Method and system for monitoring water content of port stacking surface based on multispectral remote sensing

Abstract

The present application provides a port stacking surface moisture content monitoring method and system based on multispectral remote sensing, relating to the technical field of port remote sensing monitoring. In view of the low monitoring accuracy in the prior art, the present application collects multispectral image data and thermal infrared image data through a UAV in a preset stable weather time window, and collects preset calibration point geographic location information and measured moisture content; the multispectral image data is preprocessed to obtain stacking surface reflectivity data containing short-wave infrared band, and a water-sensitive spectral index is calculated; the stacking surface temperature distribution data is obtained according to the thermal infrared image data; taking the measured moisture content as a reference standard, the water-sensitive spectral index and the stacking surface temperature distribution data as characteristic variables, an improved bidirectional LSTM recurrent neural network is trained to construct a moisture content fusion inversion model; the global water-sensitive spectral index and the stacking surface temperature distribution data are input into the inversion model to generate a moisture content spatial distribution map. The present application has high monitoring accuracy.

Description

Technical Field

[0001] This invention relates to the field of port remote sensing monitoring technology, and in particular to a method and system for monitoring the surface moisture content of port stacks based on multispectral remote sensing. Background Technology

[0002] The surface moisture content of port stacks is a key physical parameter for port production, safety, and environmental management. Currently, ports widely use manual handheld or fixed sensor contact measurement methods, which can only obtain data at discrete points and cannot reflect the spatial heterogeneity of the moisture content distribution across the entire large stack surface. Manual inspection is time-consuming and labor-intensive, and cannot achieve high-frequency monitoring, making it difficult to provide real-time data support for dust suppression spraying. To achieve large-scale monitoring, a large number of fixed sensor networks need to be deployed, which are costly to build and maintain, lack flexibility, and are difficult to adjust to changes in the position and shape of the stack.

[0003] In recent years, multispectral remote sensing technology has provided an effective way for non-destructive, large-area inversion of land surface parameters. Multispectral imaging can capture reflectance information of ground objects in multiple specific narrow bands, which are closely related to land cover type, vegetation health status, and moisture content. Currently, multispectral remote sensing technology has been successfully applied in fields such as agricultural soil moisture monitoring and soil moisture estimation. However, directly applying multispectral remote sensing technology to monitor the surface moisture content of port cargo stacks still faces a series of unique challenges. The types of materials stacked in ports are diverse, and their spectral characteristics vary greatly. Existing moisture inversion models are mostly designed for vegetation or homogeneous soil, and their accuracy is significantly insufficient when directly applied to port bulk cargo stacks with complex composition and variable conditions.

[0004] Therefore, there is an urgent need to develop a dedicated method and system for monitoring the moisture content of stacked surfaces rapidly, non-destructively, over a large area, and with high spatial resolution, specifically designed for complex port scenarios and integrating the advantages of UAV mobility and multispectral information. This would enhance intelligent port management and improve monitoring accuracy. Summary of the Invention

[0005] To address the issue of low monitoring accuracy in existing technologies, this invention proposes a method for monitoring the surface moisture content of port stacks based on multispectral remote sensing, specifically including the following steps: S1. Within a preset stable weather time window, multispectral image data and thermal infrared image data of the target port stack surface are collected by drone, and the measured moisture content and corresponding geographical location information are collected simultaneously at multiple preset calibration points in the stack area. S2. Preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and calculate the moisture-sensitive spectral index based on the stack surface reflectance data. S3. Preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface; S4. For each preset calibration point, with the measured moisture content as the reference standard, and with the moisture-sensitive spectral index and the temperature distribution data of the stack surface as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model. S5. Input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of moisture content on the stacking surface.

[0006] Furthermore, in S1, the preset stable meteorological time window includes: a preset noon time period in the target port, and the meteorological parameters during the preset noon time period meet a preset threshold.

[0007] Furthermore, in step S2, the multispectral image data is preprocessed to obtain stack surface reflectance data that includes at least the shortwave infrared band. This includes performing radiometric calibration and atmospheric correction processing on the multispectral image data sequentially to obtain stack surface reflectance data that includes at least the shortwave infrared band.

[0008] Furthermore, the radiometric calibration and atmospheric correction processing specifically includes: simultaneously capturing images of a ground radiometric calibration reference board during UAV aerial photography; converting the original pixel values ​​of the multispectral image data into surface radiation energy data based on the reflectivity parameters of the radiometric calibration reference board; performing atmospheric correction on the surface radiation energy data to remove errors caused by atmospheric scattering and absorption, and obtaining stacked surface reflectivity data that includes at least the shortwave infrared band.

