Desert wetland ecological effect automatic monitoring method

Abstract

The invention provides the technical field of ecological monitoring, and particularly relates to a desert wetland ecological effect automatic monitoring method, which comprises the following steps: acquiring space-based remote sensing data, near-earth aerial photography data and surface sensor data, generating a target data set through image registration and space-time alignment, and generating a desert wetland ecological effect data set; and constructing a dynamically coupled ecological system model to simulate a hydrological process and dynamic vegetation changes, and evaluating ecological risks to generate early warning information. According to the method, deep fusion and high-precision spatial alignment of multi-modal data can be realized, the problem of separation of a space-air-ground observation system is solved, and scientificity and accuracy of ecological risk early warning are remarkably improved.

Description

Technical Field

[0002] This invention relates to the field of evolution mechanism and synergistic evolution technology of groundwater-river-lake composite system in arid areas under changing environments, and in particular to an automatic monitoring method for the ecological effects of desert wetlands. Background Technology

[0004] Existing desert wetland ecological monitoring technologies generally suffer from a fragmented "space-air-ground" observation system, making it difficult to effectively integrate and comprehensively analyze multimodal data such as space-based remote sensing, near-ground aerial photography, and surface IoT. While space-based remote sensing has a wide coverage area, its spatial resolution is limited, making it difficult to characterize detailed hydrological and ecological processes in wetlands. Near-ground aerial photography, while capable of acquiring high-resolution images, lacks large-scale consistency verification capabilities. Although surface sensor data has high timeliness, it is difficult to integrate into the remote sensing semantic framework, resulting in isolated multi-source information and an inability to collaboratively analyze ecological dynamics. Furthermore, the lack of sub-pixel-level collaborative mechanisms for spatial registration and scale alignment among various modal data leads to significant uncertainties in the fusion results when characterizing the groundwater-river-lake transformation interface. More critically, existing methods mostly employ static or linear models, failing to dynamically couple the nonlinear disturbances of climate change and human water use behavior on wetland ecosystems, thus hindering the scientific rigor of ecological risk early warning and control decisions.

[0005] Therefore, there is an urgent need for an automatic monitoring method for the ecological effects of desert wetlands that can deeply integrate multimodal observation data, possess high-precision spatial alignment capabilities, and support dynamic human-land interaction simulation. Summary of the Invention

[0007] In view of the above problems, the present invention is proposed to provide an automatic monitoring method for the ecological effects of desert wetlands that overcomes or at least partially solves the above problems. This method can solve the problem that it is difficult to effectively integrate multimodal data such as space-based remote sensing, near-ground aerial photography and surface Internet of Things in existing desert wetland ecological monitoring technologies.

[0008] Specifically, the present invention provides an automatic monitoring method for the ecological effects of desert wetlands, comprising:

[0009] Acquire space-based data, near-Earth data, and surface data, wherein the space-based data describes the ecological environment data of the target area detected by space-based satellite remote sensing technology, the near-Earth data describes the ecological environment data of the target area detected by aerial photography technology, and the surface data describes the ecological environment data of the target area detected by multiple sets of surface sensors;

[0010] Image registration and / or spatiotemporal alignment are performed on the space-based data, the near-Earth data, and the surface data to obtain the target dataset;

[0011] Based on the target dataset, a dynamically coupled ecosystem model is constructed; wherein the ecosystem model is used to infer the impact of human activities on hydrological processes and / or vegetation dynamics in the target area.

[0012] Based on the ecosystem model, early warning of ecological risks in the target area is provided.

[0013] Furthermore, based on the ecosystem model, early warning of ecological risks in the target area is provided, including:

[0014] Current data is acquired through the aforementioned space-based satellite remote sensing technology, and / or the aforementioned aerial photography technology, and / or multiple sets of the aforementioned surface sensors;

[0015] The current data is introduced into the ecosystem model to obtain inference results;

[0016] If the predicted result is found to have reached the warning threshold, a warning message is generated and broadcast.

[0017] Furthermore, image registration is performed on the space-based data, the near-Earth data, and the surface data to obtain the target dataset, which includes:

[0018] Remote sensing images are generated from the space-based data, and aerial images are generated from the near-Earth data.

[0019] The aerial image is rasterized into a pixelated image with discrete pixels;

[0020] By performing an affine transformation on the pixelated image, a pixel-aligned image that is aligned with the pixels of the remote sensing image is obtained.

[0021] A Gaussian pyramid is constructed for the remote sensing image and the pixel-aligned image, and the normalized cross-correlation matrix of the remote sensing image and the pixel-aligned image is calculated layer by layer in the Gaussian pyramid to achieve sub-pixel localization through peak neighborhood two-dimensional parabolic interpolation, thereby obtaining the displacement vector of the corresponding level; wherein, the resolution of each level in the Gaussian pyramid decreases layer by layer.

[0022] The target displacement vector is calculated using the displacement vectors of each layer, and the pixel-aligned image is translated according to the target displacement vector to obtain the target aerial image;

[0023] The target dataset is obtained based on the aerial image of the target, the remote sensing image, and the surface data.

[0024] Further, based on the aerial image of the target, the remote sensing image, and the surface data, the target dataset is obtained, including:

[0025] Obtain the actual location of each group of surface sensors;

[0026] The spatial transformation model of the remote sensing image is used to obtain the relative mapping relationship between each group of surface sensors in the remote sensing image; wherein, the spatial transformation model describes the georegistration process between the image plane coordinates and the geospatial coordinates in the remote sensing image;

[0027] The target surface image corresponding to the surface data is generated through the relative mapping relationship;

[0028] The target aerial image, the remote sensing image, and the target map image are combined according to their geographical location mapping relationship to obtain the target dataset.

