Greening plant screening and configuration method in coal mining subsidence area

Abstract

The application discloses a coal mining subsidence area greening plant screening and configuration method, and belongs to the technical field of mine area ecological restoration. The method obtains the global pollutant distribution and the soil physical and chemical parameters by the unmanned aerial vehicle carrying the hyperspectral imaging equipment and the X-ray fluorescence sensor; a plant function matching model is constructed, and a partition plant combination scheme is generated based on a machine learning algorithm in combination with plant pollution tolerance threshold, enrichment coefficient and other characteristics; a restoration unit is divided according to the terrain, and a spatial configuration atlas is generated to define planting parameters; a thing networking network is deployed to monitor the rhizosphere microenvironment, and data space correlation is realized in combination with near-ground unmanned aerial vehicle root scanning; a dynamic feedback mechanism is established, and the target rhizosphere area is accurately regulated when the monitoring data deviates from the threshold. The application solves the problem of plant configuration deviation caused by the environmental heterogeneity of the subsidence area, realizes closed-loop management of the coal mining subsidence area treatment, significantly improves the vegetation reconstruction success rate and the restoration efficiency, and greatly reduces the artificial intervention cost.

Description

Technical Field

[0001] This invention relates to the field of ecological restoration technology in mining areas, and in particular to a method for selecting and configuring greening plants in coal mining subsidence areas. Background Technology

[0002] Large-scale surface subsidence areas formed by coal mining are challenging areas for ecological restoration due to their characteristics such as damaged soil structure, complex pollutant migration, and fragmented topography. These areas generally suffer from strong spatial heterogeneity of heavy metal pollution (e.g., cadmium and lead concentration gradients can vary by tens of times), abrupt changes in soil physicochemical properties (pH values ​​can fluctuate by 3 units within a 10-meter range), and diverse micro-topography (coexistence of slopes, waterlogged areas, and gangue piles).

[0003] Currently, greening efforts in subsidence areas mainly rely on manual surveys combined with literature review and experience in selecting plant species, which has several limitations:

[0004] At the environmental assessment level: relying on discrete sampling points (usually 3-5 points per hectare) makes it difficult to capture spatial gradient changes in pollutants. For example, the lead concentration at the edge of a subsidence and waterlogged area is 6 times higher than that in the center, but fixed sampling points fail to identify this gradient, resulting in an imbalance in the configuration of lead-tolerant plants.

[0005] At the plant configuration level: static schemes cannot respond to dynamic environmental changes. For example, in the initial stage of a project, alfalfa was configured with a pH of 7.5, but weathering of gangue caused the local pH to drop to 5.2 after six months, resulting in large-scale plant death. Existing technologies lack real-time feedback mechanisms for the rhizosphere microenvironment (such as heavy metal speciation and secretion interactions).

[0006] At the engineering implementation level: terrain obstacles lead to monitoring blind spots (the missing rate of drone scans on the back of steep slopes exceeds 40%), and traditional irrigation / pesticide application is difficult to accurately reach the active root layer (the depth error of manual operation often exceeds 30 cm).

[0007] While recent studies have attempted to incorporate remote sensing technology, its resolution is insufficient to detect changes at the rhizosphere scale. Other approaches use fixed sensor networks to monitor the soil, but the density of these sensors is insufficient (one per 50 square meters) and they cannot correlate with three-dimensional root growth data. These shortcomings often lead to problems such as excessive maintenance costs and vegetation degradation in remediation projects. Therefore, there is an urgent need to develop a dynamic and precise remediation technology system adapted to the complex environmental characteristics of subsidence areas. Summary of the Invention

[0008] This invention overcomes the problem of inaccurate plant configuration caused by the environmental heterogeneity of the subsidence area in the prior art, realizes closed-loop management of "monitoring-decision-execution-optimization", significantly improves the success rate of vegetation reconstruction and restoration efficiency, and greatly reduces the cost of artificial intervention.

