Rat hole and vegetation coverage monitoring method for grassland ecology

Abstract

The invention relates to the technical field of remote sensing and ecological monitoring, and discloses a grassland ecology-oriented mousehole and vegetation coverage monitoring method, which integrates multi-temporal optical remote sensing, synthetic aperture radar and ground sensing data, realizes high-precision recognition of a mousehole by improving a U-shaped convolutional neural network, and improves the reliability of the mousehole. The accuracy is improved by combining interference coherence attenuation characteristics and microtopography verification; synchronously inverting the multi-exponential weighted vegetation coverage, constructing a time-space aligned double-variable time series data set, calculating a Granger causal coefficient by using a vector autoregression model, and quantifying the driving intensity of the mousehole density on vegetation degradation; and when the causal coefficient exceeds a threshold value and the vegetation coverage continuously decreases, triggering first-level ecological early warning. According to the method, accurate quantification and early warning of the rodent damage-vegetation causal chain are achieved, and scientificity and timeliness of grassland ecological monitoring are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing and ecological monitoring technology, specifically relating to a method for monitoring rodent burrows and vegetation coverage in grassland ecology. Background Technology

[0002] Grassland ecosystems, as vital terrestrial ecological barriers, directly impact soil and water conservation, biodiversity maintenance, and the stability of carbon sequestration. In recent years, rodent activity has increasingly damaged grassland vegetation, becoming a significant contributing factor to grassland degradation and desertification. Against this backdrop, rodent burrows, as a core indicator of rodent activity, are considered key indicators for assessing grassland ecological risk, with their spatial distribution density and dynamic trends being crucial metrics. Meanwhile, vegetation cover has long been a core parameter for measuring grassland productivity and resilience. Together, these two elements constitute the fundamental observational dimensions of a grassland ecological monitoring system.

[0003] Monitoring methods for rodent burrows and vegetation cover in grassland ecosystems focus on quantifying and tracking vegetation evolution under the influence of rodent damage through the fusion of remote sensing, ground sensing, and computational models. This technical direction aims to establish a spatiotemporal correlation mechanism between rodent burrow distribution and vegetation degradation, thereby providing data support for ecological early warning and governance decisions. Its basic principle lies in using multi-source observation data to capture the location of rodent burrows and their expansion patterns over time, while simultaneously acquiring the temporal variation characteristics of vegetation cover, and then inferring the potential causal relationship between the two.

[0004] Using separate modeling methods to handle rodent burrow identification and vegetation cover estimation lacks the ability to systematically model the spatiotemporal coupling relationship between the two. On the one hand, rodent burrow detection often relies on manual interpretation of high-resolution imagery or simple target detection algorithms, making it difficult to achieve high-precision, large-scale dynamic positioning against complex terrain backgrounds. On the other hand, vegetation cover analysis is often based on single-temporal or multi-temporal remote sensing indices (such as NDVI), failing to fully consider the nonlinear impact of rodent burrow disturbance on local vegetation succession paths. More importantly, current monitoring systems generally lack the ability to jointly model changes in rodent burrow density with vegetation degradation rates, failing to construct a quantitative expression of the key causal chain of "rodent burrow expansion → vegetation suppression → soil exposure," resulting in a severe lag in ecological risk early warning and hindering the effective implementation of early intervention measures. Summary of the Invention

[0005] This invention provides a method for monitoring rodent burrows and vegetation cover in grassland ecosystems, aiming to address the technical problems in existing technologies, such as the lack of spatiotemporal correlation analysis between rodent burrow distribution and vegetation cover changes, the inability to quantify the causal chain of "rodent burrow density → vegetation degradation," and the delay in ecological early warning response. This invention constructs a multi-source remote sensing data fusion framework, combining ground-based sensor networks and space-based observation platforms, to establish a dynamic identification model of rodent burrow spatial distribution and a temporal evolution model of vegetation cover. Based on these, a causal inference mechanism is introduced to achieve a quantitative assessment of the impact of rodent burrow density on vegetation degradation and early ecological risk warning.

[0006] As one embodiment of the present invention, the method for monitoring rodent burrows and vegetation coverage in grassland ecology includes the following steps:

[0007] Acquire multi-temporal high-resolution optical remote sensing images, synthetic aperture radar remote sensing data, and micro-topography and soil moisture data collected by ground sensor nodes for the target grassland area;

[0008] Based on the multi-temporal high-resolution optical remote sensing images, an improved U-shaped convolutional neural network model is used to perform pixel-level semantic segmentation of mouse holes and output a spatial distribution map of mouse holes.

[0009] Based on the synthetic aperture radar remote sensing data, the surface micro-deformation area is extracted using the interferometric coherence attenuation feature to help verify the accuracy of the mouse hole identification results; based on the micro-topography and soil moisture data collected by the ground sensor node, the mouse hole identification results are locally corrected to eliminate false detections caused by wind erosion pits or animal footprints.

[0010] The spatial density of the mouse holes is calculated to generate a mouse hole density raster map with preset grid cells as units.