[0029] Furthermore, the space-based data, the near-Earth data, and the surface data are spatiotemporally aligned to obtain the target dataset, including:

[0030] The timestamps in the space-based data, near-Earth data, and surface data are all converted to Coordinated Universal Time (UTC) format, and the timestamps of the three are aggregated according to a set time to obtain time-aligned space-based data, near-Earth data, and surface data.

[0031] Further, acquiring the near-ground data includes:

[0032] Acquire local near-ground data to describe the local ecological environment of the target area;

[0033] The target near-ground data at the same time as the local near-ground data is inferred by a trained deep learning model, and the target near-ground data is used as the near-ground data; wherein the target near-ground data describes the overall ecological environment data of the target area.

[0034] Furthermore, the deep learning model is a deep convolutional neural network based on an attention mechanism. Its input is the standardized local near-ground data. The network structure includes a spatial-channel dual attention module. It achieves spectral fidelity constraint through a multi-scale adversarial loss function and outputs a high-resolution ecological state reconstruction map containing water distribution, vegetation coverage, soil moisture index and surface temperature. The high-resolution ecological state reconstruction map is used as the target near-ground data.

[0035] Furthermore, the training steps of the deep learning model include:

[0036] Using historical surface data of the target area as the training sample set and the spatiotemporally matched remote sensing images as the validation sample set, the model parameters are optimized through backpropagation algorithm until the mean square error between the output layer prediction value and the validation sample set reaches the convergence threshold, thereby obtaining the trained deep learning model.

[0037] The beneficial effects of this invention are:

[0038] This invention effectively solves the problem of difficult coordination of multi-source heterogeneous data in desert wetland ecological monitoring by using multi-level data fusion and dynamic coupling modeling technology. First, it constructs a spatiotemporally consistent target dataset through image registration / spatiotemporal alignment, overcoming the inherent differences in observation scale, accuracy, and modality between satellite remote sensing (hundred-meter resolution), aerial photography (sub-meter resolution), and surface sensors (point data). Then, the constructed dynamic coupling ecosystem model achieves physical mechanism-level fusion of multimodal data by associating hydrological processes with human activity influencing factors of vegetation dynamics (such as sudden changes in groundwater level and abnormal livestock density). Finally, by analyzing the deviation between the current monitoring data and the model prediction results in real time, an early warning is triggered when ecological parameters exceed the warning threshold, so that discrete satellite observation data, local aerial images, and fragmented IoT data form a complete "macro-monitoring-meso-verification-micro-validation" closed loop, significantly improving the spatiotemporal continuity and mechanistic interpretability of the monitoring results.

[0039] The above and other objects, advantages and features of the present invention will become more apparent to those skilled in the art from the following detailed description of specific embodiments of the invention in conjunction with the accompanying drawings. Attached Figure Description

[0041] The following sections will describe some specific embodiments of the invention in detail by way of example and not limitation, with reference to the accompanying drawings. The same reference numerals in the drawings denote the same or similar parts or portions. Those skilled in the art should understand that these drawings are not necessarily drawn to scale. In the drawings:

[0042] Figure 1 This is a schematic diagram of the overall technical solution architecture of the automatic monitoring method for the ecological effects of desert wetlands proposed in this invention;

[0043] Figure 2 This is a schematic diagram of the core principle framework of the sub-pixel level collaborative registration and spatiotemporal alignment mechanism for multi-source heterogeneous observation data in this invention;

[0044] Figure 3 This is a flowchart illustrating the logical process of deep learning-based global ecological state reconstruction and boundary recognition in this invention.

[0045] Figure 4This is a schematic diagram of the multi-module interaction relationship and data flow of hydrology, vegetation and human activities in the dynamic coupled ecosystem model of this invention;

[0046] Figure 5 This is a logical flowchart of the multi-scenario ecological risk prediction and hierarchical early warning mechanism in this invention. Detailed Implementation

[0048] Currently, in the field of desert wetland ecological monitoring, existing technologies generally suffer from a fragmented "space-air-ground" observation system, making it difficult to effectively integrate and comprehensively analyze multimodal data such as space-based remote sensing, near-ground aerial photography, and surface IoT. While space-based remote sensing has a wide coverage area, its spatial resolution is limited, making it difficult to characterize detailed hydrological and ecological processes in wetlands. Near-ground aerial photography, while capable of acquiring high-resolution images, lacks large-scale consistency verification capabilities. Surface sensor data, although highly timely, is difficult to integrate into the remote sensing semantic framework, resulting in isolated multi-source information and an inability to collaboratively analyze ecological dynamics. Furthermore, the lack of sub-pixel-level collaborative mechanisms for spatial registration and scale alignment among different modalities leads to significant uncertainties in the fusion results when characterizing the groundwater-river-lake transformation interface. More critically, existing methods mostly employ static or linear models, failing to dynamically couple the nonlinear disturbances to wetland ecosystems caused by climate change and human water use behavior, thus hindering the scientific rigor of ecological risk early warning and control decisions.

