Collaborative observation method for bamboo forest ecological environment based on multi-source internet of things sensing

Abstract

The present application relates to the technical field of bamboo forest ecological environment monitoring, in particular, the present application relates to a bamboo forest ecological environment collaborative observation method based on multi-source internet of things sensing, the present application synchronously collects multi-source collaborative observation parameter set of bamboo forest slope surface through internet of things sensor network, then constructs point-surface scale fusion bamboo forest water dynamic model, generates terrain weighted evapotranspiration by using random forest gradient boosting algorithm to correct evapotranspiration point cluster data, obtains runoff subarea total amount according to terrain weight aggregation runoff data, then executes dynamic water balance constraint calculation, combines variational assimilation algorithm to optimize infiltration capacity, when residual error exceeds threshold value, redistributes infiltration capacity spatial proportion according to soil water conductivity, finally, according to residual error spatial distribution thermodynamic diagram, the evapotranspiration point cluster is migrated to the representative area of microtopography, the observation network is self-adaptively optimized, the present application solves the problems of traditional technology data scale disconnection, insufficient observation accuracy and network difficulty in adapting to changes, and improves the accuracy and reliability of bamboo forest ecological observation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bamboo forest ecological environment monitoring, in particular to a bamboo forest ecological environment collaborative observation method based on multi-source Internet of Things sensing. BACKGROUND

[0002] Bamboo forest ecological environment monitoring is an important technology. Under the background that bamboo forest ecological protection, soil and water conservation and carbon sink function research are increasingly valued, this technology is an important support to break through the limitations of traditional bamboo forest observation. It can not only obtain multi-dimensional data synchronously and avoid ecological cognitive bias caused by single parameter observation, but also provide scientific data support for bamboo forest ecosystem health assessment, degradation early warning and sustainable management, and help balance bamboo resource development and ecological protection, and adapt to various bamboo forest management scenarios of artificial forest cultivation and ecological restoration.

[0003] The existing bamboo forest ecological environment observation technology faces the core problems of data scale disconnection, insufficient observation accuracy and lack of network self-adaptation in actual application. Traditional observation relies on scattered single-point sensors and regional-scale runoff observation equipment. The collaborative fusion mechanism of point-scale data and area-scale data has not been established. Neither the disturbance of complex micro-topography of bamboo forest on observation data is corrected, nor the spatial matching of different scale data is realized through a unified model. This technical defect will form a chain effect. First, single-point observation data cannot represent the real ecological conditions of the region due to the lack of correction of terrain undulations and soil heterogeneity disturbance, which leads to deviation in water dynamic analysis based on these data. Second, direct mixing of point-scale and area-scale data will cause imbalance in water balance calculation, such as low matching degree of actual precipitation, evapotranspiration, runoff and estimated infiltration rate, which cannot accurately reflect the water cycle law of bamboo forest. Finally, the observation points are fixed after being laid out and cannot be dynamically adjusted according to residual abnormal areas. With the changes of bamboo micro-topography and soil hydrological state, the observation network gradually loses its representativeness, and the long-term observation accuracy continues to decline. In the long run, it is difficult to accurately grasp the dynamic changes of bamboo forest ecosystem, and it also affects the scientific nature of ecological management decisions based on observation data. In order to solve this technical problem, we provide a bamboo forest ecological environment collaborative observation method based on multi-source Internet of Things sensing. SUMMARY

[0004] The purpose of the present application is to provide a bamboo forest ecological environment collaborative observation method based on multi-source Internet of Things sensing to solve the problems raised in the background technology.

[0005] 1. Due to the disconnection of point-scale and area-scale data and the uncorrected terrain disturbance, the present application can realize data scale fusion and improve observation accuracy by collecting multi-source parameters, correcting evapotranspiration data and aggregating runoff data to build a water dynamic model.

[0006] 2、Due to the observation network is fixed, it is difficult to adapt to the change of terrain, so the accuracy is reduced, therefore, the case can adaptively optimize the observation network by calculating the balance residual and migrating the evaporation point cluster to the representative area according to the balance residual heat map, and can guarantee the reliability of long-term observation.

[0007] To achieve the above object, one of the objects of the present application is to provide a bamboo forest ecological environment collaborative observation method based on multi-source Internet of Things sensing, comprising the following steps:

[0008] S1, synchronously collecting a multi-source collaborative observation parameter set of a bamboo forest slope surface by an Internet of Things sensor network, the parameter set comprising microtopographic elevation data, micro runoff unit data, evaporation point cluster data, soil moisture gradient data and microclimate data;

[0009] S2, constructing a point-surface scale fused bamboo forest water dynamic model, performing terrain representative correction on the evaporation point cluster data based on the microtopographic elevation data and the soil moisture gradient data by using a machine learning algorithm, generating terrain weighted evapotranspiration matching the spatial resolution of the micro runoff unit data, and aggregating the micro runoff unit data according to the terrain weight to the total amount of the runoff sub-area corresponding to the evaporation point cluster data;

[0010] S3, performing dynamic water balance constraint calculation, real-time fusing the precipitation data, the terrain weighted evapotranspiration and the total amount of the runoff sub-area, and iteratively optimizing the infiltration amount estimate value by a variational assimilation algorithm, so that the infiltration amount estimate value satisfies the sum of the precipitation data, the total amount of the runoff sub-area, the terrain weighted evapotranspiration and the infiltration amount, and when the balance residual of the infiltration amount and the infiltration amount estimate value exceeds a preset threshold, the spatial contribution proportion of the infiltration amount is dynamically redistributed according to the soil hydraulic conductivity;

[0011] S4, feeding back and adjusting the layout position of the evaporation point cluster to the microtopographic representative area based on the spatial distribution characteristics of the balance residual, and realizing adaptive optimization of the observation network.

[0012] Compared with the prior art, the present application has the following advantages:

[0013] 1. The point-surface fused water dynamic model is constructed by synchronously collecting multi-source parameters and correcting the evaporation point cluster data by using a random forest gradient boosting algorithm, and aggregating the runoff data according to the terrain weight, which has the technical effect of multi-source data scale collaboration, solves the problem of disconnection of traditional point-surface data and large observation deviation, and has the advantages of accurately reflecting the bamboo forest water distribution and laying a reliable data foundation for subsequent analysis.

[0014] 2. The variational assimilation algorithm is combined to iteratively optimize the infiltration amount, and the spatial proportion of the infiltration amount is redistributed according to the soil hydraulic conductivity when the balance residual exceeds the threshold, which has the technical effect of accurate water balance calculation, solves the problem of low matching degree between traditional infiltration amount estimation and actual water cycle, and has the advantages of accurately presenting the bamboo forest water cycle rule and improving the scientific nature of ecological analysis.

[0015] 3. By extracting the residual spatial distribution heat map, anomaly areas are identified and representative terrain areas are selected. The hydraulic translation device of the lysimeter is controlled to migrate the evapotranspiration point clusters, achieving the technical effect of adaptive optimization of the observation network. This solves the problem of traditional fixed observation points and difficulty in adapting to terrain changes, which leads to a decrease in accuracy. It has the advantages of ensuring long-term observation reliability and adapting to the dynamic changes of bamboo forest ecology. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall workflow of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Please see Figure 1 As shown, this embodiment provides a collaborative observation method for the bamboo forest ecological environment based on multi-source IoT sensing, including the following steps:

[0019] S1. A multi-source collaborative observation parameter set for bamboo forest slopes is synchronously collected through an Internet of Things sensor network. The parameter set includes micro-topographic elevation data, micro-runoff unit data, evapotranspiration cluster data, soil moisture gradient data, and micro-meteorological data.

[0020] S2. Construct a point-area scale fusion bamboo forest water dynamic model. Based on micro-topographic elevation data and soil moisture gradient data, use machine learning algorithms to perform topographic representativeness correction on evapotranspiration cluster data, generate topographic weighted evapotranspiration that matches the spatial resolution of micro-runoff unit data, and aggregate micro-runoff unit data according to topographic weight into the total amount of runoff sub-regions corresponding to evapotranspiration cluster data.

[0021] S3. Perform dynamic water balance constraint calculation, integrate precipitation data, topographic weighted evapotranspiration and total runoff in sub-regions in real time, and iteratively optimize the infiltration estimate through variational assimilation algorithm to make it satisfy the sum of precipitation data, total runoff in sub-regions, topographic weighted evapotranspiration and infiltration. When the balance residual between infiltration and the infiltration estimate exceeds the preset threshold, the spatial contribution ratio of infiltration is dynamically redistributed according to soil hydraulic conductivity.

[0022] S4. Based on the feedback of the spatial distribution characteristics of the equilibrium residual, adjust the layout of the evapotranspiration point cluster to the representative area of ​​micro-topography to achieve adaptive optimization of the observation network.

[0023] The specific steps for constructing a point-area scale fusion model of bamboo forest water dynamics are as follows:

[0024] A gridded moisture field with micro-topographic elevation data as the spatial framework was established, and evapotranspiration cluster data was used as the core observation node. A spatial interpolation surface of soil hydrological attributes was constructed through soil moisture gradient data. Machine learning algorithms were used to learn the nonlinear mapping relationship between evapotranspiration cluster data, micro-topographic elevation data, and soil moisture gradient data. An evapotranspiration topographic correction weight coefficient field covering the entire observation area was generated, and the spatial scale of micro-runoff unit data and evapotranspiration cluster data was dynamically fused based on this coefficient field.

