) - θ(t - 1)],

[0099] Where the first term is the adaptive gradient descent update, which is corrected by the seasonal modulation factor S(t) so that the update step size changes with the environment;

[0100] The second term is the momentum term, which helps the model overcome local minima and improve the convergence speed.

[0101] Specifically, by introducing the adaptive learning rate and momentum mechanism, it can quickly overcome local minima and accelerate convergence. At the same time, combining the seasonal modulation factor makes the optimization process of the model efficient and stable under different climate and season conditions. And through the real-time verification and feedback mechanism, it ensures the high precision and reliability of the model during long-term operation, and can adjust the model parameters immediately according to the changes in on-site data to further optimize the monitoring effect.

[0102] The present invention provides a soil and water conservation monitoring method based on remote sensing images. It has the following beneficial effects:

[0103] 1. Through the adaptive learning algorithm and the model parameter regular update mechanism of the present invention, it effectively overcomes the problems in the prior art such as seasonal changes, overlapping characteristic spectra of wet soil and vegetation, and bare soil. After introducing the seasonal modulation factor, the system can accurately capture the influence of natural factors such as precipitation, snowmelt, and evaporation on soil reflectance, adaptively adjust the model parameters, thereby eliminating the interference of reflectance fluctuations in wet and dry states on the monitoring results and ensuring the stability of long-term monitoring.

[0104] 2. By integrating ground sensing data, the present invention realizes dynamic correction of soil humidity and reflectivity data, thereby effectively improving the automatic classification accuracy between moist soil, vegetation, and bare soil. Using continuously collected remote sensing data and real-time IoT data, through an adaptive parameter update mechanism, the system can maintain high-precision monitoring under different climate and season conditions, significantly enhancing the accuracy and reliability of soil and water conservation monitoring.

[0105] 3. The present invention adopts an adaptive learning rate and momentum mechanism, which accelerates the model optimization efficiency and convergence speed, effectively overcomes the local minimum problem. At the same time, the real-time feedback mechanism continuously adjusts the model parameters according to the changes in on-site data, ensuring that the monitoring system always maintains efficient, flexible, and stable performance in complex environments, comprehensively enhancing the overall effectiveness of the soil and water conservation monitoring scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0106] Figure 1 It is a flowchart of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0107] To enable those skilled in the art to understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the scope of protection of the present invention.

[0108] The present invention will be described in detail below with reference to the accompanying drawings:

[0109] Embodiment:

[0110] Please refer to the attached Figure 1 , the embodiment of the present invention provides a soil and water conservation monitoring method based on remote sensing images, including:

[0111] Step 1: Integrate optical, infrared, and microwave remote sensing data to construct a soil information map;

[0112] Step 2: Use long-term data to capture the soil humidity changes in each season according to the soil information, and clarify the influence of precipitation, snowmelt, and evaporation on reflectivity;

[0113] Step 3: Train the data of each band through a deep neural network according to the information of the influence of reflectivity to realize automatic differentiation of moist soil, vegetation, and bare soil;

[0114] Step 4: According to the data of each band, combine the real-time humidity data of ground IoT sensors and physical models to correct the remote sensing data, and timely correct the spectral deviation caused by humidity changes;

[0115] Step 5: Regularly update the model parameters according to the calibrated information to ensure that the monitoring system can be continuously optimized according to different climate and season conditions.

[0116] The optical remote sensing data includes image data from high-resolution satellites and aerial platforms, which is used to provide surface detail information; the infrared remote sensing data includes thermal infrared images, which are used to capture the surface temperature distribution, thereby assisting in reflecting the soil moisture status; the microwave remote sensing data is synthetic aperture radar (SAR) data, and its all-weather cloud-penetrating ability is used to detect soil moisture changes.