[0049] To address the aforementioned technical problems, this invention proposes an intelligent monitoring system that integrates multi-source heterogeneous observation data, possesses high-precision spatial alignment capabilities and dynamic ecological simulation capabilities, and applies it to an automatic monitoring method for the ecological effects of desert wetlands. This method includes:

[0050] First, in the data acquisition phase, space-based remote sensing data is collected via satellite platforms. This data primarily includes macroscopic features of the target area's ecological environment, such as vegetation cover, soil moisture index, and surface temperature. Simultaneously, near-ground aerial photography data is collected using drones or low-altitude aircraft. This data captures microscopic details of the target area's local ecological environment, such as vegetation species distribution and water boundary changes. Furthermore, surface sensor data is collected through a network of sensors deployed in the target area. These sensors include temperature and humidity sensors, soil moisture sensors, and wind speed sensors, used for real-time monitoring of the target area's ecological environment. These three data sources correspond to different spatial and temporal resolutions, providing a foundation for subsequent data fusion.

[0051] Next, the image registration module is entered into its processing stage. In this stage, remote sensing images are first generated from space-based remote sensing data, and aerial images are generated from near-Earth aerial photography data. To facilitate subsequent processing, the aerial images are rasterized into pixelated images with discrete pixels. Then, an affine transformation is performed on the pixelated images to generate a pixel-aligned image that is pixel-aligned with the remote sensing images. The specific implementation of the affine transformation includes calculating rotation matrices and translation vectors to ensure that the spatial coordinate systems of the aerial images and remote sensing images are consistent. Further, a Gaussian pyramid is constructed, and the normalized cross-correlation matrix between the remote sensing images and the pixel-aligned images is calculated layer by layer at each level of the pyramid. Sub-pixel localization is achieved through peak neighborhood two-dimensional parabolic interpolation to obtain the displacement vectors for the corresponding levels. Finally, the target displacement vector is calculated using the displacement vectors at each level, and the pixel-aligned image is translated according to the target displacement vector to generate the target aerial image. This process ensures high-precision spatial alignment between the space-based remote sensing data and the near-Earth aerial photography data.

[0052] Subsequently, the spatiotemporal alignment module is processed. In this stage, the timestamps of space-based remote sensing data, near-Earth aerial imagery data, and surface sensor data are uniformly converted to Coordinated Universal Time (UTC) format to eliminate time discrepancies between different data sources. Then, the data is aggregated according to set time intervals to generate a time-aligned dataset. For example, all data can be grouped according to hourly or daily time intervals for easier subsequent analysis. Furthermore, the actual location of the surface sensor data is mapped to its relative position in the remote sensing image using a spatial transformation model, generating a target surface image. This process converts the geographic coordinates of the surface sensor data into pixel coordinates in the remote sensing image by establishing a spatial transformation model, thus achieving spatial alignment between the surface sensor data and the remote sensing image. Finally, the target aerial image, remote sensing image, and target surface image are combined according to their geographic location mapping relationships to form the target dataset. The target dataset contains spatially and temporally aligned multimodal data, providing a high-quality data foundation for subsequent ecosystem modeling.

[0053] After generating the target dataset, the model is constructed using a dynamically coupled ecosystem model. The core function of this model is to simulate hydrological processes and vegetation dynamics in the target area, and to analyze the impact of human activities on the ecosystem. The model is built using deep learning technology, employing historical surface data as the training sample set and spatiotemporally matched remote sensing imagery as the validation sample set. During training, backpropagation is used to optimize model parameters until the mean squared error between the output layer predictions and the validation sample set reaches a convergence threshold. The model structure includes a spatial-channel dual attention module, which effectively extracts multi-scale features and enhances the expressive power of key regions. Furthermore, a multi-scale adversarial loss function is used to achieve spectral fidelity constraints, ensuring high spectral consistency in the model's output ecological state reconstruction map. The model's output includes high-resolution ecological state reconstruction maps of water distribution, vegetation cover, soil moisture index, and surface temperature, comprehensively reflecting the ecological environment status of the target area.

[0054] After constructing the dynamic coupled ecosystem model, the analysis phase of the ecological risk assessment module begins. In this phase, the ecological risk of the target area is assessed using the dynamic coupled ecosystem model. The assessment process mainly includes the following steps: First, based on the ecological state reconstruction map output by the model, the hydrological processes and vegetation dynamics of the target area are extracted. Then, combined with human activity data, the degree of disturbance to the ecosystem by human activities is analyzed. For example, by comparing historical data with current data, areas of ecological degradation caused by overgrazing or excessive water resource exploitation can be identified. Finally, based on the above analysis results, an ecological risk assessment report is generated. The assessment report includes the classification of ecological risk levels, the distribution of risk areas, and an analysis of potential impacts.

[0055] After the ecological risk assessment is completed, the early warning information generation phase begins. This generation process is based on the output of the ecological risk assessment module, using set thresholds to determine whether an ecological risk exists in the target area. For example, when the vegetation cover in a certain area decreases beyond a set threshold, the system automatically generates an early warning message, indicating the risk level and specific location of the area. Early warning information can be disseminated in various ways, such as notifying relevant department heads via SMS or displaying it to decision-makers through a visualization platform. Furthermore, early warning information can be integrated with other systems, such as linking with weather forecasting systems, to provide more comprehensive support for ecological risk management.