[0011] Acquire vegetation cover remote sensing inversion data that is time-aligned with the mouse burrow density raster map. The vegetation cover remote sensing inversion data is obtained by weighted fusion of normalized vegetation index, enhanced vegetation index and soil-adjusted vegetation index.

[0012] The mouse hole density raster map and the vegetation cover remote sensing inversion data are registered using the same spatial grid to form a spatiotemporally aligned bivariate time series dataset.

[0013] Based on the bivariate time-series dataset, a vector autoregression model was constructed to calculate the Granger causality coefficient of mouse burrow density on vegetation cover.

[0014] Based on the magnitude and statistical significance of the Granger causality coefficient, the driving force of rat burrow density on changes in vegetation cover is determined.

[0015] When the Granger causality coefficient exceeds a preset threshold and the vegetation coverage decreases by more than 5% for three consecutive monitoring cycles, a Level I early warning signal for grassland ecological degradation is triggered.

[0016] Furthermore, the improved U-shaped convolutional neural network model introduces a hollow spatial pyramid pooling module in the encoder part to capture mouse hole morphological features at different scales; in the decoder part, skip connections are used to fuse multi-level features, and an attention weight mechanism is embedded in the last output channel to enhance the ability to identify small mouse hole edge regions; the loss function of the improved U-shaped convolutional neural network model is composed of a weighted cross-entropy loss and a boundary-aware loss, wherein the boundary-aware loss generates boundary labels by performing a morphological dilation operation on the mouse hole mask and assigns higher penalty weights to boundary pixels.

[0017] Furthermore, the process for extracting the interferometric coherence attenuation features of the synthetic aperture radar remote sensing data includes: selecting two synthetic aperture radar images of the same area with a time interval of no more than 15 days, performing precise registration and de-flattening processing; calculating the complex interferogram of the two images and performing multi-view processing to reduce speckle noise; performing phase filtering and phase unwrapping on the interferogram to obtain the surface deformation phase; converting the surface deformation phase into the deformation in the line-of-sight direction; performing time series analysis on the deformation to identify areas with continuous deformation less than 0.5 cm but significant spatial clustering, and marking them as potential rodent burrowing activity areas.

[0018] Furthermore, the ground sensing node includes a laser ranging micro-topography scanner and a soil dielectric constant sensor deployed in typical grassland sample plots; the laser ranging micro-topography scanner performs surface elevation scanning at a density of no less than 10 sampling points per square meter to generate a digital elevation model with centimeter-level accuracy; the soil dielectric constant sensor is buried 10 centimeters below the surface to monitor soil volumetric water content in real time; when a depression structure with a diameter between 5 centimeters and 30 centimeters and a depth greater than 3 centimeters appears in the digital elevation model, and the soil volumetric water content at the corresponding location is less than 20% of the regional average, the depression structure is confirmed as a valid mouse burrow.

[0019] Furthermore, the grid cell side length of the rat hole density grid map is 100 meters; the rat hole density is calculated by counting the number of verified valid rat holes in each grid cell, dividing it by the effective monitoring area of ​​that grid cell, and obtaining the effective rat hole density value, in units of holes per hectare.

[0020] Furthermore, the process of acquiring the vegetation cover remote sensing inversion data includes: performing atmospheric correction and geometric fine correction on the multispectral remote sensing images; calculating the normalized vegetation index, enhanced vegetation index, and soil-adjusted vegetation index respectively; dynamically adjusting the weighting coefficients of the three indices according to the grassland vegetation type and soil background brightness; the weighting coefficients are determined based on the correlation coefficient between the historical measured vegetation cover and each index, with higher correlation coefficients resulting in greater weights; and the final vegetation cover value is calculated using a weighted linear combination formula, with a value range of 0 to 1.

[0021] Furthermore, the order of the vector autoregression model is determined by the Akaike Information Criterion, and the maximum lag order does not exceed 6 monitoring periods. Before constructing the model, the stationarity of the mouse burrow density sequence and the vegetation cover sequence are tested separately. If they are not stationary, first-order differencing is performed. After the model is fitted, the Wald test is used to determine whether the mouse burrow density is a Granger cause of the vegetation cover, and the significance level of the test is set at 5%.

[0022] Furthermore, the preset threshold is determined based on a retrospective analysis of historical degradation events in grassland ecological zones; in typical grassland areas, the Granger causality coefficient threshold is set to 0.35; in desert steppe areas, the Granger causality coefficient threshold is set to 0.28; and in meadow steppe areas, the Granger causality coefficient threshold is set to 0.42.

[0023] Furthermore, after the first-level early warning signal for grassland ecological degradation is triggered, the system automatically calls up historical meteorological data and grazing intensity data for the same period to eliminate interference from the decline in vegetation coverage caused by drought or overgrazing. If the causal relationship still holds after eliminating external interference factors, a comprehensive early warning report is generated, which includes a heat map of rat burrow density, a vegetation coverage change trend map, and a causal intensity distribution map, and is pushed to the grassland ecological protection and management platform.