[0117] S1: The long-term time-series data covers at least one year or a longer time period to fully reflect the influence changes of factors such as precipitation, snowmelt, and evaporation on soil reflectance. The specific steps include:

[0118] Step 1.1: Obtain multi-source remote sensing data covering at least one year or a longer time period, and simultaneously collect the corresponding meteorological data to ensure that the data can fully reflect the environmental changes in different seasons.

[0119] Step 1.2: Perform radiometric calibration, geometric calibration, and spatio-temporal alignment on each phase of data to eliminate the differences between sensors and over time, and perform normalization processing on the meteorological data to ensure comparability between data sources. The normalization processing algorithm formula is:

[0120]

[0121] where x represents the original data value, x min and x max are the minimum and maximum values in the dataset respectively, α represents a hyperparameter, and ln represents the natural logarithm function;

[0122] Step 1.3: According to the timestamp of the data, divide the long-term time-series data by month, and separately extract the statistical characteristics of soil reflectance and the corresponding precipitation, snowmelt, and evaporation indicators within each time period to construct a feature set reflecting the influence changes of the environment. The feature algorithm formula is;

[0123]

[0124] where R i represents the feature value at time i, N represents the total number of times in the time-series data, ε represents a small constant to avoid division by zero, the first term represents the relative change amplitude between consecutive times, capturing one change, |R i+1 -2R i +R i-1 | is the second-order difference, which is used to measure the acceleration or curvature of the change, σ R represents the standard deviation of R in the entire time-series data, and β represents the weight coefficient;

[0125] In Steps 1 and 4, the extracted temporal features are used to quantitatively calculate the soil reflectance fluctuations caused by natural factor changes between adjacent time periods by using the Seasonal Influence Weighted Reflectance Change Index (SIWRCI), so as to reveal the comprehensive effects of precipitation, snowmelt, and evaporation on soil reflectance. The algorithm formula is as follows:

[0126]

[0127] where A i is the average soil reflectance of the i-th time period, and A i+1 represents the average soil reflectance of the subsequent time period. P i , S i and E i represent precipitation, snowmelt, and evaporation indicators within the i-th time period respectively, and w P , w S and w E represent the weight parameters of natural factors.

[0128] In Step 1.5, a trend analysis model is constructed based on the above reflectance change index, and the model is verified and parameter-corrected using on-site monitoring data to achieve accurate description and prediction of soil reflectance changes under long-term environmental influences. Specifically, it includes:

[0129] In Step 1.51, a preliminary trend model is constructed using the reflectance change index. The formula is as follows:

[0130]

[0131] where represents the predicted soil reflectance at time t;

[0132] t represents the time variable, and TCWFI(t) represents the Temporal Change Weighted Feature Index, which is used to quantify the mutation and trend information in temporal data;

[0133] θ0, θ1, and θ2 represent the model parameters to be estimated, representing the baseline, linear trend, and index weight respectively;

[0134] In Step 1.52, to improve the prediction accuracy of the model, the on-site monitoring data R obs (t) is used to verify the model in real time, and the prediction error is fed back for parameter update. First, calculate the prediction error:

[0135]

[0136] Then, the self-correction update formula is used to dynamically adjust the parameters:

[0137] θ t+1 = θ t-η·[Δ(t)·X(t)+λ t ·(θ t -θ ref )],

[0138] wherein, represents the parameter vector at the t-th moment;

[0139] is the feature vector, η represents the basic learning rate, and λ t represents the adaptive penalty factor, and θ ref represents the reference parameter vector;

[0140] To make the penalty factor adaptively adjusted, an update formula is further defined:

[0141] λ t+1 = λ t + γ·(Δ(t) 2 - δ),

[0142] where γ represents the penalty factor update step size, and δ represents the preset target error threshold.

[0143] Specifically, the present invention introduces a seasonal modulation factor to enable the model to adaptively adjust parameters according to seasonal factors such as precipitation, snowmelt, and evaporation, avoid the influence of soil reflectance fluctuations in wet and dry states on the monitoring results, and ensure the long-term stability of the monitoring system.