[0009] To achieve the above objectives, the present invention adopts the following solution:

[0010] A method for selecting and configuring greening plants in coal mining subsidence areas, comprising the following steps:

[0011] S1: Using a drone equipped with a hyperspectral imaging device and an X-ray fluorescence sensor, spatial distribution data of pollutants and soil physicochemical parameters of the entire subsidence area are obtained. The spatial distribution data of pollutants includes information on heavy metal types and concentration gradients, and the soil physicochemical parameters include pH value, organic matter content and soil moisture content.

[0012] S2: Construct a plant function matching model. Input the spatial distribution data of pollutants and soil physicochemical parameter data into a pre-trained machine learning model. The machine learning model outputs a target plant combination scheme based on a plant function database. The plant function database includes the pollution tolerance threshold, heavy metal enrichment coefficient, rhizosphere exudate type and ecological niche characteristics of candidate plants. The target plant combination scheme includes specific plant species and their functional ratios for different pollution zones.

[0013] S3: Divide the subsidence area into restoration units based on the topographic features, and generate a spatial configuration map in combination with the target plant combination scheme. The restoration unit includes the slope area, water accumulation area and coal gangue accumulation area. The spatial configuration map records the plant species combination, planting density and hierarchical structure in each restoration unit.

[0014] S4: Deploy an IoT monitoring network to periodically collect plant rhizosphere microenvironment data through buried soil multi-parameter sensors, and simultaneously control a near-ground UAV equipped with a micro rhizosphere probe to dynamically scan the three-dimensional coordinates of plant roots, generate a root growth trajectory map, and spatially correlate it with the rhizosphere microenvironment data.

[0015] S5: Establish a dynamic feedback mechanism. When the associated rhizosphere microenvironment data deviates from the preset threshold range, locate the target rhizosphere region based on the current root growth trajectory map, trigger the control command execution system, and perform at least one of the following operations: inject pH adjustment solution into the target rhizosphere region, apply more rhizosphere growth-promoting bacteria agent, and adjust the irrigation amount.

[0016] Preferably, the machine learning model in step S2 is trained using an incremental learning mechanism. After planting, it continuously receives rhizosphere microenvironment data collected in step S4 and plant growth status assessment indicators obtained by UAV hyperspectral imaging equipment. The plant growth status assessment indicators include chlorophyll content index, canopy nitrogen content, and biomass accumulation rate. When the actual growth index of any species in the target plant combination scheme is found to be lower than the preset growth threshold three times in a row, the model parameter update module is triggered. The weight allocation of pollution tolerance threshold and heavy metal enrichment coefficient is recalculated using the newly added data, and the corrected target plant combination scheme is output. The model parameter update module generates a version record after each run and uses the corrected target plant combination scheme to dynamically adjust the planting scheme of subsequent batches of plants through spatial configuration maps.

[0017] Preferably, the method for obtaining the plant growth status assessment indicators specifically includes:

[0018] The vegetation canopy reflectance spectrum is collected by the drone multispectral camera deployed in step S4, and the reflectance spectrum is converted into chlorophyll content index based on the normalized vegetation index calculation formula.

[0019] A convolutional neural network model was used to analyze the visible and near-infrared images of the canopy, outputting a pixel-level canopy nitrogen content distribution map and calculating the regional average value.

[0020] A three-dimensional model of the plant was reconstructed by combining lidar point cloud data, and the biomass accumulation rate was inverted based on the volume growth rate.

[0021] The method for setting the preset growth threshold for each plant is as follows: retrieve the historical growth data of the species in the non-polluted area from the plant function database, and take the average decrease of 30% in chlorophyll content index, 25% in canopy nitrogen content, and 40% in biomass accumulation rate as the dynamic judgment benchmark values.

[0022] The time interval between three consecutive monitoring cycles is automatically adjusted based on the plant growth stages recorded in the spatial configuration map, with a 7-day interval for fast-growing herbaceous plants and a 30-day interval for woody plants.