[0009] Furthermore, the moisture-sensitive spectral index includes the normalized differential moisture index, which is calculated using the following formula: ; Indicates the normalized differential moisture index. Indicates near-infrared reflectivity. This indicates the reflectivity of the shortwave infrared band.

[0010] Furthermore, in step S3, the thermal infrared image data is preprocessed to obtain the stack surface temperature distribution data, including: radiometric calibration of the original pixel values ​​of the thermal infrared image data to convert them into infrared radiation energy data; temperature inversion of the infrared radiation energy data to obtain preliminary surface temperature data; outlier removal and noise smoothing of the preliminary surface temperature data; and spatial registration with multispectral image data to obtain the stack surface temperature distribution data.

[0011] Furthermore, in step S4, using the measured moisture content as a reference standard and the moisture-sensitive spectral index and the temperature distribution data of the stack surface as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model. This includes: performing time-series standardization on the feature variables to obtain standardized feature data; constructing a time window input sequence of a preset length based on the standardized feature data; and using an improved bidirectional LSTM network to extract features from the time window input sequence. Specifically, this includes: performing long short-term memory network calculations along the forward and backward time directions respectively to obtain the forward hidden state and the backward hidden state; based on... A temporal attention mechanism is constructed for the forward and backward hidden states. By calculating the attention weights at each time step, the forward and backward hidden states are weighted and fused to generate a temporal fusion feature vector. A residual connection structure is introduced into the output layer of the moisture content inversion model to perform a nonlinear transformation on the temporal fusion feature vector. The nonlinearly transformed temporal fusion feature vector is added to the original temporal fusion feature vector to output the predicted value of the moisture content on the stack surface. The difference between the predicted value and the measured moisture content at the corresponding calibration point is used as the optimization objective to train the parameters of the improved bidirectional LSTM network, thus obtaining the moisture content fusion inversion model.

[0012] Furthermore, the parameters of the improved bidirectional LSTM network are trained, specifically including: constructing a hybrid loss function that includes mean squared error loss, Huber loss, and physical constraint loss, wherein the physical constraint loss is used to limit the mutation amplitude of the moisture content prediction values ​​between adjacent time steps to not exceed the daily maximum change threshold of the stack moisture content; using the AdamW optimizer with weight decay to update the weight matrix and bias term parameters in the improved bidirectional LSTM network, and using a cosine annealing strategy to dynamically adjust the learning rate during the training process of the improved bidirectional LSTM network.

[0013] Furthermore, after generating the spatial distribution map of moisture content on the stack surface, the method also includes: identifying and marking dust-generating risk areas with moisture content below the safe threshold in the spatial distribution map based on the preset safe threshold of moisture content, generating early warning information; generating control instructions based on the early warning information, and sending the control instructions to the yard spray dust suppression system to perform automatic spraying operations on the dust-generating risk areas.

[0014] The present invention also provides a port stack surface moisture content monitoring system based on multispectral remote sensing, the system being used to execute the port stack surface moisture content monitoring method based on multispectral remote sensing described in any of the preceding claims, the system comprising: The data acquisition module is used to collect multispectral and thermal infrared image data of the surface of the target port stacks via UAV within a preset stable meteorological time window, and simultaneously collect measured moisture content and corresponding geographical location information at multiple preset calibration points in the stack area. The first processing module is used to preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and to calculate the moisture-sensitive spectral index based on the stack surface reflectance data. The second processing module is used to preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface; The model training module is used to train an improved bidirectional LSTM recurrent neural network for each preset calibration point, with the measured moisture content as the reference standard and the moisture-sensitive spectral index and the temperature distribution data of the stack surface as the feature variables, to build a moisture content fusion inversion model. The distribution map generation module is used to input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of the moisture content on the stacking surface.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention, within a preset stable meteorological time window, uses a drone to collect multispectral and thermal infrared image data of the surface of a target port's stockpiles. Simultaneously, measured moisture content and corresponding geographical location information are collected at multiple preset calibration points within the stockpiling area. The multispectral image data is preprocessed to obtain stockpile surface reflectance data containing at least the shortwave infrared band, and a moisture-sensitive spectral index is calculated based on this data. The thermal infrared image data is preprocessed to obtain stockpile surface temperature distribution data. For each preset calibration point, using the measured moisture content as a reference standard, and the moisture-sensitive spectral index and stockpile surface temperature distribution data as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model. The moisture-sensitive spectral index and stockpile surface temperature distribution data for the entire stockpile area are input into the moisture content fusion inversion model to generate a spatial distribution map of the stockpile surface moisture content. By simultaneously collecting multispectral and thermal infrared images within a stable meteorological window and combining them with measured moisture content and location information at calibration points for modeling, environmental interference can be effectively eliminated and data consistency improved. By calculating the moisture-sensitive spectral index using short-wave infrared and fusing it with temperature distribution data, the feature discrimination and anti-interference capability of moisture content retrieval can be enhanced. An improved bidirectional LSTM recurrent neural network is used to train the fusion retrieval model, which can fully exploit the nonlinear correlation and temporal characteristics between spectrum, temperature, and moisture content, achieving high retrieval accuracy and improving monitoring precision. The resulting spatial distribution map of moisture content on the entire stack surface meets the requirements for large-area, non-contact, and high-resolution monitoring of port storage yards. Attached Figure Description