[0056] Throughout the process, the collaboration between the various modules is crucial. The data acquisition module is responsible for collecting multimodal data, providing raw data support for subsequent processing. The image registration module and the spatiotemporal alignment module work together to ensure high-precision alignment of the multimodal data in space and time. The target dataset, as an intermediate product, provides a high-quality data foundation for the construction of the dynamically coupled ecosystem model. The dynamically coupled ecosystem model achieves accurate simulation of the ecological environment status of the target area through deep learning technology, providing a scientific basis for the ecological risk assessment module. The ecological risk assessment module generates detailed ecological risk assessment reports by analyzing the model output results, providing decision support for the generation of early warning information. Finally, the release of early warning information enables timely response and scientific management of ecological risks in the target area.

[0057] As can be seen from the above specific implementation methods, the automatic monitoring method for desert wetland ecological effects provided by the present invention solves the problem of the fragmentation of the "space-air-ground" observation system in the prior art by combining deep fusion of multimodal data and high-precision spatial alignment technology with a dynamic human-land interaction model, and significantly improves the scientificity and accuracy of ecological risk early warning.

[0058] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0059] In practical applications of desert wetland ecological monitoring, space-based remote sensing data is first acquired via satellite platforms. For example, in a desert wetland reserve, high-resolution satellites are used to collect macroscopic ecological and environmental information such as vegetation cover, soil moisture index, and surface temperature in the target area. Simultaneously, drones equipped with multispectral cameras collect near-ground aerial data to capture local ecological and environmental details, such as vegetation species distribution and water boundary changes. Furthermore, a surface sensor network is deployed within the target area, including temperature and humidity sensors, soil moisture sensors, and wind speed sensors, to collect surface sensor data in real time. These sensors are deployed in a grid pattern to ensure coverage of the entire monitoring area, and data is transmitted to a central processing system via wireless communication modules. These three data sources correspond to different spatial and temporal resolutions, providing a foundation for subsequent data fusion.

[0060] The process then proceeds to the image registration module. In this phase, space-based remote sensing data is first converted into remote sensing images, and near-Earth aerial data is converted into aerial images. To facilitate subsequent processing, the aerial images are rasterized into pixelated images with discrete pixels. Then, an affine transformation is performed on the pixelated images to generate a pixel-aligned image that is pixel-aligned with the remote sensing images. The specific implementation of the affine transformation includes calculating rotation matrices and translation vectors to ensure that the spatial coordinate systems of the aerial images and remote sensing images are consistent. Further, a Gaussian pyramid is constructed, and the normalized cross-correlation matrix between the remote sensing images and the pixel-aligned images is calculated layer by layer at each level of the pyramid. Sub-pixel localization is achieved through peak neighborhood two-dimensional parabolic interpolation to obtain the displacement vectors for the corresponding levels. Finally, the target displacement vector is calculated using the displacement vectors at each level, and the pixel-aligned image is translated according to the target displacement vector to generate the target aerial image. This process ensures high-precision spatial alignment between the space-based remote sensing data and the near-Earth aerial data. For example, at the boundary of a wetland, this method enabled the precise overlay of aerial and remote sensing images, thus clearly demonstrating the dynamic changes of the water boundary.

[0061] Next, the spatiotemporal alignment module is processed. In this stage, the timestamps of space-based remote sensing data, near-Earth aerial photography data, and surface sensor data are uniformly converted to Coordinated Universal Time (UTC) format to eliminate time discrepancies between different data sources. For example, suppose the space-based remote sensing data is collected daily at 1 AM, the near-Earth aerial photography data is collected daily at 1 PM, and the surface sensor data is collected continuously at hourly intervals. By uniformly converting the timestamps of all data to UTC format and aggregating the data according to a set time interval (e.g., daily), a time-aligned dataset is generated. Furthermore, the actual location of the surface sensor data is mapped to its relative position in the remote sensing image using a spatial transformation model, generating a target surface image. This process establishes a spatial transformation model to convert the geographic coordinates of the surface sensor data into pixel coordinates in the remote sensing image, thus achieving spatial alignment between the surface sensor data and the remote sensing image. For example, at a certain monitoring point, this method accurately maps the soil moisture data collected by the surface sensor to the corresponding position in the remote sensing image, thereby forming the target surface image. Finally, the aerial images, remote sensing images, and surface images of the target are combined according to their geographic location mapping relationships to form the target dataset. The target dataset contains spatially and temporally aligned multimodal data, providing a high-quality data foundation for subsequent ecosystem modeling.

[0062] After generating the target dataset, the dynamic coupled ecosystem model is constructed. The core function of this model is to simulate hydrological processes and vegetation dynamics in the target area, and to analyze the impact of human activities on the ecosystem. The model construction process is based on deep learning technology, using historical surface data as the training sample set and spatiotemporally matched remote sensing imagery as the validation sample set. During model training, the backpropagation algorithm is used to optimize the model parameters until the mean squared error between the output layer predictions and the validation sample set reaches a convergence threshold. For example, in a wetland reserve, historical surface data from the past five years is used as the training sample set, and remote sensing imagery from the same period each year is used as the validation sample set. Through model training, accurate simulations of wetland hydrological processes and vegetation dynamics are achieved. The model structure includes a spatial-channel dual attention module, which effectively extracts multi-scale features and enhances the expressive power of key areas. Furthermore, a multi-scale adversarial loss function is used to implement spectral fidelity constraints, ensuring that the ecological state reconstruction map output by the model has high spectral consistency. The model outputs high-resolution ecological state reconstruction maps of water distribution, vegetation cover, soil moisture index, and surface temperature, which comprehensively reflect the ecological environment status of the target area. For example, during a dry season, the model successfully predicted the water shrinkage trend in the core wetland area and identified potential ecological degradation zones.