[0023] Preferably, in step S2, the model parameter update module performs pollutant migration early warning analysis simultaneously during operation. Based on the spatial distribution data of pollutants obtained in step S1 and the rate of change of specific heavy metal ion concentrations collected in real time in step S4, the abnormal diffusion area of ​​pollutants is identified by spectral analysis. When the rate of change of specific heavy metal ion concentrations exceeds three standard deviations of the background value of the area, the incremental learning process is immediately interrupted and the emergency model optimization thread is started to prioritize the reconstruction of the target plant combination scheme in the abnormal diffusion area. The corrected scheme output by the emergency model optimization thread is updated to the spatial configuration map of the corresponding area in real time through an independent communication channel, and at the same time, the control command execution system in step S5 is triggered to perform rhizosphere microenvironment pre-regulation operation in the area.

[0024] As a preferred option, when the emergency model optimization thread reconstructs the target plant combination scheme, it temporarily adds rhizosphere microenvironment monitoring points in the abnormal pollutant diffusion area through the control command execution system. The newly added monitoring points are arranged in a concentric circle array, with the center point located at the location of the maximum abnormal diffusion concentration gradient, and the array radius covering the pollutant migration prediction boundary. The dynamic data of rhizosphere redox potential and the data of plant root exudate components collected by the newly added points are transmitted back to the emergency model optimization thread in real time to verify the feasibility of the plant rhizosphere interaction effect in the reconstruction scheme. After the verification is successful, the temporary monitoring points are immediately removed, and the collected data are incorporated into the abnormal working condition dataset of the plant function database for subsequent model training.

[0025] As a preferred option, the temporarily added rhizosphere microenvironment monitoring points are arranged using retrievable sensor modules. These modules include a multi-parameter soil sensor with a height-adjustable support and a drone hoisting interface. When the emergency model optimization thread issues an addition command, the control center automatically schedules idle fixed monitoring point sensor modules and hoists them to the area of ​​abnormal pollutant diffusion via a multi-rotor drone. The support adjusts the implantation depth according to the coordinates of the concentric circle array to position the sensor probe in the active rhizosphere layer of the target plant. After successful verification, the drone hoists and retrieves the sensor module to its original position for standby. During the support resetting process, the soil adhering to the probe surface is removed. The hoisting interface matches the electromagnetic lock of the drone's cargo compartment, and the reset coordinates are determined by the original burial position recorded in the spatial configuration map.

[0026] Preferably, the control command execution system in step S5 includes a micro-injection device carried by a multi-rotor drone. The micro-injection device includes a pH adjustment liquid storage tank, a bacterial agent storage tank, and a retractable injection needle. When the control command is triggered, the multi-rotor drone flies to the target rhizosphere region directly above it according to the three-dimensional coordinates recorded on the root growth trajectory map. It then injects the pH adjustment liquid or rhizosphere growth-promoting bacterial agent into the soil at a depth of 20 to 30 centimeters below the soil surface through the retractable injection needle. The horizontal distance error between the injection position and the center point of the dense root zone marked on the root growth trajectory map is controlled to be no more than 15 centimeters. The operation of adjusting the irrigation volume is performed through an independently laid drip irrigation network. The outlet position of the drip irrigation network is based on the planting density distribution set by the spatial configuration map.

[0027] As a preferred method, the specific method for dynamically scanning the three-dimensional coordinates of the plant root system in step S4 includes:

[0028] A ground-based mobile robot equipped with a penetrating radar cruises along a preset path. When a near-ground drone encounters a scanning blind spot due to terrain obstacles, the ground-based mobile robot automatically plans a path to enter the blind spot area. Within a 1-meter distance from the plant trunk, it collects root depth distribution data in a fan-shaped scanning mode, generating supplementary trajectory image segments. The penetrating radar operates at a frequency of 200MHz to 800MHz. The scanning data and the root growth trajectory map collected by the near-ground drone are merged using a spatial coordinate transformation algorithm to form a complete three-dimensional dynamic model of the root system in the collapsed area. The cruise path of the ground-based mobile robot is generated based on the coordinates of the plant planting sites recorded in the spatial configuration map.