[0016] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0017] Figure 1 This is a flowchart of a method for monitoring the surface moisture content of port stacks based on multispectral remote sensing, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of a port stack surface moisture content monitoring system based on multispectral remote sensing provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0019] The specific embodiments of the present invention will be described below.

[0020] To address the issue of low monitoring accuracy in existing technologies, this invention collects multispectral and thermal infrared image data via drones within a preset stable meteorological time window, and simultaneously collects measured moisture content and corresponding geographical location information at multiple preset calibration points within the stacking area. The multispectral image data is preprocessed to obtain stack surface reflectance data including shortwave infrared bands, and a moisture-sensitive spectral index is calculated. The stack surface temperature distribution data is obtained based on the thermal infrared image data. Using the measured moisture content as a reference standard, and the moisture-sensitive spectral index and stack surface temperature distribution data as feature variables, an improved bidirectional LSTM recurrent neural network is trained to construct a moisture content fusion inversion model. The global moisture-sensitive spectral index and stack surface temperature distribution data are input into the inversion model to generate a spatial distribution map of moisture content. This invention offers high monitoring accuracy.

[0021] Example 1 This invention provides a method for monitoring the surface moisture content of port stacks based on multispectral remote sensing. Figure 1 This is a flowchart of a method for monitoring the surface moisture content of port stacks based on multispectral remote sensing, as provided in an embodiment of the present invention. Figure 1 As shown, the specific steps include the following: S1. Within a preset stable weather time window, multispectral image data and thermal infrared image data of the target port stack surface are collected by drone, and the measured moisture content and corresponding geographical location information are collected simultaneously at multiple preset calibration points in the stack area.

[0022] A stable meteorological time window refers to a specific observation period that ensures the stability of remote sensing data acquisition quality. For example, a preset stable meteorological time window might include a pre-set noon time period at the target port, where meteorological parameters meet preset thresholds. The pre-set noon time period is based on noon at the target port location, selecting a two-hour period before and after noon to ensure stable illumination angles and uniform shadows. Meteorological parameters meeting preset thresholds typically include: clear skies, no clouds, no precipitation, and wind speeds below a set upper limit. Setting a stable meteorological time window can eliminate meteorological interference and ensure the consistency and reliability of remote sensing data.

[0023] The drone is equipped with a multispectral imaging module, a thermal infrared imaging module, and positioning equipment to conduct non-contact, large-area, rapid aerial photography of the entire bulk cargo stack area in the port. Multispectral imagery data is multi-band reflectance information of ground features simultaneously collected by the drone's multispectral camera, including surface reflectance data in six narrow bands: blue light (approximately 450nm), green light (approximately 560nm), red light (approximately 650nm), red edge (approximately 730nm), near-infrared (approximately 840nm), and short-wave infrared (approximately 1600nm). Different bands have different absorption responses to moisture, making this core foundational data for moisture content retrieval. Thermal infrared imagery data is collected by an uncooled thermal infrared camera on the drone, acquiring radiation information in the 8-14μm band. After calibration, this data can be calculated as the surface temperature distribution of the stack. Temperature changes are directly related to surface moisture evaporation and can serve as an important auxiliary feature for moisture content retrieval. Preset calibration points refer to representative locations selected on the surface of the stack, such as dry areas, wet areas, edge areas, and central areas, used to correlate ground-measured values ​​with remote sensing features. Geographic location information refers to high-precision latitude, longitude, and elevation coordinates provided by positioning equipment mounted on the UAV, used to achieve spatial matching between remote sensing data and ground-measured data. Measured moisture content refers to the actual moisture content of the stacked material measured on-site at the calibration points using a portable moisture content detection probe, used to establish a quantitative mapping relationship between remote sensing features and moisture content.

[0024] During a stable meteorological window characterized by clear skies, low wind speeds, and stable lighting, a UAV equipped with a six-channel multispectral camera, an 8-14μm thermal infrared camera, and an RTK-GNSS positioning module was controlled to fly along a planned route, simultaneously acquiring multispectral reflectance and thermal infrared radiation images of the stack surface in six bands. High-precision latitude and longitude coordinates were matched to each frame of the image using the positioning module. At multiple pre-set calibration points within the stack area, the measured moisture content was obtained using a moisture content probe, and the geographical location information of each pre-set calibration point was recorded, achieving spatiotemporal synchronous acquisition of remote sensing data, positioning information, and ground-based measured values.