[0144] S2. The deep neural network is a convolutional neural network (CNN), which is used to automatically extract features from multi-band data and achieve precise classification of surface types. The specific steps include:

[0145] Step 2.1: For data in different bands, perform fusion through a weighted entropy fusion algorithm to form a high-dimensional feature space, enabling the CNN to make full use of the information in each band. The fusion algorithm includes:

[0146]

[0147] where F represents the fused feature representation, w i represents the weight of the i-th band, and Entropy(X i ) represents the entropy value of the i-th band;

[0148] Step 2.2: Use the CNN to extract features from multi-band data. The CNN automatically learns local features through multiple convolutional layers and reduces the spatial dimension of the features through pooling layers. To enhance the model's ability, an adaptive convolutional kernel size is adopted to dynamically adjust the size of the convolutional kernel according to the different characteristics of the input data, so as to better capture local and global features:

[0149]

[0150] Among them, k adapt represents the size of the adaptive convolution kernel, and k0 represents the size of the initial convolution kernel.

[0151] Var(X) represents the variance of the input data, and μ X represents the mean of the input data.

[0152] Through this adaptive adjustment, the convolutional neural network can more flexibly adapt to the changes of different data and extract more effective features.

[0153] Preferably, in step 2.3, by training a convolutional neural network model and using a multi-task learning framework, combining a classification task and a regression task, the generalization ability and accuracy of the model are improved. The algorithm formula of the multi-task learning framework is:

[0154] L total =B·L classification +β·L regression ,

[0155] Among them, L total represents the total loss function, L classification represents the classification loss, and the cross-entropy or softmax loss function is adopted. L regression represents the regression loss, and B and β represent the weight coefficients that control the importance of the two tasks.

[0156] The CNN can not only perform land surface type classification but also provide additional regression information for each pixel point.

[0157] In step 2.4, a dynamic error adjustment mechanism is adopted. According to the loss difference between the training set and the validation set, the learning rate and optimization strategy of the model are dynamically adjusted. The algorithm formula of the dynamic error adjustment is:

[0158] η t+1 =η t ·(1 + λ·|L train (t)-L val (t)|),

[0159] Among them, η t+1 represents the updated learning rate, L train (t) and L val (t) are the losses of the training set and the validation set respectively. After the model training in step 2.5 is completed, cross-validation and an adaptive parameter optimization method are used to verify the model to ensure its generalization ability:

[0160]

[0161] Among them, denotes the gradient of the loss function, and δ(t) denotes the correction term related to the validation error.

[0162] Specifically, by combining ground sensing data with adaptive parameter updates, the present invention can dynamically adjust model parameters, reduce the spectral overlap problem between wet soil and vegetation or bare soil, significantly improve the accuracy of automatic classification, and thus improve the accuracy of monitoring results.

[0163] S3. The ground IoT sensor includes soil moisture sensors deployed in key areas, which are used to collect soil water content data in real time and compare and correct it with remote sensing data. The specific steps are as follows:

[0164] Step 3.1: According to the difference between the IoT data M LoT (t) and the remote sensing estimated value S rs (t), use the dynamic correction algorithm to update the correction factor C(t);

[0165]

[0166] where M LoT (t) represents the soil moisture directly measured by the IoT sensor at time t, S rs (t) represents the soil moisture preliminarily estimated by the remote sensing data at time t, and ζ represents the momentum factor;

[0167] Step 3.2: The calculation formula for the corrected soil moisture:

[0168] S corr (t) = S rs (t) · (1 + C(t)),

[0169] By multiplying by the correction factor of 1 + C(t), the remotely sensed estimated soil moisture is dynamically corrected, so that S corr (t) is closer to the on-site IoT data.