[0063] After constructing the dynamic coupled ecosystem model, the analysis phase of the ecological risk assessment module begins. In this phase, the ecological risk of the target area is assessed using the dynamic coupled ecosystem model. The assessment process mainly includes the following steps: First, based on the ecological state reconstruction map output by the model, the hydrological processes and vegetation dynamics of the target area are extracted. For example, in a wetland edge area, the vegetation cover map output by the model reveals a significant decrease in vegetation cover. Then, combined with human activity data, the degree of disturbance to the ecosystem by human activities is analyzed. For example, by comparing historical and current data, areas of ecological degradation caused by overgrazing or excessive water resource exploitation are identified. Finally, based on the above analysis results, an ecological risk assessment report is generated. The assessment report includes ecological risk level classification, risk area distribution, and potential impact analysis. For example, in a high-risk area, the assessment report shows that the vegetation cover has decreased by more than 30%, and the groundwater level has significantly decreased, indicating a serious risk of ecological degradation.

[0064] After the ecological risk assessment is completed, the early warning information generation phase begins. This generation process is based on the output of the ecological risk assessment module, using set thresholds to determine whether an ecological risk exists in the target area. For example, when the vegetation cover of a certain area decreases by more than a set threshold (e.g., 20%), the system automatically generates an early warning, indicating the risk level and specific location of the area. Early warning information can be disseminated in various ways, such as notifying relevant department heads via SMS or displaying it to decision-makers through a visualization platform. Furthermore, early warning information can be integrated with other systems, such as linking with weather forecasting systems, to provide more comprehensive support for ecological risk management. For instance, during a dry season, the system, in conjunction with weather forecasting systems, can issue early warnings of ecological risks in core wetland areas, providing a scientific basis for relevant departments to take emergency measures.

[0065] Throughout the process, the collaboration between the various modules is crucial. The data acquisition module is responsible for collecting multimodal data, providing raw data support for subsequent processing. The image registration module and the spatiotemporal alignment module work together to ensure high-precision alignment of the multimodal data in space and time. The target dataset, as an intermediate product, provides a high-quality data foundation for the construction of the dynamically coupled ecosystem model. The dynamically coupled ecosystem model achieves accurate simulation of the ecological environment status of the target area through deep learning technology, providing a scientific basis for the ecological risk assessment module. The ecological risk assessment module generates detailed ecological risk assessment reports by analyzing the model output results, providing decision support for the generation of early warning information. Finally, the release of early warning information enables timely response and scientific management of ecological risks in the target area.

[0066] As can be seen from the above specific application scenarios, the automatic monitoring method for desert wetland ecological effects provided by the embodiments of the present invention solves the problem of the fragmentation of the "space-air-ground" observation system in the existing technology by combining deep fusion of multimodal data and high-precision spatial alignment technology with a dynamic human-land interaction model, and significantly improves the scientificity and accuracy of ecological risk early warning.

[0067] The following reference Figures 1 to 5 To describe another embodiment of the present invention, an automatic monitoring method for the ecological effects of desert wetlands is described, the steps of which include:

[0068] S100 collaboratively acquires multi-scale remote sensing and sensor data covering the target area, including space-based satellite remote sensing data, near-ground UAV aerial imagery, and ecological environment parameters collected by the ground-based Internet of Things sensor network.

[0069] Specifically, in step S100, space-based satellite remote sensing data is acquired through a multi-satellite collaborative remote sensing strategy, fusing high spatial resolution panchromatic imagery with high temporal resolution multispectral data. The panchromatic imagery has a spatial resolution of 0.5 meters, the multispectral data has a time revisit period of 1 day, and the spectral bands cover 8 bands including visible light, near-infrared, and short-wave infrared, with center wavelengths of 485 nm, 560 nm, 660 nm, 830 nm, 1650 nm, 2100 nm, 1375 nm, and 2200 nm, respectively, and a quantization depth of 12 bits.

[0070] The near-ground drone aerial imagery is captured using a hybrid fixed-wing and multi-rotor drone platform. The flight altitude is set between 100 and 300 meters, and the ground surface sampling distance is controlled between 2 and 10 centimeters. It is equipped with a multispectral camera and a thermal infrared camera. The multispectral channels are strictly aligned with the space-based remote sensing bands. The center wavelength of the thermal infrared band is 10.7 micrometers, and the temperature sensitivity is better than 0.05 degrees Celsius.

[0071] The surface IoT sensor network consists of self-organizing wireless sensor nodes. The node deployment density is dynamically adjusted according to the ecological gradient of the wetland-desert transition zone, with no fewer than 20 nodes per square kilometer in the core area and no fewer than 5 nodes per square kilometer in the edge area. Each node integrates a soil moisture sensor, a soil temperature sensor, an air temperature and humidity sensor, an anemometer, and a groundwater level gauge. The soil moisture sensor has a measurement range of 0% to 60% with an accuracy of ±2%, the soil temperature measurement range is -20°C to 60°C with an accuracy of ±0.2°C, and the groundwater level gauge uses a pressure sensor with a range of 0 meters to 20 meters and an accuracy of ±1 centimeter. All surface nodes are networked via the LoRaWAN protocol, with a communication distance of up to 10 kilometers. Data is uploaded every 5 minutes, and time synchronization is achieved through a BeiDou / GNSS dual-mode timing module, with timestamp errors controlled within 100 milliseconds.