[0024] As another embodiment of the present invention, a mouse burrow and vegetation coverage monitoring system for grassland ecology is provided, which includes a multi-source data acquisition unit, a mouse burrow intelligent identification unit, a vegetation coverage inversion unit, a spatiotemporal registration unit, a causal inference analysis unit, and an ecological early warning decision unit.

[0025] The multi-source data acquisition unit is used to acquire multi-temporal high-resolution optical remote sensing images of the target grassland area, synthetic aperture radar remote sensing data, and micro-topography and soil moisture data collected by ground sensing nodes.

[0026] The intelligent mouse hole identification unit is used to perform pixel-level semantic segmentation of mouse holes based on the multi-temporal high-resolution optical remote sensing images using an improved U-shaped convolutional neural network model, and to verify and correct the results by combining the synthetic aperture radar interferometric coherence attenuation characteristics and ground sensing data, and output a spatial distribution map of mouse holes.

[0027] The vegetation cover inversion unit is used to acquire vegetation cover remote sensing inversion data that is time-aligned with the spatial distribution map of mouse holes;

[0028] The spatiotemporal registration unit is used to register the spatial distribution map of the mouse holes with the remote sensing inversion data of vegetation coverage according to the same spatial grid, forming a spatiotemporally aligned bivariate time series dataset;

[0029] The causal inference analysis unit is used to construct a vector autoregression model based on the bivariate time series dataset, calculate the Granger causality coefficient of mouse burrow density on vegetation cover, and determine the driving intensity.

[0030] The ecological early warning decision unit is used to trigger grassland ecological degradation early warning signals of corresponding levels based on the Granger causality coefficient and the magnitude of vegetation cover change, and to generate a comprehensive early warning report.

[0031] The intelligent mouse hole identification unit includes an optical image processing subunit, a radar feature extraction subunit, a ground data fusion subunit, and a mouse hole density calculation subunit. The optical image processing subunit performs forward inference of an improved U-shaped convolutional neural network model and outputs an initial mouse hole mask. The radar feature extraction subunit performs interferometric coherence attenuation analysis and outputs a mask of potential mouse hole activity areas. The ground data fusion subunit receives data from a laser ranging micro-topography scanner and a soil dielectric constant sensor to generate a ground verification mask. The three masks are used to generate a final mouse hole spatial distribution map through a logical AND operation. The mouse hole density calculation subunit performs gridded statistics on the final mouse hole spatial distribution map to generate a mouse hole density raster map.

[0032] The vegetation cover retrieval unit includes an atmospheric correction subunit, a vegetation index calculation subunit, a weight dynamic adjustment subunit, and a cover fusion subunit. The atmospheric correction subunit performs radiometric calibration and atmospheric correction on the multispectral remote sensing image. The vegetation index calculation subunit calculates the normalized vegetation index, enhanced vegetation index, and soil-adjusted vegetation index, respectively. The weight dynamic adjustment subunit dynamically allocates weighting coefficients based on the correlation between historical measured data and each index. The cover fusion subunit performs a weighted linear combination and outputs the final vegetation cover remote sensing retrieval data.

[0033] The causal inference analysis unit includes a sequence stabilization subunit, a model building subunit, a Granger test subunit, and a causal strength determination subunit. The sequence stabilization subunit performs a unit root test on the input time series and performs differencing operations when necessary. The model building subunit determines the optimal lag order based on the Akaike information criterion and fits a vector autoregressive model. The Granger test subunit performs a Wald test to determine the statistical significance of the causal relationship. The causal strength determination subunit compares the Granger causality coefficient with a preset threshold and outputs the driving strength level.

[0034] The ecological early warning decision-making unit includes an interference factor elimination subunit, an early warning level determination subunit, and a report generation subunit. The interference factor elimination subunit calls meteorological and grazing data, performs multivariate regression residual analysis, and eliminates the influence of non-rodent factors. The early warning level determination subunit determines whether a level one early warning is triggered based on the causal strength and the rate of vegetation cover decline. The report generation subunit integrates spatial distribution maps, time series curves, and causal maps to generate a structured early warning report.

[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0036] 1. By integrating optical remote sensing, synthetic aperture radar and ground sensor data, high precision and robustness of mouse hole identification were achieved, effectively distinguishing mouse holes from other surface depression features; by constructing a multi-index weighted inversion model of vegetation cover, the accuracy and timeliness of vegetation status monitoring were improved.

[0037] 2. By establishing a spatiotemporally aligned bivariate time-series dataset of mouse burrow density and vegetation cover, and introducing the Granger causal inference mechanism, we have technically achieved a quantitative assessment of the causal chain of "mouse burrow density → vegetation degradation", which breaks through the limitation of traditional correlation analysis that cannot distinguish the causal direction.