[0025] Further explanation is needed. After synchronously collecting the multi-source collaborative observation parameter set of the bamboo forest slope through the Internet of Things sensor network, it is necessary to construct a bamboo forest water dynamic model that integrates point and area scales. This is because the collected parameters have differences between point scale (evapotranspiration cluster data) and area scale (micro-runoff unit data). If used directly, the water dynamic analysis is prone to deviation due to scale mismatch. The core function of the model is to achieve the collaborative fusion of data at different scales through a unified spatial framework and data correction. The specific implementation method is as follows:

[0026] First, a gridded moisture field was established using micro-topographic elevation data as the spatial framework. This data was acquired by scanning the bamboo forest slope using a drone equipped with a lidar device, capturing subtle topographic undulations such as ridges, valleys, and gentle slopes. During the construction process, the southwest corner of the bamboo forest slope was used as the origin, and the observation area was divided into 5m × 5m grids. The elevation value of the center point of each grid cell was labeled, and the topographic features within the grid were recorded, including whether it was a ridge and the slope gradient, forming a gridded spatial framework covering the entire observation area. Based on this framework, the soil moisture and evapotranspiration data of each grid were correlated. Upon reaching the corresponding grid, a gridded moisture field is formed. The core function of this field is to provide a unified spatial positioning benchmark for all subsequent multi-source data, avoiding scale conflicts caused by chaotic data collection locations, and allowing point data and area data to be analyzed in the same spatial dimension. Evapotranspiration cluster data is used as the core observation node. Evapotranspiration cluster data is point-scale data reflecting the total transpiration of bamboo forest vegetation and soil evaporation. It is obtained by deploying a cluster of lysimeters in the bamboo forest: 3-5 lysimeters are deployed according to terrain type (convex ridge, flat slope, concave valley) (each with a monitoring range of 1m² and an accuracy of ±0.1mm / d). Lysometry data of different terrain types form a cluster (e.g., three lysometry meters at a ridge constitute a cluster of evapotranspiration points). The average value within the cluster is taken as the representative evapotranspiration value for that terrain type. This data accurately reflects the evapotranspiration characteristics of the local terrain, but it only covers a limited number of points. It needs to be extended to the regional scale by the model, thus becoming the core observation node of the model, providing a baseline true value for subsequent regional evapotranspiration estimation. A spatial interpolation surface for soil hydrological properties is constructed using soil moisture gradient data. Soil moisture gradient data reflects the differences in soil moisture distribution in the vertical and horizontal directions. In the vertical direction, at each evapotranspiration point... Layered soil moisture sensors (buried at depths of 10cm, 30cm, and 50cm) are deployed around the scattered data points. Moisture content (in %) at each depth is collected hourly. The difference in moisture content between adjacent depths is calculated (e.g., the difference between 10cm and 30cm is 5%) to obtain the vertical moisture gradient. Horizontally, surface soil moisture sensors are deployed at 10m intervals, and the difference in moisture content between adjacent sensors is calculated to obtain the horizontal moisture gradient. Both types of sensors together constitute the soil moisture gradient data, which reflects the spatial uneven distribution of soil moisture. When constructing the spatial interpolation surface for soil hydrological attributes, the Kriging spatial interpolation method is used.

[0027] Using moisture data collected by each sensor as known sample points, and combining the spatial coordinates of these sample points (based on the grid positions of the gridded moisture field), the soil moisture content and hydraulic conductivity hydrological properties of all grid cells are estimated by fitting the spatial correlation between sample points (e.g., samples closer together have smaller moisture differences). This ultimately generates a continuous spatial interpolation surface for soil hydrological properties. This surface spatially completes the discrete sensor data, expanding the soil hydrological data, which originally only existed at sensor points, into a continuous field covering the entire observation area. This provides comprehensive data support for analyzing the correlation between soil moisture and evapotranspiration and topography. Machine learning algorithms are used to learn the nonlinear mapping relationship between evapotranspiration cluster data, micro-topographic elevation data, and soil moisture gradient data. The random forest gradient boosting algorithm (adapted to nonlinear data fitting) is selected. The input feature layer contains two types of data: micro-topographic elevation data... The model employs two main methods: first, it generates topographic parameters (topographic curvature index, slope vector, and aspect-light intensity coefficient, reflecting the influence of topography on water distribution); second, it generates parameters derived from soil moisture gradient data (vertical stratified water content difference rate and horizontal water migration trend parameters, reflecting soil moisture dynamics). The output layer is a topographic-soil correction factor for each evapotranspiration cluster. During the learning process, using the evapotranspiration cluster data as a benchmark, the algorithm repeatedly learns the correspondence between a certain topographic parameter, a certain soil moisture parameter, and evapotranspiration values. Through iterative optimization of model parameters, the model can accurately output correction factors matching actual evapotranspiration values ​​based on the input topographic and soil data, thereby mastering the nonlinear relationship between the three and generating a topographic correction weight coefficient field covering the entire observation area. Based on the correction factors output by the above machine learning model, topographic correction weight coefficients are assigned to each grid of the gridded water field.

[0028] For grids containing evapotranspiration clusters, the correction factor of that cluster is directly used. For grids without clusters, the correction factor is obtained by interpolation based on the similarity between the grid's topographic parameters (slope, elevation) and the surrounding clusters. This results in an evapotranspiration topographic correction weight coefficient field covering all grids. The purpose of this field is to eliminate topographic bias in the evapotranspiration data. For example, evapotranspiration clusters in concave valleys are affected by water accumulation, resulting in lower original data. The coefficient field will assign a correction coefficient greater than 1 to them, improving the data to the true level. At the same time, it is also a key bridge for scale fusion: micro-runoff unit data is at the surface scale (corresponding to multiple grids), while evapotranspiration cluster data is at the point scale (corresponding to a single grid). The coefficient field expands the point-scale evapotranspiration data to the surface scale according to the grid weights, and then matches it with the grid size of the micro-runoff unit data. For example, the corrected data of multiple evapotranspiration grids can be aggregated into a single runoff unit grid data, realizing the dynamic fusion of the two spatial scales. This allows water data at different scales to be used collaboratively for subsequent water balance calculations.

[0029] The machine learning algorithm adopts a random forest gradient boosting architecture. Its input feature layer includes topographic curvature index, slope vector, aspect light intensity coefficient derived from micro-topographic elevation data, as well as vertical stratified water content difference rate and horizontal water migration trend parameters of soil moisture gradient data. The output layer is a topographic representativeness correction factor for each evapotranspiration cluster data unit. This factor is used to quantify the representativeness bias of single-point evapotranspiration observations under micro-topographic variation conditions, and optimizes the spatial continuity of the evapotranspiration topographic correction weight coefficient field through a backpropagation mechanism.

[0030] Further explanation is needed: After constructing the spatial interpolation surface for soil hydrological properties, to accurately establish the correlation between evapotranspiration cluster data and micro-topography and soil moisture, the model employs a random forest gradient boosting architecture—a machine learning algorithm adapted for fitting nonlinear data. This architecture combines the advantages of random forest multi-decision tree ensemble and gradient boosting iterative error optimization. By constructing multiple decision trees and gradually correcting prediction biases according to the gradient direction, it can efficiently learn complex multi-factor coupling relationships, avoiding the inadequacy of a single algorithm in fitting the nonlinear characteristics of bamboo forest micro-topography and soil moisture. The specific implementation method is as follows:

[0031] The random forest gradient boosting architecture selected in this study first generates multiple independent decision trees using random forests. Each tree is trained based on different feature subsets and sample subsets to initially fit the relationship between input and output. Then, a gradient boosting strategy is used to construct new decision trees to compensate for the prediction errors of the previous decision trees. After 100-200 iterations, the prediction results of all decision trees are weighted and fused to output the final result. This architecture retains the robustness of random forests to multiple features while improving prediction accuracy through gradient boosting. It is suitable for the complex and variable topography and soil moisture characteristics of bamboo forest ecosystems. The input feature layer of the architecture includes three types of parameters derived from micro-topographic elevation data and soil moisture. Two types of parameters derived from soil moisture gradient data were designed specifically for the dynamic analysis of bamboo forest moisture. The first is the topographic curvature index, calculated from micro-topographic elevation data, reflecting the degree of surface curvature. Positive curvature represents ridges, negative curvature represents valleys, and zero curvature represents flat slopes. Its function is to quantify the impact of topographic undulation on water accumulation / dispersion. The second is the slope vector, which includes the slope magnitude, surface inclination angle, and direction. Its function is to determine the speed of water transport along the slope. The third is the aspect-light intensity coefficient, calculated by combining slope aspect and local solar radiation data. Its function is to reflect the impact of sunlight on vegetation transpiration. The fourth is the vertical stratified moisture content difference rate, derived from soil moisture gradient data at different depths (10cm, 30cm, etc.). The parameters are calculated based on the moisture content difference between adjacent surface soil moisture sensors (50cm and 50cm), which reflects the uneven distribution of soil moisture vertically. The fifth parameter is the horizontal moisture migration trend parameter, calculated by the moisture content difference and distance between adjacent surface soil moisture sensors, which determines the direction and rate of water flow horizontally. These five parameters comprehensively cover the key factors affecting evapotranspiration from both topographic and soil moisture dimensions, providing sufficient feature support for accurate fitting. The output layer of the architecture is the topographic representativeness correction factor for each evapotranspiration cluster data unit. The topographic representativeness correction factor is a numerical value that quantifies the representativeness deviation of single-point evapotranspiration observations under micro-topographic variations, ranging from 0.8 to 1.2. A value greater than 1 indicates that the observation value is lower due to topographic limitations. The value needs to be magnified. A value less than 1 indicates that the observed value is too high due to topographic advantages and needs to be reduced. For example, evapotranspiration clusters at ridges are greatly affected by wind and the soil dries easily, so the observed value is often lower than the true level of the region. The correction factor is set to 1.15. After correction, the evapotranspiration value = original observed value × correction factor, eliminating the observation bias caused by topography and allowing point-scale data to accurately reflect the regional evapotranspiration characteristics. To further improve data reliability, the spatial continuity of the evapotranspiration topographic correction weight coefficient field needs to be optimized through the backpropagation mechanism. The previously generated correction weight coefficient field may have abrupt changes in coefficients between adjacent grids due to the lack of evapotranspiration cluster data in some grids, resulting in data connection discontinuities during subsequent scale fusion. The specific optimization process is as follows:

[0032] The coefficients of each grid in the evapotranspiration topographic correction weight coefficient field are used as the values ​​to be optimized and input into the backpropagation mechanism. The optimization objective is to minimize the difference in coefficients between adjacent grids. The error between the current coefficient field and the spatially continuous ideal field is calculated. Assuming that the difference in coefficients between adjacent grids is ≤0.05, the feature weights of each decision tree in the random forest gradient boosting architecture are adjusted in the reverse direction through gradient descent. The architecture is rerun to generate new correction factors and update the weight coefficient field. The above process is repeated until the difference in coefficients between adjacent grids is ≤0.05. At this time, the weight coefficient field shows a continuous and smooth spatial distribution, such as from ridge to flat slope to valley, with the coefficients gradually transitioning from 1.15 to 1.0 and then to 0.95. This provides a continuous and seamless correction basis for the scale fusion of subsequent micro-runoff unit data and evapotranspiration cluster data, ensuring the spatial consistency of data at different scales.

[0033] The specific steps for performing terrain representativeness correction on evapotranspiration cluster data using machine learning algorithms are as follows:

[0034] Using evapotranspiration cluster data as the baseline, and local topographic complexity index generated from micro-topographic elevation data and soil dry-wet spatial variability rate generated from soil moisture gradient data as driving variables, the topographic-soil coupling correction model is iteratively trained through a random forest gradient boosting architecture. The dynamic correction weight of each evapotranspiration cluster data unit is output. The evapotranspiration cluster data is multiplied by the corresponding dynamic correction weight to obtain the topographically corrected evapotranspiration that eliminates the interference of micro-topographic undulation and soil heterogeneity.

[0035] Further explanation is needed: after completing the design of the feature layer and output layer of the random forest gradient boosting architecture, it is necessary to perform terrain representativeness correction on the evapotranspiration cluster data based on this architecture. This step is crucial to solving the interference of micro-topography and soil heterogeneity on point-scale evapotranspiration data. By introducing driving variables and iterative training, the corrected evapotranspiration data can truly reflect the water consumption patterns at the regional scale. The specific implementation method is as follows:

[0036] When using machine learning algorithms to perform topographic representativeness correction on evapotranspiration cluster data, the core data relationships are first clarified. The evapotranspiration cluster data is used as the baseline true value. This data is the raw data measured by a cluster of lysimeters and averaged within the cluster. Although it is point-scale data, the lysimeters directly monitor the vegetation-soil evapotranspiration process, resulting in high data accuracy. Therefore, it serves as the benchmark for model training. The local topographic complexity index generated from micro-topographic elevation data and the soil moisture gradient rate generated from soil moisture gradient data are the driving variables. These two variables quantify the key factors affecting evapotranspiration observation bias from the topographic and soil dimensions, respectively. The local topographic complexity index is calculated from the micro-topographic elevation data. A 20m × 20m local area is defined centered on each evapotranspiration cluster. The elevation values ​​of all grids within this area are extracted, and the elevation standard deviation (reflecting the degree of elevation fluctuation) and the slope change rate (the average of the slope differences between adjacent grids) are calculated. These two values ​​are then weighted and summed with the standard deviation accounting for 60% and the slope change rate accounting for 40% to obtain the local topographic complexity index (values ​​0-1). 0 represents gentle terrain, and 1 represents drastic terrain undulation. For example, in a local area of ​​a cluster of points at a ridge, the elevation standard deviation is 0.8m, the slope variation rate is 0.6, and the index = 0.8 × 0.6 + 0.6 × 0.4 = 0.72. This indicates that the terrain of this cluster is highly complex, and evapotranspiration observation is easily affected by terrain. The spatial variability of soil moisture is generated from soil moisture gradient data. For each evapotranspiration cluster, the soil moisture sensors (including vertically stratified and horizontally distributed sensors) around the cluster extract the moisture content data of each sensor over 24 hours and calculate the horizontal direction. The ratio of the maximum difference in water content between adjacent sensors to the average value (horizontal variability), and the ratio of the difference in water content between the surface and deep layers to the average value in the vertical stratification (vertical variability), are taken as the spatial variability of soil moisture (values ​​range from 0 to 1, where 0 represents uniform soil moisture distribution and 1 represents extreme differences). For example, at the point cluster in the valley, the maximum difference in water content in the horizontal direction is 12%, and the average value is 20% (horizontal variability 0.6), while the vertical difference is 8%, and the average value is 18% (vertical variability 0.44). The final variability is taken as 0.6 indicates that the significant differences in soil moisture levels around the point cluster cause evapotranspiration observations to deviate from the true regional level. Subsequently, a topography-soil coupling correction model is iteratively trained using a random forest gradient boosting architecture. The training process aims to eliminate topography and soil interference. The evapotranspiration cluster data (benchmark true value), local topography complexity index, and soil moisture spatial variability rate are mapped one-to-one to each point cluster to form a training dataset (containing 60 point cluster samples, covering different topography and soil conditions). The number of decision trees in the random forest is set to 150 (balancing accuracy and computational efficiency). The gradient boosting iteration runs for 200 rounds, with each iteration correcting the prediction error of the previous round. In the first round of training, the architecture uses the local topography complexity index... The soil moisture and humidity spatial variability rate is used as input, and evapotranspiration cluster data is used as output. An initial decision tree cluster is constructed to obtain preliminary predictions. The error between the predicted value and the baseline true value is calculated (e.g., the predicted value of a certain cluster is 4.5 mm / d, the true value is 4.2 mm / d, and the error is 0.3 mm / d). The second round of training aims to minimize the error by adjusting the feature weights of the decision tree and constructing a new decision tree to compensate for the error. Each subsequent iteration repeats the prediction-error calculation-error compensation process until 200 iterations are completed, or the average error between two adjacent iterations is less than 0.05 mm / d (meeting the accuracy requirement). Training is then stopped. The model obtained at this point is the topography-soil coupled correction model. After the model training is completed, the dynamic data of each evapotranspiration cluster data unit is output. Dynamic correction weights are personalized correction coefficients (ranging from 0.9 to 1.3) output by the model based on the driving variables of the point cluster (topographic complexity index and soil moisture spatial variability rate). Their output logic is directly related to the topography and soil condition of the point cluster. For point clusters with high topographic complexity (index > 0.7) and large soil moisture variability (rate > 0.6) (such as ridge point clusters), the model determines that their evapotranspiration observations are low due to rapid water loss and dry soil, and outputs a weight greater than 1 (such as 1.15). For point clusters with gentle topography (index < 0.3) and uniform soil moisture (rate < 0.3) (such as flat slope point clusters), the model determines that their observations are close to the true level of the region, and outputs a weight close to 1 (such as 1). 02) For point clusters in the valley with high topographic complexity (index 0.4-0.6) and soil moisture variation (rate 0.4-0.6), a weight of 0.95-1.05 is output. The weight of each point cluster is obtained by nonlinear fitting of its driving variables to ensure that the weight accurately matches the degree of interference of the point cluster. Finally, the evapotranspiration data of the point cluster is multiplied by the corresponding dynamic correction weight to obtain the topographically corrected evapotranspiration that eliminates the interference of micro-topographic undulation and soil heterogeneity. For example, the original evapotranspiration data of the ridge point cluster is 4.2 mm / d, the dynamic correction weight is 1.15, and the corrected evapotranspiration is 4.2 × 1.15 = 4.83 mm / d; the original data of the flat slope point cluster is 5.0 mm / d, the weight is 1.02, and the corrected evapotranspiration is 5.0 mm / d.At 1 mm / d, the corrected evapotranspiration data has been freed from the interference of topographic relief and soil heterogeneity. It retains the accuracy advantages of point-scale data while also possessing representativeness in reflecting regional evapotranspiration characteristics, laying the foundation for subsequent generation of topographically weighted evapotranspiration data that matches micro-runoff unit data.

