Space-air-ground collaborative ecological quality multi-source data fusion method and system

Abstract

The present application relates to the field of ecological environment monitoring and multi-source data fusion, and discloses a space-air-ground coordinated ecological quality multi-source data fusion method and system, comprising: collecting multi-source data, at least including satellite remote sensing image data, unmanned aerial vehicle aerial image data and ground observation time series data; standardizing and spatiotemporal aligning the multi-source data to obtain a multi-modal data set; constructing a multi-modal prediction model and training the model and data assimilation fusion using the multi-modal data set; inputting the real-time collected multi-source data of the to-be-measured region into the trained multi-modal prediction model to output the evaluation and prediction results of the ecological quality core indicators, the ecological quality core indicators including at least one of the ecosystem integrity, water conservation capacity and permafrost degradation influence degree; and the present application can improve the accuracy and generalization ability of the ecological quality index inversion as a whole.

Description

Technical Field

[0001] This invention relates to the field of ecological environment monitoring and multi-source data fusion technology, specifically to a method and system for ecological quality multi-source data fusion that integrates air, space, and ground. Background Technology

[0002] With the increasing demand for renewable energy development and ecological protection, the importance of ecological quality monitoring in high-altitude and cold mountainous areas, as ecologically fragile zones and core water conservation areas, is becoming increasingly prominent. Currently, ecological quality monitoring in these areas mainly relies on three types of data: ground-based station observations, satellite remote sensing, and drone aerial photography. However, significant technical bottlenecks exist:

[0003] 1. Heterogeneous characteristics of multi-source data make fusion difficult: Satellite remote sensing data has a wide coverage but limited resolution, UAV aerial photography data is detailed but has a small coverage, and ground observation data is accurate but the stations are sparsely distributed. The three types of data present different modal characteristics such as images and time series. The scale differences between modalities, information redundancy and noise interference make it difficult for traditional fusion methods to effectively integrate them.

[0004] 2. Poor cross-modal feature integration: Existing fusion technologies mostly use simple splicing or weighted summation methods, which do not fully explore the potential correlations between modalities. In particular, they lack a deep adaptation mechanism for image features (spatial dimension) and temporal features (temporal dimension), resulting in low quality of fused features that cannot support high-precision ecological assessment.

[0005] 3. Modal bias and loss imbalance exist in the training process: In the training of multimodal models, image modalities are easy to optimize due to their rich feature structure, while temporal modalities (ground-based measured data), although core references, are easily ignored. Furthermore, traditional loss functions only focus on a single prediction task and cannot balance the effects of single-modal feature learning and multimodal fusion, resulting in weak model generalization ability and insufficient prediction accuracy.

[0006] 4. Lack of a unified air-space-ground collaborative framework: Existing technologies have not formed a collaborative mechanism for the entire process of "data acquisition-standardization-fusion-evaluation". Remote sensing data lacks verification by ground measurement data, and the ability to dynamically evaluate and predict trends is insufficient, making it difficult to meet the needs of comprehensive, accurate and dynamic monitoring of ecological quality in high-altitude and cold mountainous areas. Summary of the Invention

[0007] To address the shortcomings of existing technologies, the core objective of this invention is to provide a method and system for the fusion of multi-source ecological quality data in a space-air-ground collaborative manner. This method and system are suitable for comprehensive monitoring of ecological quality in complex terrains such as high-altitude and cold mountainous areas, accurate inversion of core indicators, and dynamic trend prediction. It can be widely applied in scenarios such as ecological environmental protection, natural resource management, and climate change response.

[0008] To achieve the above objectives, the following technical solution is adopted:

[0009] In a first aspect, embodiments of the present invention provide a method for fusing multi-source ecological quality data in a space-air-ground coordinated manner, comprising the following steps:

[0010] Collect multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data, and ground observation time series data;

[0011] The multi-source data is standardized and spatiotemporally aligned to obtain a multimodal dataset that includes image modalities and temporal modalities with temporal-spatial dual-dimensional alignment.

[0012] A multimodal prediction model is constructed, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics.

[0013] The model training and data assimilation fusion using the multimodal dataset specifically include: employing a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, using a non-negative tensor decomposition method with applied dynamic sparsity constraints to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; and optimizing the training process using a joint loss function, which includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term.

[0014] The multi-source data of the area to be tested collected in real time is input into the trained multimodal prediction model, and the evaluation and prediction results of the core ecological quality indicators are output. The core ecological quality indicators include at least one of the following: ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

[0015] Furthermore, the ecosystem integrity is quantified based on at least one of vegetation cover or NDVI, the ratio of bare / disturbed surface, and landscape fragmentation, and output as a score of 0–100. The water conservation capacity is calculated using water balance on a monthly or annual scale, satisfying the condition that water conservation capacity (W) = precipitation (P) - evapotranspiration (ET) - runoff (R) + soil moisture content (ΔS). The degree of permafrost degradation is calculated based on at least one of active layer thickness (ALT), ground temperature change trend, and freeze-thaw period length / freeze-thaw days. The weighted sum of the normalized active layer thickness (ALT), ground temperature change trend, and freeze-thaw period length / freeze-thaw days is mapped to a level of 0–10.

[0016] Furthermore, the differentiable architecture search algorithm is used to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure, specifically including:

[0017] The preset discrete search spaces of the image feature extraction branch and the temporal feature extraction branch are each constructed into a continuously differentiable supernetwork; wherein the structure of the supernetwork is defined by nodes and directed edges, and each edge is associated with multiple candidate operations in the preset discrete search space;

[0018] By assigning trainable architecture parameters to each candidate operation on each edge and normalizing the architecture parameters using the Softmax function, a continuous mixed operation representing the weights of each candidate operation is obtained.

[0019] By constructing and solving a two-layer optimization problem with network weights as inner optimization variables and the architecture parameters as outer optimization variables, the network weights and the architecture parameters are jointly optimized.

[0020] After the search is completed, the candidate operation with the largest weight is selected for each edge according to the optimized architecture parameters, thereby determining the final network architecture of the image feature extraction branch and the temporal feature extraction branch.

[0021] Furthermore, the image feature extraction branch employs a convolutional neural network containing normal convolutional units and dimensionality-reducing convolutional units; the temporal feature extraction branch employs a recurrent neural network.