[0038] 3. By setting dual criteria of causal intensity threshold and vegetation change rate based on ecological zoning, the false alarm rate and missed alarm rate of ecological early warning are significantly reduced, and the early warning time of grassland rodent infestation-induced ecological degradation is advanced by at least two monitoring cycles. The entire methodology is fully automated, scalable, and deployable at the regional scale, providing scientific, accurate, and operable technical support for grassland ecological protection. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method for monitoring mouse burrows and vegetation coverage in grassland ecology proposed in this invention.

[0040] Figure 2 This is a schematic diagram of the core principle framework of the causal inference mechanism of the influence of rat burrow density on vegetation coverage in this invention;

[0041] Figure 3 This is a flowchart of the logical process for intelligent identification of mouse holes based on the fusion of multi-source remote sensing and ground sensing data in this invention.

[0042] Figure 4 This is a flowchart illustrating the logical process of vegetation cover multi-index weighted inversion and spatiotemporal registration in this invention.

[0043] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow between mouse hole density raster generation and bivariate time series dataset construction in this invention;

[0044] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between ecological degradation early warning decision-making and interference factor elimination in this invention. Detailed Implementation

[0045] Please refer to Figures 1 to 6 This invention provides a method for monitoring rodent burrows and vegetation cover in grassland ecosystems, aiming to address the technical problems in existing technologies, such as the lack of spatiotemporal correlation analysis between rodent burrow distribution and vegetation cover changes, the inability to quantify the causal chain of "rodent burrow density → vegetation degradation," and the delay in ecological early warning response. This method constructs a multi-source remote sensing data fusion framework, combining ground-based sensor networks and airborne observation platforms, to establish a dynamic identification model of rodent burrow spatial distribution and a temporal evolution model of vegetation cover. Based on these, a causal inference mechanism is introduced to achieve a quantitative assessment of the impact of rodent burrow density on vegetation degradation and early ecological risk warning.

[0046] The method for monitoring rodent burrows and vegetation cover in grassland ecology includes the following steps:

[0047] Acquire multi-temporal high-resolution optical remote sensing images, synthetic aperture radar remote sensing data, and micro-topography and soil moisture data collected by ground sensor nodes for the target grassland area;

[0048] Based on the multi-temporal high-resolution optical remote sensing images, an improved U-shaped convolutional neural network model is used to perform pixel-level semantic segmentation of mouse holes and output a spatial distribution map of mouse holes.

[0049] Based on the synthetic aperture radar remote sensing data, the surface micro-deformation area is extracted using the interferometric coherence attenuation feature to help verify the accuracy of the mouse hole identification results;

[0050] Based on the micro-topography and soil moisture data collected by the ground sensing nodes, the mouse hole identification results are locally corrected to eliminate false detections caused by wind erosion pits or animal footprints.

[0051] The spatial density of the mouse holes is calculated to generate a mouse hole density raster map with preset grid cells as units.

[0052] Acquire vegetation cover remote sensing inversion data that is time-aligned with the mouse burrow density raster map. The vegetation cover remote sensing inversion data is obtained by weighted fusion of normalized vegetation index, enhanced vegetation index and soil-adjusted vegetation index.

[0053] The mouse hole density raster map and the vegetation cover remote sensing inversion data are registered using the same spatial grid to form a spatiotemporally aligned bivariate time series dataset.

[0054] Based on the bivariate time-series dataset, a vector autoregression model was constructed to calculate the Granger causality coefficient of mouse burrow density on vegetation cover.

[0055] Based on the magnitude and statistical significance of the Granger causality coefficient, the driving force of rat burrow density on changes in vegetation cover is determined.

[0056] When the Granger causality coefficient exceeds a preset threshold and the vegetation coverage decreases by more than 5% for three consecutive monitoring cycles, a Level I early warning signal for grassland ecological degradation is triggered.

[0057] In step S1, multi-temporal high-resolution optical remote sensing images, synthetic aperture radar (SAR) remote sensing data, and micro-topography and soil moisture data collected by ground-based sensor nodes are acquired for the target grassland area. The multi-temporal high-resolution optical remote sensing images are acquired by a multispectral imager mounted on a low-Earth orbit satellite platform, with a spatial resolution of 0.5 meters to 2 meters, a revisit period of 3 to 7 days, and coverage bands including blue, green, red, and near-infrared bands. The SAR remote sensing data are acquired by a C-band or X-band spaceborne SAR system, with an incident angle range of 30 degrees to 45 degrees, a time baseline not exceeding 15 days, and a spatial resolution of 1 meter to 3 meters.

[0058] The ground-based sensing nodes are deployed in a grid pattern across typical grassland plots, with a node spacing of 500 to 1000 meters. Each node integrates a laser ranging micro-topographic scanner and a soil dielectric constant sensor. The laser ranging micro-topographic scanner scans the surface elevation at a density of no less than 10 sampling points per square meter, generating a digital elevation model with centimeter-level accuracy. The soil dielectric constant sensor is buried 10 centimeters below the surface, monitoring soil volumetric moisture content in real time, with a sampling frequency of once per hour. All data is uploaded to the central data processing platform via a wireless communication module, along with precise timestamps and geographic coordinates to ensure the accuracy of subsequent spatiotemporal alignment.