[0072] The overall architecture of the above multi-source data acquisition system is described in [reference]. Figure 1 The space-based remote sensing data stream is labeled 101, the near-ground aerial photography data stream is labeled 102, and the surface sensing data stream is labeled 103. All three are integrated into the data fusion hub 104.

[0073] S200, based on a deep learning model, uses high-resolution near-Earth imagery to spatially extrapolate and reconstruct the overall ecological state of a region, and uses space-based remote sensing data as a verification benchmark to ensure the fidelity of the extrapolation results in terms of spatial structure and spectral characteristics.

[0074] Specifically, in step S200, the deep learning model employs a deep convolutional neural network architecture incorporating an attention mechanism. Its backbone network is a variant of ResNet-101. The input layer receives multispectral and thermal infrared channel data from near-ground UAV aerial images, which are then input into the network after standardization. A spatial-channel dual attention module is embedded in the network's intermediate layers. This module first generates channel weight vectors through global average pooling, and then generates a spatial attention map through convolutional layers. The two are multiplied and applied to the feature map, enabling the model to focus on the key ecological boundaries of the wetland-desert transition zone. The model's output layer generates a global ecological state reconstruction map, containing four core layers: water distribution, vegetation cover, soil moisture index, and surface temperature, with a spatial resolution of 2 centimeters. To ensure the spectral fidelity of the reconstruction results, space-based remote sensing data was introduced as a supervisory signal, and a multi-scale adversarial loss function was constructed: at the pixel level, L1 loss was used to constrain the spectral consistency between the reconstructed image and the space-based remote sensing image at the same geographical location; at the feature level, a pre-trained VGG network was used to extract high-level semantic features, and the mean square error between feature maps was calculated; at the discriminant level, a PatchGAN discriminator was trained to distinguish the local authenticity of the reconstructed image and the real space-based remote sensing image. During training, a sliding window strategy was used to process large-scale images, with a window size of 1024×1024 pixels and a stride of 512 pixels. Overlapping areas were fused using a weighted average. The model training dataset contained 5000 pairs of paired samples of historical UAV and satellite images from the same period. The training epochs were 200, with an initial learning rate of 0.001, and a cosine annealing strategy was used for decay. The spatial structural integrity of the reconstructed results was evaluated using the Edge Preservation Index (EPI), with a target value of no less than 0.92; spectral fidelity was evaluated using Spectral Angle Mapping (SAM), with the average angular deviation controlled within 5 degrees. The reconstruction logic framework is described in [link to relevant documentation]. Figure 3 The input image 301 is processed by the deep learning model 302 to output a global ecological status map 303, which is then verified for fidelity by the space-based remote sensing verification module 304.

[0075] The S300 achieves refined coordination of multi-source data in terms of spatial location and timestamp through a sub-pixel-level image registration algorithm and a high-precision spatiotemporal alignment mechanism. The registration accuracy reaches 0.1 pixels, and the spatiotemporal alignment error is controlled within 1 meter and 5 minutes, respectively.

[0076] Specifically, in step S300, spatial registration employs a combination of a multi-scale pyramid registration strategy and a phase correlation algorithm. Wherein:

[0077] S310 performs initial coarse registration of space-based remote sensing images, near-ground aerial images, and geographic coordinates of surface sensors. It then performs affine transformation correction using ground control points (GCPs). There are no fewer than 20 GCPs, which are evenly distributed to cover the entire monitoring area, and the positioning accuracy is better than 0.5 meters.

[0078] S320: Construct a Gaussian pyramid with 5 levels, where the resolution of each level is reduced to half that of the previous level. At each pyramid level, calculate the normalized cross-correlation (NCC) matrix of the two images and perform sub-pixel localization around the peak using two-dimensional parabolic interpolation to obtain the displacement vector of that level.

[0079] S330 weights and fuses the displacement vectors of each level, with the weight increasing as the level increases. The weight of the bottom layer is 0.1 and the weight of the top layer is 0.5, thus obtaining the final sub-pixel level registration parameters.

[0080] For surface sensor data, its spatial location is directly mapped to the image coordinate system through GNSS positioning coordinates. The mapping process uses bilinear interpolation to ensure that the sensor observations can accurately correspond to the image pixel positions.

[0081] The S300 ring employs a time alignment mechanism based on a high-precision time synchronization system. Time stamps for space-based remote sensing data are provided by satellite-borne atomic clocks with an accuracy better than 1 millisecond; time stamps for near-Earth aerial imagery are recorded by an airborne GNSS time synchronization module with an accuracy better than 10 milliseconds; and the surface sensor network is synchronized via BeiDou time synchronization, with the overall network time deviation controlled within 50 milliseconds. The timestamps of multi-source data are uniformly converted to Coordinated Universal Time (UTC), and data aggregation is performed in 5-minute time windows. Data within each window is fused using a time-weighted average method, with the weight inversely proportional to the absolute value of the time distance from the center of the window. The registered and aligned multi-source data form a spatiotemporally consistent data cube with a spatial dimension of X×Y, a temporal dimension of T, and a feature dimension of F, where F includes at least 12 feature channels such as spectral, thermal infrared, soil moisture content, and groundwater level. The core principle framework of this collaborative registration and spatiotemporal alignment mechanism is described in [link to relevant documentation]. Figure 2 The multi-scale pyramid module 201, phase correlation calculation module 202, surface control point correction module 203, and spatiotemporal aggregation module 204 together constitute a complete alignment pipeline.

