Multi-index comprehensive evaluation method for restoration effect of degraded haloxylon ammodendron forest

Abstract

The present application relates to the technical field of ecological restoration effect evaluation, and specifically discloses a multi-index comprehensive evaluation method for the restoration effect of degraded haloxylon ammodendron forest, comprising: for the restoration of degraded haloxylon ammodron forest, fusing digital twin and improved fruit fly algorithm, constructing a root-water spatiotemporal mapping model and dynamically optimizing twin parameters, synchronously capturing the mutation boundary of root activity and water field, and outputting a collaborative perception field; taking the collaborative perception field as the state space base for multi-index comprehensive evaluation; by fusing digital twin and improved fruit fly algorithm, constructing a root-water spatiotemporal mapping model and introducing a self-adaptive sampling strategy driven by microtopographic heterogeneity, the mutation boundary of root activity and water field under different microtopographies can be synchronously captured, a high-resolution collaborative perception field is outputted, and the problem of evaluation distortion caused by traditional single-point sampling is solved, thus providing an accurate environment-individual collaborative base for multi-index comprehensive evaluation.

Description

Technical Field

[0001] This invention relates to the field of ecological restoration effect evaluation technology, and in particular to a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests. Background Technology

[0002] With the intensification of global climate change and human activities, many Haloxylon ammodendron forests are facing degradation and destruction, leading to a decline in ecological functions and a reduction in biodiversity. Therefore, the evaluation of their restoration effects has become particularly important. By integrating multiple indicators from ecological, social and economic perspectives, a comprehensive evaluation framework can be provided for the restoration effects of Haloxylon ammodendron forests. This framework can not only reflect the health status of the ecosystem and the effectiveness of restoration, but also provide a scientific basis for relevant policy formulation and promote the rational allocation and management of resources.

[0003] For example, the evaluation method for forest and grassland ecological restoration effect in Chinese patent publication number CN115293473A provides standardized guidance for the evaluation and acceptance of forest and grassland ecological restoration projects, improves the standardization and scientific nature of forest and grassland ecological restoration effect evaluation, and is conducive to accelerating the evaluation of ecological restoration effect after the implementation of national forest and grassland governance projects, and scientifically promoting the high-quality development of land greening.

[0004] In existing technologies, it is difficult to simultaneously capture the root vitality and soil moisture heterogeneity of Haloxylon ammodendron under different micro-topographical conditions in desert areas, which easily leads to distortion of single-point evaluation. Moreover, without obtaining the true water-root vitality synergy field, it is difficult to distinguish the short-term stimulation effect and long-term survival contribution of remediation measures, which may lead to misjudgment of effectiveness to some extent. At the same time, without the root-water synergy field and continuous gain value, it is difficult to quantify the nonlinear compensation threshold between the lag in lateral branch sprouting and the competition for carbon sinks in the main stem. The remediation evaluation cannot be closed-loop to the physiological integration of the whole plant. Therefore, how to form a synergistic evaluation path that runs through the environment-individual-physiological integration chain from capturing micro-topographical heterogeneity, to the continuous separation of effectiveness, and then to the calculation of the compensation threshold of the whole plant is the problem that this invention aims to solve. Summary of the Invention

[0005] To overcome the shortcomings of the prior art, the present invention provides a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests, which can effectively solve the problems involved in the prior art.

[0006] The objective of this invention can be achieved through the following technical solution: This invention provides a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests, comprising the following steps: Step 1: For the restoration of degraded Haloxylon ammodendron forests, we integrate digital twins and improved fruit fly algorithms to construct a root-water spatiotemporal mapping model and dynamically optimize the twin parameters. We introduce an adaptive sampling strategy driven by micro-topographic heterogeneity to simultaneously capture the abrupt boundary of root vitality and water field, and output a high-resolution collaborative sensing field to achieve high-precision collaborative sensing of water and root vitality at the micro-topographic scale. Step 2: Using the collaborative sensing field as the state space basis for multi-index comprehensive evaluation, micro-topographic topology coding and root vitality spatiotemporal ridge line extraction are introduced to construct a deep alignment layer of environmental-individual heterogeneous indicators, realize the structured expression of the spatial heterogeneity characteristics of the water-root collaborative field in the restoration effect of degraded Haloxylon ammodendron forest, and complete the cross-domain feature alignment and compact expression of environmental and individual indicators. Step 3: Using the collaborative sensing field as the state space of the deep Q-network, the causal path pruning and counterfactual reasoning of the knowledge graph are combined to remove short-term moisture pulse interference and random fluctuation noise, and to strip away the persistent survival gain value caused by the remediation measures, effectively separating the long-term gain and short-term disturbance of the remediation measures. Step 4: Using the sustained survival gain as the core time-efficiency indicator, the time series of the collaborative perception field is dynamically weighted by integrating the long short-term memory network to quantify the contribution of restoration measures to the long-term survival of degraded Haloxylon ammodendron forests. Furthermore, a causal inference counterfactual control is introduced to distinguish between the sustained restoration effect driven by short-term water pulse stimulation and structural root recovery, thus accurately differentiating between short-term stimulation and structural sustained restoration effects. Step 5: Using the collaborative sensing field as the environmental boundary and the survival gain value as the reward signal, construct a hybrid of two population fruit flies and a deep Q network. The two populations respectively perform global carbon sink competition threshold search and local lateral compensation time-series decoupling. Through adaptive odor concentration weight transfer and experience pool adversarial sampling mechanism, calculate the nonlinear carbon sink competition threshold required to start lateral compensation, and calculate the nonlinear threshold of carbon sink competition to start lateral compensation. Step 6: Based on the carbon sink competition threshold, drive the integrated physiological evaluation of the whole plant, and introduce a nonlinear compensation time-varying model between the lag in lateral branch germination and the carbon sink in the main stem. Dynamically invert the compensation start time and intensity, and correct the boundary conditions of the collaborative sensing field in a closed loop to complete the collaborative diagnosis of the environment-individual-physiological chain, forming a closed-loop collaborative diagnosis of the entire environment-individual-physiological chain.

[0007] Preferably, step 1 specifically includes: Soil moisture time-series data and root electrical conductivity profiles of typical micro-topographic units in the degraded Haloxylon ammodendron forest restoration area were collected. An initial digital twin containing the root topology of Haloxylon ammodendron was constructed, and a coarse-grained chaotic mapping relationship between water field and root vitality was established. A high-fidelity digital mirror was constructed, laying the data foundation for collaborative perception. An improved fruit fly algorithm was used to dynamically optimize twin parameters. The odor concentration function was defined as the weighted sum of the water field prediction error and the spatiotemporal smoothness of root vitality, and the micro-topography curvature tensor was introduced to adaptively adjust the odor search step size of individual fruit flies, thereby improving the parameter optimization efficiency and enhancing the model's adaptability to complex terrain. The optimized twin parameters are reinjected into the water conduction layer and root water absorption layer of the digital twin at different scales. The unsteady coupling mechanism of water and root system during the restoration of degraded Haloxylon ammodendron forest is iteratively inverted to generate a dynamic parameterized expression of the water-root vitality synergy field that can characterize the restoration process, so as to realize the dynamic tracking and parameterized characterization of the water-root coupling process.

[0008] Preferably, step 1 further includes: We constructed a micro-topographic heterogeneity weight tensor based on the slope, aspect, runoff accumulation, and topographic humidity index of the degraded Haloxylon ammodendron forest restoration area. We calculated the heterogeneity information gain of each candidate sampling unit and generated an adaptive sampling prior probability density distribution. We accurately located the heterogeneous areas through the weight tensor, avoiding the waste of sampling resources in homogeneous areas. Based on the prior probability density, progressively encrypted sampling is automatically triggered in regions where the heterogeneity weight exceeds the chaos threshold. Dynamic tracking and critical point marking of mutation boundaries are implemented when the root vitality gradient change rate is greater than the set Lyapunov exponent threshold, thereby achieving precise spatial location locking of the water-root mutation boundary and improving the ability to capture key interfaces. The high-frequency water-root data from encrypted sampling is fed back in real time to the improved fruit fly optimization process, dynamically correcting the nonlinear response function of root water absorption in the twin parameters, and outputting a collaborative sensing field with spatial resolution adapted to the degree of micro-topographic heterogeneity. This allows the resolution of the collaborative sensing field to be dynamically matched with the degree of local heterogeneity, enhancing adaptability to complex terrain.

[0009] Preferably, step 2 specifically includes: Dynamic topological reconnection is performed on each micro-topographic unit in the collaborative sensing field. Delaunay triangulation is driven by a water-root vitality joint gradient field to encode the adaptive graph mapping between topographic position and root vitality ridge node, ensuring that the abrupt interface is accurately captured by the triangle boundary and avoiding cross-unit information aliasing. With the constraint of maximizing cross-scale mutual information in the ridge neighborhood, a deep residual graph convolutional network is introduced to project the multimodal heterogeneous features of water nodes and root vitality nodes in the degraded Haloxylon ammodendron forest environment to the shared latent space of the restoration state, thereby achieving cross-domain alignment of micro-topography and individual physiological features, and improving the consistency of heterogeneous data fusion. The output is a structured matrix for the evaluation of degraded Haloxylon ammodendron forest restoration. This matrix uniformly encodes micro-topographic topological entropy, water field heterogeneity, spatiotemporal ridge curvature of root activity, and restoration response delay spectrum. It completes the hierarchical heterogeneity feature compact expression of the water-root synergistic field, forming a compact structured feature expression to support the input of subsequent multi-index comprehensive evaluation.