[0037] Topographic representativeness correction further includes:

[0038] Based on micro-topographic elevation data, the terrain feature categories of the location of evapotranspiration point cluster data are identified, and three terrain response modes are divided into ridges, flat slopes, and valleys. For each mode, a pre-trained soil moisture-topography coupled response function is loaded. This function is jointly calibrated by the rate of change of water content in the vertical stratification and the intensity of water redistribution in the horizontal direction using soil moisture gradient data. Finally, an adaptive correction coefficient matrix is ​​generated that dynamically adjusts with the terrain category and soil hydrological state. The terrain representativeness is corrected by using the adaptive correction coefficient matrix.

[0039] Further explanation is needed: after obtaining dynamic correction weights and completing the initial correction of evapotranspiration cluster data through the random forest gradient boosting architecture, to further improve the correction accuracy, a more refined topographic representativeness correction needs to be performed based on the unique hydrological response patterns of different terrains. This is because the water collection, loss, and storage characteristics of the three types of terrain—convex ridges, flat slopes, and concave valleys—differ significantly, and a single correction mode is difficult to adapt to all scenarios. Therefore, classification response and function calibration are required to make the correction process more closely reflect the actual ecological environment. The specific implementation method is as follows:

[0040] The in-depth step of topographic representativeness correction first involves identifying the topographic feature category of the evapotranspiration cluster data based on micro-topographic elevation data. This identification process relies on the previously constructed gridded moisture field. Centered on each evapotranspiration cluster, elevation data of its own grid and surrounding 3×3 grids are extracted. Topographic relief (the average elevation difference between the central grid and surrounding grids) and surface curvature (the curvature value of the surface fitted by the elevations of adjacent grids) are calculated. Based on these two indicators, the topographic category is classified: if the topographic relief is >0.5m... Furthermore, if the surface curvature is positive (the surface bulges upwards), it is classified as a ridge (water easily flows away along the slope, and the soil is relatively dry). If the terrain relief is <0.2m and the surface curvature is close to 0 (the surface is gentle), it is classified as a flat slope (water is relatively evenly distributed, with no obvious accumulation or loss). If the terrain relief is >0.5m and the surface curvature is negative (the surface sinks downwards), it is classified as a valley (rainwater and slope runoff easily accumulate, and the soil moisture content is relatively high). For example, the grid where a certain evapotranspiration cluster is located has an elevation of 126.2m, and the surrounding grid... If the grid elevation is below 125.7m, the undulation is 0.6m, and the curvature value is 0.03 (positive), it is identified as a ridge terrain. Another point cluster has an elevation difference of only 0.15m and a curvature value of 0.002, and is identified as a flat slope terrain. This quantitative identification ensures that each point cluster can accurately match the corresponding terrain feature category, classifying three terrain response modes: ridge, flat slope, and valley. Each mode corresponds to a unique water transport pattern: the ridge mode is characterized by rapid water loss and frequent soil wetting and drying, with evapotranspiration significantly limited by soil moisture content; the flat slope mode exhibits slow water infiltration and uniform distribution, with evapotranspiration mainly driven by vegetation transpiration; and the valley mode is characterized by water accumulation and retention, with long-term soil moisture, and evapotranspiration is more affected by light and ventilation conditions. For these three modes, a pre-trained soil moisture-terrain coupled response function needs to be loaded. This function is a mathematical relationship model constructed through long-term observation and calibration, which quantifies the correlation between soil moisture dynamics and evapotranspiration under different terrains. Its construction process is as follows:

[0041] Over the past three years of observation, soil moisture gradient data and evapotranspiration cluster data have been continuously recorded for each type of terrain (ridge, flat slope, valley), and the relationship between the two has been analyzed (e.g., in ridge terrain, for every 1% decrease in soil moisture content, evapotranspiration decreases by 0.2 mm / d). Through statistical fitting, this relationship has been transformed into a computable function (e.g., in the ridge pattern, evapotranspiration correction value = original evapotranspiration value × (1 + 0.05 × soil moisture content deviation), and stored as pre-training parameters for the model. The core function is to provide targeted correction criteria for different terrain patterns, avoiding bias caused by a one-size-fits-all correction. The calibration of the function depends on the rate of change of soil moisture gradient data in the vertical stratification and the intensity of water redistribution in the horizontal direction, where the vertical... The rate of change of soil moisture content in stratified soil layers refers to how quickly soil moisture changes in the vertical direction over time. It is calculated using continuous data collected by stratified soil moisture sensors (10cm, 30cm, and 50cm depths). For example, if the moisture content at a depth of 10cm decreases from 20% to 15% in 24 hours, the rate of change is -0.21% / h (a negative value indicates a decrease in moisture content). This indicator reflects the vertical water consumption process of root water absorption and soil evaporation. The intensity of horizontal water redistribution refers to the degree of water flow and transfer in the horizontal direction of the soil surface. It is calculated using the difference in moisture content between adjacent surface soil moisture sensors and the time difference. For example, if the moisture content of the sensor on the east side is 5% higher than that on the west side, and the difference decreases by 2% within 2 hours, the redistribution intensity is 1% / h. This model reflects the horizontal water movement of slope runoff and lateral infiltration driven by topography. During calibration, these two indicators are substituted into the function as variables, and the function parameters are adjusted (e.g., in the ridge model, the function correction coefficient increases by 0.02 for every 0.1% / h increase in the vertical change rate) to ensure that the function can accurately reflect the evapotranspiration response under different water movement states, thereby improving the correction accuracy. Finally, an adaptive correction coefficient matrix is ​​generated that dynamically adjusts according to the topography type and soil hydrological state. The matrix is ​​generated by constructing a two-dimensional structure with topography feature categories as rows and soil hydrological state as columns: the row dimension includes three types of topography: ridge, flat slope, and valley; the column dimension is divided into dry, slow-wetting, and fast-wetting categories based on the vertical stratified water content change rate and the horizontal water redistribution intensity. Six soil hydrological states (e.g., vertical change rate < -0.1% / h and horizontal redistribution intensity < 0.5% / h, corresponding to arid-slow state) are generated. During the process, for each combination of topography and hydrology, the corresponding soil moisture-topography coupled response function is called to calculate the correction coefficient for that combination (e.g., ridge-arid-slow state, correction coefficient 1.2; flat slope-wet-fast state, correction coefficient 1.05). All coefficients are filled into a matrix according to rows and columns to form an adaptive correction coefficient matrix. This matrix can dynamically retrieve the corresponding correction coefficient based on the real-time monitored topography type and soil hydrological state (e.g., for valley topography in wet-fast state, the coefficient 0 corresponding to valley-wet-fast in the matrix is ​​automatically selected).(95) Further optimize the correction effect of evapotranspiration cluster data to provide more accurate basic data for subsequent generation of terrain-weighted evapotranspiration.

[0042] The generation of terrain-weighted evapotranspiration that matches the spatial resolution of micro-runoff unit data is achieved through spatial scale dimensionality reduction mapping, and the specific steps are as follows:

[0043] Using the spatial grid corresponding to the micro-runoff unit data as the target resolution unit, the topographically corrected evapotranspiration of all evapotranspiration cluster data within the unit is extracted. The spatial contribution weight is assigned based on the topographic similarity distance between each cluster and the center point of the target grid. The topographic similarity distance is composed of the elevation difference, slope difference, and curvature difference calculated from the micro-topographic elevation data. After weighted summation, the topographically weighted evapotranspiration that matches the spatial resolution of the micro-runoff unit data is generated.

[0044] Further explanation is needed: after completing the topographic representativeness correction of the evapotranspiration cluster data and obtaining the topographically corrected evapotranspiration that eliminates topographic and soil interference, these point-scale data need to be converted into topographically weighted evapotranspiration that matches the surface-scale micro-runoff unit data. This is because the micro-runoff unit data uses a specific spatial grid as the statistical unit, while the evapotranspiration cluster data corresponds to only a single point. If directly used for water balance calculations, errors will occur due to scale mismatch. Therefore, spatial scale dimensionality reduction mapping is used to achieve resolution unification between the two. The specific implementation method is as follows:

[0045] The core idea behind generating terrain-weighted evapotranspiration that matches the spatial resolution of micro-runoff unit data is to use the spatial grid of micro-runoff units as a reference and integrate point-scale evapotranspiration data into area-scale data based on terrain similarity. This is carried out through the following process:

[0046] First, the spatial grid corresponding to the micro-runoff unit data is used as the target resolution unit. The micro-runoff unit data is runoff data collected from a 20m×20m spatial grid divided according to the runoff characteristics on the bamboo forest slope. Each grid is a micro-runoff unit, recording the surface runoff and runoff time parameters of the area. These grids constitute the basic statistical unit of the area-scale data. Therefore, setting it as the target resolution unit means that the subsequently generated topographically weighted evapotranspiration must completely correspond to the grid size to ensure that the two can be integrated within the same spatial framework. Next, the topographically corrected evapotranspiration of all evapotranspiration clusters within the unit is extracted. The topographically corrected evapotranspiration is the evapotranspiration data corrected by the random forest gradient lifting architecture and optimized by the topographic response model (e.g., the corrected evapotranspiration of a ridge cluster is 4.83mm / d, and that of a flat slope cluster is 5.1mm / d). It has eliminated the interference of micro-topographic undulations and soil heterogeneity and has the necessary characteristics. For regional representativeness, for each target resolution unit (20m×20m grid), evapotranspiration clusters that fall entirely within the grid or whose center point is located within the grid are first selected through coordinate matching of the gridded water field (e.g., a grid contains two evapotranspiration clusters: one with a ridge and the other with a flat slope). Then, the terrain-corrected evapotranspiration corresponding to these clusters is extracted as the basic data for calculating the evapotranspiration of that grid. If no evapotranspiration clusters are distributed within a target grid, the correction data of the closest clusters with the most similar terrain features from adjacent grids is extracted to ensure that each target grid has corresponding basic evapotranspiration data. Spatial contribution weights are assigned based on the terrain similarity distance between each cluster and the center point of the target grid. The terrain similarity distance is an indicator that quantifies the degree of terrain difference between the evapotranspiration cluster and the center point of the target grid. The smaller the distance, the more similar the terrain is, and the stronger the representativeness of the cluster data for the grid. It is weighted by the elevation difference, slope difference, and curvature difference calculated from the micro-topographic elevation data.