[0022] The preset discrete search space of the image feature extraction branch includes at least 3×3 and 5×5 dilated convolution, depthwise separable convolution, 3×3 max pooling, 3×3 average pooling, and various activation functions; the preset discrete search space of the temporal feature extraction branch includes at least various activation functions.

[0023] Furthermore, the dynamic sparsity constraint is implemented in the following way:

[0024] A dynamic sparse normalization loss term is introduced into the joint loss function. The dynamic sparse normalization loss term constrains feature sparsity by calculating the sum of the standardized L1 norms of each modality feature, so as to balance the contribution weights of different modalities in the fusion process and suppress redundant features.

[0025] Furthermore, the fusion module adopts a step-by-step fusion strategy, specifically including: first, fusing satellite remote sensing image features and UAV aerial image features to obtain preliminary fused features; then, fusing the preliminary fused features with ground observation time series features to obtain the final cross-modal fused features.

[0026] Furthermore, the prediction head is a multilayer perceptron containing at least two fully connected layers; wherein, the first fully connected layer maps the fused features to a high-dimensional feature space, the second fully connected layer maps the high-dimensional features to an output dimension consistent with the number of core ecological quality indicators, and nonlinear activation functions are used for transformation between layers.

[0027] Furthermore, the multi-source data undergoes standardization processing, including:

[0028] Image data standardization employs Z-Score normalization; ground time-series monitoring data undergoes logarithmic normalization, including logarithmic transformation based on the ratio of observations at adjacent times to eliminate dimensions and enhance data stability.

[0029] Secondly, embodiments of the present invention also provide a space-air-ground integrated ecological quality multi-source data fusion system for implementing the space-air-ground integrated ecological quality multi-source data fusion method as described in the first aspect, the system comprising:

[0030] The data acquisition module is used to acquire multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data, and ground observation time series data.

[0031] The data preprocessing module is used to standardize and spatiotemporally align the multi-source data to obtain a multimodal dataset that includes image modalities and temporal modalities with temporal-spatial dual-dimensional alignment.

[0032] The model building module is used to build a multimodal prediction model, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics.

[0033] The model training module is used for model training and data assimilation fusion using the multimodal dataset. Specifically, it includes: using a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, a non-negative tensor decomposition method with dynamic sparsity constraints is used to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; during training, a joint loss function is used for optimization, which includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term;

[0034] The ecological quality assessment module is used to input multi-source data of the area to be tested collected in real time into the trained multimodal prediction model and output the assessment and prediction results of the core ecological quality indicators. The core ecological quality indicators include at least one of ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

[0035] Compared with the prior art, the present invention achieves the following beneficial effects:

[0036] 1. Significantly improved cross-modal fusion accuracy: Through non-negative CANDECOMP / PARAFAC decomposition and factor fusion technology, deep integration of image and temporal features is achieved, effectively mining potential correlations between modalities. Compared with traditional splicing fusion methods, the information entropy of fused features is effectively improved, providing high-quality data support for ecological indicator inversion.

[0037] 2. Enhanced model training stability and generalization ability: By designing a joint loss function that takes into account the needs of multiple tasks, the multimodal training and evaluation tasks are balanced, as well as the single-modal loss and multimodal fusion loss are balanced. This not only constrains the accuracy of single-modal feature extraction such as images and time series, but also ensures the predictive performance of fused features, thus alleviating the modality bias problem. This improves the feature extraction accuracy of the core modality (ground time series data), effectively reduces the prediction error of the model in different high-altitude and cold mountainous scenarios, and demonstrates significantly better robustness than existing technologies.

[0038] 3. Maximize the effectiveness of air-space-ground collaborative monitoring: By integrating the advantages of satellite full coverage, UAV fine supplementation, and ground precise verification, a unified data platform is built to realize the transformation of ecological quality monitoring from "point observation" to "full coverage" and from "static assessment" to "dynamic prediction". The monitoring cycle is significantly shortened and the accuracy of core indicator inversion is improved.

[0039] 4. Highly practical and low deployment cost: The method and process are standardized and modular, supporting data access from different satellites, drones and ground observation equipment. No complex hardware upgrades are required, and it can be adapted to existing ecological monitoring networks. At the same time, the network structure is automatically optimized through Differentiable Architecture Search (DARTS), reducing manual design costs and facilitating its application in ecological monitoring of various complex terrains.

[0040] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0041] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0042] Figure 1 This is a schematic diagram of a multi-source data fusion method for ecological quality that integrates air, space, and ground, provided in an embodiment of the present invention.

[0043] Figure 2 This is a framework diagram of the air-space-ground collaborative data fusion system according to an embodiment of the present invention;

[0044] Figure 3 This is a schematic diagram of the architecture of the multimodal prediction model according to an embodiment of the present invention;

[0045] Figure 4 This is a schematic diagram of the dynamic sparse nonnegative tensor decomposition and fusion process according to an embodiment of the present invention;

[0046] Figure 5 This is a schematic diagram of a multi-source data fusion system for ecological quality that integrates air, space, and ground, according to an embodiment of the present invention.

[0047] Figure 6 This is a performance comparison chart between the algorithm of this invention and traditional algorithms. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0050] The core of this invention lies in the following: First, by adaptively optimizing the structure of image and temporal feature extraction networks through Differentiable Architecture Search (DARTS) technology to adapt to heterogeneous data sources; second, by employing Non-Negative Tensor Coefficient Decomposition (NNCP) with dynamic sparsity constraints to achieve deep fusion of image and temporal features at the factor level, effectively mining cross-modal correlations; and finally, by designing a joint loss function to collaboratively optimize multimodal fusion prediction, single-modal feature learning, and dynamic sparsity constraints, thereby alleviating modality bias, balancing the training process, and ultimately improving the overall accuracy and generalization ability of ecological quality index inversion.