[0029] Preferably, 72 hours after the completion of the regulation operation in step S5, the effect verification program is initiated. Chlorophyll fluorescence imaging data of vegetation in the target rhizosphere region is collected by a multispectral camera mounted on a near-ground UAV, while the stem flow rate data of the corresponding region in step S4 is retrieved. When the chlorophyll fluorescence parameter Fv / Fm value is lower than 0.7 and the stem flow rate does not recover to 120% of the pre-regulation level for 24 consecutive hours, it is determined that the regulation has failed and the target rhizosphere region is automatically marked as a high-risk site. The coordinates of the high-risk site are transmitted in real time to the machine learning model in step S2, triggering a secondary analysis of pollutant concentration and reconstruction of plant functional ratio for the site, generating a supplementary remediation plan to be embedded in the next round of spatial configuration map update cycle.

[0030] As a preferred embodiment, when constructing the plant function matching model in step S2, the plant root exudate interaction effect analysis module is set to perform the following operations:

[0031] Data on root exudate composition of all species in the target plant combination scheme were collected, and the concentration ratios of organic acids, phenolic acids, and amino acids in the exudates were determined by gas chromatography-mass spectrometry.

[0032] The rhizosphere chemical compatibility index of plant combinations was calculated based on exudate composition data. The rhizosphere chemical compatibility index was derived by weighting the promoting or inhibiting effects of each species' exudates on the heavy metal accumulation capacity of coexisting plants.

[0033] When the rhizosphere chemical compatibility index is lower than the preset index threshold, species with significant inhibitory effects are removed and companion plants that secrete mutually beneficial substances are added.

[0034] When the updated target plant combination scheme is output to the spatial configuration map, the planting distance between the companion plants that secrete mutually beneficial substances and the target enrichment plants is adjusted to be less than the sum of the average root radius of the two plants.

[0035] The present invention includes at least the following beneficial effects: (1) Through the closed-loop architecture of “full-domain environmental perception - intelligent model decision-making - zoned precise configuration - rhizosphere dynamic monitoring - targeted regulation and optimization”, the traditional static restoration mode is broken through, and the adaptability of plant configuration to the spatiotemporal heterogeneity of the collapsed area environment is significantly improved, and the success rate of vegetation planting is very high; (2) The incremental learning mechanism combined with multi-source growth index monitoring (chlorophyll index / canopy nitrogen content / biomass rate) realizes the dynamic iteration of plant combination scheme. Compared with the traditional scheme, which requires manual re-examination and adjustment, the response speed is accelerated, and the risk of restoration failure is suppressed at the bud stage; (3) The spectrum analysis method is used to warn of abnormal diffusion of pollutants. Combined with the mobile monitoring network and emergency model optimization thread, “hour-level” risk response is realized, which effectively solves the chain vegetation degradation problem caused by sudden pollution such as acid rain leaching of heavy metals from gangue piles; (4) The air-ground collaborative root scanning (drone + ground robot) provides sub-meter-level positioning data, and combined with the drone injection device, the rhizosphere microenvironment is regulated at the “centimeter level”. The utilization rate of the remediation agent was increased to near the theoretical limit, and soil disturbance was reduced to a negligible level; (5) the dual-indicator verification of the regulation effect (chlorophyll fluorescence / stem flow rate) and the prediction of the interaction between root exudates ensured the scientific nature of the remediation plan. The secondary remediation needs caused by interspecific inhibition or environmental mutations were reduced to an economically acceptable range. Attached Figure Description

[0036] Figure 1 This is a flowchart of one method of the present invention. Detailed Implementation

[0037] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0038] like Figure 1 As shown, the method for screening and configuring greening plants in coal mining subsidence areas provided by the present invention includes the following steps:

[0039] S1: Using a drone equipped with a hyperspectral imaging device and an X-ray fluorescence sensor, spatial distribution data of pollutants and soil physicochemical parameters of the entire subsidence area are obtained. The spatial distribution data of pollutants includes information on heavy metal types and concentration gradients, and the soil physicochemical parameters include pH value, organic matter content and soil moisture content.