[0047] The elevation difference is the difference in elevation between the location of the point cluster and the center point of the grid (e.g., the elevation of the point cluster is 125.3m, and the center of the grid is 125.5m, the difference is 0.2m). The slope difference is the difference in slope angle between the two (e.g., the slope of the point cluster is 15°, and the slope of the grid center is 14°, the difference is 1°). The curvature difference is the difference in surface curvature between the two (e.g., the curvature of the point cluster is 0.03, and the curvature of the grid center is 0.02, the difference is 0.01). In the calculation, the elevation difference accounts for 50%, the slope difference accounts for 30%, and the curvature difference accounts for 20% (verified by the hydrological model, this weight can best reflect the influence of topography on evapotranspiration). After converting the three to a unified dimension (e.g., converting them all to standardized values ​​of 0-1), they are weighted and summed to obtain the topographic similarity distance, which takes a value of 0-1, where 0 represents completely identical topography and 1 represents extremely large differences. The allocation of spatial contribution weights follows the principle that the smaller the topographic similarity distance, the larger the weight.

[0048] First, take the reciprocal of the terrain similarity distance for all point clusters within the same target grid (e.g., a distance of 0.2 corresponds to a reciprocal of 5, and a distance of 0.3 corresponds to a reciprocal of 3.33). Then, calculate the proportion of the reciprocal of each point cluster to the sum of the reciprocals of all point clusters. This proportion is the spatial contribution weight of that point cluster. For example, if there are two point clusters in a target grid, point cluster A has a terrain similarity distance of 0.2 (reciprocal of 5), and point cluster B has a distance of 0.3 (reciprocal of 3.33), with a total reciprocal of 8.33, the weight of point cluster A is 5 ÷ 8.33 ≈ 0.6, and the weight of point cluster B is 3.33 ÷ 8.33 ≈ 0.4. This means that the data of point cluster A contributes 60% to the evapotranspiration of the grid, and point cluster B contributes 40%. Finally, after weighted summation, a topographically weighted evapotranspiration that matches the spatial resolution of the micro-runoff unit data is generated. The topographically corrected evapotranspiration of each point cluster within each target grid is multiplied by its spatial contribution weight, and all products are summed to obtain the topographically weighted evapotranspiration of that grid. For example, in the grid above, the corrected evapotranspiration of point cluster A is 4.83 mm / d × weight 0.6 ≈ 2.9 mm / d, and that of point cluster B is 5.1 mm / d × weight 0.4 ≈ 2.04 mm / d. After weighted summation, the topographically weighted evapotranspiration is approximately 4.94 mm / d. Through this process, the originally scattered point-scale evapotranspiration data is integrated into area-scale data that perfectly matches the grid size of the micro-runoff unit. Each target grid corresponds to a unique topographically weighted evapotranspiration, which not only retains the accuracy advantage of evapotranspiration point cluster data but also achieves spatial scale unification with area-scale runoff data, providing basic parameters for scale coordination in subsequent dynamic water balance constraint calculations.

[0049] The specific steps for aggregating micro-runoff unit data into total runoff sub-regions corresponding to evapotranspiration cluster data based on topographic weights are as follows:

[0050] Based on the spatial distribution of evapotranspiration cluster data, the micro-runoff unit data is spatially assigned according to the confluence paths divided by micro-topographic elevation data. For micro-runoff units belonging to the same evapotranspiration cluster influence sub-region, aggregation weights are assigned based on their topographic connectivity index with the evapotranspiration cluster data center. The topographic connectivity index is calculated by combining slope continuity, elevation difference, and surface roughness. The weighted micro-runoff unit data is accumulated to generate the total runoff sub-region corresponding to the evapotranspiration cluster data.

[0051] Further explanation is needed: after matching the spatial resolution of topographically weighted evapotranspiration and micro-runoff unit data, to achieve deep fusion of the two at the point-area scale, it is also necessary to aggregate the area-scale micro-runoff unit data according to topographic weights into the total runoff sub-region corresponding to the point-scale evapotranspiration cluster data. This is because the evapotranspiration cluster data is centered on discrete points; only by ensuring that the runoff data also corresponds to the influence range of each cluster can accurate coupling of evapotranspiration and runoff be achieved in subsequent water balance calculations. The specific implementation method is as follows:

[0052] Firstly, the spatial distribution of evapotranspiration cluster data is used as the benchmark. The spatial distribution of evapotranspiration cluster data refers to the specific geographical location of all evapotranspiration clusters on the bamboo forest slope, determined based on the coordinates of the gridded water field. For example, the ridge cluster is located at grids (5,3) and (5,4), and the flat slope cluster is located at (6,3) and (6,4). Each cluster, through prior topographic representativeness correction, can accurately reflect the evapotranspiration characteristics within a certain range around it, forming the influence range of the point and the surrounding area. For example, the influence radius of a single cluster is 10-15 meters. Using it as the benchmark is primarily because the evapotranspiration cluster data is high-precision point-scale data that has undergone multiple rounds of correction and covers the main topographic types of bamboo forests (ridges, flat slopes, valleys). Using its spatial distribution as the anchor point allows the aggregation of runoff data to better match the actual water movement patterns of the terrain, avoiding sub-region division deviations caused by benchmark confusion. The micro-runoff unit data is spatially assigned according to the confluence paths divided by micro-topographic elevation data. The division of confluence paths relies on micro-topographic elevation data.

[0053] First, elevation data of the entire observation area is extracted using a gridded water field. The D8 confluence algorithm is then employed to determine the flow direction by comparing the elevation difference between each grid and its eight surrounding grids. Flow paths are drawn from high-elevation grids to low-elevation grids, forming a dendritic confluence network. Based on the influence range of evapotranspiration clusters, the confluence network is divided into several confluence path segments, each corresponding to a potential influence area of ​​an evapotranspiration cluster. For spatial attribution, for each micro-runoff unit (20m×20m grid), its position within the confluence network determines its corresponding confluence path segment, thus identifying which evapotranspiration cluster's influence sub-region it belongs to. For example, if a micro-runoff unit's confluence path segment ultimately merges into a convex... The main runoff around a ridge cluster is determined to belong to the sub-region influenced by the convex ridge evapotranspiration cluster. If a unit is located at the boundary between the influence ranges of two clusters, its affiliation is determined based on the distance from its center point to the two clusters (clusters closer are given priority), ensuring that each micro-runoff unit uniquely corresponds to a sub-region influenced by an evapotranspiration cluster. For micro-runoff units belonging to the same sub-region influenced by an evapotranspiration cluster, aggregation weights are assigned based on the topographic connectivity index between the micro-runoff unit and the cluster center. The topographic connectivity index is an indicator that quantifies the topographic connectivity between the micro-runoff unit and the cluster center. The higher the index, the less obstruction of the terrain to the water flow and the stronger the runoff correlation. It is calculated jointly by slope continuity, elevation difference, and surface roughness.

[0054] Slope continuity is determined by the slope change rate between the unit and the cluster center. Elevation difference is the difference in elevation between the two. Surface roughness is calculated using vegetation cover and litter thickness within the unit. The calculation is weighted by 40% for slope continuity, 30% for elevation difference, and 30% for surface roughness to obtain the terrain connectivity index, which ranges from 0 to 1. For example, if the three indices for a unit are 0.8, 0.7, and 0.6, the index = 0.8 × 0.4 + 0.7 × 0.3 + 0.6 × 0.3 = 0.71. Aggregate weight allocation is based on the terrain connectivity index.