[0051] like Figure 1 and Figure 2 As shown, a multi-source data fusion method for ecological quality that integrates air, space, and ground includes the following steps:

[0052] S1: Collect multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data, and ground observation time series data;

[0053] Step S1 is used to achieve multi-source data acquisition:

[0054] 1. Satellite remote sensing data acquisition: 12 high-resolution satellites, 2 medium-resolution satellites and supporting remote sensing observation systems were selected to acquire multi-temporal and multispectral remote sensing images of high-altitude and cold mountainous areas from 2018 to 2025. Key ecological quality indicators such as vegetation coverage, water area and land use type were extracted, covering the entire area including glaciers and water conservation areas, high-altitude and cold endorheic areas. The data format is 3-channel (RGB).

[0055] 2. UAV Aerial Photography Data Acquisition: For areas with insufficient satellite remote sensing resolution (such as alpine wetlands and shrub core areas), high-resolution images are acquired through low-altitude UAV aerial photography to supplement detailed data such as vegetation growth, surface micro-topography, and traces of human activities, forming a supplementary verification data layer for satellite remote sensing. The data format is 3-channel RGB images, with a cumulative sample size of ≥5000 images.

[0056] 3. Ground Observation Data Acquisition: Relying on observations across four dimensions—hydrology, soil, atmosphere, and ecology—measured data on air pressure, temperature, humidity, precipitation, soil temperature and humidity, soil moisture content, evapotranspiration, latent heat flux, runoff, soil and vegetation nitrogen and phosphorus concentrations, and vegetation biomass were collected from 2018 to 2025 using instruments at comprehensive observation fields and evapotranspiration observation fields. The sampling interval was 15 minutes, with a cumulative sample size of ≥36,000 data points. Based on this ground observation data, quantitative values ​​for ecosystem integrity and water conservation capacity were calculated using well-known ecological quality assessment methods or in accordance with relevant ecological and environmental technical specifications. Ecosystem integrity was quantified based on at least one of the following: vegetation cover or NDVI, surface bare / disturbed ratio, and landscape fragmentation. Preferably, each component was first normalized to obtain normalization coefficients, and the weights of each indicator were preset or determined through learning from historical samples. Water conservation capacity is preferably estimated using a combination of water balance and storage change, calculated monthly or annually. The specific formula is: Water conservation capacity (W) = Precipitation (P) - Evapotranspiration (ET) - Runoff (R) + Soil moisture content change (ΔS). Based on soil temperature profile data and freeze-thaw process data from the permafrost hydrothermal monitoring system, the active layer thickness (ALT), interannual variation trend of ground temperature, and freeze-thaw period length / number of freeze-thaw days are calculated. After normalizing the above indicators, the weighted summation results are mapped to the permafrost degradation impact level G (0–10) as a supervision label for model training.

[0057] S2: Standardize and spatiotemporally align the multi-source data to obtain a multimodal dataset with temporal-spatial dual-dimensional alignment, including image modalities and temporal modalities;

[0058] Step S2 is used to achieve multi-source data standardization:

[0059] 1. Image data standardization: Satellite remote sensing images and UAV aerial images are uniformly resized to... Pixels are used to eliminate feature extraction bias caused by differences in image size; Z-Score normalization is applied to pixel values ​​to ensure consistent image feature scale, and the standardized values ​​are then used as the model image modal input, with the formula:

[0060] ;

[0061] The raw value of a single pixel in a satellite remote sensing image or drone aerial image, covering pixel data of the RGB three channels; Z-score normalized pixel values ​​are used to eliminate brightness scale differences between different images. The average value of all pixels in a single image, calculated over all channels and pixels of the image, reflects the overall brightness level of the image; Standard deviation of all pixels in a single image measures the dispersion of pixel values ​​in the image and is used to balance differences in image contrast.

[0062] 2. Time-series data standardization: Logarithmic normalization is applied to the time-series data from ground observations. Specifically, a logarithmic transformation is performed based on the ratio of observations at adjacent times to eliminate dimensional differences and enhance data stability. This data serves as the time-series modal input for the model. The formula is as follows:

[0063] ;

[0064] : The current observation time of the time series data, corresponding to the 15-minute sampling interval node of the ground observation network; : Real-time ground-based data, including time-series indicators related to ecological quality such as air pressure, temperature, humidity, soil temperature and humidity, nitrogen and phosphorus concentrations in water bodies, and vegetation biomass; The previous observation time (i.e.) The raw ground measurement data (minutes) and... For the same observation index; Log-normalized time series characteristic values ​​are used to eliminate dimensional differences between different observed indicators (such as temperature units). (Humidity unit %), enhancing the stability and comparability of time-series data; The natural logarithm function is used to transform the trend of time series data into a linearly modelable form, thus mitigating the impact of extreme values ​​on model training.

[0065] 3. Data alignment processing: Time-series alignment of the three types of data is achieved through timestamp matching, and spatial alignment of the entire domain data is achieved through geographic coordinate mapping, forming a multimodal dataset with "time-space" dual-dimensional alignment. The input dimensions satisfy: image modality 112×112×3, time-series modality (time step × number of features) (time step ≥ 30, number of features ≥ 8).

[0066] The spatial resolution of the image modality refers to the number of rows and columns of pixels in the standardized satellite remote sensing image and UAV aerial image, ensuring consistency in feature extraction. The image modality has three channels, corresponding to the core RGB channels (red, green, and blue), used to carry the image's color and spectral information. The time step is the length of the input sequence for the temporal modality, i.e., the number of consecutive observations for a single input to the model, requiring... (correspond minute Hourly continuous time-series data is used to capture long-term time-series dependencies; Feature number: the number of input indicators for the time-series mode, i.e., the types of ground observation indicators input simultaneously, requiring... (Covering core eco-hydrological indicators such as air pressure, temperature, humidity, precipitation, and soil temperature and humidity) to ensure the comprehensiveness of time-series information.

[0067] S3: Construct a multimodal prediction model, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics;

[0068] Step S3 is used to construct the multimodal prediction model, such as... Figure 3 The diagram shown is a schematic diagram of the architecture of the multimodal prediction model according to an embodiment of the present invention.