[0010] Preferably, step 3 specifically includes: We constructed a time-sensitive causal knowledge graph for restoration measures of degraded Haloxylon ammodendron forests. The nodes cover intermediate variables such as irrigation pulse, water-retaining agent, root pruning and Haloxylon ammodendron-specific physiological response. The edges are labeled with the causal path of the measure-effect and the restoration time window to realize the explicit causal relationship between restoration measures and physiological response and improve interpretability. The current water-root activity joint distribution of the collaborative sensing field is encoded into a hierarchical state embedding vector of a deep Q network. At the same time, the dynamic causal intensity coefficient of each restoration path is calculated using a causal graph attention mechanism. Short-term pulse paths with restoration time windows smaller than the physiological response cycle of roots in degraded Haloxylon ammodendron forests are pruned, transient invalid interferences are filtered out, key causal chains are retained, and the robustness of state representation is enhanced. By constraining the state-action space topology of a deep Q-network based on the causal graph of the restoration of degraded Haloxylon ammodendron forests after pruning, transitional states that do not satisfy the causal chain of continuous restoration are eliminated, while fully causal driving trajectories that can be traced back to the original restoration measures are retained. This approach compresses the invalid state space, focuses on the continuous restoration trajectory, and improves learning efficiency.

[0011] Preferably, step 3 further includes: In the digital twin of the degraded Haloxylon ammodendron forest restoration area, a multi-hypothesis counterfactual control scenario was constructed. The target restoration measures were virtually removed while keeping the boundary conditions of micro-topographic heterogeneity unchanged. The counterfactual evolution trajectory of water-root activity under the untreated condition was generated. A pure control was constructed by virtual removal to achieve quantitative separation of the contribution of restoration measures. The structured residual fields of the actual restoration observation trajectory and counterfactual trajectory of degraded Haloxylon ammodendron forest at each state node of the deep Q network are calculated. Time-frequency domain differential causal estimation is used to remove sub-physiological periodic scale interference introduced by occasional precipitation pulses and random fluctuations of root conductivity sensors, and to remove environmental noise and sensor disturbances, thereby improving the purity of the residual field response to the real restoration signal. The positive offset component whose duration exceeds the minimum response window of the root structure restoration of degraded Haloxylon ammodendron forest is extracted from the structured residual field and used as a persistent survival gain value driven independently by restoration measures. This value is then output and injected into the restoration evaluation structured matrix. The persistent positive offset is locked, and the short-term impulse contribution is removed to form an independently driven survival gain index.

[0012] Preferably, step 4 specifically includes: The collaborative sensing field of the degraded Haloxylon ammodendron forest restoration area, along with the continuous survival gain value, is input into the long short-term memory network. The root physiological reversibility time window is set as the forgetting gate time constant. The interference weight of short-term water pulses on the Haloxylon ammodendron root vitality memory unit is suppressed, effectively filtering out short-term water fluctuation interference and improving the identification accuracy of the long-term evolution trend of root vitality. Using the difference-in-differences method in causal inference, the degraded Haloxylon ammodendron forest in the restoration area was paired with the Haloxylon ammodendron in the control area with the same microtopography but without restoration. The average treatment effect of restoration measures on the recovery of Haloxylon ammodendron root structure and its attenuation slope along the growing season were calculated. The interference of microtopographic heterogeneity was eliminated, and the net treatment effect and timeliness of restoration measures were accurately quantified. The effect type is determined based on the attenuation slope combined with the biological threshold of lateral root germination in Haloxylon ammodendron: those with a negative slope and an absolute value greater than the minimum duration of root recovery are short-term water pulse stimulation, while those with a slope close to zero or positive are continuous repair effects driven by the structural recovery of the deep root system of Haloxylon ammodendron. This distinguishes between stimulation and recovery effects and avoids misjudging the long-term repair evaluation based on short-term results.

[0013] Preferably, step 5 specifically includes: We constructed a global population of fruit fly individuals, with each individual encoding a candidate solution for the carbon sink competition threshold of the main stem of a degraded Haloxylon ammodendron forest. We used the long-term survival contribution of root vitality output by the deep Q-network as the global odor concentration evaluation function to conduct a targeted search for the carbon allocation critical point of Haloxylon ammodendron main stem and lateral branches, thereby improving the physiological relevance of the carbon allocation critical point search and avoiding blind optimization. We constructed local populations of fruit flies, and each individual encoded candidate solutions for the lag time and compensation intensity of lateral branch germination in degraded Haloxylon ammodendron forests. We used the marginal compensation rate of lateral branch biomass accumulation to carbon sink loss in the main stem as the local odor concentration to quantify the initiation conditions of Haloxylon ammodendron lateral branch compensation after restoration, directly quantified the economics of lateral branch compensation, and linked the initiation conditions to carbon gains. The current water-root distribution state of the collaborative sensing field is defined as the environmental boundary of the hybrid of two population fruit flies and deep Q network. The persistent survival gain value is mapped as the instantaneous reward signal of the deep Q network for the carbon allocation strategy of the whole plant of Haloxylon ammodendron, so that the reward signal is directly related to long-term survival and guides the strategy to prioritize root recovery.

[0014] Preferably, step 5 further includes: When the global population search finds a candidate interval for carbon sink competition of Haloxylon ammodendron main stem with odor concentration higher than the migration threshold, the carbon allocation ratio of the interval is used as a priori weight to migrate to the local population, thereby narrowing the initial range of the time series search for lag compensation of lateral branches in degraded Haloxylon ammodendron forests, avoiding blind search, and improving the convergence speed and stability of the local population. In the experience pool of the deep Q network, both high-survival trajectories and low-survival trajectories of Haloxylon ammodendron are stored simultaneously. An adversarial sampling mechanism is adopted to prioritize sampling the experience with the greatest difference from the current carbon sink competition state of Haloxylon ammodendron. This avoids the lateral compensation strategy from falling into the shallow root adaptation mode of local optima, forces the model to pay attention to extreme carbon competition states, and avoids inefficient compensation strategies. Iteratively perform a collaborative search of global and local populations until the odor concentration converges, calculate the critical allocation ratio of the main stem carbon sink required to initiate lateral branch compensation in degraded Haloxylon ammodendron forests and the corresponding nonlinear carbon sink competition threshold, accurately quantify the carbon allocation critical point, and support dynamic decision-making for lateral branch compensation.

[0015] Preferably, step 6 specifically includes: Based on the calculated nonlinear carbon sink competition threshold and lateral branch germination lag time, this study constructs a step response function for lateral branch compensation initiation to address the typical imbalance state of excessive encroachment of carbon allocation on lateral branches by the main stem carbon sink in the degraded Haloxylon ammodendron forest restoration area. It also establishes a dynamic switching mechanism of inhibition-release for the main stem-lateral branch carbon sink competition to avoid frequent oscillations of carbon allocation strategy near the threshold and improve the stability of carbon utilization of the whole plant. The inhibition-release dynamic switching mechanism is embedded into the whole-plant physiological integration model of degraded Haloxylon ammodendron forest. The measured lateral branch germination time and germination intensity in the restoration area are used as inversion constraints to dynamically invert the water-carbon flow redistribution trajectory after compensation is initiated. By iteratively approximating the actual response delay and compensation efficiency of lateral branch compensation to carbon sink competition in the main stem, the timing and intensity of lateral branch compensation are accurately quantified, and carbon flow allocation error is reduced. The compensation initiation parameters obtained from the inversion are fed back to the collaborative sensing field. With the goal of dynamic boundary calibration of the collaborative sensing field, the root water absorption parameters and water conduction boundary conditions are updated to complete the closed-loop collaborative diagnosis of the environment-individual-physiology chain for the restoration evaluation of degraded Haloxylon ammodendron forests. This forms a full-chain closed-loop self-calibration, improving the physiological consistency and field applicability of the restoration evaluation.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. This method for comprehensive evaluation of the restoration effect of degraded Haloxylon ammodendron forests by integrating digital twins and improved fruit fly algorithms, constructing a root-water spatiotemporal mapping model and introducing an adaptive sampling strategy driven by micro-topographic heterogeneity, can simultaneously capture the abrupt boundary of root vitality and water field under different micro-topographic conditions, and output a high-resolution collaborative sensing field. This solves the problem of evaluation distortion caused by traditional single-point sampling and provides an accurate environment-individual collaborative basis for comprehensive evaluation of multiple indicators.

[0017] 2. This multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests uses a collaborative perception field as the state space of a deep Q-network, and combines causal path pruning and counterfactual reasoning of knowledge graphs to eliminate short-term water pulse interference and random noise. It can independently extract the continuous survival gain value caused by restoration measures from the time series, clearly distinguish the continuous restoration effect driven by short-term water pulse stimulation and structural root recovery, and avoid misjudgment of restoration effectiveness.

[0018] 3. This multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests uses a collaborative sensing field as the environmental boundary and survival gain value as the reward signal to construct a hybrid of two populations of fruit flies and a deep Q network. By decoupling the global carbon sink competition threshold search with the local lateral branch compensation time sequence, and combining adaptive odor concentration weight transfer and experience pool adversarial sampling mechanism, the nonlinear carbon sink competition threshold required to initiate lateral branch compensation is calculated, so that the critical conditions for whole-plant carbon allocation that are originally difficult to quantify can be accurately expressed.