[0025] By simultaneously collecting multispectral and thermal infrared remote sensing data with ground-measured moisture content and location information under stable meteorological conditions, data errors caused by environmental factors such as light, wind speed, and precipitation can be effectively reduced, ensuring the accuracy of matching between remote sensing image quality and measured values.

[0026] S2. Preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and calculate the moisture-sensitive spectral index based on the stack surface reflectance data.

[0027] The shortwave infrared band refers to the near-infrared band with a center wavelength in the range of 1550-1650 nm. Water molecules have a strong absorption characteristic in this band of electromagnetic waves, making it a core sensitive band for identifying the water content of the earth's surface and used to accurately reflect changes in the moisture content of stockpiles. Stockpile surface reflectance data is obtained through preprocessing multispectral image data and represents the ratio of solar radiation energy reflected by the material on the stockpile surface to the total incident energy, with a value ranging from 0 to 1. The moisture-sensitive spectral index is a dimensionless index constructed based on reflectance in different bands.

[0028] Specifically, the multispectral image data is preprocessed to obtain stack surface reflectance data containing at least the shortwave infrared band. This includes: sequentially performing radiometric calibration and atmospheric correction on the multispectral image data to obtain stack surface reflectance data containing at least the shortwave infrared band. The radiometric calibration and atmospheric correction processes specifically include: simultaneously capturing images of a ground radiometric calibration reference board during UAV aerial photography; converting the raw pixel values ​​of the multispectral image data into surface radiant energy data based on the reflectance parameters of the radiometric calibration reference board; and performing atmospheric correction on the surface radiant energy data to remove errors caused by atmospheric scattering and absorption, thereby obtaining stack surface reflectance data containing at least the shortwave infrared band.

[0029] Radiometric calibration refers to the calibration process of converting raw pixel values ​​in multispectral image data into true surface radiation energy. Atmospheric correction refers to the correction process of eliminating the scattering and absorption of solar radiation by gas molecules and aerosols in the atmosphere, restoring the true reflectance characteristics of ground objects. A radiometric calibration reference board is a standard board with known and uniformly stable surface reflectivity, placed on the ground in the aerial photography area. The radiometric calibration reference board is photographed simultaneously with the drone collecting stacked data, and is used to establish the correspondence between the raw pixel values ​​of the multispectral camera and the true radiation energy.

[0030] While drones collect multispectral images of the port's stacking area, images of a pre-deployed radiometric calibration reference board on the ground are simultaneously captured. Using the known standard reflectance of this reference board as a benchmark, the raw pixel values ​​output by the multispectral camera are converted and calculated to obtain surface radiation energy data that accurately reflects the physical characteristics of the stacking surface. Atmospheric correction is then applied to the radiation energy data to eliminate interference and deviations caused by environmental factors such as atmospheric scattering and absorption, resulting in accurate and reliable stacking surface reflectance data that includes at least the shortwave infrared band. By simultaneously capturing images of the calibration reference board and sequentially performing radiometric calibration and atmospheric correction, systematic errors caused by data acquisition equipment, lighting, and atmospheric environment can be effectively eliminated, improving the accuracy of the stacking surface reflectance data.

[0031] Moisture-sensitive spectral indices include the normalized differential moisture index, which is calculated using the following formula: ; Indicates the normalized differential moisture index. Indicates near-infrared reflectivity. This indicates the reflectivity of the shortwave infrared band.

[0032] A normalized differential moisture index was constructed based on the reflectance of near-infrared and short-wave infrared bands. This fully utilizes the water absorption characteristics of the short-wave infrared band and eliminates the influence of factors such as topography, illumination, and surface roughness of the stack through normalization. Compared with single-band reflectance, it can more accurately characterize the difference in moisture content of the stack surface and effectively distinguish the spectral characteristics of dry and humid areas.

[0033] S3. Preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface.

[0034] Specifically, this includes: radiometric calibration of the raw pixel values ​​of thermal infrared image data to convert them into infrared radiation energy data; temperature inversion of the infrared radiation energy data to obtain preliminary surface temperature data; outlier removal and noise smoothing of the preliminary surface temperature data; and spatial registration with multispectral image data to obtain stack surface temperature distribution data.

[0035] Infrared radiation energy data refers to a physical quantity obtained after radiometric calibration, representing the actual radiance of a ground feature in the thermal infrared band and reflecting the intensity of infrared radiation emitted by the feature itself. Temperature inversion refers to the process of calculating the actual surface temperature based on the physical correspondence between thermal infrared radiation energy and temperature. Spatial registration refers to the process of aligning temperature data with multispectral imagery geographically, ensuring that data at the same location correspond.