[0069] Step S3.1: Backbone Network Construction

[0070] Image feature extraction branch: A convolutional neural network (CNN) containing normal and reduced-dimensional convolutional units is used, including 4 normal cells and 2 reduced-dimensional convolutional cells. The kernel sizes cover 3×3, 5×5 dilated convolution, and depthwise separable convolution, combined with 3×3 max pooling and average pooling operations. The activation function is searched for optimal combinations from a candidate set (including Tanh, ReLU, Sigmoid, etc.) using the DARTS algorithm. DARTS stands for Differentiable Architecture Search, a highly efficient neural network architecture search method that transforms the discrete search space into a continuous space and uses gradient descent to jointly optimize the architecture and network weights, eliminating the need to train candidate architectures one by one, significantly reducing search costs. During training, the Adam optimizer is used with an initial learning rate of 5e-5, weight decay of 1e-4, and a dropout probability of 0.2. The image input size is fixed at 112×112 to ensure consistency in feature extraction.

[0071] Temporal Feature Extraction Branch: The temporal feature extraction branch employs a recurrent neural network (RNN) containing four recurrent units. The candidate activation functions are Tanh, ReLU, Sigmoid, Softplus, ELU, PReLU, and SeLU. The optimal activation combination is searched using the DARTS algorithm to capture temporal dependencies. The initial learning rate is 1e-4, the weight decay is 1e-4, and the dropout probability is 0.5. The temporal input time step is ≥30, and the number of features is ≥8, meeting the requirements for long-term temporal dependency modeling.

[0072] Step S3.2: Construction of the Fusion Module

[0073] like Figure 4 The diagram illustrates the dynamic sparse nonnegative tensor decomposition fusion process according to an embodiment of the present invention. It integrates the dynamic sparse NNCP fusion operation. NNCP, short for "Non-negative CANDECOMP / PARAFAC decomposition," is a variant of tensor decomposition. By imposing nonnegative constraints on tensor factors, it avoids mutual cancellation between factors, making it easier to identify matching components in multi-source data. Simultaneously, combined with a dynamic sparse strategy, it dynamically filters key features of each modality, suppressing noise and redundancy. In the fusion module, the tensor decomposition rank is set to 32, and the number of iterations is 25; the L1 norm weight κ of the dynamic sparse normalization is 0.1 to balance the contribution weights of multiple modalities; the dropout probability of the fusion branch is 0.5 to further reduce the risk of overfitting.

[0074] Step S3.3: Prediction Head Construction

[0075] The prediction head is an MLP (Multilayer Perceptron), which contains two fully connected layers. The first layer has an output dimension of 128, and the second layer has an output dimension consistent with the number of core ecological quality indicators (ecosystem integrity, water conservation capacity, degree of impact of permafrost degradation, etc.). The activation function is ReLU to ensure the nonlinear transformation capability of the features.

[0076] Specifically, in this embodiment, the prediction head is a multilayer perceptron containing at least two fully connected layers; wherein, the first fully connected layer maps the fused features to a high-dimensional feature space, and the second fully connected layer maps the high-dimensional features to an output dimension consistent with the number of core ecological quality indicators, and a non-linear activation function is used for transformation between layers.

[0077] In this invention, ecosystem integrity is used to characterize the integrity of the structure and function of the ecosystem in the monitoring area, preferably with an output score of 0–100. Ecosystem integrity is calculated based on at least one of the following quantities: vegetation cover (VC) or NDVI retrieved from remote sensing / UAVs, the ratio of bare land to human disturbance, and landscape fragmentation; optionally, ground plot biomass or species richness may be included. Each component is normalized and weighted to obtain ecosystem integrity. Water conservation capacity is used to characterize the ability of precipitation input to be intercepted, infiltrated, stored, and released as delayed runoff / baseflow on the underlying surface, preferably with an output unit of mm. The degree of permafrost degradation involves processes such as increased active layer thickness, increased permafrost temperature, prolonged freeze-thaw period, and changes in freeze-thaw frequency, which can manifest as eco-hydrological responses such as changes in surface water content, changes in surface runoff / baseflow processes, and changes in moisture conditions during the vegetation growing season.

[0078] Step S3.4 Initialization of key parameters

[0079] During training, the batch size was set to 64, with a total of 250 training epochs. Cosine annealing was used for learning rate scheduling, with a minimum learning rate of 1e-6 and a maximum learning rate of 1e-3. The optimizer was consistently Adam, with weight decay of 1e-4 to ensure training stability. In the architecture search phase, the architecture search learning rate for the image feature extraction branch was 5e-5, and the network training learning rate was also 5e-5. The architecture search learning rate for the temporal feature extraction branch was 1e-4, and the network training learning rate was 3e-4. Differentiated learning rates were used to adapt to the training needs of different modules. Furthermore, tensor decomposition parameters were initialized according to the above settings to ensure stable model startup.

[0080] S4: Model training and data assimilation fusion are performed using the multimodal dataset, specifically including: using a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, a non-negative tensor decomposition method with applied dynamic sparsity constraints is used to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; during training, a joint loss function is used for optimization, the joint loss function including at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term;

[0081] Step S4 is used to achieve multi-source data assimilation and fusion:

[0082] S4.1: DARTS-based backbone network architecture search

[0083] The DARTS algorithm achieves joint optimization of the architecture and network weights by transforming the discrete architecture search space into a continuous space. Specifically, a differentiable architecture search algorithm is used to search the architecture of the image feature extraction branch and the temporal feature extraction branch to optimize the network structure. This includes: for the image feature extraction branch, within its preset search space, jointly searching and optimizing the types of convolution operations, pooling operations, and activation functions. The search space includes at least 3×3 and 5×5 dilated convolutions, 3×3 / 5×5 depthwise separable convolutions, 3×3 max pooling, 3×3 average pooling, and various activation functions; for the temporal feature extraction branch, within its preset search space, searching and optimizing the types of activation functions in the recurrent units.

[0084] Furthermore, a differentiable architecture search algorithm is employed to search the architecture of the image feature extraction branch and the temporal feature extraction branch respectively, in order to optimize the network structure. Specifically, this includes: constructing a continuously differentiable supernetwork for each of the preset discrete search spaces of the image feature extraction branch and the temporal feature extraction branch; wherein the structure of the supernetwork is defined by nodes and directed edges, with each edge associated with multiple candidate operations in the preset discrete search space; assigning trainable architecture parameters to each candidate operation on each edge, and normalizing the architecture parameters using the Softmax function to obtain a continuous mixed operation representing the weights of each candidate operation; jointly optimizing the network weights and the architecture parameters by constructing and solving a two-layer optimization problem with network weights as inner optimization variables and architecture parameters as outer optimization variables; after the search is completed, selecting the candidate operation with the largest weight for each edge based on the optimized architecture parameters, thereby determining the final network architecture of the image feature extraction branch and the temporal feature extraction branch. The core process is as follows:

[0085] S4.1.1: Search Space Construction

[0086] The search space of the image feature extraction branch includes operations such as 3×3 / 5×5 dilated convolution, 3×3 / 5×5 depthwise separable convolution, 3×3 max / average pooling, Zero, and Identity.