[0019] 4. This multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests is based on the calculated carbon sink competition threshold to drive the whole-plant physiological integration evaluation. It introduces a nonlinear compensation time-varying model between the lag in lateral branch germination and the carbon sink in the main stem, dynamically inverts the compensation initiation time and intensity, and feeds back the inversion results to correct the boundary conditions of the collaborative sensing field. This realizes the collaborative diagnosis of the environment-individual-physiological chain, so that the restoration evaluation is no longer limited to a single index, but runs through the whole-plant physiological integration process. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the workflow of a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to the present invention; Figure 2 This is a schematic diagram of the process for a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0022] Example 1, please refer to Figure 1 , Figure 2 This invention provides a technical solution: a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests, comprising the following steps: Step 1: For the restoration of degraded Haloxylon ammodendron forests, a root-water spatiotemporal mapping model is constructed by integrating digital twins and an improved fruit fly algorithm. The twin parameters are dynamically optimized, and an adaptive sampling strategy driven by micro-topographic heterogeneity is introduced to simultaneously capture the abrupt boundary of root activity and water field, outputting a high-resolution collaborative sensing field. This achieves high-precision collaborative sensing of water and root activity at the micro-topographic scale. Soil moisture time-series data and root electrical conductivity profiles of typical micro-topographic units within the degraded Haloxylon ammodendron forest restoration area are collected. An initial digital twin containing the topological structure of Haloxylon ammodendron roots is constructed, and a coarse-grained chaotic mapping relationship between the water field and root activity is established. A high-fidelity digital mirror is constructed, laying the foundation for collaborative sensing data. Based on this, an improved fruit fly algorithm is used to dynamically optimize the twin parameters. The odor concentration function is defined as the weighted sum of the prediction error of the water field and the spatiotemporal smoothness of root vitality, and the information entropy is introduced. The micro-topography curvature tensor is introduced to adaptively adjust the odor search step size of individual fruit flies, thereby improving the parameter optimization efficiency and enhancing the model's adaptability to complex terrain. The optimized twin parameters are then reinjected into the water conduction layer and root water absorption layer of the digital twin at different scales. The non-steady-state coupling mechanism of water and root system during the restoration of degraded Haloxylon ammodendron forest is iteratively inverted to generate a dynamic parameterized expression of the water-root vitality synergistic field that can characterize the restoration process, thereby realizing the dynamic tracking and parameterized characterization of the water-root coupling process. It should be noted that within the degraded Haloxylon ammodendron forest restoration area, three typical micro-topographic units—hilltop, inter-hill, and transitional slope—were selected based on the micro-topographic classification results. Three monitoring quadrats were set up for each unit, with an EC-5 soil moisture sensor buried in the center of each quadrat at monitoring depths of 10 cm, 30 cm, 60 cm, and 100 cm, respectively. The sampling interval was set to 30 minutes, continuously recording soil moisture time-series data for a complete growing season (April to October). Simultaneously, a DC-TT root conductivity probe was used to acquire root conductivity profile data at the same depth, with a measurement frequency of once a week. Based on the collected data, an initial digital twin containing the topology of Haloxylon ammodendron taproot, primary lateral roots, and secondary lateral roots was constructed in the ANSYS TwinBuilder platform. The root topology parameters were set based on measured root anatomy data from the restoration area: taproot diameter 8-15 mm, lateral root branching angle 45°. o -70 oBased on this, a coarse-grained chaotic mapping relationship between water field and root activity is established. The phase space reconstruction method is used to map the soil moisture temporal sequence and electrical conductivity sequence to a three-dimensional chaotic attractor, with an embedding dimension of 6 and a time delay of 3. An improved fruit fly algorithm is used to dynamically optimize the water conduction coefficient and root water absorption rate parameters in the digital twin. The population size is set to 50, and the maximum number of iterations is 200. The odor concentration function is defined as the weighted sum of the water field prediction error and the spatiotemporal smoothness of root activity, with the information entropy being calculated using the root mean square error. The spatiotemporal smoothness of root activity is represented by the standard deviation of adjacent sampling points within a 72-hour time window, with weighting coefficients of 0.6 and 0.4, respectively. A micro-topographic curvature tensor is introduced to adaptively adjust the odor search step size of individual fruit flies: based on the slope of each sampling unit (0... o -15 o ) and plane curvature (-0.05 to 0.05 m) -1 The principal components of the curvature tensor are mapped to a search step size range of 0.02-0.15. The larger the absolute value of curvature, the smaller the step size. After each iteration, the global optimal odor concentration value is recorded. If the global optimal value is not updated after 15 consecutive iterations, a step size restart mechanism is activated, resetting the search step size to 1.5 times the initial value to escape local extrema. The water conduction layer is inoculated with water diffusivity and hydraulic conductivity according to a horizontal grid of 5m×5m and a vertical layer of 4 layers (0-20cm, 20-50cm, 50-80cm, 80-120cm). The root water absorption layer is inoculated with root water absorption rate parameters for each root according to the root topology nodes, with the water absorption rate of the main root being 0.08-0.12 cm. 3 ·g -1 ·d -1 Lateral roots should be 0.03-0.06 cm long. 3 ·g -1 ·d -1 Using a finite difference scheme with a time step of 1 hour and a spatial step of 0.5 m, the non-steady-state coupling mechanism of water and root system during the restoration of degraded Haloxylon ammodendron forests was iteratively inverted. The convergence condition of the iteration was set as the maximum change rate of water field between two adjacent iterations being less than 0.5% and the change rate of root vitality field being less than 0.8%. Finally, a dynamic parameterized expression of the water-root vitality synergistic field that can characterize the restoration process was generated. The output format is a structured data file containing timestamps, spatial coordinates, water content and root vitality index, with a time resolution of daily and a spatial resolution of 5m×5m grid. Furthermore, step 1 also includes: constructing a micro-topographic heterogeneity weight tensor based on the slope, aspect, runoff accumulation, and topographic humidity index of the degraded Haloxylon ammodendron forest restoration area; calculating the heterogeneity information gain of each candidate sampling unit; generating an adaptive sampling prior probability density distribution; accurately locating heterogeneous areas through the weight tensor to avoid wasting sampling resources in homogeneous areas; automatically triggering progressively denser sampling in areas where the heterogeneity weight exceeds the chaos threshold based on the prior probability density; and implementing dynamic tracking and critical point marking of mutation boundaries when the root vitality gradient change rate is greater than the set Lyapunov exponent threshold to accurately lock the spatial location of the water-root mutation boundary, improve the ability to capture key interfaces, and feed back the high-frequency water-root data from the dense sampling to the improved fruit fly optimization process in real time to dynamically correct the nonlinear response function of root water absorption in the twin parameters, and outputting a collaborative sensing field with spatial resolution adapted to the degree of micro-topographic heterogeneity, so that the resolution of the collaborative sensing field is dynamically matched with the degree of local heterogeneity, thereby enhancing the adaptability to complex terrain. It should be noted that after completing the initial parameterization expression of the water-root synergistic field, sampling units were divided into 5m×5m grids within the degraded Haloxylon ammodendron forest restoration area, and the slope (0) of each unit was extracted. o -15 o ), slope aspect (according to 45) oThe sampling intervals are divided into 8 directional categories, runoff accumulation (calculated based on DEM flow direction analysis, unit: raster number), and topographic humidity index (calculated using the formula ln(α / tanβ), where α is runoff accumulation and β is slope radian value). After normalizing these four categories, a three-dimensional weight tensor is constructed. The weight coefficients for each dimension are set according to the principal component analysis contribution rate: slope 0.25, aspect 0.10, runoff accumulation 0.35, and topographic humidity index 0.30. The heterogeneity information gain of each candidate sampling unit is calculated based on the weight tensor, with an information gain threshold of 0.35. Units exceeding this value are considered significantly heterogeneous. An adaptive sampling prior probability density distribution is generated based on the information gain distribution. The probability density is linearly proportional to the information gain and, after normalization, serves as the spatial allocation basis for the next round of encrypted sampling, ensuring higher sampling density in highly heterogeneous areas. Based on the prior probability density, progressively encrypted sampling is automatically triggered in areas where the heterogeneity weight exceeds the chaos threshold. The chaos threshold is defined as the 75th quantile of the information gain. The first encryption will... The original 5m grid was refined to 2.5m. If the local information gain was still higher than 0.40 after refinement, it was refined again to 1.25m. The root activity gradient change rate was calculated simultaneously. The ratio of the difference in root activity indices between adjacent grids to the spatial distance was obtained using the spatial difference method. When the gradient change rate exceeded the set Lyapunov index threshold of 0.25 (based on statistical calibration of root activity fluctuations in the remediation area), dynamic tracking of the abrupt boundary was implemented. Critical points were marked point by point along the direction of maximum gradient, with a spacing of no more than 1m between critical points. The critical point marking map was updated after each tracking to avoid duplicate sampling. Tracking continued until the gradient change rate dropped below the threshold or no new critical points appeared within three consecutive tracking steps. The water-root high-frequency data (30-minute temporal resolution, up to 1.25m spatial resolution) obtained from the refined sampling was fed back to the improved Drosophila optimization process in real time. Parameter correction was triggered every 100 sets of high-frequency data collected, dynamically correcting the nonlinear response function of root water absorption in the twin parameters. The correction target was the baseline value of the main root water absorption rate (0.08-0.12 cm). 3 ·g -1 ·d -1 ) and the baseline value for lateral root water absorption rate (0.03-0.06cm) 3 ·g -1 ·d -1 The nonlinear coefficients of the response function are updated using a sliding window (window width 48 hours, step size 6 hours), where the nonlinear coefficients are limited to between 0.7 and 1.3. The final output of the collaborative sensing field spatial resolution is adapted to the degree of micro-topographic heterogeneity: regions with heterogeneity weight less than 0.20 maintain a 5m grid, regions with heterogeneity weight between 0.20 and 0.35 are refined to a 2.5m grid, and regions with heterogeneity weight greater than 0.35 are refined to a 1.25m grid. Step 2: Using the collaborative sensing field as the state space basis for multi-index comprehensive evaluation, micro-topographic topological encoding and spatiotemporal ridge line extraction of root vitality are introduced to construct a deep alignment layer for environmental-individual heterogeneous indicators. This achieves a structured expression of the spatial heterogeneity characteristics of the water-root collaborative field in the restoration effect of degraded Haloxylon ammodendron forests, completing the cross-domain feature alignment and compact expression of environmental and individual indicators. Dynamic topological reconnection is performed on each micro-topographic unit in the collaborative sensing field. Delaunay triangulation is driven by the water-root vitality joint gradient field to encode the adaptive graph mapping between topographic order and root vitality ridge line nodes, ensuring that abrupt change interfaces are accurately captured by the triangle boundaries and avoiding cross-unit information mixing. By maximizing cross-scale mutual information in the ridge neighborhood, a deep residual graph convolutional network is introduced to project the multimodal heterogeneous features of water nodes and root vitality nodes in degraded Haloxylon ammodendron forests onto the shared latent space of the restoration state. This achieves cross-domain feature alignment between micro-topography and individual physiology, as well as cross-domain alignment between environmental and individual features, improving the consistency of heterogeneous data fusion. The resulting structured matrix is ​​used to evaluate the restoration of degraded Haloxylon ammodendron forests. This matrix uniformly encodes micro-topographic topological entropy, water field heterogeneity, spatiotemporal ridge curvature of root vitality, and restoration response delay spectrum, completing a compact expression of the hierarchical heterogeneity features of the water-root synergistic field and forming a compact structured feature expression to support subsequent multi-index comprehensive evaluation input. It should be noted that after constructing the adaptive resolution collaborative sensing field, dynamic topological reconnection is implemented for each micro-topography unit. A two-dimensional joint gradient vector is constructed using the soil moisture content and root activity index at each 5m, 2.5m, and 1.25m grid node. The gradient magnitude and orientation angle between adjacent nodes are calculated, with a joint gradient magnitude greater than 0.15 (moisture content unit: m). 