[0036] Radiometric calibration is performed on the raw pixel values ​​of thermal infrared image data to convert them into infrared radiation energy data. Preliminary surface temperature data is obtained through temperature inversion calculation. Outlier removal and noise smoothing are performed on the temperature data. The processed temperature data is then spatially registered with multispectral image data to obtain the stack surface temperature distribution data corresponding to the spectral data location.

[0037] By radiometric calibration and temperature inversion, the raw thermal infrared signal is converted into directly usable physical temperature data. Outlier removal and smoothing processes eliminate interference from equipment and the environment. Spatial registration enables precise fusion of temperature and spectral information.

[0038] S4. For each preset calibration point, with the measured moisture content as the reference standard, and the moisture-sensitive spectral index and the temperature distribution data of the stack surface as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model.

[0039] The reference standard directly reflects the true moisture content of the stacked materials and is used for supervised learning and accuracy calibration. The feature variables are the input parameters during the training of the bidirectional LSTM recurrent neural network. The improved bidirectional LSTM recurrent neural network is a deep learning model oriented towards time-series data. Based on the traditional bidirectional LSTM, a temporal attention mechanism and residual connection output layer are added to adapt to the time-series inversion scenario of port stack moisture content. The moisture content fusion inversion model is a dedicated machine learning model that takes moisture-sensitive spectral index and stack surface temperature distribution data as input and measured moisture content as output, used to estimate the overall moisture content of the stacked area.

[0040] Specifically, using measured moisture content as a reference standard and moisture-sensitive spectral index and stack surface temperature distribution data as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model. This includes: performing time-series standardization on the feature variables to obtain standardized feature data; constructing a time window input sequence of a preset length based on the standardized feature data; and using an improved bidirectional LSTM network to extract features from the time window input sequence. Specifically, this includes: performing Long Short-Term Memory (LSTM) network calculations along the forward and backward time axes to obtain the forward hidden state and the backward hidden state; and based on the forward hidden state... A temporal attention mechanism is constructed using the forward and backward hidden states. By calculating the attention weights at each time step, the forward and backward hidden states are weighted and fused to generate a temporal fusion feature vector. A residual connection structure is introduced into the output layer of the moisture content inversion model to perform a nonlinear transformation on the temporal fusion feature vector. The nonlinearly transformed temporal fusion feature vector is added to the original temporal fusion feature vector to output the predicted value of the moisture content on the stack surface. The difference between the predicted value and the measured moisture content at the corresponding calibration point is used as the optimization objective to train the parameters of the improved bidirectional LSTM network, thus obtaining the moisture content fusion inversion model.

[0041] Temporal standardization refers to the process of eliminating differences in the dimensions and numerical distributions of different feature variables, mapping the feature variables to a uniform scale. A time window input sequence refers to an input sample composed of continuous temporal feature data truncated to a fixed duration prior to the current time. The forward hidden state is the temporal feature propagated forward in time by a bidirectional LSTM recurrent neural network, while the backward hidden state is the temporal feature propagated backward in time by the same network. The forward and backward hidden states capture historical and future information, respectively. Temporal attention mechanisms automatically assign higher weights to important time steps, strengthening key information and weakening noise through a feature weighting structure. Temporal fusion feature vectors are the comprehensive temporal features obtained by fusing the forward and backward hidden states after attention weighting. Residual connection structures are network structures that directly add the nonlinearly transformed feature vector to the original input feature vector, used to alleviate the gradient vanishing problem during model training, retaining the original effective temporal feature information, and improving model training stability and inversion accuracy.

[0042] For example, time-series standardization is performed on feature variables to eliminate the influence of dimensions and accelerate model convergence. The standardization formula is: ; in, This represents the standardized value of the i-th feature at time t. Let represent the original value of the i-th feature at time t. Let represent the mean of the i-th feature. Let represent the standard deviation of the i-th feature.

[0043] For example, if the preset length of the input sequence for the time window is τ, the input sequence for the time window is: ; in, Let be the eigenvector at time t.

[0044] The time window input sequence is fed into an improved bidirectional LSTM network, and the forward hidden state is calculated along the forward and backward time directions, respectively. With backward hidden state Based on the forward and backward hidden states, a temporal attention mechanism is constructed, and the attention weights at each time step are calculated. The calculation formula is: ; ; ; in, This refers to attention energy value. Let v represent the attention energy at the k-th time step, where T is the total number of time steps. a W a b represents the attention layer parameters. a denoted by bias, and s represents the temporal fusion feature vector.