[0170] S4. The physical model is based on the calibration curve between soil moisture and reflectance in each band, and is used to correct the spectral deviation caused by humidity changes. The specific steps include:

[0171] Step 4.1: Use the IoT sensor to collect soil moisture data M LoT (t) in different regions in real time, and collect multi-band remote sensing reflectance data R band (t), covering visible light, infrared, and microwave bands;

[0172] Step 4.2: Establish a calibration curve model between soil moisture and reflectance in each band through long-term monitoring data:

[0173] Assume that the soil moisture M IoT (t) and the reflectance Rband There is a non-linear relationship between them, and a piecewise function is used to represent the relationship in different humidity ranges:

[0174]

[0175] Among them, a1, b1, c1 and a2, b2, c2 represent the parameters fitted according to experimental data, which are used to represent the relationship between reflectivity and humidity in different humidity ranges, M threshold represents the critical value of humidity;

[0176] Step 4.3 combines the real-time collected soil humidity M IoT (t) and the reflectivity R band (t) data, and dynamically corrects the reflectivity through the calibration curve to correct the spectral deviation caused by humidity changes:

[0177] Δθ(t) = η · [R band (t) - f(M IoT (t))] · X(t),

[0178] Among them, Δθ(t) represents the parameter update amount at the current moment, and f(M IoT (t)) represents the predicted reflectivity calculated through the current soil humidity;

[0179] The updated reflectivity is calculated through the following formula:

[0180] R corr (t) = R band (t) + Δθ(t),

[0181] R corr (t) represents the corrected reflectivity value, which can more accurately reflect the influence of soil humidity on reflectivity after dynamic adjustment.

[0182] Preferably, a meteorological influence correction factor C met (t) is constructed based on the change of meteorological data and integrated into the dynamic correction model. The relationship between this correction factor and factors such as rainfall, temperature, and wind speed can be represented by a weighted model:

[0183] C met (t) = α · P(t) + β · T(t) + γ · V(t),

[0184] Among them, P(t) represents rainfall, T(t) represents temperature, and V(t) represents wind speed;

[0185] Based on the basic factor of the traditional dynamic correction formula of soil humidity and reflectivity, the correction process of reflectivity is adjusted. The new correction formula is:

[0186]

[0187] Calculate the corrected reflectance R by combining the dynamic correction factor and the meteorological correction factor corr (t):

[0188] R corr (t) = R band (t) + Δθ(t) + λ·C met (t),

[0189] where λ·C met (t) represents the correction brought by meteorological data.

[0190] Specifically, through continuous learning and optimization of remote sensing and ground sensing data, the present invention can adaptively adjust model parameters under different climate conditions, ensuring that the monitoring system always maintains high precision and long-term stability, and improving the reliability of the system.

[0191] S5. The regular update of the model parameters adopts an adaptive learning algorithm, which can be continuously optimized according to the latest remote sensing and ground data to ensure the long-term stability and accuracy of the monitoring system under different climate and season conditions. The specific steps include:

[0192] Step 5.1 Update the cumulative gradient square using the exponential weighted moving average method to obtain G(t):

[0193] G(t) = D·G(t - 1) + (1 - D)·[g(t)] 2 ,

[0194] where D is the attenuation coefficient;

[0195] Step 5.2 To adapt to different climate and season changes, construct a seasonal modulation factor S(t) to make the parameter update environmentally sensitive:

[0196]

[0197] where E represents the modulation amplitude, T season represents the seasonal period, and φ represents the phase shift;

[0198] Step 5.3 Adaptive learning rate calculation:

[0199]

[0200] Step 5.4 Combining the above items, the update formula of the model parameters is:

[0201] θ(t + 1) = θ(t) - η(t)·g(t)·S(t) + ζ·[θ(t) - θ(t - 1)],

[0202] Among them, the first item is the adaptive gradient descent update, which is corrected by the seasonal modulation factor S(t) so that the update step size changes with the environment;

[0203] The second item is the momentum term, which helps the model overcome local minima and improve the convergence speed.