3 ·m -3 Root vitality unit mV·s -1 The node pairs are used as forced edges for Delaunay triangulation, ensuring that the water-root abrupt change interface is accurately captured by the triangular mesh boundary. The terrain positional encoding is updated in real time during the triangulation process: the centroid slope (0) of each triangular unit is... o -15 o ) is divided into three levels (0 o -5 o 5 o -10 o 10 o -15 oThe micro-topographic category (hilltop, inter-hill, transitional slope) of each unit is recorded. Root vitality ridge node extraction is based on local maximum search, with the search window radius adaptively adjusted according to spatial resolution: 2 nodes for a 5m grid window, 3 nodes for a 2.5m grid window, and 4 nodes for a 1.25m grid window. Ridge nodes are forced to be the vertices of triangulation, and triangulation edges are prohibited from intersecting within a 1m radius around them to ensure spatial connectivity. After triangulation, each micro-topographic unit corresponds to a set of graph nodes. Node features include mean water content, mean root vitality, joint gradient principal direction, and topographic positional encoding. To achieve multimodal feature alignment between environmental water content nodes and root vitality nodes, a deep residual graph convolutional network is constructed. The network consists of three graph convolutional layers, each followed by a residual connection. The hidden layer dimensions are set to 64, 128, and 64, respectively. The input layer receives two types of node features: a 5-dimensional feature vector for water nodes (containing water content, water variation coefficient, runoff accumulation, topographic humidity index, and slope), and a 4-dimensional feature vector for root vitality nodes (containing conductivity amplitude, daily conductivity fluctuation, root water absorption rate, and root diameter). The network is constrained by maximizing cross-scale mutual information in the ridge neighborhood. The mutual information calculation window extends by three nodes on each side along the ridge extension direction. Cross-scale pairings include three combinations: 5m-2.5m, 5m-1.25m, and 2.5m-1.25m. The threshold for the mutual information objective function is set to 0.45. When the mutual information of a node pair is lower than 0.45, the pair is considered to have a lower mutual information value. At the threshold, the weight decay coefficient of the node pair in the adjacency matrix of the graph convolutional network is set to 0.3. The network training uses the Adam optimizer with a learning rate of 0.001, a batch size of 32, and 50 early stopping rounds. After training, the features of the water node and the root vitality node are projected into a 128-dimensional shared latent space. Nodes with an Euclidean distance less than 0.2 in the latent space are considered successfully aligned, and alignment confidence labels (continuous values ​​from 0 to 1, with values ​​above 0.6 considered reliable) are output. The alignment results are used to correct the local parameters of the collaborative sensing field in step 1. In particular, when the alignment confidence of the water node and the root vitality node is lower than 0.5, the twin parameters of that local region are recalibrated. A structured matrix for restoration evaluation is output, with each row corresponding to a triangulation. Each graph node (i.e., spatial sampling unit) after partitioning contains four types of encoded features. The first type is micro-topographic topological entropy: based on the adjacency relationship of the triangular unit where the node is located, the micro-topographic category distribution entropy of the node in the third-order neighborhood is calculated. The entropy value ranges from 0 to 1.5. The higher the entropy, the more complex the micro-topographic transition zone. The second type is water field heterogeneity: the ratio of the standard deviation of water content of the node and its first-order neighboring nodes to the mean is calculated. Combined with the local variation coefficient of the runoff accumulation (window radius of 2 grids), the two are weighted to synthesize the heterogeneity index. The weight coefficients are 0.6 and 0.4, respectively, and the output range is 0 to 0.8. The third type is the spatiotemporal ridge curvature of root vitality: the rate of change of the tangential direction is calculated every 3 nodes along the ridge direction. The curvature value is greater than 0.3 rad·m.-1 The segment was marked as a strongly curved ridge, and the average root vigor decay rate of that segment was recorded (in days). -1 The fourth category is the repair response delay spectrum: taking the moment when the first repair measure is applied as the zero point, the number of days (in days) required for the root activity at each node to reach its peak and the number of days required for the root activity to decay to 80% of the peak value are extracted. The two constitute a two-dimensional vector of the delay spectrum. This structured matrix is ​​output in CSV format, while retaining the spatial coordinates and grid resolution labels of each node. Step 3: Using the collaborative sensing field as the state space of a deep Q-network, causal path pruning and counterfactual reasoning from the knowledge graph are combined to eliminate short-term water pulse interference and random fluctuation noise, and to remove the persistent survival gain value caused by restoration measures. This effectively separates the long-term gain and short-term disturbance of restoration measures, constructing a time-sensitive causal knowledge graph for the restoration of degraded Haloxylon ammodendron forests. Nodes cover intermediate variables such as irrigation pulses, water-retaining agents, root pruning, and Haloxylon ammodendron-specific physiological responses. Edges are labeled with the measure-effect causal path and restoration time window, realizing explicit causal association between restoration measures and physiological responses, improving interpretability, and integrating the current water-root... The joint distribution encoding of system vitality is used as a hierarchical state embedding vector of a deep Q-network. At the same time, a causal graph attention mechanism is used to calculate the dynamic causal intensity coefficient of each restoration path. Short-term pulse paths with restoration time windows smaller than the physiological response cycle of the degraded Haloxylon ammodendron root system are pruned to filter out transient invalid interference, retain key causal chains, and enhance the robustness of state representation. The state-action space topology of the deep Q-network is constrained by the restoration causal graph of the pruned degraded Haloxylon ammodendron. Transitional states that do not meet the continuous restoration causal chain are eliminated, and the fully causal driving trajectory that can be traced back to the original restoration measures is retained. The invalid state space is compressed, the continuous restoration trajectory is focused, and the learning efficiency is improved. It should be noted that within the degraded Haloxylon ammodendron forest restoration area, a time-sensitive causal knowledge graph for restoration measures was first constructed. The graph nodes cover four types of entities: irrigation pulses (30-50 L / plant per irrigation), water-retaining agents (50-80 g / hole, buried 20-30 cm), root pruning (removal of no more than 15% of necrotic lateral roots), and intermediate variables of Haloxylon ammodendron-specific physiological responses (including daily increase in root vitality, rate of water potential recovery in branches, and delay in stomatal conductance regulation). The graph edges are labeled with the causal path of the measure-effect and the restoration time window. The time window for irrigation pulses is set to 7-14 days, for water-retaining agents to 30-60 days, and for root pruning to 60-90 days. The time windows are determined based on the restoration area's connectivity. Based on three years of observational data, the following time window criteria were established: the number of consecutive days with root activity response exceeding the baseline value by more than 10% was used as the lower limit of the time window; the response decaying to within 5% of the baseline value was used as the upper limit of the time window. The map was constructed using triplet storage, with each causal edge accompanied by time window parameters and a confidence score (fitted from historical data, ranging from 0.65 to 0.92). The joint distribution of water and root activity at the current time step in the collaborative sensing field was encoded into a hierarchical state embedding vector of a deep Q-network. The state embedding consisted of three layers: the top layer was the micro-topographic category encoding (hilltop, inter-hill, transition slope, using one-hot 3D encoding); the middle layer was the node-level water and root activity features (water content 0.05-0.25 m³). 3 ·m -3 Root vitality index: 0.3-1.2 mV·s -1The underlying layer consists of temporal difference features (differences between the current value and values ​​from 24 hours ago, 72 hours ago, and 168 hours ago). A multi-head attention mechanism (4 heads) is used for the causal graph attention mechanism to calculate the dynamic causal strength coefficient of each restoration path. The strength coefficient is defined as the cosine similarity between the current state embedding vector and the anchor vector of the causal path in the graph, multiplied by the time window matching degree. Short-term pulse paths with restoration time windows shorter than the root physiological response cycle of degraded Haloxylon ammodendron forests are pruned. The root physiological response cycle is defined as 21 days based on the anatomical data of Haloxylon ammodendron roots in the restoration area (the minimum number of days required for the main root to perceive changes in water and for root vitality to significantly improve). Causal edges with time windows shorter than 21 days are removed from the graph, with the removal threshold set to an upper limit of the time window of less than 21 days. The state-action space topology of the deep Q-network is constrained by the causal graph of the restored degraded Haloxylon ammodendron forest after pruning. The state space dimension is compressed from 128 dimensions before pruning to 96 dimensions after pruning. The action space contains three types of discrete actions: maintaining the current repair measure, switching to another repair measure, and stopping the application of the measure. Transitional states that do not meet the continuous repair causal chain are eliminated. The continuous repair causal chain is defined as a path that starts from the original repair measure, passes through all intermediate nodes in the graph, satisfies the time window of more than 21 days and the causal edge confidence of more than 0.70, and finally reaches the root vitality enhancement state, which can be maintained for at least 30 days. The fully causal driven trajectory that can be traced back to the original repair measure is retained. Each trajectory is stored as a state-action-reward sequence with a sequence length of no less than 5 time steps and a time step size of 7 days (consistent with the minimum observation interval of the irrigation pulse). The pruned causal graph is stored in the form of an adjacency matrix with a matrix dimension of 96×96. Non-zero elements represent allowed state transitions. During the training of the deep Q network, the state transitions corresponding to zero elements in the graph adjacency matrix are prohibited from being sampled or having their weights updated. Furthermore, step 3 also includes: constructing a multi-hypothesis counterfactual control scenario in the digital twin of the degraded Haloxylon ammodendron forest restoration area; virtually removing the target restoration measures while keeping the micro-topographical heterogeneity boundary conditions unchanged; generating the water-root vitality counterfactual evolution trajectory under the untreated condition; constructing a pure control through virtual removal to achieve quantitative separation of the contribution of restoration measures; calculating the structured residual field of the actual restoration observation trajectory and the counterfactual trajectory of the degraded Haloxylon ammodendron forest at each state node of the deep Q network; using time-frequency domain differential causal estimation to remove the sub-physiological cycle scale interference introduced by occasional water pulses and random fluctuations of root conductivity sensors; removing environmental noise and sensor disturbances; improving the purity of the response of the residual field to the real restoration signal; extracting the positive offset component whose duration exceeds the minimum response window of the root structural recovery of the degraded Haloxylon ammodendron forest from the structured residual field as the continuous survival gain value independently driven by the restoration measures; outputting and injecting it into the restoration evaluation structured matrix; locking the continuous positive offset; stripping the short-term pulse contribution; and forming an independently driven survival gain index. It should be noted that, for target units where remediation measures have been implemented, the irrigation pulse flag (single irrigation volume) is forcibly set to zero, the water-retaining agent mixing coefficient is set to zero, and the root pruning ratio is set to zero in the twin parameter configuration interface. Simultaneously, the initial water field, initial root vitality field, and boundary water flux conditions of this unit are kept completely consistent with the original scenario. The micro-topographic heterogeneity boundary conditions remain unchanged, including slope, aspect, runoff accumulation, and topographic humidity index. Twin parameters optimized using the improved fruit fly algorithm are adopted, including a baseline value of 0.10 cm for water diffusivity, hydraulic conductivity, and taproot water absorption rate. 3 ·g -1 ·d -1 The baseline value for lateral root water absorption rate is 0.045 cm. 3 ·g -1 ·d -1 A digital twin was used to perform a finite-difference simulation with a time step of 1 hour and a spatial step of 0.5 m, covering the entire growing season from April to October. The simulation output daily water content and root activity index under counterfactual conditions, forming a control dataset corresponding to the measured observation trajectory in time and space. At each output triangulation node, the actual observed value was subtracted from the counterfactual simulation value to obtain the water content residual and root activity residual, respectively. The time resolution remained consistent with the output from step 1 at a daily scale. Time-frequency domain differential causal estimation was used to remove interference components: the residual time series of each node was decomposed using db4 wavelet decomposition with 5 decomposition levels, retaining low-frequency components with a period greater than 3 days to remove sub-physiological cycle-scale fluctuations; the amplitude cutoff threshold was set to 0.005 m of the nominal noise floor of the water sensor. 3 ·m -3 With conductivity probe noise substrate 0.02 mV·s -1 Residual fluctuations below a threshold are zeroed out. Then, Savitzky-Golay filtering is applied to the noise-removed residual field, with a window width of 7 days and a polynomial order of 2, to smooth out residual daily-scale random fluctuations in the residual time series. The final output is a structured residual field purified by the time-frequency domain. Positive offset components whose duration exceeds the minimum response window for structural restoration of the root system in degraded Haloxylon ammodendron forests are extracted from the structured residual field. The minimum response window is calibrated to 21 days based on the anatomical data of Haloxylon ammodendron roots in the restoration area. The extraction rules are as follows: scan the root vitality residual ΔR time series of each node throughout the growing season, identifying those continuously greater than 0.05 mV·s. -1 Within a time window, positive offset segments with a window length greater than or equal to 21 days are retained. For retained segments, it is further determined whether their peak value exceeds 15% of the sum of the counterfactual simulated root vitality value and the upper bound of the sensor noise at that node. Those that meet the condition are marked as structural recovery segments independently driven by remediation measures, and their residual value is defined as the persistent survival gain value, with an output range of 0.05 to 0.35 mV·s. -1The daily continuous survival gain value sequence for each node is injected into the structured matrix of the repair evaluation as a new field and stored in parallel with micro-topographic topological entropy, water field heterogeneity, spatiotemporal ridge curvature of root vitality, and repair response delay spectrum. Step 4: Using the sustained survival gain as the core time-efficiency indicator, the time series of the collaborative perception field is dynamically weighted by integrating the long short-term memory network to quantify the contribution of restoration measures to the long-term survival of degraded Haloxylon ammodendron forests. Furthermore, a causal inference counterfactual control is introduced to distinguish between the sustained restoration effect driven by short-term water pulse stimulation and structural root recovery, thus accurately differentiating between short-term stimulation and structural sustained restoration effects. Step 5: Using the collaborative sensing field as the environmental boundary and the survival gain value as the reward signal, construct a hybrid of two population fruit flies and a deep Q network. The two populations respectively perform global carbon sink competition threshold search and local lateral compensation time-series decoupling. Through adaptive odor concentration weight transfer and experience pool adversarial sampling mechanism, calculate the nonlinear carbon sink competition threshold required to start lateral compensation, and calculate the nonlinear threshold of carbon sink competition to start lateral compensation. Step 6: Based on the carbon sink competition threshold, drive the integrated physiological evaluation of the whole plant, and introduce a nonlinear compensation time-varying model between the lag in lateral branch germination and the carbon sink in the main stem. Dynamically invert the compensation start time and intensity, and correct the boundary conditions of the collaborative sensing field in a closed loop to complete the collaborative diagnosis of the environment-individual-physiological chain, forming a closed-loop collaborative diagnosis of the entire environment-individual-physiological chain.