[0059] In step S2, based on the multi-temporal high-resolution optical remote sensing images, an improved U-shaped convolutional neural network model is used to perform pixel-level semantic segmentation of mouse holes, outputting a spatial distribution map of the mouse holes. The improved U-shaped convolutional neural network model includes an encoder-decoder structure. The encoder part consists of 5 downsampling stages, each stage containing two convolutional layers and one max-pooling layer. The convolutional kernel size is 3×3, and the activation function is a modified linear unit. After the fourth downsampling stage, a dilated spatial pyramid pooling module is introduced. This module uses dilated convolutional operations with dilation rates of 1, 6, 12, and 18 in parallel to capture the morphological features of mouse holes at different scales, including small isolated mouse holes, dense groups of mouse holes, and degenerate mouse holes with blurred edges.

[0060] The decoder consists of four upsampling stages. Each stage doubles the resolution of the feature map through transposed convolution and then performs skip connections to fuse it with the feature map of the corresponding layer in the encoder. The fusion method is channel concatenation followed by a convolutional layer. An attention weight mechanism is embedded in the last output channel. This mechanism generates a channel attention vector through global average pooling and then multiplies it with spatial location features to enhance the ability to recognize small mouse hole edge regions.

[0061] The model training employs a composite loss function consisting of a weighted average of cross-entropy loss and boundary-aware loss. The boundary-aware loss generates boundary labels by performing morphological dilation on the mouse hole mask and assigns a higher penalty weight to boundary pixels, with a weight coefficient of 2.5. The training dataset contains 5000 manually labeled mouse hole sample images, covering different seasons, lighting conditions, and grassland types. The training process uses a stochastic gradient descent optimizer with an initial learning rate of 0.001, a batch size of 8, and 150 training epochs.

[0062] In step S3, based on the synthetic aperture radar remote sensing data, the surface micro-deformation region is extracted using interferometric coherence attenuation characteristics to help verify the accuracy of the mouse burrow identification results. This process includes: selecting two synthetic aperture radar images of the same area with a time interval of no more than 15 days, performing precise registration and de-flattening processing; calculating the complex interferogram of the two images and performing multi-view processing to reduce speckle noise, with 4 and 2 multi-views in the azimuth and range directions, respectively; performing phase filtering and phase unwrapping on the interferogram to obtain the surface deformation phase; converting the surface deformation phase into deformation in the line-of-sight direction; performing time-series analysis on the deformation to identify areas with continuous deformation less than 0.5 cm but significant spatial clustering, marking them as potential mouse burrow activity areas.

[0063] Spatial clustering is determined using the local Moran's index test. A region is considered clustered when the local Moran's index is greater than 0.3 and the p-value is less than 0.05. The potential mousehole activity region mask output from this step is used to perform a logical intersection operation with the optical image segmentation results, preserving regions simultaneously identified as mouseholes by both modalities, thus improving the reliability of the recognition results.

[0064] In step S4, based on the micro-topography and soil moisture data collected by the ground sensing nodes, the mouse burrow identification results are locally corrected to eliminate false detections caused by wind erosion pits or animal footprints. Specifically, when a depression structure with a diameter between 5 cm and 30 cm and a depth greater than 3 cm appears in the digital elevation model, and the soil volumetric water content at the corresponding location is less than 20% of the regional average, the depression structure is confirmed as a valid mouse burrow.

[0065] The regional mean is calculated using a sliding window, with a window size of 50 meters × 50 meters. If a pixel is jointly identified as a mouse burrow by optical and radar sensors, but does not meet the aforementioned micro-topography and soil moisture conditions in the ground sensor data, it is removed from the final mouse burrow distribution map. This correction process is performed pixel-by-pixel after spatial resolution matching to ensure that ground verification information is accurately applied to the corresponding location.

[0066] In step S5, the spatial density of the mouse burrows is calculated, generating a mouse burrow density raster map using preset grid units. Each grid unit has a side length of 100 meters and covers the entire target grassland area. The mouse burrow density is calculated by counting the number of verified valid mouse burrows within each grid unit and dividing this number by the effective monitoring area of ​​that grid unit to obtain the effective mouse burrow density value, expressed as burrows per hectare. The effective monitoring area refers to the actual monitorable area after excluding non-grassland land types such as water bodies, roads, and buildings; this information is provided by the land use classification map. The density calculation results are stored in a floating-point raster format, with the spatial reference system being the WGS84 geographic coordinate system.

[0067] In step S6, vegetation cover remote sensing inversion data time-aligned with the mouse burrow density raster map is acquired. This vegetation cover remote sensing inversion data is obtained through weighted fusion of the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), and Soil-Adjusted Vegetation Index (SDI). The specific process includes:

[0068] Atmospheric correction and geometric fine correction were performed on the multispectral remote sensing images; the normalized vegetation index, enhanced vegetation index and soil-adjusted vegetation index were calculated respectively; the weighting coefficients of the three indices were dynamically adjusted according to the grassland vegetation type and soil background brightness.