[0045] A residual connection structure is introduced into the output layer of the moisture content inversion model to perform a nonlinear transformation on the temporal fusion feature vector: ; After adding z and s, the predicted water content is obtained by applying an activation function: ; Where z is the result of the nonlinear transformation, ReLU is the linear rectified activation function, W1 and W2 are the output layer weights, and b1 and b2 are the biases. σ represents the predicted moisture content, and σ is the Sigmoid activation function.

[0046] By eliminating feature dimension differences through temporal standardization, combining bidirectional LSTM to fully capture temporal dependencies from both positive and negative directions, utilizing temporal attention mechanism to enhance key time step features and improve information utilization, effectively mitigating gradient vanishing through residual connections, preserving original effective features, and completing network training with measured moisture content as the supervised target, the model can improve the fitting accuracy, temporal modeling ability, and generalization stability of the moisture content fusion inversion model, providing a more accurate and reliable inversion model for predicting the surface moisture content of port stacks.

[0047] Based on the above embodiments, the parameters of the improved bidirectional LSTM network are trained, specifically including: constructing a hybrid loss function that includes mean squared error loss, Huber loss and physical constraint loss, wherein the physical constraint loss is used to limit the mutation amplitude of the moisture content prediction value between adjacent time steps to not exceed the daily maximum change threshold of the stack moisture content; using the AdamW optimizer with weight decay to update the weight matrix and bias term parameters in the improved bidirectional LSTM network, and using the cosine annealing strategy to dynamically adjust the learning rate during the training process of the improved bidirectional LSTM network.

[0048] The hybrid loss function is a weighted combination of mean squared error loss, Huber loss, and physical constraint loss. The mean squared error loss is a loss term that uses the squared mean of the error between the predicted and measured values ​​as an indicator, focusing on improving the overall fitting accuracy. The Huber loss is a robust loss term that combines the advantages of squared error and absolute error, and can suppress outlier interference. The physical constraint loss is used to constrain the variation of the water content prediction value over time, so that the water content prediction value does not exceed the actual physical law.

[0049] For example, using the difference between the predicted moisture content and the measured moisture content at the corresponding calibration point as the optimization objective, a hybrid loss function is used for training. The hybrid loss function is as follows: ; Where λ1, λ2, and λ3 are weights. This represents the mean square error.

[0050] Huber's loss term is: ; δ is the outlier threshold, and y is the actual moisture content measured using the probe method.

[0051] The physical constraints are: ; in, This represents the maximum daily variation in moisture content at the storage yard. This represents the predicted moisture content at time t. This represents the predicted moisture content at time t-1.

[0052] Employs an improved AdamW optimizer: ; This represents the model parameters after the nth iteration. This represents the model parameters after the (n+1)th iteration update. This represents the dynamic learning rate at the nth iteration. This represents the bias correction estimate of the first moment. This represents the bias correction estimate of the second moment. For smoothing, λ represents the weight decay coefficient.

[0053] Learning rate The adjustment method is as follows: ; The upper bound of the learning rate, The lower bound of the learning rate, where N represents the total number of training iterations.

[0054] By employing a hybrid loss function consisting of mean squared error loss, Huber loss, and physical constraint loss during training, the overall fitting accuracy of the model can be guaranteed, outlier interference can be suppressed, and the prediction results can be made consistent with the actual physical change patterns. Simultaneously, the AdamW optimizer with weight decay and the cosine annealing learning rate strategy can effectively improve the convergence speed of model training, parameter update stability, and generalization ability, avoiding overfitting and training oscillations, making the moisture content fusion inversion model more closely aligned with the actual monitoring scenario of port stacking.

[0055] S5. Input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of moisture content on the stacking surface.

[0056] The entire stacking area refers to the entire area of ​​stacked goods within the port that needs to be monitored. A spatial distribution map of moisture content is a visual distribution map using spatial location as coordinates and moisture content values ​​as characterizing information; it can intuitively reflect the distribution of moisture content across the entire area.

[0057] The preprocessed and calculated global moisture-sensitive spectral index and stack surface temperature distribution data are simultaneously input into the trained moisture content fusion inversion model. The moisture content fusion inversion model performs point-by-point moisture content prediction calculations for each spatial location in the entire region, generating a spatial distribution map that can completely reflect the moisture content distribution of the entire stack surface.

[0058] By inputting global remote sensing features into a trained inversion model, a spatial distribution map of moisture content on the surface of a stack can be generated quickly, non-destructively, and over a large area, realizing the transformation from discrete point monitoring to global area monitoring and intuitively presenting spatial differences in moisture content.

[0059] Based on the above embodiments, after generating the spatial distribution map of moisture content on the stack surface, the method further includes: identifying and marking dust-generating risk areas with moisture content below the safety threshold in the spatial distribution map according to a preset moisture content safety threshold, generating early warning information; generating control instructions based on the early warning information, and sending the control instructions to the yard spray dust suppression system to perform automatic spraying operations on the dust-generating risk areas.