[0023] Example 2, as Figure 1 , Figure 2 As shown, based on Example 1, the present invention provides a technical solution: Step 4 specifically includes: inputting the synergistic sensing field of the degraded Haloxylon ammodendron forest restoration area along with the persistent survival gain value into a long short-term memory network; setting the root physiological reversibility time window as the forgetting gate time constant; suppressing the interference weight of short-term water pulses on the Haloxylon ammodendron root vitality memory unit; effectively filtering out short-term water fluctuation interference; improving the identification accuracy of the long-term evolution trend of root vitality; and using the difference-in-difference method in causal inference to compare the degraded Haloxylon ammodendron forest in the restoration area with the unrestored but similarly micro-topographical control area. Pairing was performed to calculate the average treatment effect of restoration measures on the root structure recovery of Haloxylon ammodendron and its decay slope along the growing season. The interference of micro-topographic heterogeneity was eliminated to accurately quantify the net treatment effect and timeliness of restoration measures. Based on the decay slope and the biological threshold of Haloxylon ammodendron lateral root germination, the effect type was determined: those with a negative slope and an absolute value greater than the minimum duration of root recovery were short-term water pulse stimulation, and those with a slope close to zero or positive were continuous restoration effects driven by the structural recovery of Haloxylon ammodendron deep roots. The two types of effects, stimulation and recovery, were distinguished to avoid misjudging the long-term restoration evaluation by short-term results. It should be noted that the co-sensing field of the degraded Haloxylon ammodendron forest restoration area, along with the persistent survival gain value, was input into the Long Short-Term Memory (LSTM) network. The co-sensing field was input in diurnal time-series form, including soil moisture content and root activity index for each triangulation node. The persistent survival gain value was used as a parallel input feature channel. The forgetting gate time constant of the LTM network was set to 21 days based on the physiological reversibility time window of Haloxylon ammodendron roots. This value was obtained from three consecutive years of statistical observations on the attenuation of Haloxylon ammodendron root activity response to water pulses in the restoration area: when the water pulse stopped, root activity could recover to 90% of the baseline level within 21 days. Above this window, changes exceeding this range are considered irreversible structural changes. The forgetting gate outputs forgetting weights between 0 and 1 through the sigmoid activation function, selectively discarding historical memory units older than 21 days. This suppresses the interference weights of short-term moisture pulses (time window of 7 to 14 days) on root vitality memory units, ensuring the network focuses on long-term evolutionary features driven by persistent survival gains. The hidden layer dimension of the Long Short-Term Memory network is set to 64, the time step is maintained at 7 days, and training uses the mean squared error loss function and the Adam optimizer with a learning rate of 0.001. The difference-in-differences method in causal inference is used to average the treatment of remedial measures. To quantify the effect, quadrats within the remediation area that had undergone remediation were first designated as the treatment group, while control quadrats with the same microtopography but without remediation were selected as the control group. The pairing criteria were that the difference in microtopographic topographic entropy output from step 2 was less than 0.1 and the difference in water field heterogeneity was less than 0.05. Using the moment of the first application of remediation as the baseline time zero, the mean differences in root activity between the treatment group and the control group were calculated at each time point before and after remediation (7-day intervals, covering the entire growing season from April to October). The difference-in-differences estimator was defined as the difference in the treatment group minus the difference in the control group. A linear mixed-effects model was used to fit the treatment effect. The model uses the attenuation slope along the growing season, with micro-topographical categories (hilltop, inter-hill, transitional slope) as the random intercept and time as the fixed effect. A significant treatment effect is determined when the 95% confidence interval of the difference-in-differences estimator does not include zero. The attenuation slope and its standard error are also output. The effect type is determined based on the attenuation slope combined with the biological threshold for lateral root germination in *Haloxylon ammodendron*. This biological threshold is defined using root anatomy data from the remediation area: lateral root growth requires root activity to be at least 12% above baseline for more than 14 consecutive days to trigger a germination signal. The determination rule is as follows: when the attenuation slope is negative and its absolute value is greater than 0.015 mV·s... -1 ·d -1 (Corresponding to a minimum root recovery period of 21 days with vigor decline exceeding 0.315 mV·s) -1 When the effect is short-term, it is considered a pulse effect of water stimulation. This type of effect decays to baseline within 21 days after irrigation stops and does not trigger lateral root structural compensation; when the decay slope approaches zero (absolute value less than 0.005 mV·s), it is considered short-term. -1·d -1 When the value is positive, it is determined to be a continuous repair effect driven by the structural restoration of the deep root system of Haloxylon ammodendron. This type of effect is accompanied by the baseline value of the taproot water absorption rate increasing from 0.10 cm. 3 ·g -1 ·d -1 To 0.12 cm 3 ·g -1 ·d -1 The gradual improvement, which can be maintained for at least 60 days, is marked in the structured matrix with labels; Step 5 specifically includes: constructing global population fruit fly individuals, each encoding candidate solutions for carbon sink competition threshold of degraded Haloxylon ammodendron main stem, using the long-term survival contribution of root vitality output by the deep Q network as the global odor concentration evaluation function, and conducting targeted search for the carbon allocation critical point of Haloxylon ammodendron main stem and lateral branches to improve the physiological relevance of the carbon allocation critical point search and avoid blind optimization; constructing local population fruit fly individuals, each encoding candidate solutions for lateral branch germination lag time and compensation intensity of degraded Haloxylon ammodendron, using the marginal compensation rate of lateral branch biomass accumulation for carbon sink loss of main stem as the local odor concentration, quantifying the initiation condition of Haloxylon ammodendron lateral branch compensation after restoration, directly quantifying the economics of lateral branch compensation, linking the initiation condition with carbon gain, defining the current water-root distribution state of the collaborative sensing field as the environmental boundary of the dual population fruit fly-deep Q network hybrid, and mapping the persistent survival gain value as the immediate reward signal of the deep Q network for the whole plant carbon allocation strategy of Haloxylon ammodendron, so that the reward signal is directly related to long-term survival and guides the strategy to prioritize root recovery; It should be noted that the global population of fruit flies is set to 30 individuals. Each individual encodes a candidate solution for the carbon sink competition threshold of the main stem in a degraded Haloxylon ammodendron forest. The encoding form is a one-dimensional real-valued vector, and the value range is 0.55 to 0.85, which is determined based on the carbon-13 isotope ratio of the assimilated branches of Haloxylon ammodendron in the restoration area, to determine the physiologically permissible range of the carbon allocation ratio of the main stem. The spatial location of each fruit fly individual corresponds to a candidate carbon sink competition threshold of the main stem, that is, the upper limit of the carbon allocation ratio of the whole plant obtained by the main stem. The root vitality long-term survival contribution output by the deep Q network is used as the global odor concentration evaluation function. This contribution value has been obtained from step 4, and the value ranges from 0 to 1, representing the expected duration of root vitality that can be maintained under the current carbon allocation strategy. In calculating odor concentration, the carbon allocation ratio encoded by each fruit fly individual is input into a deep Q-network. This network drives forward propagation and outputs the corresponding predicted long-term survival contribution value, which is directly used as the odor concentration for that fruit fly individual. The global population performs odor concentration optimization within a five-dimensional search space. The five dimensions of the search space correspond to the main stem diameter, the number of lateral branches, the root water absorption rate, the water availability coefficient, and the carbon allocation ratio adjustment step size, respectively. When searching for the carbon allocation critical point of the main stem and lateral branches of Haloxylon ammodendron, the carbon allocation ratio corresponding to the globally optimal odor concentration is recorded. When the globally optimal value has not been updated for 10 consecutive iterations, a step size contraction mechanism is triggered, and the search step size is reduced from the initial 0.05 to 0.01 to ensure convergence to the carbon allocation. The critical point was set at 20 individuals in the local fruit fly population. Each individual was encoded as a candidate solution for the lag time and compensation intensity of lateral branch sprouting in the degraded Haloxylon ammodendron forest. The encoding format was a two-dimensional real-valued vector. The first dimension was the lag time of lateral branch sprouting, with a value range of 15 to 60 days based on three consecutive years of lateral branch phenological observation data in the restoration area. The second dimension was the compensation intensity, which was the ratio of the rate of biomass accumulation per unit time after lateral branch sprouting to the undamaged state, with a value range of 0.2 to 1.2. The marginal compensation rate of lateral branch biomass accumulation to the carbon sink loss of the main stem was used as the evaluation function for local odor concentration. The marginal compensation rate was calculated as the daily increase in lateral branch biomass divided by the daily loss of carbon sink in the main stem. A ratio greater than 1 indicated complete compensation. When calculating the initiation conditions for lateral branch compensation after quantification and restoration of the cover loss, the optimal main stem carbon allocation ratio searched by the global population is used as the input boundary condition for the local population. Under this boundary condition, the local population searches for the combination of lag time and compensation intensity that maximizes the marginal compensation rate. When calculating the odor concentration, the candidate solution is input into the time-varying model of lateral branch compensation pre-embedded in step 6. The model outputs the lateral branch biomass accumulation curve and the main stem carbon sink loss curve. The average marginal compensation rate of the two is calculated within the time window from day 30 to day 90 after lateral branch sprouting. This average value is used as the odor concentration of local fruit fly individuals. The local population adopts the same step size adaptive adjustment mechanism as the global population, but the initial step size is set to 2 days and 0.05, corresponding to the adjustment granularity of lag time and compensation intensity respectively; the current water-root distribution state of the collaborative sensing field is defined as the environmental boundary of the hybrid of two population fruit flies and deep Q network. The collaborative sensing field includes the soil moisture content, root vitality index and persistent survival gain value output in step 3 at each triangulation node. The spatial range is the sampling unit of all applied remediation measures in the remediation area, and the temporal range is a sliding window 90 days before the current evaluation time. The real-time update frequency of the environmental boundary is set to once every 7 days, consistent with the time step in step 4. When the persistent survival gain value is mapped to the immediate reward signal of the deep Q network for the carbon allocation strategy of the whole plant of Haloxylon ammodendron, a piecewise mapping function is used: when the gain value is lower than 0.10 mV·s. -1 The time reward is set to a negative value of -0.2, indicating that the current carbon allocation strategy is not conducive to root recovery; the gain value is between 0.10 and 0.20 mV·s. -1 When the gain is between 0 and 0.5, the reward is linearly mapped to the range of 0 to 0.5; the gain value is higher than 0.20 mV·s. -1 The reward is set to 1.0, indicating that the carbon allocation strategy has triggered effective root structure recovery. The collaborative mechanism between the two populations is achieved through odor concentration weighted migration: after the global population completes every 50 iterations, the current optimal carbon allocation ratio is passed to the local population as prior knowledge. The local population adjusts the search center of the lag time accordingly. The migration weight is set to 0.3, that is, the search center of the local population moves 30% of the current distance towards the global optimal direction. The experience pool capacity of the deep Q network is set to 5000 trajectories. After each iteration of the two populations, 64 trajectories are randomly sampled from the experience pool for network parameter updates. The learning rate is set to 0.0005, and the target network update frequency is once every 100 steps. Furthermore, step 5 also includes: when the global population finds a candidate interval for carbon sink competition in the main stem of Haloxylon ammodendron with an odor concentration higher than the migration threshold, the carbon allocation ratio of the interval is used as a priori weight to migrate to the local population, narrowing the initial range of the time-series search for lag compensation of lateral branches in degraded Haloxylon ammodendron forests, avoiding blind search, improving the convergence speed and stability of the local population, storing both high-survival and low-survival trajectories of Haloxylon ammodendron in the experience pool of the deep Q network, and using an adversarial sampling mechanism to prioritize sampling the experience with the greatest difference from the current carbon sink competition state of Haloxylon ammodendron, avoiding the lateral branch compensation strategy from falling into the shallow root adaptation mode of local optima, forcing the model to focus on extreme carbon competition states, avoiding inefficient compensation strategies, iteratively executing the collaborative search of the global and local populations until the odor concentration converges, solving the critical allocation ratio of carbon sink in the main stem required to start lateral branch compensation in degraded Haloxylon ammodendron forests and the corresponding nonlinear carbon sink competition threshold, accurately quantifying the carbon allocation critical point, and supporting the dynamic decision-making of lateral branch compensation; It should be noted that when the global population finds a candidate interval for carbon sink competition on the main stem of *Haloxylon ammodendron* with an odor concentration higher than the migration threshold of 0.75, the carbon allocation ratio within this interval is used as a priori weight and transferred to the local population. The migration threshold is determined according to the criteria for determining the sustained repair effect in step 4: when the long-term survival contribution value is greater than 0.75, it indicates that the current carbon allocation strategy has the potential to trigger lateral branch compensation. After every 50 iterations, the global population extracts the carbon allocation ratio corresponding to the current global optimal solution and uses it as the central reference value for the lag time search of the local population. The migration weight coefficient is set to 0.3, that is, the lag time candidate solution of each fruit fly individual in the local population is updated according to the formula: new value = original value × 0.7 + The global optimal carbon allocation ratio is mapped to a value of ×0.3. The mapping relationship between the carbon allocation ratio and the lag time is fitted as a linear function based on three years of phenological data of the remediation area, with a slope of 0.8. Through the above migration mechanism, the initial range of the lateral branch lag compensation time series search is narrowed from 15-60 days to a window of ±10 days from the center value after migration, significantly improving the convergence efficiency of the local population. High survival trajectories are defined as sequences that are determined to have a continuous remediation effect in step 4 and have a long-term survival contribution value greater than 0.7, while low survival trajectories are defined as sequences with a short-term impulse stimulation effect and a contribution value less than 0.3. The two types of trajectories are stored in a 1:1 ratio, and the total capacity of the experience pool is maintained at 5000. The adversarial sampling mechanism prioritizes sampling the current shuttle. The experience with the greatest difference in carbon sink competition states is as follows: Calculate the Euclidean distance between the current carbon allocation ratio and the average carbon allocation ratio of each trajectory in the experience pool, sort them from largest to smallest distance, and randomly sample 64 trajectories from the 20% with the largest distance for network updates. This mechanism forces the deep Q network to pay attention to boundary situations where the carbon allocation strategy deviates significantly, avoiding the shallow-rooted adaptation mode of lateral compensation strategies falling into local optima, i.e., the inefficient compensation path that relies only on surface water absorption and ignores the structural recovery of deep roots. The convergence condition is set as follows: the change in the global optimal odor concentration is less than 0.01 in 20 consecutive iterations, and the change in the local population optimal marginal compensation rate is less than 0.02. In each iteration, After the global population completes 100 individual fruit fly position updates, a migration is triggered. The local population performs 80 position updates on this basis. The deep Q network updates its network parameters after each complete iteration of the two populations (approximately 180 individual evaluations). The target network performs a soft update every 100 steps with a soft update coefficient τ of 0.01. After convergence, the critical allocation ratio of the main stem carbon sink required to initiate lateral branch compensation in the degraded Haloxylon ammodendron forest is calculated. The output value is the median of the convergence interval, rounded to two decimal places. At the same time, the corresponding nonlinear carbon sink competition threshold (range 0.62-0.78) is recorded as the input boundary conditions for the whole-plant physiological integration model in step 6, which is used to drive the dynamic inversion of the initiation time and intensity of lateral branch compensation. Step 6 specifically includes: based on the calculated nonlinear carbon sink competition threshold and lateral branch germination lag time, to address the typical imbalance state of excessive carbon sequestration by the main stem in the degraded Haloxylon ammodendron forest restoration area, a step response function for lateral branch compensation initiation is constructed. A dynamic switching mechanism of inhibition-release in the main stem-lateral branch carbon sink competition is established to avoid frequent oscillations of the carbon allocation strategy near the threshold, thereby improving the stability of whole-plant carbon use. This dynamic switching mechanism is embedded into the whole-plant physiological integration model of the degraded Haloxylon ammodendron forest, using the measured lateral branch germination time and intensity in the restoration area as inversion constraints for dynamic inversion compensation. The water-carbon flow redistribution trajectory after initiation is used to iteratively approximate the actual response delay and compensation efficiency of lateral branch compensation to carbon sink competition in the main stem. This enables precise quantification of the timing and intensity of lateral branch compensation, reduces carbon flow distribution errors, and feeds back the compensation initiation parameters obtained from the inversion to the collaborative sensing field. With the goal of dynamic boundary calibration of the collaborative sensing field, the root water uptake parameters and water conduction boundary conditions are updated. This completes the closed-loop collaborative diagnosis of the environment-individual-physiological chain for the restoration evaluation of degraded Haloxylon ammodendron forests, forms a full-chain closed-loop self-calibration, and improves the physiological consistency and field applicability of the restoration evaluation. It should be noted that the calculated nonlinear carbon sink competition threshold (median of the convergence interval, ranging from 0.62 to 0.78) and the lateral branch germination lag time (15-60 days, locked to the central value ± 10 days after migration and contraction) are used as input boundary conditions. In the degraded Haloxylon ammodendron forest restoration area, a step response function triggers lateral branch compensation when the carbon allocation ratio of the main stem exceeds the threshold. The step response function uses an S-shaped growth curve to describe the carbon allocation and release process after lateral branch germination: when the carbon sink competition intensity of the main stem exceeds the threshold, the carbon allocation ratio obtained by the lateral branches increases stepwise from an inhibited state (accounting for less than 0.10% of the total plant carbon allocation) to a released state. The carbon allocation ratio of the main stem (0.25-0.35% of the total plant carbon allocation) was set to 7 days based on root anatomy data of the repair area, representing the transition period from the inhibition state to the release state. A dynamic switching mechanism for carbon sink competition between the main stem and lateral branches was established: when the carbon allocation ratio of the main stem is above the threshold for 14 consecutive days, the system enters the release state, and the carbon allocation weight of the lateral branches is increased by 0.02 every 3 days; when the carbon allocation ratio of the main stem falls below the threshold by 0.05 for 10 consecutive days, the system reverts to the inhibition state, and the carbon allocation weight of the lateral branches is decreased by 0.01 every 5 days, forming a competitive switching logic with a hysteresis loop to avoid carbon allocation strategies. The model exhibits frequent oscillations near the threshold. A comprehensive physiological model of the degraded Haloxylon ammodendron forest uses a synergistic sensing field as the environmental driving boundary, a persistent survival gain value as a correction factor for root water absorption capacity, and measured lateral branch sprouting times (defined as the date when lateral branch length first exceeds 2 cm) and sprouting intensity (defined as the average daily increase in lateral branch biomass within 30 days after sprouting) measured every 7 days in the restoration area as inversion constraints. A particle swarm optimization algorithm is used to iteratively approximate the response delay and compensation efficiency parameters in the model. The particle swarm size is set to 30, the number of iterations to 200, and the search range for response delay is set to 5-20 days after the step response. The search range for compensation efficiency is... The range was set to 0.4-1.1 (the compensation ratio of lateral branch biomass accumulation to carbon sink loss in the main stem). In each iteration, the model ran for 90 days with the current parameter combination (covering the entire cycle of lateral branch germination to stable compensation), outputting the simulated lateral branch germination time and germination intensity. The root mean square error between the calculated and measured values ​​was used as the fitness function. Convergence was determined when the fitness value was below 0.05 for 25 consecutive iterations. The actual response delay and compensation efficiency of lateral branch compensation were calculated. Parameter update operations were performed with the goal of dynamic calibration of the boundary of the collaborative sensing field. The specific calibration objects included: the baseline value of the taproot water absorption rate in the root water absorption layer (originally 0.10 cm). 3 ·g -1 ·d -1 The compensation efficiency coefficient is adjusted linearly to 0.10-0.12 cm. 3 ·g -1 ·d -1 (Interval), lateral root water absorption rate benchmark value (original 0.045 cm) 3 ·g-1 ·d -1 Adjust to 0.045-0.06cm 3 ·g -1 ·d -1 The range), and the hydraulic conductivity parameters of the soil at a depth of 80-120cm in the water conduction layer (originally calibrated to 0.15cm·d). -1 Adjust the value by 0.02-0.05 cm·d, depending on the length of the response delay days. -1 The calibrated boundary conditions are re-injected into the iterative inversion process of step 1, triggering a new round of collaborative sensing field generation and cascade calculations of steps 2 to 5. The loop converges when the relative difference between the compensation initiation parameters output in step 6 and the previous inversion result is less than 5%. After convergence, an environmental-individual-physiological chain collaborative diagnostic report for the evaluation of degraded Haloxylon ammodendron forest restoration is output. The report includes a comprehensive score of the restoration effect of each micro-topographic unit, a marker of the lateral branch compensation initiation status, and suggestions for carbon allocation strategy optimization.