[0069] The weighting coefficients are determined based on the correlation coefficients between the measured vegetation cover in the same period of history and each index. The higher the correlation coefficient, the greater the weight. The final vegetation cover value is calculated by a weighted linear combination formula, and its value ranges from 0 to 1.

[0070] Atmospheric correction employs a dark target subtraction model, while geometric fine correction uses ground control points and cubic convolution interpolation, resampling to the same resolution as the mouse burrow density raster. Historical measured vegetation cover data are obtained from UAV multispectral surveys conducted in fixed quadrats, collected four times annually during the growing season, and used to establish an exponential weighting mapping table. The weighted dynamic adjustment sub-unit queries this mapping table before each inversion to obtain the optimal weight combination under the current grassland type and soil brightness conditions.

[0071] In step S7, the rat burrow density raster map and the vegetation cover remote sensing inversion data are registered using the same spatial grid to form a spatiotemporally aligned bivariate time-series dataset. The registration process includes spatial resampling, coordinate system one, and timestamp alignment. Spatial resampling uses nearest-neighbor interpolation to ensure that the rat burrow density values ​​are not smoothed; coordinate system one is aligned to the universal transverse Mercator projection for easier subsequent spatial analysis; timestamp alignment is based on monitoring periods, with each period being 15 days. If there is no vegetation cover data in a certain period, linear interpolation between the preceding and following periods is used to fill in the gaps. The final bivariate time-series dataset is stored in NetCDF format, with each grid cell containing two time series: a rat burrow density series and a vegetation cover series, with a length of no less than 12 periods.

[0072] In step S8, based on the bivariate time-series dataset, a vector autoregressive model is constructed to calculate the Granger causality coefficient of mouse burrow density on vegetation cover. Before model construction, stationarity tests are performed on the mouse burrow density and vegetation cover sequences respectively using the augmented Dickey-Fuller test. If the p-value is greater than 0.05, it is determined to be non-stationary, and first-order differencing is performed. The order of the vector autoregressive model is determined by the Akaike information criterion, and the maximum lag order does not exceed 6 monitoring periods. The model form is as follows: ;

[0073] in, This represents the vegetation cover at time t. express The density of mouse holes at any given time The lag order is... This is the residual term. The Granger causality coefficient is defined as the proportion by which the rat burrow density reduces the variance of the vegetation cover prediction error, and is calculated using the following formula:

[0074] ;

[0075] in, To constrain the residuals of the model (excluding the mouse hole density lag term), This represents the residuals of the complete model.

[0076] In step S9, the driving strength of mouse burrow density on vegetation cover change is determined based on the magnitude and statistical significance of the Granger causality coefficient. A Wald test is used to determine whether mouse burrow density is a Granger cause of vegetation cover change, with a significance level set at 5%. If the p-value is less than 0.05 and the Granger causality coefficient is greater than 0, a one-way causal relationship is considered to exist. The driving strength levels are divided into: weak driving (coefficient less than 0.2), moderate driving (coefficient between 0.2 and 0.35), and strong driving (coefficient greater than 0.35). The preset threshold is determined based on a retrospective analysis of historical degradation events in grassland ecological zones; in typical grassland areas, the Granger causality coefficient threshold is set at 0.35; in desert steppe areas, the Granger causality coefficient threshold is set at 0.28; and in meadow steppe areas, the Granger causality coefficient threshold is set at 0.42.

[0077] In step S10, when the Granger causality coefficient exceeds a preset threshold and the vegetation cover decreases by more than 5% for three consecutive monitoring periods, a Level I early warning signal for grassland ecological degradation is triggered. Upon triggering, the system automatically retrieves historical meteorological data and grazing intensity data for the same period to eliminate interference from vegetation cover decline caused by drought or overgrazing. The meteorological data includes precipitation, temperature, and evapotranspiration, while the grazing intensity data includes livestock numbers and the number of grazing days.

[0078] The relationship between vegetation cover change and meteorological and grazing factors was fitted using a multiple linear regression model, and the residual sequence was calculated. If the residual sequence still showed a significant downward trend and maintained a strong causal relationship with rodent burrow density, then the degradation was confirmed to be dominated by rodent infestation. Finally, a comprehensive early warning report was generated, including a heat map of rodent burrow density, a vegetation cover change trend map, and a causal intensity distribution map, and was pushed to the grassland ecological protection and management platform.

[0079] The monitoring system for rodent burrows and vegetation coverage in grassland ecology includes a multi-source data acquisition unit, a rodent burrow intelligent identification unit, a vegetation coverage inversion unit, a spatiotemporal registration unit, a causal inference analysis unit, and an ecological early warning decision-making unit.