[0082] S400 is a dynamically coupled ecosystem model built based on fused multimodal data. It integrates three core modules: hydrological processes, vegetation dynamics, and human activities, and introduces data assimilation and scenario prediction mechanisms to achieve real-time simulation of wetland ecological status, prediction of future trends, and early warning of ecological risks.

[0083] Specifically, the hydrological process module describes the movement of water in the unsaturated zone based on the Richards equation, which is as follows:

[0084]

[0085] in, This refers to the volumetric water content. For time, For soil hydraulic conductivity, For soil water potential, The equation is solved by discretization on a two-dimensional horizontal plane with a grid size of 10 m × 10 m and a time step of 1 hour. Groundwater flow is described using Darcy's law, and the dynamics of unconfined aquifers are simulated using the Businesk equation.

[0086] The vegetation dynamics module employs an improved logistic growth model, incorporating water and temperature stress factors, and vegetation cover. The evolution equation is:

[0087]

[0088] in, The intrinsic growth rate For environmental carrying capacity, The water stress factor is defined as the ratio of current soil moisture content to field capacity. The temperature stress factor is calculated based on a Gaussian function, with a peak value at 25 degrees Celsius.

[0089] The human activity module quantifies water use behavior, including agricultural irrigation, industrial water use and domestic water use. The data comes from the local water affairs department's real-time reporting system and irrigation area obtained by remote sensing inversion. Water use is allocated to the model grid according to administrative division and land use type.

[0090] The three modules interact through a coupling interface. The hydrology module outputs soil moisture content and groundwater level as input to the vegetation module; the vegetation module outputs evapotranspiration as feedback to the upper boundary conditions of the hydrology module; the water consumption of the human activity module is directly deducted from the groundwater storage of the hydrology module. To improve model accuracy, an ensemble Kalman filter (EnKF) data assimilation component is embedded. The state vector contains three core variables: soil moisture content profile, groundwater level, and vegetation cover. The observation vector is the fused multimodal data. The ensemble size is set to 100, and the assimilation period is 6 hours. The prediction and early warning mechanism supports multi-scenario simulation, including a baseline scenario (maintaining current water consumption intensity), a water-saving scenario (reducing water consumption by 20%), and an expansion scenario (increasing water consumption by 30%). The simulation time span is 1 to 5 years, and the time step is 1 day. The comprehensive calculation formula for the Ecological Risk Index (ERI) is:

[0091]

[0092] in, Due to the depth of groundwater, For vegetation coverage, For water consumption, , , The weighting coefficients are 0.4, 0.4, and 0.2, respectively. A yellow alert is triggered when the ERI exceeds 0.6, and a red alert is triggered when it exceeds 0.8. The alert information is pushed to the management department through the automated platform within 30 minutes. See the diagram illustrating the multi-module interaction relationship and data flow of this dynamic coupling model. Figure 4 The hydrology module 401, vegetation module 402, and human activity module 403 are tightly coupled with the multi-scenario prediction engine 405 through the data assimilation interface 404.

[0093] In the aforementioned automatic monitoring method for the ecological effects of desert wetlands, the method is integrated into an automated monitoring platform, featuring low-latency data processing, high-concurrency user support, and a modular system architecture to meet operational needs. Specifically, the platform adopts a microservice architecture, deployed in a distributed cloud computing environment, and includes four core service modules: data access service, preprocessing service, model computation service, and early warning release service. The data access service receives space-based, near-ground, and surface data via gigabit-level fiber optic links, with a throughput of no less than 1Gbps; the preprocessing service performs data unpacking, format conversion, quality control, and preliminary fusion, with a processing latency of less than 2 minutes; the model computation service calls a GPU cluster to perform deep learning reconstruction and ecological model simulation, with a single full-domain simulation taking no more than 3 minutes; the early warning release service pushes early warning information to web, mobile, and SMS platforms via message queues, with end-to-end latency controlled within 5 minutes. The platform supports access from no fewer than 100 concurrent users and provides API interfaces for third-party systems to call. The system's operational status is tracked in real time through a health monitoring module, including CPU utilization, memory usage, disk I / O, and network bandwidth. Abnormal conditions automatically trigger alarms and activate the disaster recovery backup mechanism. For the platform's business processes and multi-scenario early warning logic, please refer to [link to relevant documentation]. Figure 5 The data input 501 is processed by the engine 502 to generate a prediction result 503, the risk assessment module 504 calculates the ERI value, and the early warning decision module 505 triggers the corresponding level of early warning 506 according to the threshold.

[0094] To verify the effectiveness of the method of this invention, a field application was conducted in the Taitma Lake wetland on the northern edge of the Taklamakan Desert. The monitoring area covered 500 square kilometers, with 120 surface sensor nodes deployed. Drone aerial photography was conducted twice a week, and space-based remote sensing data was updated daily. After six months of continuous operation, the results showed that: the spatial registration accuracy of multi-source data reached 0.08 pixels, with spatiotemporal alignment errors of 0.8 meters and 3.2 minutes, respectively; the F1-score for water body boundary identification in the global ecological state reconstruction map reached 0.94, and the root mean square error for vegetation cover inversion was 4.2%; the Nash efficiency coefficient for simulating groundwater level changes using the dynamic coupling model was 0.87; under the simulated water-saving scenario, the model predicted that vegetation cover would increase by 12% and groundwater depth would rise by 1.5 meters in five years; the system successfully provided an early warning of a local wetland degradation event caused by agricultural expansion 45 days in advance, with a warning response time of 22 minutes. These examples fully demonstrate the technical advantages of the method of this invention in multi-source fusion, high-precision reconstruction, dynamic simulation, and risk early warning.