[0087] The search space of the temporal feature extraction branch includes activation functions such as Tanh, ReLU, Sigmoid, Softplus, ELU, PReLU, and SeLU, and defines the combination of operations between nodes in the form of a directed acyclic graph (DAG).

[0088] S4.1.2: Continuous Relaxation and Mixed Operations

[0089] The operations on each edge (i,j) in the directed acyclic graph (DAG) are continuously relaxed. A weighted combination of all candidate operations is modeled using mixed operations, with the following formula:

[0090] ;

[0091] : Edges in DAG The output of the hybrid operation is a weighted fusion of all candidate operations, used to continue the architecture search space; A directed edge in a directed acyclic graph (DAG) connects node i (input node) and node j (output node), representing the operational association between the two nodes; : Input node i's feature vector / tensor, which is the image feature tensor for the CNN backbone and the temporal feature vector for the RNN backbone; Candidate operation set A specific candidate operation in the process; This means that this specific operation is applied to the feature vector / tensor x of the input node i; : A predefined set of candidate operations, for the image feature extraction branch, May include Operations such as dilated convolution, depthwise separable convolution, and pooling; for the temporal feature extraction branch. It can include 7 activation functions such as Tanh and ReLU; :side Candidate operation The corresponding architecture parameters are used to measure the weight of this operation in the hybrid combination; Represents the set of candidate operations Any candidate operation in the process; :side Any candidate operation The corresponding architecture parameters; exp: exponential function, used to convert architecture parameters into non-negative weights, in preparation for Softmax normalization; The sum of the exponentialized architecture parameters corresponding to all candidate operations is used as the denominator of the Softmax normalization to ensure that the total weights are 1.

[0092] All edges All candidate operations Corresponding architecture parameters Together, they constitute the set of learnable architectural parameters for the entire network, denoted as . .

[0093] S4.1.3: Set a two-level optimization target

[0094] This step aims to jointly optimize the set of architecture parameters. With network weight set The goal is to minimize the validation set loss. Specifically, this two-layer optimization problem can be solved using an approximate gradient descent method, thus efficiently updating both the architecture parameters and network weights simultaneously. This involves jointly optimizing the set of architecture parameters. With network weight set Minimize the validation set loss, the formula is:

[0095] ;

[0096] ;

[0097] : A set of architecture parameters, containing all edges in the DAG. corresponding Define the network topology; : A set of network weights, containing learnable parameters such as convolutional kernels and recurrent units in the backbone of a CNN / RNN; The validation set loss function, using RMSE (root mean square error), is used to evaluate the generalization performance of the current architecture and weight combination. The training set loss function, RMSE, is used to fit the feature-label mapping relationship of the training data. Given a set of architecture parameters The optimal set of network weights that minimizes the loss of the training set; The argmin function is used to solve for... The set of network weights that achieves the global minimum .

[0098] Because solving the two-layer optimization problem precisely requires a huge amount of computation, a single training step is used to approximate the optimal network weight set. The gradient approximation formula is:

[0099] ;

[0100] : Architecture parameters The gradient operator is used to update the architecture parameters to reduce the validation set loss; : Network weights The gradient operator is used to calculate the gradient of the training set loss with respect to the weights; The learning rate of the inner layer is optimized to control the network weights. The update step size is consistent with the training learning rate of the backbone network (CNN backbone). RNN backbone 1e-4). The network weights updated through a single gradient descent are used to approximate the optimal set of network weights. This reduces computational complexity.

[0101] S4.1.4: Architecture Discretization

[0102] After the search converges, the architecture parameters for each edge are... After obtaining the Softmax, the operation with the largest weight is selected as the final operation for that edge to construct the optimal backbone network structure.

[0103] S4.2: Dynamic Sparse Tensor Fusion

[0104] Furthermore, such as Figure 4As shown, in the fusion module, a non-negative tensor decomposition method with dynamic sparsity constraints is used to perform cross-modal factor-level deep fusion of features extracted by the image feature extraction branch and the temporal feature extraction branch. This includes: performing non-negative CANDECOMP / PARAFAC decomposition on the image features extracted by the image feature extraction branch and the temporal features extracted by the temporal feature extraction branch to obtain non-negative feature factors; performing dynamic sparse normalization on the feature factors to screen key features and suppress noise and redundancy; and fusing and reconstructing the screened feature factors to obtain cross-modal fused features. The specific implementation is as follows:

[0105] S4.2.1: Dynamic Sparse Normalization

[0106] Furthermore, dynamic sparsity constraints are achieved by introducing a dynamic sparse normalization loss term into the joint loss function. The dynamic sparse normalization loss term constrains feature sparsity by calculating the sum of the standardized L1 norms of each modality feature, so as to balance the contribution weights of different modalities in the fusion process and suppress redundant features.

[0107] After standardizing the features of each modality, the L1 norm is calculated and incorporated into the loss function to balance the contribution weights of multiple modalities and avoid a single modality dominating the fusion process.

[0108] The specific calculation process is performed for a single training batch:

[0109] (1) Obtain batch features:

[0110] Image feature batch Dimensions are ,in, For batch size, , These are the space height and width, respectively. This represents the number of image feature channels.

[0111] Time-series feature batch In a preferred embodiment, It is a 2D feature matrix with dimension . ,in This is the temporal feature dimension. It is used in calculating the dynamic sparse normalization loss. At that time, in order to be with To carry out unified standardization processing, Adjusted to a 4D tensor by adding dimensions. .