[0204] Specifically, by introducing the adaptive learning rate and momentum mechanism, it is possible to quickly overcome local minima and accelerate convergence. At the same time, combining the seasonal modulation factor makes the optimization process of the model efficient and stable under different climate and season conditions. And through the real-time verification and feedback mechanism, the high precision and reliability of the model during long-term operation are ensured, and it can instantly adjust the model parameters according to the changes in on-site data to further optimize the monitoring effect.

[0205] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A soil and water conservation monitoring method based on remote sensing images, characterized in that: include: Step 1: Integrate optical, infrared and microwave remote sensing data to construct soil information maps; Step 2: Based on soil information, long-term time series data is used to capture seasonal soil moisture changes and clarify the impact of precipitation, snowmelt, and evaporation on reflectivity; Step 3: Based on the information about the influence of reflectivity, the data of each band is trained through a deep neural network to automatically distinguish between wet soil, vegetation and bare soil; Step 4: According to the data of each band, combined with the real-time humidity data of the ground IoT sensor and the physical model, the remote sensing data is corrected to correct the spectral deviation caused by humidity changes in a timely manner; Step 5: Based on the correction information, the model parameters are updated regularly to ensure that the monitoring system can be continuously optimized according to different climate and seasonal conditions.

2. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The optical remote sensing data includes image data from high-resolution satellites and aerial platforms, which are used to provide detailed surface information; the infrared remote sensing data includes thermal infrared images, which are used to capture the surface temperature distribution, thereby assisting in reflecting the soil moisture status; the microwave remote sensing data is synthetic aperture radar (SAR) data, which uses its all-weather cloud penetration capability to detect soil moisture changes.

3. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The long time series data covers at least one year or longer to fully reflect the impact of factors such as precipitation, snowmelt and evaporation on soil reflectivity. The specific steps include: Step 1.1 Obtain multi-source remote sensing data covering at least one year or longer, and collect corresponding meteorological data to ensure that the data can fully reflect environmental changes in different seasons; Step 1.2: Perform radiation correction, geometric correction, and time-space alignment on the data of each phase to eliminate the differences between sensors and time, and normalize the meteorological data to ensure comparability between data sources. The normalization algorithm formula is: Among them, x represents the original data value, x min and x max are the minimum and maximum values ​​in the data set, α represents a hyperparameter, and ln represents the natural logarithm function; Step 1.3: According to the timestamp of the data, the long time series data is divided into months, and the statistical characteristics of soil reflectivity and the corresponding precipitation, snowmelt and evaporation indicators in each period are extracted to construct a feature set reflecting the changes in environmental impacts. The feature algorithm formula is: Among them, R i represents the eigenvalue at time i, N represents the total number of time series data, ε represents a small constant to avoid zero division, and the first term Indicates the relative change between consecutive moments, capturing a change, |R i+1 -2R i +R i-1 | is the second-order difference, which is used to measure the change in acceleration or curvature, σ R represents the standard deviation of R in the entire time series data, and β represents the weight coefficient; Steps 1 and 4 use the extracted time series features and the seasonal impact weighted reflectivity change index (SIWRCI) to quantitatively calculate the soil reflectivity fluctuations caused by changes in natural factors between adjacent time periods, thereby revealing the comprehensive impact of precipitation, snowmelt and evaporation on soil reflectivity. The algorithm formula is: Among them, A i is the average soil reflectivity in the i-th period, A i+1 represents the average soil reflectivity during the subsequent period, P i , S i and E i represent the precipitation, snowmelt and evaporation indexes in the i-th period, respectively, and w P 、w S and w E Represents the weight parameter of the natural factor.