[0024] The following section details the workflow of a multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests.

[0025] First, within the degraded Haloxylon ammodendron forest restoration area, soil moisture sensors and root conductivity probes were deployed to acquire temporal data on water availability and root conductivity profiles of typical micro-topographic units. An initial digital twin incorporating root topology was constructed, establishing a chaotic mapping relationship between the water field and root vitality. An improved fruit fly algorithm was used to dynamically optimize the water conduction coefficient and root water absorption rate. The optimized parameters were then reinjected into the water conduction layer and root water absorption layer of the digital twin, iteratively retrieving the unsteady-state coupling mechanism between water and roots to generate a collaborative sensing mechanism for water and root vitality. Based on this, using the collaborative sensing field as the basis, the heterogeneity information gain is calculated through the micro-topography weight tensor, triggering progressively encrypted sampling and dynamic tracking of abrupt boundary, and outputting a collaborative sensing field with spatial resolution adapted to the degree of micro-topography heterogeneity. Furthermore, dynamic topological reconnection and triangulation are implemented on the collaborative sensing field, and a deep residual graph convolutional network is introduced to achieve multimodal feature alignment between environmental water nodes and root vitality nodes, outputting a structured matrix, and uniformly encoding micro-topographic topological entropy, water field heterogeneity, spatiotemporal ridge curvature of root vitality, and repair response delay spectrum. Secondly, a time-sensitive causal knowledge graph for remediation measures is constructed. The co-perception field is encoded as a hierarchical state embedding vector of a deep Q-network. A causal graph attention mechanism is used to prune short-term pulse paths. The state-action space topology is constrained by the pruned causal graph. A counterfactual comparison scenario is constructed in a digital twin. The structured residual field of actual observation and counterfactual trajectory is calculated. Random fluctuations and sub-physiological cycle interferences are eliminated by time-frequency domain differential causal estimation. The persistent survival gain value is extracted and injected into the structured matrix. The co-perception field and the persistent survival gain value are jointly input into the long short-term memory network. The root physiological reversible time window is used as the forgetting gate constant to suppress short-term water pulse interference. The average treatment effect and decay slope of the remediation measures are quantified by the double difference method. The persistent remediation effect driven by short-term water pulse stimulation and structural root recovery is determined by combining the biological threshold of lateral root germination. Finally, a hybrid system of two populations of fruit flies and a deep Q-network was constructed. Using a collaborative sensing field as the environmental boundary and a continuous survival gain value as the reward signal, the global population searched for the carbon sequestration competition threshold of the main stem. Local populations decoupled the lag time and compensation intensity of lateral branch germination. The nonlinear carbon sequestration competition threshold required for lateral branch compensation initiation was calculated through odor concentration weighted transfer and an adversarial sampling mechanism using an experience pool. Based on the calculated threshold and lag time, a dynamic switching mechanism of inhibition-release for carbon sequestration competition between the main stem and lateral branches was constructed and embedded into a whole-plant physiological integration model. Using the measured lateral branch germination time and intensity as inversion constraints, the response delay and compensation efficiency were dynamically inverted. The compensation initiation parameters were fed back to the collaborative sensing field to update the boundary conditions for root water absorption and water conduction, forming a closed-loop convergent collaborative diagnosis. The final output was a diagnostic report containing a comprehensive score of the restoration effect of each micro-topographic unit, a lateral branch compensation initiation status marker, and carbon allocation strategy optimization suggestions.