[0112] (2) Calculate the loss: Standardize the two batches of 4D features mentioned above (subtract the mean and divide the standard deviation), and then calculate the feature tensors of each standardized feature. Norm and summation, the formula is:

[0113] ;

[0114] : Dynamic sparse normalization loss term, used to constrain the sparsity of features of each modality and balance the contribution weights of multiple modalities; Modal index, Corresponding image modal features, Corresponding temporal modal features; Summing the sparsity loss for both image and temporal modalities comprehensively constrains multimodal features; : No. Primitive characteristics of class modalities; : No. The mean of the modal features is obtained by statistical analysis of the training set features in the batch dimension B, and is used to standardize the feature mean. : No. The standard deviation of the modal features is obtained through statistical analysis of the training set features in the batch dimension B and is used to standardize the feature scale. : No. The standardization results of modal features eliminate scale differences between different modalities; The L1 norm operator calculates the sum of the absolute values ​​of standardized features, used to introduce sparsity and suppress redundant and noisy features.

[0115] S4.2.2: Dynamic Sparse NNCP Fusion

[0116] NNCP fusion operations are performed at the single-sample level. For a single sample:

[0117] (1) Extracting features from a single sample:

[0118] Single-sample image features From image feature batches Take a sample from the sample to obtain a 3D tensor with dimension . .

[0119] Single-sample time series features From the time-series characteristics of batch (Before adjusting dimensions) Take a sample from the given data to obtain a 1D feature vector with dimension 1. .

[0120] (2) Dimensional adaptation and decomposition: In order to perform 3D tensor decomposition, the dimensionality adaptation and decomposition are... Expanded into a 3D tensor with dimensions of [dimensionality missing] using the unsqueeze operation. Subsequently, regarding and the expanded Perform non-negative CANDECOMP / PARAFAC decompositions respectively to obtain their respective factor vector groups.

[0121] (3) Perform fusion: Perform fusion on single sample image features and single-sample time series features Cross-modal feature integration is achieved through NNCP decomposition and factor fusion. The fusion formula is as follows:

[0122] ;

[0123] Cross-modal fusion features, specifically a 3D tensor, single-sample cross-modal fusion feature, integrates key information from images and temporal modalities for subsequent ecological quality assessment and prediction; The rank of the tensor decomposition is fixed at 32, representing the dimension of the factor matrix after decomposition, which controls the complexity of the fused features. : Sum the fusion factors corresponding to the 32 decomposition ranks to construct the complete fusion feature; : No. Tensors ( correspond correspond The first one obtained after NNCP decomposition The column factor matrix and column vectors correspond to the three dimensions of the tensor, respectively. : Element-wise multiplication operator, which performs element-wise multiplication on the factor vectors corresponding to two modes to achieve factor-level fusion between modes; The outer product operator performs an outer product operation on the three-dimensional factor vectors after the element-wise product to reconstruct cross-modal fusion features.

[0124] S4.3: Modal Bias Mitigation

[0125] The fusion module employs a step-by-step fusion strategy. First, it fuses satellite remote sensing image features with UAV aerial image features (auxiliary modalities) to obtain preliminary fused features. Then, it fuses these preliminary fused features with ground observation time-series features (core modalities) to increase the weight of the core modalities in the search space. Finally, it uses a joint loss function... Simultaneously constrain the multimodal fusion loss and the backbone loss of each modality to ensure effective learning of core modal features.

[0126] S4.4: Joint Loss Function Optimization

[0127] During training, a joint loss function is used for optimization to balance multimodal training and evaluation tasks. In one specific implementation, the joint loss function includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term, specifically expressed as follows:

[0128] ;

[0129] The joint loss function comprehensively constrains multimodal fusion prediction, single-modal feature learning, and dynamic sparse normalization, guiding the overall optimization of the model. The evaluation and prediction loss after multimodal fusion is based on MAE (mean absolute error), which is suitable for regression prediction tasks of ecological quality indicators and measures the prediction accuracy of fusion features. The single-mode loss of the ground-based time-series backbone network is determined by using RMSE (root mean square error) to constrain the feature extraction capability of the RNN backbone for time-series data such as air pressure, temperature, and humidity. The single-modal loss of the satellite image backbone network is calculated using RMSE to constrain the extraction accuracy of spatial features of satellite remote sensing images by the CNN backbone. The single-modal loss of the UAV image backbone network is RMSE, which constrains the extraction accuracy of the CNN backbone for detailed features of UAV aerial images. The scaling factor for single-modal loss is set to 0.2, which is used to balance the weight ratio of single-modal loss and multimodal fusion loss. The weight of the dynamic sparse normalization loss term is set to 0.1 to adjust the contribution strength of the dynamic sparse normalization loss and avoid excessive sparsity leading to feature loss. : Dynamic sparse normalization loss term, i.e., the sparsity constraint loss defined in step S4.2.1, is used to suppress multimodal feature redundancy and noise.

[0130] In step S4, this invention proposes a cross-modal feature integration scheme that combines nonnegative CANDECOMP / PARAFAC decomposition (NNCP) with factor fusion. Addressing the heterogeneity between image features (satellite / UAV images) and temporal features (ground observation data), it extracts key factors within each modality through tensor decomposition, and then achieves deep factor-level fusion through element-wise multiplication and outer product operations. This eliminates redundancy and noise between modalities and strengthens cross-modal correlation information. Furthermore, a joint loss function is designed to meet the needs of multiple tasks, balancing multimodal training and evaluation. By integrating multimodal fusion prediction loss, single-modal backbone network loss, and dynamic sparse normalization loss, it constrains the extraction accuracy of single-modal features such as images and temporal series while ensuring the predictive performance of fused features. Simultaneously, it alleviates modal bias and highlights the importance of the core modality (ground temporal data).