[0040] S2: Construct a plant function matching model. Input the spatial distribution data of pollutants and soil physicochemical parameter data into a pre-trained machine learning model. The machine learning model outputs a target plant combination scheme based on a plant function database. The plant function database includes the pollution tolerance threshold, heavy metal enrichment coefficient, rhizosphere exudate type and ecological niche characteristics of candidate plants. The target plant combination scheme includes specific plant species and their functional ratios for different pollution zones.

[0041] S3: Divide the subsidence area into restoration units based on the topographic features, and generate a spatial configuration map in combination with the target plant combination scheme. The restoration unit includes the slope area, water accumulation area and coal gangue accumulation area. The spatial configuration map records the plant species combination, planting density and hierarchical structure in each restoration unit.

[0042] S4: Deploy an IoT monitoring network to periodically collect plant rhizosphere microenvironment data through buried soil multi-parameter sensors, and simultaneously control a near-ground UAV equipped with a micro rhizosphere probe to dynamically scan the three-dimensional coordinates of plant roots, generate a root growth trajectory map, and spatially correlate it with the rhizosphere microenvironment data.

[0043] S5: Establish a dynamic feedback mechanism. When the associated rhizosphere microenvironment data deviates from the preset threshold range, locate the target rhizosphere region based on the current root growth trajectory map, trigger the control command execution system, and perform at least one of the following operations: inject pH adjustment solution into the target rhizosphere region, apply more rhizosphere growth-promoting bacteria agent, and adjust the irrigation amount.

[0044] The key to successful greening plant selection and configuration in coal mining subsidence areas lies in establishing a highly intelligent, data-driven, dynamic closed-loop system to precisely address the complex and ever-changing environmental challenges of these areas. This method begins with a comprehensive environmental data acquisition phase. Utilizing the high mobility and wide coverage of unmanned aerial vehicle (UAV) platforms, hyperspectral imaging equipment is used to conduct continuous spectral scanning of the subsidence area surface. Hyperspectral imaging captures the reflectance characteristics of ground features across hundreds of narrow, continuous spectral bands. These characteristics, like "fingerprints," reflect vegetation health and stress responses caused by soil pollution. Simultaneously, the UAV carries an X-ray fluorescence (XRF) sensor. This sensor non-destructively and in situ identifies heavy metal elements (such as lead, cadmium, arsenic, chromium, and mercury) in the soil and measures their concentration gradients by emitting X-rays to excite elemental atoms and detecting the characteristic X-ray fluorescence energy they release. In addition to pollutant information, key soil physicochemical parameters are also acquired simultaneously, including pH values ​​(characterizing soil acidity / alkalinity), organic matter content (reflecting soil fertility), and soil moisture content (affecting plant water absorption). These data were spatially registered using a Geographic Information System (GIS) to ultimately create a high-resolution spatial distribution map of pollutants (including heavy metal types and concentration gradients) and a spatial dataset of soil physicochemical parameters covering the entire subsidence area. This data acquisition provides detailed spatial environmental baseline information for subsequent precise decision-making.

[0045] After obtaining the environmental baseline data of the subsidence area, the intelligent decision-making stage begins. The constructed plant function matching model then comes into play. The core of this model is a pre-trained machine learning algorithm (e.g., a model based on random forest, support vector machine, or neural network). This model is fed with the spatial distribution data of pollutants across the entire area and the soil physicochemical parameter data obtained in the preceding steps. The core of the model comes from a large plant function database, which can be a self-constructed database or obtained from existing public databases. This database not only contains the names of a large number of candidate plant species, but more importantly, it records their key functional characteristics, including: the tolerance limits of each plant to different pollutants (pollution tolerance thresholds), the ability to absorb and accumulate specific heavy metal elements (heavy metal enrichment coefficients), the types of chemical substances secreted by the roots into the soil (such as organic acids, phenols, amino acids, etc., which may affect the solubility and bioavailability of heavy metals), and the plant's role and adaptive characteristics in the ecosystem (niche characteristics, such as drought tolerance, flood tolerance, nitrogen fixation, soil stabilization, etc.). The machine learning model uses complex algorithms to analyze the input environmental data and the characteristic data in the plant function database to find the optimal match. It doesn't simply select a single plant; instead, it intelligently calculates a target plant combination scheme with specific functional ratios (e.g., one plant is responsible for enriching heavy metal A, another for tolerating high salinity, and a third for improving soil structure) based on the specific environmental stress combinations in different areas of the subsidence zone (e.g., high-pollution, moderate-pollution, low-pollution, different pH levels, and different water levels). This scheme is a customized plant community configuration recommendation for spatial zoning based on the environmental heterogeneity within the subsidence zone.