[0095] Therefore, those skilled in the art should recognize that although numerous exemplary embodiments of the present invention have been shown and described in detail herein, many other variations or modifications conforming to the principles of the present invention can be directly determined or derived from the disclosure of the present invention without departing from the spirit and scope of the invention. Thus, the scope of the present invention should be understood and construed as covering all such other variations or modifications.

Claims

1. An automatic monitoring method for the ecological effects of desert wetlands, characterized in that, include: Acquire space-based data, near-Earth data, and surface data, wherein the space-based data describes the ecological environment data of the target area detected by space-based satellite remote sensing technology, the near-Earth data describes the ecological environment data of the target area detected by aerial photography technology, and the surface data describes the ecological environment data of the target area detected by multiple sets of surface sensors; Image registration and / or spatiotemporal alignment are performed on the space-based data, the near-Earth data, and the surface data to obtain the target dataset; Based on the target dataset, a dynamically coupled ecosystem model is constructed; wherein the ecosystem model is used to infer the impact of human activities on hydrological processes and / or vegetation dynamics in the target area. Based on the ecosystem model, early warning of ecological risks in the target area is provided.

2. The automatic monitoring method for the ecological effects of desert wetlands according to claim 1, characterized in that, Based on the ecosystem model, early warning of ecological risks in the target area is provided, including: Current data is acquired through the aforementioned space-based satellite remote sensing technology, and / or the aforementioned aerial photography technology, and / or multiple sets of the aforementioned surface sensors; The current data is introduced into the ecosystem model to obtain inference results; If the predicted result is found to have reached the warning threshold, a warning message is generated and broadcast.

3. The automatic monitoring method for the ecological effects of desert wetlands according to claim 1, characterized in that, Image registration is performed on the space-based data, the near-Earth data, and the surface data to obtain the target dataset, which includes: Remote sensing images are generated from the space-based data, and aerial images are generated from the near-Earth data. The aerial image is rasterized into a pixelated image with discrete pixels; By performing an affine transformation on the pixelated image, a pixel-aligned image that is aligned with the pixels of the remote sensing image is obtained. A Gaussian pyramid is constructed for the remote sensing image and the pixel-aligned image, and the normalized cross-correlation matrix of the remote sensing image and the pixel-aligned image is calculated layer by layer in the Gaussian pyramid to achieve sub-pixel localization through peak neighborhood two-dimensional parabolic interpolation, thereby obtaining the displacement vector of the corresponding level; wherein, the resolution of each level in the Gaussian pyramid decreases layer by layer. The target displacement vector is calculated using the displacement vectors of each layer, and the pixel-aligned image is translated according to the target displacement vector to obtain the target aerial image; The target dataset is obtained based on the aerial image of the target, the remote sensing image, and the surface data.

4. The automatic monitoring method for the ecological effects of desert wetlands according to claim 3, characterized in that, The target dataset is obtained based on the aerial image of the target, the remote sensing image, and the surface data, including: Obtain the actual location of each group of surface sensors; The spatial transformation model of the remote sensing image is used to obtain the relative mapping relationship between each group of surface sensors in the remote sensing image; wherein, the spatial transformation model describes the georegistration process between the image plane coordinates and the geospatial coordinates in the remote sensing image; The target surface image corresponding to the surface data is generated through the relative mapping relationship; The target aerial image, the remote sensing image, and the target map image are combined according to their geographical location mapping relationship to obtain the target dataset.

5. The automatic monitoring method for the ecological effects of desert wetlands according to claim 1, characterized in that, The space-based data, near-Earth data, and surface data are spatiotemporally aligned to obtain a target dataset, including: The timestamps in the space-based data, near-Earth data, and surface data are all converted to Coordinated Universal Time (UTC) format, and the timestamps of the three are aggregated according to a set time to obtain time-aligned space-based data, near-Earth data, and surface data.

6. The automatic monitoring method for the ecological effects of desert wetlands according to claim 1, characterized in that, Acquiring the near-Earth data includes: Acquire local near-ground data to describe the local ecological environment of the target area; The target near-ground data at the same time as the local near-ground data is inferred by a trained deep learning model, and the target near-ground data is used as the near-ground data; wherein the target near-ground data describes the overall ecological environment data of the target area.

7. The automatic monitoring method for the ecological effects of desert wetlands according to claim 6, characterized in that, The deep learning model is a deep convolutional neural network based on an attention mechanism. Its input is the standardized local near-ground data. The network structure includes a spatial-channel dual attention module. It achieves spectral fidelity constraint through a multi-scale adversarial loss function and outputs a high-resolution ecological state reconstruction map containing water distribution, vegetation coverage, soil moisture index and surface temperature. The high-resolution ecological state reconstruction map is used as the target near-ground data.

8. The automatic monitoring method for the ecological effects of desert wetlands according to claim 6, characterized in that, The training steps of the deep learning model include: Using historical surface data of the target area as the training sample set and the spatiotemporally matched remote sensing images as the validation sample set, the model parameters are optimized through backpropagation algorithm until the mean square error between the output layer prediction value and the validation sample set reaches the convergence threshold, thereby obtaining the trained deep learning model.

Images