[0026] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests, characterized in that, Includes the following steps: Step 1: For the restoration of degraded Haloxylon ammodendron forests, we integrate digital twins and improved fruit fly algorithms to construct a root-water spatiotemporal mapping model and dynamically optimize the twin parameters. Simultaneously, we capture the abrupt boundary of root vitality and water field mutation and output a collaborative sensing field. Step 2: Using the collaborative sensing field as the state space basis for multi-index comprehensive evaluation, micro-topographic topology coding and root vitality spatiotemporal ridge line extraction are introduced to construct a deep alignment layer of environment-individual heterogeneous indicators; Step 3: Using the collaborative perception field as the state space of the deep Q-network, we combine causal path pruning and counterfactual reasoning from the knowledge graph to remove the persistent survival gain value caused by the remedial measures. Step 4: Using the sustained survival gain value as the core time-efficiency indicator, the time series of the collaborative sensing field is dynamically weighted by integrating the long short-term memory network to quantify the contribution of restoration measures to the long-term survival of degraded Haloxylon ammodendron forests, and to distinguish the sustained restoration effect driven by short-term water pulse stimulation and structural root restoration. Step 5: Using the collaborative sensing field as the environmental boundary and the survival gain value as the reward signal, construct a hybrid of two population fruit flies and a deep Q network, and calculate the nonlinear carbon sink competition threshold required to initiate lateral branch compensation. Step 6: Based on the carbon sink competition threshold, drive the whole-plant physiological integrated evaluation, close the loop to correct the boundary conditions of the collaborative sensing field, and complete the collaborative diagnosis of the environment-individual-physiological chain.

2. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 1 specifically includes: Soil moisture time-series data and root electrical conductivity profiles of typical micro-topographic units in the degraded Haloxylon ammodendron forest restoration area were collected. An initial digital twin containing the root topology of Haloxylon ammodendron was constructed, and a coarse-grained chaotic mapping relationship between water field and root activity was established. An improved fruit fly algorithm was used to dynamically optimize twin parameters. The odor concentration function was defined as the weighted sum of the water field prediction error and the spatiotemporal smoothness of root vitality, and the micro-topography curvature tensor was introduced to adaptively adjust the odor search step size of individual fruit flies. The optimized twin parameters were reinjected at different scales into the water conduction layer and root water absorption layer of the digital twin. The unsteady coupling mechanism of water and root system during the restoration of degraded Haloxylon ammodendron forest was iteratively inverted to generate a dynamic parameterized expression of the water-root vitality synergy field that can characterize the restoration process.

3. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 2, characterized in that: Step 1 further includes: We constructed a micro-topographic heterogeneity weight tensor based on the slope, aspect, runoff accumulation, and topographic humidity index of the degraded Haloxylon ammodendron forest restoration area, calculated the heterogeneity information gain of each candidate sampling unit, and generated an adaptive sampling prior probability density distribution. Based on the prior probability density, progressively encrypted sampling is automatically triggered in regions where the heterogeneity weight exceeds the chaos threshold, and dynamic tracking of mutation boundaries and critical point marking are implemented when the root vitality gradient change rate is greater than the set Lyapunov exponent threshold. The high-frequency water-root data from encrypted sampling is fed back in real time to the improved fruit fly optimization process, dynamically correcting the nonlinear response function of root water absorption in the twin parameters, and outputting a collaborative sensing field with spatial resolution adapted to the degree of micro-topographic heterogeneity.

4. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 2 specifically includes: Dynamic topological reconnection is performed on each micro-topographic unit in the collaborative sensing field. Delaunay triangulation is driven by a water-root vitality joint gradient field to encode the adaptive graph mapping between topographic position and root vitality ridge node. With the constraint of maximizing cross-scale mutual information in the ridge neighborhood, a deep residual graph convolutional network is introduced to project the multimodal heterogeneous features of water nodes and root vitality nodes in degraded Haloxylon ammodendron forest environment into the shared latent space of the restoration state. The output is a structured matrix for evaluating the restoration of degraded Haloxylon ammodendron forests. This matrix uniformly encodes the micro-topographic topological entropy, water field heterogeneity, spatiotemporal ridge curvature of root activity, and restoration response delay spectrum, thus completing a compact expression of the hierarchical heterogeneity characteristics of the water-root synergistic field.

5. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 3 specifically includes: A time-sensitive causal knowledge graph for restoration measures of degraded Haloxylon ammodendron forests was constructed. The nodes cover intermediate variables such as irrigation pulse, water-retaining agent, root pruning and Haloxylon ammodendron-specific physiological response. The edges are labeled with the causal path of the measure-effect and the restoration time window. The current water-root activity joint distribution of the collaborative sensing field is encoded into a hierarchical state embedding vector of a deep Q network. At the same time, the dynamic causal intensity coefficient of each restoration path is calculated using a causal graph attention mechanism, and short-term pulse paths with restoration time windows smaller than the physiological response cycle of roots in degraded Haloxylon ammodendron forests are pruned. Using the state-action space topology of a deep Q-network constrained by the causal graph of the restoration of degraded Haloxylon ammodendron forests after pruning, transitional states that do not satisfy the causal chain of continuous restoration are eliminated, while the fully causal driving trajectory that can be traced back to the original restoration measures is retained.

6. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 5, characterized in that: Step 3 also includes: In the digital twin of the degraded Haloxylon ammodendron forest restoration area, a multi-hypothesis counterfactual control scenario was constructed. The target restoration measures were virtually removed while keeping the micro-topographic heterogeneity boundary conditions unchanged, and the counterfactual evolution trajectory of water-root vitality under the condition of no measures was generated. The structured residual fields of the actual restoration observation trajectory and counterfactual trajectory of degraded Haloxylon ammodendron forest at each state node of the deep Q network were calculated. Time-frequency domain differential causal estimation was used to eliminate sub-physiological periodic scale interference introduced by occasional precipitation pulses and random fluctuations of root conductivity sensors. The positive offset component whose duration exceeds the minimum response window of root structure restoration of degraded Haloxylon ammodendron forest is extracted from the structured residual field and used as a persistent survival gain value driven independently by restoration measures. This value is then output and injected into the restoration evaluation structured matrix.

7. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 4 specifically includes: The collaborative sensing field of the degraded Haloxylon ammodendron forest restoration area, along with the continuous survival gain value, is input into the long short-term memory network. The root physiological reversibility time window is set as the forgetting gate time constant to suppress the interference weight of short-term water pulses on the Haloxylon ammodendron root vitality memory unit. Using the difference-in-differences method in causal inference, the degraded Haloxylon ammodendron forest in the restoration area was paired with the Haloxylon ammodendron in the control area with the same microtopography but no restoration, and the average treatment effect of restoration measures on the recovery of Haloxylon ammodendron root structure and its attenuation slope along the growing season were calculated. The effect type was determined based on the attenuation slope combined with the biological threshold of lateral root germination in Haloxylon ammodendron: a negative slope with an absolute value greater than the minimum duration of root recovery indicates a short-term water pulse stimulus, while a slope approaching zero or being positive indicates a sustained repair effect driven by the structural recovery of the deep root system of Haloxylon ammodendron.

8. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 5 specifically includes: We constructed a global population of fruit fly individuals, and each individual encoded a candidate solution for the carbon sink competition threshold of the main stem of a degraded Haloxylon ammodendron forest. We used the long-term survival contribution of root vitality output by the deep Q network as the global odor concentration evaluation function to search for the carbon allocation critical point of the main stem and lateral branches of Haloxylon ammodendron. We constructed local populations of fruit flies, and each individual encoded candidate solutions for the lag time and compensation intensity of lateral branch germination in degraded Haloxylon ammodendron forests. We used the marginal compensation rate of lateral branch biomass accumulation to carbon sink loss in the main stem as the local odor concentration to quantify the initiation conditions for Haloxylon ammodendron lateral branch compensation after restoration. The current water-root distribution state of the collaborative sensing field is defined as the environmental boundary of the dual-population fruit fly-deep Q network hybrid, and the persistent survival gain value is mapped as the instantaneous reward signal of the deep Q network for the whole-plant carbon allocation strategy of Haloxylon ammodendron.

9. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 8, characterized in that: Step 5 further includes: When the global population search finds a candidate interval for carbon sink competition of Haloxylon ammodendron main stem with odor concentration higher than the migration threshold, the carbon allocation ratio of the interval is used as a priori weight to migrate to the local population, thereby narrowing the initial range of the time series search for lag compensation of degraded Haloxylon ammodendron lateral branches. In the experience pool of the deep Q network, both high-survival trajectories and low-survival trajectories of Haloxylon ammodendron are stored simultaneously. An adversarial sampling mechanism is used to prioritize sampling the experience that differs the most from the current carbon sink competition state of Haloxylon ammodendron. The global and local populations are searched iteratively until the odor concentration converges. The critical allocation ratio of the main stem carbon sink required to initiate the compensation of lateral branches in the degraded Haloxylon ammodendron forest and the corresponding nonlinear carbon sink competition threshold are calculated.

10. The multi-index comprehensive evaluation method for the restoration effect of degraded Haloxylon ammodendron forests according to claim 1, characterized in that: Step 6 specifically includes: Based on the calculated nonlinear carbon sink competition threshold and lateral branch germination lag time, a step response function for lateral branch compensation is constructed to address the typical imbalance state of excessive encroachment of carbon allocation on lateral branches by the main stem carbon sink in the degraded Haloxylon ammodendron forest restoration area. A dynamic switching mechanism of inhibition-release of carbon sink competition between the main stem and lateral branches is established. The inhibition-release dynamic switching mechanism is embedded into the whole-plant physiological integration model of degraded Haloxylon ammodendron forest. The measured lateral branch germination time and germination intensity in the restoration area are used as inversion constraints to dynamically invert the water-carbon flow redistribution trajectory after compensation is initiated. The actual response delay and compensation efficiency of lateral branch compensation to carbon sink competition of the main stem are calculated through iterative approximation. The compensation initiation parameters obtained by inversion are fed back to the collaborative sensing field. With the goal of dynamic boundary calibration of the collaborative sensing field, the root water absorption parameters and water conduction boundary conditions are updated to complete the closed-loop collaborative diagnosis of the environment-individual-physiological chain for the evaluation of degraded Haloxylon ammodendron forest restoration.

Images