[0055] First, the indices of all micro-runoff units within the same influence sub-region are normalized, with the sum of all unit indices equal to 1. The normalized index of a unit is its aggregate weight. For example, in the influence sub-region of the ridge evapotranspiration cluster, there are 3 micro-runoff units with indices of 0.71, 0.65, and 0.64, totaling 2.0. The weights of each unit are 0.71 ÷ 2.0 ≈ 0.355, 0.65 ÷ 2.0 = 0.325, and 0.64 ÷ 2.0 = 0.32, respectively. This means that the unit with the best connectivity contributes 35.5% to the total runoff in the sub-region. Finally, the weighted micro-runoff unit data are summed to generate the total runoff in the sub-region corresponding to the evapotranspiration cluster data. During the weighting process, the measured runoff of each micro-runoff unit (e.g., the runoff of unit 1 is 20 m³ / s) is used. The weighted runoff is obtained by multiplying the unit data of units 1, 218m³, and 317m³ with their aggregation weights (Unit 1: 20×0.355=7.1m³, Unit 2: 18×0.325=5.85m³, Unit 3: 17×0.32=5.44m³). The weighted runoff of all units within the same influence sub-region is summed to obtain the total runoff of that sub-region (7.1+5.85+5.44≈18.39m³), which is the total runoff sub-region corresponding to the evapotranspiration cluster data. Through this process, the surface-scale micro-runoff unit data is aggregated into point-scale runoff data corresponding one-to-one with each evapotranspiration cluster, realizing the accurate matching of runoff and evapotranspiration at the point scale, and providing scale-coordinated runoff parameters for subsequent dynamic water balance constraint calculations.

[0056] The specific steps of the variational assimilation algorithm for iteratively optimizing the infiltration estimate are as follows:

[0057] A water balance cost function is constructed with precipitation data as input and topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration as state variables. A time continuity constraint is introduced through a four-dimensional variational assimilation framework to force the physical relationship between the rate of change of infiltration and soil hydraulic conductivity to be consistent between adjacent time steps. Finally, a quasi-Newton iteration method is used to minimize the cost function so that the residual of the precipitation data and the sum of the estimated values ​​of topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration converges to a preset noise level.

[0058] Further explanation is needed: after completing the scale fusion of topographically weighted evapotranspiration and total runoff in sub-regions, the infiltration estimate needs to be iteratively optimized using a variational assimilation algorithm. Infiltration, as a core component of bamboo forest water balance, connects precipitation, evapotranspiration, and runoff; its estimation accuracy directly affects the reliability of water balance calculations. The variational assimilation algorithm can achieve precise optimization of infiltration by fusing multi-source observation data with physical constraints. The specific implementation method is as follows:

[0059] The core of the variational assimilation algorithm for optimizing infiltration begins with constructing a water balance cost function. This function takes precipitation data as input and topographically weighted evapotranspiration, total runoff in a sub-region, and infiltration as state variables. Precipitation data is acquired through tipping bucket rain gauges deployed on bamboo forest slopes, serving as the sole source of water input. Topographically weighted evapotranspiration (reflecting water consumption), total runoff in a sub-region (reflecting water loss), and infiltration (reflecting water penetration into deeper soil layers) are the state variables, collectively determining the water distribution process. The cost function is constructed based on the principle of water balance, with the core logic being to minimize the difference between the input precipitation data and the sum of the output evapotranspiration, runoff, and infiltration. This is achieved by first dividing the variables into time series (every 30...). The data is processed in minutes (with each time step as a unit). For each time step, the difference between precipitation data and (topographically weighted evapotranspiration + total runoff in the sub-region + estimated infiltration) is calculated. The differences from all time steps are then squared and summed to form a cost function. The smaller this function value, the closer the current estimated infiltration is to the actual water balance, providing a quantitative target for subsequent optimization. Next, a four-dimensional variational assimilation framework is introduced to introduce temporal continuity constraints. This framework is an assimilation method that integrates spatial and temporal data. Its core principle is not only to consider the variable relationships within a single time step but also to ensure that variable changes conform to physical laws through temporal constraints, avoiding unreasonable abrupt changes in infiltration between adjacent time steps. In this context, the time continuity constraint requires that the change in infiltration between adjacent time steps must be smooth and consistent with the soil's own hydrological characteristics (such as soil hydraulic conductivity), and must not violate the physical mechanism of water infiltration. In practice, the framework introduces the rate of change of infiltration over time (the difference in infiltration between adjacent time steps) as a constraint term into the cost function. If the infiltration between two adjacent time steps suddenly increases from 5 mm to 15 mm (the rate of change is too large), the constraint term will increase the cost function value, indicating that the change is unrealistic. Simultaneously, the rate of change of infiltration between adjacent time steps is forced to maintain consistency with the physical relationship of soil hydraulic conductivity. Soil hydraulic conductivity, derived from soil moisture gradient data, reflects the soil's ability to allow water infiltration. The maximum possible variation in infiltration is determined by adding a matching index between the rate of change of infiltration and the soil hydraulic conductivity to the constraints. If the rate of change of infiltration exceeds the reasonable range corresponding to the hydraulic conductivity (e.g., in clay soil with low hydraulic conductivity, the rate of change of infiltration exceeds 3 mm / 30 minutes), the framework will automatically penalize the cost function, forcing the change in infiltration back to the range allowed by physical laws. Finally, the quasi-Newton iteration method is used to minimize the cost function. The quasi-Newton iteration method is an optimization algorithm that does not require calculating the second derivative of the function and can quickly approximate the minimum value of the function using only first-order information. Compared with the traditional Newton method, it is more suitable for the complex characteristics of bamboo forest moisture data and avoids iteration stagnation due to the complexity of derivative calculation. The specific optimization process is as follows:

[0060] First, set an initial estimate for the infiltration rate based on the average value of historical observation data, such as 6 mm / 30 minutes. Substitute this value into the cost function to calculate the initial function value. Then, calculate the gradient of the cost function at the current infiltration rate using an algorithm. This gradient reflects the trend of the function value changing with the infiltration rate, determining the direction for adjusting the infiltration rate. If the gradient is positive, it means that increasing the infiltration rate will increase the function value, requiring a decrease in the infiltration rate. Conversely, if the gradient is negative, the infiltration rate is adjusted according to the gradient direction. For example, if the initial value is 6 mm and the gradient is negative, adjust it to 7 mm. Recalculate the cost function value, repeating the iterative process of calculating the gradient, adjusting the infiltration rate, and calculating the function value. After each iteration, compare the difference between the current function value and the previous value. If the difference is less than 0.01, it indicates that the function... When the value is close to the minimum, the iteration stops. Through this process, the residual of the sum of precipitation data and the topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration estimates converges to the preset noise level. The residual is the difference between the precipitation data and the sum of evapotranspiration, runoff, and infiltration. The preset noise level is a reasonable error range set according to the accuracy of the observation equipment. When the quasi-Newton iteration method makes the cost function reach the minimum value, the residual will naturally shrink to within this noise level. This means that the infiltration estimate at this time can make the water balance equation (precipitation = evapotranspiration + runoff + infiltration) hold within the error tolerance range. It fits the multi-source observation data and conforms to physical laws, providing accurate infiltration parameters for subsequent dynamic water balance constraint calculations.

[0061] The specific steps for dynamically redistributing the spatial contribution ratio of infiltration based on soil hydraulic conductivity are as follows:

[0062] When the balance residual exceeds the threshold, the spatial redistribution module based on soil moisture gradient data is triggered. It uses the vertical stratified soil moisture content change rate to invert the actual infiltration flux. Combined with the potential infiltration path network generated by micro-topographic elevation data, the residual exceeding the threshold is redistributed according to the relative proportion of the saturated hydraulic conductivity of each sub-region. After redistribution, the estimated infiltration value is updated and fed back to the variational assimilation iteration process until the water balance equation is closed.

[0063] Further explanation is needed regarding the process of iteratively optimizing the infiltration estimate using the variational assimilation algorithm. If the balance residual between precipitation data and the sum of topographically weighted evapotranspiration, total runoff in sub-regions, and the estimated infiltration exceeds a preset threshold (typically 0.2 mm, set based on the accuracy of the observation equipment and the error range of the ecological model), it indicates that the current spatial distribution of infiltration does not conform to the actual water movement pattern. Therefore, it is necessary to dynamically redistribute the spatial contribution ratio of infiltration based on soil hydraulic conductivity, eliminating the residual through targeted adjustments. The specific implementation method is as follows:

[0064] When the equilibrium residual exceeds a threshold, the system automatically triggers the spatial redistribution module for infiltration based on soil moisture gradient data. The core function of this module is to infer the actual infiltration situation from the actual changes in soil moisture, providing a basis for spatial allocation. The module first uses the vertical stratified soil moisture content change rate to invert the actual infiltration flux. This rate is calculated from continuous data acquired by soil moisture sensors buried at different depths (10cm, 30cm, 50cm). For example, if the moisture content at each depth is recorded every 30 minutes, and the moisture content at 10cm depth decreases from 20% to 18%, at 30cm from 22% to 21%, and at 50cm remains constant at 23%, then the rate of change at each depth is calculated. The values ​​were -0.067% / min, -0.033% / min, and 0, respectively. When retrieving the actual infiltration flux, based on the principle of soil moisture balance, the decrease in surface soil moisture content is caused by evaporation, plant absorption, and infiltration. However, changes in deep soil moisture content are mainly affected by infiltration. If deep soil moisture content increases, it indicates water infiltration; if it decreases, it may be due to deep soil water being absorbed by the root system. However, considering the root distribution characteristics of bamboo forests, there are fewer roots below 50cm, mainly reflecting infiltration. By comparing the rate of change in moisture content at different depths, eliminating the influence of evaporation and plant absorption, and referring to the topographically weighted evapotranspiration during the same period, the water consumed by evaporation and absorption can be estimated, thus deriving the actual infiltration flux, such as the decrease in surface soil moisture. Of the total water volume, 60% is used for evapotranspiration and absorption, and 40% infiltrates into deeper soil layers. Based on soil bulk density and soil layer thickness, the actual infiltration flux is calculated to be 0.5 mm / 30 min. This value reflects the true rate of water infiltration into deeper soil layers, providing a quantitative basis for subsequent allocation. The module uses a potential infiltration path network generated from micro-topographic elevation data to perform allocation. This potential infiltration path network is a virtual network constructed based on micro-topographic elevation data, reflecting the natural infiltration direction of water. Its generation process relies on a previously divided gridded water field: first, the elevation value of each grid is extracted; then, the preferential infiltration direction of water within the grid is determined using the D8 confluence algorithm, pointing from high-elevation grids to low-elevation grids; finally, based on soil texture differences... Through soil sampling and analysis, the infiltration direction is corrected based on factors such as the faster infiltration in sandy soil areas and the slower infiltration in clay soil areas. For example, the infiltration direction of the sandy soil grid is more divergent, while that of the clay soil grid is more concentrated. This ultimately forms a dendritic potential infiltration path network that starts at high altitudes, ends at low altitudes, and dynamically adjusts along the terrain slope and soil texture. Each grid corresponds to one or more infiltration paths, clarifying the possible destination of water infiltration. Based on this, the module redistributes the residuals exceeding the threshold according to the relative proportion of the soil saturated hydraulic conductivity of each sub-region. Soil saturated hydraulic conductivity refers to the maximum rate at which water can pass through when the soil reaches saturation water content. Each sub-region is the runoff sub-region corresponding to the evapotranspiration clusters defined above. The allocation process consists of three steps:

[0065] Calculate the relative proportion of soil saturated hydraulic conductivity in each sub-region. First, obtain the average saturated hydraulic conductivity of each sub-region (e.g., sub-region A is 1.2 mm / min, sub-region B is 0.8 mm / min, and sub-region C is 0.4 mm / min). Then, calculate the total hydraulic conductivity (1.2 + 0.8 + 0.4 = 2.4 mm / min). The relative proportion of each sub-region = its own hydraulic conductivity / total hydraulic conductivity (sub-region A is 0.5, sub-region B is 0.33, and sub-region C is 0.17). Determine the total residual to be allocated. If the equilibrium residual is 0.3 mm (exceeding the threshold of 0.1 mm), then the residual to be allocated is 0.1 mm. Allocate the residual according to the relative proportion: sub-region A is allocated 0.1 × 0.5 = 0.05 mm, sub-region B is allocated 0.033 mm, and sub-region C is allocated 0.017 mm. This means that, based on the original estimated infiltration volume, sub-region A needs to increase the infiltration volume by 0.05 mm, and sub-region B needs to increase... The infiltration flux is increased by 0.033 mm, and sub-region C is increased by 0.017 mm. By tilting the distribution to sub-regions with high water conductance, it conforms to the physical law that water preferentially infiltrates to areas with strong permeability. After the distribution is completed, the estimated infiltration flux is updated and fed back to the variational assimilation iteration process. The adjusted infiltration flux of each sub-region is resubstituted into the water balance cost function to calculate the new balance residual. If the new residual still exceeds the threshold, the above process of inverting the actual infiltration flux, generating the potential path network, and distributing proportionally is repeated until the balance residual drops to within the threshold and the water balance equation is closed. That is, the difference between the precipitation data and the sum of the topographic weighted evapotranspiration, the total runoff of the sub-region, and the estimated infiltration flux is within the allowable range. At this time, the spatial distribution of infiltration flux not only conforms to the actual permeability of the soil, but also conforms to the influence of micro-topography on water movement, ensuring the accuracy of water balance calculation and providing reliable data support for the adaptive optimization of the subsequent observation network.

[0066] The specific steps for adjusting the layout of the evapotranspiration clusters based on feedback are as follows:

[0067] Extract the residual spatial distribution heat map from the dynamic water balance constraint calculation, identify micro-topographical anomaly areas with persistently high residuals, filter representative areas based on median slope using micro-topographical elevation data, generate a priority scoring matrix for evapotranspiration cluster data migration, and control the hydraulic translation device of the lyoinfiltration meter cluster through an IoT edge computing gateway to automatically migrate evapotranspiration clusters located in areas with high residual values ​​to the topographically representative areas with the highest priority scores.

[0068] Further explanation is needed: After completing the spatial redistribution of infiltration and achieving the closure of the water balance equation, in order to further improve the long-term accuracy of the observation network and avoid continuous residual bias in subsequent observation data due to the deviation of the evapotranspiration clusters from the representative areas of the terrain, it is necessary to optimize the layout of the evapotranspiration clusters through a feedback adjustment mechanism. This ensures that the observation nodes always conform to the water dynamic characteristics of the bamboo forest micro-topography. The specific implementation method is as follows:

[0069] The first step in adjusting the location of evapotranspiration clusters is to extract a residual spatial distribution heatmap from the dynamic water balance constraint calculation. The dynamic water balance constraint calculation outputs the residuals of each grid cell based on the previously constructed gridded water field balance residuals—that is, the difference between precipitation data and the sum of topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration estimates. These residual values ​​are matched to the grid according to spatial coordinates, and the residual distribution is visually presented using a color gradient. For example, dark red represents areas with residuals exceeding 0.2 mm (threshold), light yellow represents areas with residuals between 0.1 and 0.2 mm, and light blue represents areas with residuals below 0.1 mm, forming a residual spatial distribution heatmap. This heatmap clearly marks which areas have significant deviations between observed data and actual water balance patterns, providing a visual basis for subsequent anomaly area identification. It identifies micro-topographical anomalies with persistently high residuals. Sustained high residuals need to be defined in conjunction with the time dimension to avoid misjudgment due to single, accidental fluctuations, such as brief deviations caused by instantaneous heavy rainfall. Three consecutive observation periods are set, each period... If, over a period of 30 minutes (out of a total of 90 minutes), the residual values ​​of a certain grid all exceed a preset threshold (0.2 mm), and more than 50% of the grids within a 3×3 radius of that grid simultaneously exhibit elevated residual values, then that area is designated as a micro-topographical anomaly zone. During identification, grids with high residual values ​​in a single instance are first screened using a heatmap. Then, residual data from the last three periods for these grids are retrieved. Based on the criteria of continuous elevation and surrounding correlation, the anomaly zone's extent is ultimately determined. For example, at the junction of a ridge and a flat slope, if the residual values ​​reach 0.25 mm for three consecutive periods, and the surrounding six grids are simultaneously elevated, it is identified as an anomaly zone. Such areas are often due to existing evapotranspiration clusters not covering their unique micro-topography, resulting in unrepresentative observational data. Subsequently, representative areas based on the median slope are selected using micro-topographical elevation data. These representative areas are those within the anomaly zone whose slope values ​​are close to the median slope of all grids and reflect the mainstream topographical characteristics of the area. Extreme terrain with excessively steep or gentle slopes is avoided, as such terrain is difficult to represent the overall water transport pattern. The specific selection process is as follows:

[0070] First, extract the slope values ​​of all grids within the anomaly area from the micro-topographic elevation data (e.g., if the anomaly area contains 10 grids with slope values ​​of 15°, 16°, 14°, 20°, 12°, 17°, 13°, 18°, 11°, and 19°). Calculate the median of these slope values, sort the values, and take the median value, which is 15.5° in this case. Filter out grids whose slope value differs from the median by ≤1° (e.g., the grids corresponding to 15°, 16°, 14°, and 17°). Among these grids, prioritize those with terrain curvature close to 0, indicating more pronounced flat slope characteristics, more even water distribution, and distance from obstacles such as large bamboo groves or rocks. Define these as the median slope representative area, as this area best matches the overall terrain characteristics of the anomaly area. To provide ideal landing points for evapotranspiration cluster migration, the next step is to generate a migration priority scoring matrix for evapotranspiration cluster data. The scoring matrix uses the severity of residuals in anomaly areas, representing the representativeness of the regional topography, and the coverage of existing observation blind spots as core indicators. For each evapotranspiration cluster to be migrated, the matching degree between the cluster located in the high residual value area and the candidate representative area is scored. Specifically, for every 0.05mm increase in residuals in the anomaly area, the score increases by 10 points (weight 40%); for every 0.5° decrease in the difference between the slope of the representative area and the median of the anomaly area, the score increases by 8 points (weight 30%); and for every 5m increase in distance between the representative area and the existing evapotranspiration cluster (i.e., covering a new observation blind spot), the score increases by 6 points (weight 30%). The higher the total score, the higher the migration priority. For example, for a certain residual... The evapotranspiration clusters in the high-value residual areas correspond to anomaly areas with residuals exceeding the threshold by 0.1 mm (20 points), candidate representative areas with slope differences of 0.3° (48 points), and coverage blind zone distances of 8 m (9.6 points), resulting in a total score of 77.6 points. This establishes a clear priority ranking, ensuring that migration resources are prioritized for the areas most in need of optimization. Finally, the hydraulic translation device of the lysimeter cluster is controlled by an IoT edge computing gateway to automatically migrate the evapotranspiration clusters located in the high-value residual areas to the terrain representative areas with the highest priority scores. The IoT edge computing gateway is the core connecting the sensor network and the execution device, capable of receiving residual analysis results and scoring matrices in real time without relying on cloud computing, ensuring low-latency response to migration commands (latency ≤ 1 second). The hydraulic translation of the lysimeter cluster... The device is mounted on a movable guide rail, which is pre-laid according to a gridded moisture field to cover the observation area. It supports precise translation within a range of ±5m, and keeps the lysimeter level during the translation process to avoid tilting and affecting the observation accuracy. When performing the migration, the gateway sends a migration command to the evapotranspiration cluster in the high residual value area, including the coordinates of the target area and the translation path, avoiding the dense bamboo root area. After receiving the command, the hydraulic translation device moves smoothly at a speed of 5cm / s. After reaching the target position, it automatically calibrates the level and restarts evapotranspiration observation. After the migration is completed, the gateway feeds back the new position information to the data processing system, updates the coordinates of the evapotranspiration cluster in the gridded moisture field, realizes the adaptive optimization of the observation network, makes the subsequent observation data more consistent with the micro-topographic features, and reduces residual bias.