[0060] The moisture content safety threshold is a pre-set critical moisture content value used as a standard to determine whether there is a risk of dust generation in the stacking area. A dust-risk area refers to a localized area on the stack surface where the moisture content is below the preset safety threshold, making it prone to dust pollution. A yard spray dust suppression system is an intelligent equipment system deployed in port yards that automatically sprays water to suppress dust upon receiving instructions.

[0061] After generating a spatial distribution map of moisture content on the stack surface, the moisture content values ​​at each point on the map are compared and filtered against preset safety thresholds. Areas with moisture content values ​​below the safety threshold are identified and marked, generating early warning information corresponding to dust risk. Based on the location and extent of the risk areas in the early warning information, corresponding spray control commands are generated and sent to the yard's spray dust suppression system. The system then precisely targets the identified dust-generating risk areas and performs targeted spray dust suppression operations. No manual on-site inspections or operations are required; the system can accurately locate dust hazard areas and promptly implement dust suppression measures.

[0062] This embodiment collects multispectral and thermal infrared image data of the target port's stacked surface using a drone within a preset stable meteorological time window. Simultaneously, it collects measured moisture content and corresponding geographical location information at multiple preset calibration points within the stacked area. The multispectral image data is preprocessed to obtain stacked surface reflectance data containing at least the shortwave infrared band, and a moisture-sensitive spectral index is calculated based on this data. The thermal infrared image data is preprocessed to obtain stacked surface temperature distribution data. For each preset calibration point, using the measured moisture content as a reference standard and the moisture-sensitive spectral index and stacked surface temperature distribution data as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model. The moisture-sensitive spectral index and stacked surface temperature distribution data for the entire stacked area are input into the moisture content fusion inversion model to generate a spatial distribution map of the stacked surface moisture content. By simultaneously collecting multispectral and thermal infrared images within a stable meteorological window and combining them with measured moisture content and location information at calibration points for modeling, environmental interference can be effectively eliminated and data consistency improved. By calculating the moisture-sensitive spectral index using short-wave infrared and fusing it with temperature distribution data, the feature discrimination and anti-interference capability of moisture content retrieval can be enhanced. An improved bidirectional LSTM recurrent neural network is used to train the fusion retrieval model, which can fully exploit the nonlinear correlation and temporal characteristics between spectrum, temperature, and moisture content, achieving high retrieval accuracy and improving monitoring precision. The resulting spatial distribution map of moisture content on the entire stack surface meets the requirements for large-area, non-contact, and high-resolution monitoring of port storage yards.

[0063] Example 2 This invention also provides a port stack surface moisture content monitoring system based on multispectral remote sensing. Figure 2 This is a schematic diagram of the structure of a port stack surface moisture content monitoring system based on multispectral remote sensing provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the system includes: The data acquisition module is used to collect multispectral and thermal infrared image data of the surface of the target port stacks, as well as the geographical location information and measured moisture content of multiple preset calibration points within the stacking area, through a drone within a preset stable meteorological time window. The first processing module is used to preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and to calculate the moisture-sensitive spectral index based on the stack surface reflectance data. The second processing module is used to preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface; The model training module is used to train an improved bidirectional LSTM recurrent neural network for each preset calibration point, with the measured moisture content as the reference standard and the moisture-sensitive spectral index and the temperature distribution data of the stack surface as the feature variables, to build a moisture content fusion inversion model. The distribution map generation module is used to input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of the moisture content on the stacking surface.

[0064] The port stack surface moisture content monitoring system based on multispectral remote sensing in this embodiment is used to execute the port stack surface moisture content monitoring method based on multispectral remote sensing in any of the above embodiments, and has the beneficial effects of any of the above embodiments, which will not be repeated here.

[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method for monitoring the surface moisture content of port stacks based on multispectral remote sensing, characterized in that, include: S1. Within a preset stable weather time window, multispectral image data and thermal infrared image data of the target port stack surface are collected by drone, and the measured moisture content and corresponding geographical location information are collected simultaneously at multiple preset calibration points in the stack area. S2. Preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and calculate the moisture-sensitive spectral index based on the stack surface reflectance data. S3. Preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface; S4. For each preset calibration point, using the measured moisture content as the reference standard, and the moisture-sensitive spectral index and the stack surface temperature distribution data as feature variables, an improved bidirectional LSTM recurrent neural network is used for training to construct a moisture content fusion inversion model; S4 includes: Standardized feature data are obtained by performing time-series standardization on the feature variables; Construct a time window input sequence of a preset length based on standardized feature data; An improved bidirectional LSTM network is used to extract features from the input sequence within the time window; specifically, this includes performing Long Short-Term Memory network computations along the forward and backward time directions to obtain the forward hidden state and the backward hidden state. Based on the forward and backward hidden states, a temporal attention mechanism is constructed. By calculating the attention weights at each time step, the forward and backward hidden states are weighted and fused to generate a temporal fusion feature vector. A residual connection structure is introduced into the output layer of the moisture content inversion model. The temporal fusion feature vector is nonlinearly transformed, and the nonlinearly transformed temporal fusion feature vector is added to the original temporal fusion feature vector to output the predicted value of the moisture content of the stack surface. Using the difference between the predicted value and the measured moisture content at the corresponding calibration point as the optimization objective, the parameters of the improved bidirectional LSTM network are trained to obtain the moisture content fusion inversion model. S5. Input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of moisture content on the stacking surface.

2. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, In S1, the preset stable meteorological time window includes: the local noon preset period at the target port, and the meteorological parameters during the noon preset period meet the preset threshold.

3. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, In step S2, the multispectral image data is preprocessed to obtain stack surface reflectance data that includes at least the shortwave infrared band, including: Radiometric calibration and atmospheric correction were performed sequentially on the multispectral image data to obtain stack surface reflectance data that included at least the shortwave infrared band.

4. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 3, characterized in that, The radiation calibration and atmospheric correction process specifically includes: During the drone aerial photography process, images of the ground radiation calibration reference board are captured simultaneously; Based on the reflectivity parameters of the radiometric calibration reference plate, the raw pixel values ​​of multispectral image data are converted into surface radiation energy data. Atmospheric correction is performed on the surface radiation energy data to remove errors caused by atmospheric scattering and absorption, resulting in stacked surface reflectance data that includes at least the shortwave infrared band.

5. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, Moisture-sensitive spectral indices include the normalized differential moisture index, which is calculated using the following formula: ； Indicates the normalized differential moisture index. Indicates near-infrared reflectivity. This indicates the reflectivity in the shortwave infrared band.

6. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, In step S3, the thermal infrared image data is preprocessed to obtain the temperature distribution data of the stack surface, including: Radiometric calibration is performed on the raw pixel values ​​of the thermal infrared image data to convert them into infrared radiation energy data; temperature inversion is then performed on the infrared radiation energy data to obtain preliminary surface temperature data. Outlier removal and noise smoothing were performed on the preliminary surface temperature data, and spatial registration was performed with multispectral image data to obtain the surface temperature distribution data of the stack.

7. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, The parameters of the improved bidirectional LSTM network are trained, specifically including: A hybrid loss function is constructed, which includes mean squared error loss, Huber loss and physical constraint loss. The physical constraint loss is used to limit the abrupt change in the moisture content prediction value between adjacent time steps to not exceed the daily maximum change threshold of the stack moisture content. The AdamW optimizer with weight decay is used to update the weight matrix and bias parameters in the improved bidirectional LSTM network, and the learning rate during the training process of the improved bidirectional LSTM network is dynamically adjusted using a cosine annealing strategy.

8. The method for monitoring the surface moisture content of port stacks based on multispectral remote sensing according to claim 1, characterized in that, After generating the spatial distribution map of moisture content on the stack surface, the following steps are also included: Based on the preset safe threshold for moisture content, identify and mark dust-generating risk areas with moisture content below the safe threshold in the spatial distribution map of moisture content, and generate early warning information; Control commands are generated based on the early warning information and sent to the yard spray dust suppression system to perform automatic spraying operations on dust-prone areas.

9. A port stack surface moisture content monitoring system based on multispectral remote sensing, characterized in that, The system is used to execute the port stack surface moisture content monitoring method based on multispectral remote sensing as described in any one of claims 1-8, the system comprising: The data acquisition module is used to collect multispectral and thermal infrared image data of the surface of the target port stacks via UAV within a preset stable meteorological time window, and simultaneously collect measured moisture content and corresponding geographical location information at multiple preset calibration points in the stack area. The first processing module is used to preprocess the multispectral image data to obtain stack surface reflectance data that includes at least the shortwave infrared band, and to calculate the moisture-sensitive spectral index based on the stack surface reflectance data. The second processing module is used to preprocess the thermal infrared image data to obtain the temperature distribution data of the stack surface; The model training module is used to train an improved bidirectional LSTM recurrent neural network for each preset calibration point, with the measured moisture content as the reference standard and the moisture-sensitive spectral index and the temperature distribution data of the stack surface as the feature variables, to build a moisture content fusion inversion model. The distribution map generation module is used to input the moisture-sensitive spectral index of the entire stacking area and the temperature distribution data of the stacking surface into the moisture content fusion inversion model to generate a spatial distribution map of the moisture content on the stacking surface.

Images