4. The soil and water conservation monitoring method based on remote sensing images according to claim 3 is characterized in that: The step 1.5 constructs a trend analysis model based on the above reflectivity change index, and uses field monitoring data to verify and calibrate the model parameters, so as to achieve an accurate description and prediction of soil reflectivity changes under the influence of long-term environmental conditions, specifically including: Step 1.51 Use the reflectivity change index to construct a preliminary trend model, the formula is: in, represents the predicted soil reflectivity at time t; t represents the time variable, TCWFI(t) represents the time series change weighted characteristic index, which is used to quantify the mutation and trend information in time series data; θ0, θ1, and θ2 represent the model parameters to be estimated, representing the baseline, linear trend, and indicator weight, respectively; Step 1.52 To improve the prediction accuracy of the model, use the field monitoring data R obs (t) Verify the model in real time and use the prediction error feedback to update the parameters. First, calculate the prediction error: Then the self-correction update formula is used to dynamically adjust the parameters: i t+1 =θ t -η·[Δ(t)·X(t)+λ t ·(θ t -θ ref )], in, represents the parameter vector at time t; is the feature vector, η represents the basic learning rate, λ t represents the adaptive penalty factor, θ ref represents the reference parameter vector; In order to make the penalty factor adaptively adjusted, the update formula is further defined: l t+1 =λ t +γ·(Δ(t) 2 -d), Among them, γ represents the penalty factor update step size, and δ represents the preset target error threshold.

5. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The deep neural network is a convolutional neural network (CNN), which is used to automatically extract features from multi-band data and achieve accurate classification of surface types. The specific steps include: Step 2.1: The data of different bands are fused through the weighted entropy fusion algorithm to form a high-dimensional feature space, so that CNN can make full use of the information of each band. The fusion algorithm includes: Among them, F represents the fused feature representation, w i Indicates the weight of the i-th band, Entropy(X i ) represents the entropy value of the i-th band; Step 2.2 uses CNN to extract features from multi-band data. CNN automatically learns local features through multiple convolutional layers and reduces the spatial dimension of features through pooling layers. In order to enhance the model's capabilities, an adaptive convolution kernel size is used to dynamically adjust the size of the convolution kernel according to the different characteristics of the input data, thereby better capturing local and global features: Among them, k adapt represents the adaptive convolution kernel size, k0 represents the initial convolution kernel size, Var(X) represents the variance of the input data, μ X represents the mean of the input data, Through this adaptive adjustment, convolutional neural networks can more flexibly adapt to changes in different data and extract more effective features.

6. The soil and water conservation monitoring method based on remote sensing images according to claim 5 is characterized in that: The step 2.3 improves the generalization ability and accuracy of the model by training the convolutional neural network model and combining the classification task and the regression task through a multi-task learning framework. The algorithm formula of the multi-task learning framework is: L total =B·L classification +β·L regression , Among them, L total Represents the total loss function, L classification Represents classification loss, using cross entropy or soft max loss function, L regression represents the regression loss, B and β represent the weight coefficients that control the importance of the two tasks; CNN can not only classify the surface type, but also provide additional regression information for each pixel; Step 2.4 uses a dynamic error adjustment mechanism to dynamically adjust the learning rate and optimization strategy of the model according to the loss difference between the training set and the validation set. The dynamic error adjustment algorithm formula is: or t+1 =the t ·(1+λ·|L train (t)-L val (t)|), Among them, η t+1 represents the updated learning rate, L train (t) and L val (t) are the losses of the training set and the validation set, respectively. After the model training in step 2.5 is completed, the model is validated using cross-validation and adaptive parameter optimization methods to ensure its generalization ability: in, represents the gradient of the loss function and δ(t) represents the correction term related to the validation error.

7. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The ground IoT sensor includes a soil moisture sensor deployed in key areas, which is used to collect soil moisture data in real time and compare and calibrate it with remote sensing data. The specific steps are as follows: Step 3.1 Based on IoT data M LoT (t) and the remote sensing estimated value S rs (t), and use the dynamic correction algorithm to update the correction factor C(t); Among them, M LoT (t) represents the soil moisture directly measured by the IoT sensor at the time, S rs (t) represents the soil moisture estimated initially from remote sensing data at time t, ζ represents the momentum factor; Step 3.2 Calculation formula for corrected soil moisture: S corr (t)=S rs (t)·(1+C(t)), The remote sensing estimated soil moisture is dynamically corrected by multiplying it by the correction factor 1+C(t), so that S corr (t) Closer to on-site IoT data.

8. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The physical model is based on a calibration curve between soil moisture and reflectance of each band, and is used to correct spectral deviation caused by humidity changes. The specific steps include: Step 4.1 Use IoT sensors to collect soil moisture data M in different areas in real time LoT (t), collect multi-band remote sensing reflectance data R band (t), covering the visible, infrared, and microwave bands; Step 4.2 Establish a calibration curve model between soil moisture and reflectance of each band through long-term monitoring data: Assume that the soil moisture M IoT (t) and reflectivity R band (t) has a nonlinear relationship, and a piecewise function is used to represent the relationship between different humidity ranges: Among them, a1, b1, c1 and a2, b2, c2 are parameters fitted according to experimental data, which are used to express the relationship between reflectivity and humidity in different humidity ranges. threshold Indicates the critical value of humidity; Step 4.3: The real-time collected soil moisture M IoT (t) and reflectivity R band (t) Combined with the data, the reflectance is dynamically corrected through the calibration curve to correct the spectral deviation caused by humidity changes: Δθ(t)=η·[R band (t)-f(M IoT (t))]·X(t), Among them, Δθ(t) represents the parameter update amount f(M) at the current moment. IoT (t)) represents the predicted reflectivity calculated from the current soil moisture; The updated reflectivity is calculated using the following formula: R corr (t)=R band (t)+Δθ(t), R corr (t) represents the corrected reflectance value, which can more accurately reflect the impact of soil moisture on reflectance after dynamic adjustment.

9. The soil and water conservation monitoring method based on remote sensing images according to claim 1 is characterized in that: The meteorological impact correction factor C is constructed based on the change of meteorological data. met (t), and integrate it into the dynamic correction model. The relationship between the correction factor and factors such as rainfall, temperature and wind speed can be expressed by a weighted model: C met (t)=α·P(t)+β·T(t)+γ·V(t), Among them, P(t) represents rainfall, T(t) represents temperature, and V(t) represents wind speed; The correction process of reflectivity is adjusted based on the basic factors of the traditional soil moisture and reflectivity dynamic correction formula. The new correction formula is: By combining the dynamic correction factor and the meteorological correction factor, the corrected reflectivity R is calculated. corr (t): R corr (t)=R band (t)+Δθ(t)+λ·C met (t), Among them, λ·C met (t) represents the correction due to meteorological data.

10. The soil and water conservation monitoring method based on remote sensing images according to claim 1, characterized in that: The model parameters are regularly updated using an adaptive learning algorithm that can be continuously optimized based on the latest remote sensing and ground data to ensure the long-term stability and accuracy of the monitoring system under different climatic and seasonal conditions. The specific steps include: Step 5.1: Update the cumulative gradient square using the exponentially weighted moving average method to obtain G(t): G(t)=D·G(t-1)+(1-D)·[g(t)] 2 , Where D is the attenuation coefficient; Step 5.2: To adapt to different climate and seasonal changes, construct the seasonal modulation factor S(t) so that the parameter update has environmental sensitivity: Where E represents the modulation amplitude, T season represents the seasonal cycle, φ represents the phase shift; Step 5.3 Adaptive learning rate calculation: Step 5.4 Combining the above items, the update formula of the model parameters is: θ(t+1)=θ(t)-η(t)·g(t)·S(t)+ζ·[θ(t)-θ(t-1)], Among them, the first item is the adaptive gradient descent update, which is modified by the seasonal modulation factor S(t) so that the update step size changes with the environment; The second term is the momentum term, which helps the model overcome local minima and improves the convergence speed.

Images