[0071] This invention synchronously collects multi-source collaborative observation parameter sets of bamboo forest slopes through an Internet of Things (IoT) sensor network, then constructs a point-to-area scale fusion bamboo forest water dynamic model, uses the random forest gradient boosting algorithm to correct evapotranspiration cluster data to generate topographically weighted evapotranspiration, aggregates runoff data according to topographic weights to obtain the total runoff sub-region, then performs dynamic water balance constraint calculation, combines variational assimilation algorithm to optimize infiltration, and when the residual exceeds the threshold, redistributes the spatial proportion of infiltration according to soil hydraulic conductivity. Finally, based on the residual spatial distribution heat map, the evapotranspiration clusters are migrated to representative micro-topographic areas to achieve adaptive optimization of the observation network. This invention solves the problems of data scale disconnect, insufficient observation accuracy, and difficulty in adapting the network to changes in traditional technologies, and improves the accuracy and reliability of bamboo forest ecological observation.

[0072] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A collaborative observation method for bamboo forest ecological environment based on multi-source Internet of Things sensing, characterized in that: Includes the following steps: S1. A multi-source collaborative observation parameter set for bamboo forest slopes is synchronously collected through an Internet of Things sensor network. The parameter set includes micro-topographic elevation data, micro-runoff unit data, evapotranspiration cluster data, soil moisture gradient data, and micro-meteorological data. S2. Construct a point-area scale fusion bamboo forest water dynamic model. Based on micro-topographic elevation data and soil moisture gradient data, use machine learning algorithms to perform topographic representativeness correction on evapotranspiration cluster data, generate topographic weighted evapotranspiration that matches the spatial resolution of micro-runoff unit data, and aggregate micro-runoff unit data according to topographic weight into the total amount of runoff sub-regions corresponding to evapotranspiration cluster data. S3. Perform dynamic water balance constraint calculation, integrate precipitation data, topographic weighted evapotranspiration and total runoff sub-region in real time, and iteratively optimize the infiltration estimate through variational assimilation algorithm to ensure that the precipitation data is equal to the sum of the total runoff sub-region, topographic weighted evapotranspiration and infiltration. When the balance residual between infiltration and the infiltration estimate exceeds the preset threshold, the spatial contribution ratio of infiltration is dynamically redistributed according to soil hydraulic conductivity. S4. Based on the spatial distribution characteristics of the equilibrium residuals, the location of evapotranspiration point clusters is adjusted to a representative area of ​​micro-topography to achieve adaptive optimization of the observation network; The specific steps for adjusting the layout of the evapotranspiration point clusters based on feedback are as follows: Extract the residual spatial distribution heat map from the dynamic water balance constraint calculation, identify micro-topographical anomaly areas with persistently high residuals, filter representative areas based on median slope using micro-topographical elevation data, generate a priority scoring matrix for evapotranspiration cluster data migration, and control the hydraulic translation device of the lyoinfiltration meter cluster through an IoT edge computing gateway to automatically migrate evapotranspiration clusters located in areas with high residual values ​​to the topographically representative areas with the highest priority scores.

2. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 1, characterized in that: The specific steps for constructing the bamboo forest moisture dynamics model that integrates point and area scales are as follows: A gridded moisture field with micro-topographic elevation data as the spatial framework was established, and evapotranspiration cluster data was used as the core observation node. A spatial interpolation surface of soil hydrological attributes was constructed through soil moisture gradient data. Machine learning algorithms were used to learn the nonlinear mapping relationship between evapotranspiration cluster data, micro-topographic elevation data, and soil moisture gradient data. An evapotranspiration topographic correction weight coefficient field covering the entire observation area was generated, and the spatial scale of micro-runoff unit data and evapotranspiration cluster data was dynamically fused based on this coefficient field.

3. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 2, characterized in that: The machine learning algorithm adopts a random forest gradient boosting architecture. Its input feature layer includes topographic curvature index, slope vector, aspect light intensity coefficient derived from micro-topographic elevation data, as well as vertical stratified water content difference rate and horizontal water migration trend parameters of soil moisture gradient data. The output layer is a topographic representativeness correction factor for each evapotranspiration cluster data unit. This factor is used to quantify the representativeness deviation of single-point evapotranspiration observations under micro-topographic variation conditions, and optimizes the spatial continuity of the evapotranspiration topographic correction weight coefficient field through a backpropagation mechanism.

4. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 3, characterized in that: The specific steps for correcting the topographic representativeness of evapotranspiration cluster data using machine learning algorithms are as follows: Using evapotranspiration cluster data as the baseline, and local topographic complexity index generated from micro-topographic elevation data and soil dry-wet spatial variability rate generated from soil moisture gradient data as driving variables, the topographic-soil coupling correction model is iteratively trained through a random forest gradient boosting architecture. The dynamic correction weight of each evapotranspiration cluster data unit is output. The evapotranspiration cluster data is multiplied by the corresponding dynamic correction weight to obtain the topographically corrected evapotranspiration that eliminates the interference of micro-topographic undulation and soil heterogeneity.

5. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 4, characterized in that: The terrain representativeness correction further includes: Based on micro-topographic elevation data, the terrain feature categories of the location of evapotranspiration point cluster data are identified, and three terrain response modes are divided into ridge, flat slope and valley. For each mode, a pre-trained soil moisture-topography coupled response function is loaded. This function is jointly calibrated by the vertical stratified water content change rate and the horizontal water redistribution intensity of soil moisture gradient data, and finally generates an adaptive correction coefficient matrix that is dynamically adjusted according to the terrain category and soil hydrological state.

6. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 5, characterized in that: The generation of terrain-weighted evapotranspiration that matches the spatial resolution of micro-runoff unit data is achieved through spatial scale dimensionality reduction mapping, and the specific steps are as follows: Using the spatial grid corresponding to the micro-runoff unit data as the target resolution unit, the topographically corrected evapotranspiration of all evapotranspiration cluster data within the unit is extracted. The spatial contribution weight is assigned based on the topographic similarity distance between each cluster and the center point of the target grid. The topographic similarity distance is composed of the elevation difference, slope difference, and curvature difference calculated from the micro-topographic elevation data. After weighted summation, a topographically weighted evapotranspiration that matches the spatial resolution of the micro-runoff unit data is generated.

7. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 1, characterized in that: The specific steps for aggregating micro-runoff unit data into total runoff sub-region data corresponding to evapotranspiration cluster data according to terrain weights are as follows: Based on the spatial distribution of evapotranspiration cluster data, the micro-runoff unit data is spatially assigned according to the confluence paths divided by micro-topographic elevation data. For micro-runoff units belonging to the same evapotranspiration cluster influence sub-region, aggregation weights are assigned based on their topographic connectivity index with the evapotranspiration cluster data center. The topographic connectivity index is calculated by combining slope continuity, elevation difference, and surface roughness. The weighted micro-runoff unit data is accumulated to generate the total runoff sub-region corresponding to the evapotranspiration cluster data.

8. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 1, characterized in that: The specific steps of the variational assimilation algorithm to iteratively optimize the infiltration estimate are as follows: A water balance cost function is constructed with precipitation data as input and topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration as state variables. A time continuity constraint is introduced through a four-dimensional variational assimilation framework to force the physical relationship between the rate of change of infiltration and soil hydraulic conductivity to be consistent between adjacent time steps. Finally, a quasi-Newton iteration method is used to minimize the cost function so that the residual of the precipitation data and the sum of the estimated values ​​of topographically weighted evapotranspiration, total runoff in sub-regions, and infiltration converges to a preset noise level.

9. The method for collaborative observation of bamboo forest ecological environment based on multi-source Internet of Things sensing according to claim 8, characterized in that: The specific steps for dynamically redistributing the spatial contribution ratio of infiltration based on soil hydraulic conductivity are as follows: When the balance residual exceeds the threshold, the spatial redistribution module based on soil moisture gradient data is triggered. It uses the vertical stratified soil moisture content change rate to invert the actual infiltration flux. Combined with the potential infiltration path network generated by micro-topographic elevation data, the residual exceeding the threshold is redistributed according to the relative proportion of the saturated hydraulic conductivity of each sub-region. After redistribution, the estimated infiltration value is updated and fed back to the variational assimilation iteration process until the water balance equation is closed.

Images