[0131] S5: Input the multi-source data of the area to be tested collected in real time into the trained multimodal prediction model, and output the evaluation and prediction results of the core ecological quality indicators. The core ecological quality indicators include at least one of the following: ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

[0132] In step S5, multi-source data of the region to be tested is collected in real time, and the collected multi-source data is input into the trained multimodal prediction model, based on cross-modal fusion features. The core indicators of ecological quality are inverted using the MLP prediction head. Each layer of the calculation incorporates an activation function and a regularization term, as shown in the following formula:

[0133] 1. First fully connected layer (feature mapping):

[0134] ;

[0135] The output feature vector of the first fully connected layer has a dimension of . This enables high-dimensional mapping and nonlinear transformation of fused features; ReLU: modifies the linear unit activation function, introduces nonlinearity, enhances the model's ability to fit complex features, and suppresses gradient vanishing. The weight matrix of the first fully connected layer has dimensions of . ( For fusion features (the dimension of which is used to map the fused features to a 128-dimensional feature space); Cross-modal fusion features, with dimensions of Integrating key information from images and temporal modalities; The bias vector of the first fully connected layer, with dimension . It is used to adjust the offset of the feature map and improve the flexibility of model fitting; : L2 regularization coefficient, set as Used to constrain the weight matrix The parameter size should be adjusted to avoid model overfitting. Weight matrix The L2 norm is used to calculate the sum of squares of the weight parameters, and excessive weights are penalized through regularization.

[0136] 2. Second-layer fully connected (metric output):

[0137] ;

[0138] : Vector of inversion results of core ecological quality indicators, with dimensions of For the number of indicators (e.g.) ); The weight matrix of the second fully connected layer has a dimension of . Used to convert 128-dimensional mapping features into Predicted values ​​for several ecological quality indicators; The output feature vector of the first fully connected layer has a dimension of . , as input features for the second layer; The bias vector of the second fully connected layer, with dimension . This is used to fine-tune the baseline offset of the predicted values ​​of each indicator; Quantity of ecological quality indicators, such as This corresponds to the integrity of the ecosystem, water conservation capacity, and the degree of impact from permafrost degradation. The inversion values ​​of individual ecological quality indicators correspond to ecosystem integrity (0-100 points), water conservation capacity (mm), and the degree of impact of permafrost degradation (0-10 levels).

[0139] Figure 5 This is a schematic diagram of a multi-source data fusion system for ecological quality that integrates air, space, and ground data, provided in an embodiment of the present invention. Figure 5 As shown, a multi-source data fusion system for ecological quality that integrates air, space, and ground is included:

[0140] Data acquisition module 210 is used to acquire multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data and ground observation time series data;

[0141] Data preprocessing module 220 is used to standardize and spatiotemporally align the multi-source data to obtain a multimodal dataset that includes image modalities and temporal modalities with temporal-spatial dual-dimensional alignment.

[0142] The model building module 230 is used to build a multimodal prediction model, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics.

[0143] The model training module 240 is used for model training and data assimilation fusion using the multimodal dataset. Specifically, it includes: using a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, a non-negative tensor decomposition method with dynamic sparsity constraints is used to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; during training, a joint loss function is used for optimization, and the joint loss function includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term;

[0144] The ecological quality assessment module 250 is used to input multi-source data of the area to be tested collected in real time into the trained multimodal prediction model and output the assessment and prediction results of the core ecological quality indicators. The core ecological quality indicators include at least one of ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

[0145] The air-space-ground collaborative ecological quality multi-source data fusion system provided in this embodiment of the invention can execute the air-space-ground collaborative ecological quality multi-source data fusion method provided in any of the above embodiments of the invention, and has the corresponding functions and beneficial effects of executing the method. For detailed process, please refer to the relevant operations of the foregoing method embodiments, which will not be repeated here.

[0146] In summary, the air-space-ground integrated ecological quality multi-source data fusion method and system according to embodiments of the present invention innovatively integrates three core technologies: differentiable architecture search, dynamic sparse constraint nonnegative tensor decomposition, and joint loss function optimization, constructing an adaptive, high-precision air-space-ground integrated fusion framework. This invention effectively solves the core contradiction in ecological monitoring of high-altitude and cold mountainous areas where coverage and accuracy are difficult to balance. It achieves accurate inversion and dynamic prediction of core indicators such as ecosystem integrity, water conservation capacity, and the degree of permafrost degradation, providing reliable technical support for comprehensive ecological quality assessment and protection decision-making.

[0147] To verify the effectiveness of the algorithm of this invention, comparative experiments were conducted, such as... Figure 6As shown, the performance comparison results of the error rate of the algorithm of this invention and three single-modal learning algorithms, CNN+RNN+Softmax, CNN+RNN+Sigmoid, and CNN+RNN+ReLU, in multimodal prediction tasks as the number of training iterations change are presented. From the overall trend, the error rate of all four algorithms shows a monotonically decreasing trend with the increase of the number of training iterations (10 to 100), reflecting the continuous improvement of the model's gradual fitting and generalization ability to the data features during training. Among them, the algorithm of this invention maintains the lowest error rate level throughout, with an error rate of 0.35 in the initial training stage (10 iterations), significantly lower than CNN+RNN+Softmax (0.42), CNN+RNN+Sigmoid (0.40), and CNN+RNN+ReLU (0.38) at the same time. Furthermore, the advantage further expands with the increase of training iterations, with the error rate dropping to 0.05 after 100 training iterations, compared to CNN+RNN+Softmax (0.16), CNN+... RNN+Sigmoid (0.13) and CNN+RNN+ReLU (0.09) reduced the errors by 68.75%, 61.54%, and 44.44%, respectively, fully verifying the effectiveness of non-negative CANDECOMP / PARAFAC decomposition in cross-modal feature integration. By deeply fusing image features and temporal features at the factor level, the information redundancy and noise interference caused by modal heterogeneity are effectively eliminated. The joint loss function balances the single-modal feature learning loss, multi-modal fusion prediction loss, and dynamic sparse normalization loss, which not only alleviates the modal bias problem that is common in traditional algorithms, but also strengthens the representation ability of core modal information, enabling the model to achieve a significant breakthrough in feature extraction and fusion accuracy. Furthermore, among the three traditional algorithms, CNN+RNN+ReLU has a lower overall error rate than CNN+RNN+Softmax and CNN+RNN+Sigmoid due to its better fitting ability for nonlinear features. However, due to the lack of a targeted cross-modal fusion mechanism and loss balancing strategy, its performance is always inferior to the algorithm of this invention, which further highlights the applicability of this invention in multi-source ecological data fusion tasks.

[0148] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.