[0080] The multi-source data acquisition unit is used to acquire multi-temporal high-resolution optical remote sensing images of the target grassland area, synthetic aperture radar remote sensing data, and micro-topography and soil moisture data collected by ground sensor nodes. The intelligent mouse burrow identification unit is used to perform pixel-level semantic segmentation of mouse burrows based on the multi-temporal high-resolution optical remote sensing images using an improved U-shaped convolutional neural network model, and to verify and correct the results by combining the synthetic aperture radar interferometric coherence attenuation characteristics with ground sensor data, outputting a spatial distribution map of mouse burrows. The vegetation cover inversion unit is used to acquire vegetation cover remote sensing inversion data that is temporally aligned with the spatial distribution map of mouse burrows.

[0081] The spatiotemporal registration unit is used to register the spatial distribution map of mouse burrows with the remote sensing inversion data of vegetation cover using the same spatial grid, forming a spatiotemporally aligned bivariate time-series dataset. The causal inference analysis unit is used to construct a vector autoregressive model based on the bivariate time-series dataset, calculate the Granger causality coefficient of mouse burrow density on vegetation cover, and determine the driving strength. The ecological early warning decision-making unit is used to trigger a grassland ecological degradation early warning signal of the corresponding level based on the Granger causality coefficient and the magnitude of vegetation cover change, and generate a comprehensive early warning report.

[0082] The intelligent mouse hole identification unit includes an optical image processing subunit, a radar feature extraction subunit, a ground data fusion subunit, and a mouse hole density calculation subunit. The optical image processing subunit performs forward inference using an improved U-shaped convolutional neural network model, outputting an initial mouse hole mask. The radar feature extraction subunit performs interferometric coherence attenuation analysis, outputting a mask of potential mouse hole activity areas. The ground data fusion subunit receives data from a laser ranging micro-topography scanner and a soil dielectric constant sensor, generating a ground verification mask. The three masks are used in a logical AND operation to generate a final mouse hole spatial distribution map. The mouse hole density calculation subunit performs gridded statistical analysis on the final mouse hole spatial distribution map, generating a mouse hole density raster map.

[0083] The vegetation cover retrieval unit includes an atmospheric correction subunit, a vegetation index calculation subunit, a weight dynamic adjustment subunit, and a cover fusion subunit. The atmospheric correction subunit performs radiometric calibration and atmospheric correction on the multispectral remote sensing imagery. The vegetation index calculation subunit calculates the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EDI), and Soil-Adjusted Vegetation Index (SDI). The weight dynamic adjustment subunit dynamically allocates weighting coefficients based on the correlation between historical measured data and each index. The cover fusion subunit performs a weighted linear combination, outputting the final vegetation cover remote sensing retrieval data.

[0084] The causal inference analysis unit includes a sequence stabilization subunit, a model building subunit, a Granger test subunit, and a causality strength determination subunit. The sequence stabilization subunit performs a unit root test on the input time series and performs differencing operations when necessary. The model building subunit determines the optimal lag order based on the Akaike information criterion and fits a vector autoregressive model. The Granger test subunit performs a Wald test to determine the statistical significance of the causal relationship. The causality strength determination subunit compares the Granger causality coefficient with a preset threshold and outputs the driving strength level.

[0085] The ecological early warning decision-making unit includes a disturbance factor elimination subunit, an early warning level determination subunit, and a report generation subunit. The disturbance factor elimination subunit uses meteorological and grazing data to perform multivariate regression residual analysis to eliminate the influence of non-rodent factors. The early warning level determination subunit determines whether a Level 1 early warning is triggered based on causal strength and the rate of vegetation cover decline. The report generation subunit integrates spatial distribution maps, time-series curves, and causal maps to generate a structured early warning report.

[0086] This embodiment achieves a mouse hole identification accuracy of over 92%, a root mean square error of less than 0.08 for vegetation coverage inversion, a false alarm rate of less than 10% for causal inference, and an early warning lead time of over 30 days through the above-described method and system, which is significantly better than the existing technology.

Claims

1. A method for monitoring rodent burrows and vegetation cover in grassland ecology, characterized in that, include: Acquire multi-temporal high-resolution optical remote sensing images, synthetic aperture radar remote sensing data, and micro-topography and soil moisture data collected by ground sensor nodes for the target grassland area; Based on the multi-temporal high-resolution optical remote sensing images, an improved U-shaped convolutional neural network model is used to perform pixel-level semantic segmentation of mouse holes and output a spatial distribution map of mouse holes. Based on the synthetic aperture radar remote sensing data, the surface micro-deformation area is extracted using the interferometric coherence attenuation feature to help verify the accuracy of the mouse hole identification results; Based on the micro-topography and soil moisture data collected by the ground sensing nodes, the mouse hole identification results are locally corrected to eliminate false detections caused by wind erosion pits or animal footprints. The spatial density of the mouse holes is calculated to generate a mouse hole density raster map with preset grid cells as units. Obtain vegetation cover remote sensing inversion data that is time-aligned with the mouse hole density raster map; The mouse hole density raster map and the vegetation cover remote sensing inversion data are registered using the same spatial grid to form a spatiotemporally aligned bivariate time series dataset. Based on the bivariate time-series dataset, a vector autoregression model was constructed to calculate the Granger causality coefficient of mouse burrow density on vegetation cover. Based on the magnitude and statistical significance of the Granger causality coefficient, the driving force of rat burrow density on changes in vegetation cover is determined. When the Granger causality coefficient exceeds a preset threshold and the vegetation coverage decreases by more than 5% for three consecutive monitoring cycles, a Level I early warning signal for grassland ecological degradation is triggered.

2. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 1, characterized in that, The vegetation cover remote sensing inversion data were obtained by weighted fusion of normalized vegetation index, enhanced vegetation index and soil-adjusted vegetation index.

3. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 2, characterized in that, The improved U-shaped convolutional neural network model introduces a hole spatial pyramid pooling module in the encoder part to capture mouse hole morphological features at different scales. In the decoder section, skip connections are used to fuse multi-level features, and an attention weight mechanism is embedded in the last output channel to enhance the ability to identify small mouse hole edge regions. The loss function of the improved U-shaped convolutional neural network model is composed of a weighted sum of cross-entropy loss and boundary-aware loss. The boundary-aware loss generates boundary labels by performing a morphological dilation operation on the mouse hole mask and assigns higher penalty weights to boundary pixels.

4. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 3, characterized in that, The process of extracting the interferometric coherence attenuation features from the synthetic aperture radar remote sensing data includes: Two synthetic aperture radar images of the same area with an interval of no more than 15 days were selected for precise registration and flat-land effect removal. Calculate the complex interferogram of the two images and perform multi-view processing to reduce speckle noise; perform phase filtering and phase unwrapping on the interferogram to obtain the surface deformation phase; Convert the surface deformation phase into deformation along the line of sight; Time series analysis of the deformation variables identified areas with sustained deformation of less than 0.5 cm but significant spatial clustering, which were marked as potential mouse burrowing activity areas.

5. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 4, characterized in that, The ground sensing nodes include a laser ranging micro-topography scanner and a soil dielectric constant sensor deployed in typical grassland sample plots. The laser ranging micro-topography scanner performs surface elevation scanning at a density of no less than 10 sampling points per square meter to generate a digital elevation model with centimeter-level accuracy. The soil dielectric constant sensor is buried 10 centimeters below the surface to monitor the soil volumetric water content in real time. When a depression structure with a diameter between 5 cm and 30 cm and a depth greater than 3 cm appears in the digital elevation model, and the soil volumetric water content at the corresponding location is less than 20% of the regional average, the depression structure is confirmed as a valid mouse burrow.

6. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 5, characterized in that, The grid cell side length of the rat hole density raster is 100 meters; the rat hole density is calculated as follows: The number of verified effective mouse holes in each grid cell is counted, and then divided by the effective monitoring area of ​​that grid cell to obtain the effective mouse hole density value, expressed in units per hectare.

7. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 6, characterized in that, The process of acquiring the vegetation cover remote sensing inversion data includes: performing atmospheric correction and geometric fine correction on the multispectral remote sensing images; The normalized vegetation index, enhanced vegetation index, and soil-adjusted vegetation index were calculated separately; the weighting coefficients of the three indices were dynamically adjusted according to the grassland vegetation type and soil background brightness. The weighting coefficients are determined based on the correlation coefficients between the measured vegetation cover in the same historical period and each index; the higher the correlation coefficient, the greater the weight. The final vegetation coverage value is calculated using a weighted linear combination formula, and its value ranges from 0 to 1.

8. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 7, characterized in that, The order of the vector autoregressive model is determined by the Akaike Information Criterion, and the maximum lag order does not exceed 6 monitoring cycles; Before constructing the model, the stationarity of the mouse burrow density sequence and the vegetation cover sequence were tested. If they were not stationary, first-order differencing was performed. After model fitting, the Wald test was used to determine whether mouse burrow density was a Granger cause of vegetation cover, with the significance level set at 5%.

9. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 8, characterized in that, The preset threshold is determined based on a retrospective analysis of historical degradation events in grassland ecological zones; In typical grassland areas, the Granger causality coefficient threshold was set at 0.35; In the desert steppe region, the Granger causality coefficient threshold was set at 0.28; In the meadow steppe region, the Granger causality coefficient threshold was set at 0.

42.

10. The method for monitoring rodent burrows and vegetation cover in grassland ecology according to claim 9, characterized in that, After the first-level early warning signal for grassland ecological degradation is triggered, the system automatically calls up historical meteorological data and grazing intensity data for the same period to eliminate interference from the decline in vegetation cover caused by drought or overgrazing. If the causal relationship still holds after excluding external interference factors, a comprehensive early warning report is generated, which includes a heat map of rat burrow density, a vegetation coverage change trend map, and a causal intensity distribution map, and is pushed to the grassland ecological protection and management platform.

Images