[0149] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0150] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A method for fusing multi-source ecological quality data in a space-air-ground coordinated manner, characterized in that, Includes the following steps: Collect multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data, and ground observation time series data; The multi-source data is standardized and spatiotemporally aligned to obtain a multimodal dataset that includes image modalities and temporal modalities with temporal-spatial dual-dimensional alignment. A multimodal prediction model is constructed, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics. The model training and data assimilation fusion using the multimodal dataset specifically include: employing a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, using a non-negative tensor decomposition method with applied dynamic sparsity constraints to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; and optimizing the training process using a joint loss function, which includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term. In the fusion module, a non-negative tensor decomposition method with applied dynamic sparse constraints is used to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch. This includes: performing non-negative CANDECOMP / PARAFAC decomposition on the image features extracted by the image feature extraction branch and the temporal features extracted by the temporal feature extraction branch to obtain non-negative feature factors; performing dynamic sparse normalization on the feature factors to screen key features and suppress noise and redundancy; and fusing and reconstructing the screened feature factors to obtain cross-modal fused features. The dynamic sparsity constraint is achieved by introducing a dynamic sparse normalization loss term into the joint loss function. The dynamic sparse normalization loss term constrains feature sparsity by calculating the sum of the standardized L1 norms of each modality feature, so as to balance the contribution weights of different modalities in the fusion process and suppress redundant features. The multi-source data of the area to be tested collected in real time is input into the trained multimodal prediction model, and the evaluation and prediction results of the core ecological quality indicators are output. The core ecological quality indicators include at least one of the following: ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

2. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 1, characterized in that, The ecosystem integrity is quantified based on at least one of vegetation cover or NDVI, the ratio of bare / disturbed land surface, and landscape fragmentation, and output as a score of 0–100. The water conservation capacity is calculated using water balance on a monthly or annual scale, satisfying the condition that water conservation capacity (W) = precipitation (P). Evapotranspiration (ET) Runoff (R) + Soil moisture content (ΔS); The degree of permafrost degradation is calculated from at least one of the active layer thickness ALT, ground temperature change trend, and freeze-thaw period length / freeze-thaw days. The weighted sum of the active layer thickness ALT, ground temperature change trend, and freeze-thaw period length / freeze-thaw days after normalization is mapped to a level of 0-10.

3. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 1, characterized in that, The method employs a differentiable architecture search algorithm to perform architecture searches on the image feature extraction branch and the temporal feature extraction branch respectively, in order to optimize the network structure. Specifically, this includes: The preset discrete search spaces of the image feature extraction branch and the temporal feature extraction branch are each constructed into a continuously differentiable supernetwork; wherein the structure of the supernetwork is defined by nodes and directed edges, and each edge is associated with multiple candidate operations in the preset discrete search space; By assigning trainable architecture parameters to each candidate operation on each edge and normalizing the architecture parameters using the Softmax function, a continuous mixed operation representing the weights of each candidate operation is obtained. By constructing and solving a two-layer optimization problem with network weights as inner optimization variables and the architecture parameters as outer optimization variables, the network weights and the architecture parameters are jointly optimized. After the search is completed, the candidate operation with the largest weight is selected for each edge according to the optimized architecture parameters, thereby determining the final network architecture of the image feature extraction branch and the temporal feature extraction branch.

4. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 3, characterized in that, The image feature extraction branch employs a convolutional neural network containing normal convolutional units and dimensionality-reduced convolutional units; the temporal feature extraction branch employs a recurrent neural network. The preset discrete search space of the image feature extraction branch includes at least 3×3 and 5×5 dilated convolution, depthwise separable convolution, 3×3 max pooling, 3×3 average pooling, and various activation functions; the preset discrete search space of the temporal feature extraction branch includes at least various activation functions.

5. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 1, characterized in that, The fusion module adopts a step-by-step fusion strategy, which includes: first, fusing satellite remote sensing image features and UAV aerial image features to obtain preliminary fused features; then, fusing the preliminary fused features with ground observation time series features to obtain the final cross-modal fused features.

6. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 1, characterized in that, The prediction head is a multilayer perceptron containing at least two fully connected layers; wherein, the first fully connected layer maps the fused features to a high-dimensional feature space, and the second fully connected layer maps the high-dimensional features to an output dimension consistent with the number of core ecological quality indicators, and nonlinear activation functions are used for transformation between layers.

7. The method for multi-source ecological quality data fusion based on air-space-ground coordination according to claim 1, characterized in that, The multi-source data is standardized, including: Image data standardization employs Z-Score normalization; ground time-series monitoring data undergoes logarithmic normalization, including logarithmic transformation based on the ratio of observations at adjacent times to eliminate dimensions and enhance data stability.

8. A multi-source data fusion system for ecological quality that integrates air, space, and ground, characterized in that, The system is used to implement the air-space-ground integrated ecological quality multi-source data fusion method as described in any one of claims 1-7, the system comprising: The data acquisition module is used to acquire multi-source data, which includes at least satellite remote sensing image data, UAV aerial image data, and ground observation time series data. The data preprocessing module is used to standardize and spatiotemporally align the multi-source data to obtain a multimodal dataset that includes image modalities and temporal modalities with temporal-spatial dual-dimensional alignment. The model building module is used to build a multimodal prediction model, which includes an image feature extraction branch, a temporal feature extraction branch, a fusion module for cross-modal feature fusion, and a prediction head for outputting evaluation metrics. The model training module is used for model training and data assimilation fusion using the multimodal dataset. Specifically, it includes: using a differentiable architecture search algorithm to perform architecture search on the image feature extraction branch and the temporal feature extraction branch respectively to optimize the network structure; in the fusion module, a non-negative tensor decomposition method with dynamic sparsity constraints is used to perform cross-modal factor-level deep fusion of the features extracted by the image feature extraction branch and the temporal feature extraction branch; during training, a joint loss function is used for optimization, which includes at least a multimodal fusion prediction loss term, a loss term for the image feature extraction branch, a loss term for the temporal feature extraction branch, and a dynamic sparse normalization loss term; The ecological quality assessment module is used to input multi-source data of the area to be tested collected in real time into the trained multimodal prediction model and output the assessment and prediction results of the core ecological quality indicators. The core ecological quality indicators include at least one of ecosystem integrity, water conservation capacity and the degree of impact of permafrost degradation.

Images