[0046] With the target plant combination scheme in place, the next step is to spatially implement it, which requires careful consideration of the complex topographic features of the subsidence area. Based on high-resolution topographic data (usually obtained by UAV lidar or photogrammetry), the entire subsidence area is scientifically divided into restoration units with similar environmental characteristics and restoration needs. Typical unit types include: slope areas at risk of soil erosion, waterlogged areas that may accumulate water or form wetlands, and coal gangue accumulation areas with poor physical and chemical properties formed by mining waste. For each defined restoration unit, the specific plant species recommended for that unit in the target plant combination scheme generated in the previous steps, as well as their functional ratios (such as the ratio of pioneer species to late-succession species, and the ratio of enriching plants to cover plants), are further refined into an operable spatial configuration map based on the specific conditions of that unit (such as slope, aspect, water depth, and gravel content). This blueprint serves as a guide for on-site planting. It details the specific locations within each restoration unit for planting different plant species (species combination), the number of plants per unit area (planting density, e.g., how many herbaceous plants per square meter or how many shrubs / trees per hectare), and the vertical spatial arrangement of the plants (hierarchical structure, such as the combination of tree, shrub, herbaceous, and ground cover layers). This ensures that the spatial distribution of plants not only meets the needs of ecological restoration but also conforms to the actual topographical conditions and landscape design requirements.

[0047] After the planting program is implemented, continuous monitoring and data feedback are crucial for achieving dynamic optimization. To this end, a dense Internet of Things (IoT) monitoring network is deployed in the subsidence area. This network periodically (e.g., hourly or daily) collects real-time data on the plant rhizosphere microenvironment by pre-burying multi-parameter sensor nodes in the soil (these nodes may be distributed across different remediation units and key locations). The rhizosphere is a micro-region where plant roots interact closely with the soil; the monitored data is critical, including but not limited to real-time pH, humidity, temperature, conductivity (salinity indicator), redox potential, and concentration changes of specific ions (such as heavy metal ions) in the rhizosphere soil. To gain a more comprehensive understanding of the growth dynamics of plant roots in three-dimensional space and their interaction with the environment, drones (typically multi-rotor drones capable of flexible flight at low altitudes and in complex terrain) are also simultaneously controlled near the ground. These drones are equipped with miniaturized rhizosphere probes (based on electrical impedance imaging, microelectrode arrays, or other sensing technologies) that can non-destructively scan the distribution and morphology of plant roots beneath the surface, acquiring three-dimensional spatial coordinate data of the roots. By using data fusion technology, point-like rhizosphere microenvironment data collected by buried sensors are spatially correlated with three-dimensional root coordinate data covering a larger spatial area obtained by UAV scanning (generating a root growth trajectory map). This not only reveals the soil condition at a specific point but also the root growth status in the vicinity of that point, establishing a spatial correspondence between changes in environmental parameters and root responses, thus forming a dynamic and visualized understanding of underground ecological processes.

[0048] Finally, a robust dynamic feedback and regulation mechanism is established. When the correlated rhizosphere microenvironment data (e.g., pH, specific heavy metal concentrations, and moisture) deviates from the preset safe or optimal threshold range (these thresholds are based on tolerance thresholds and environmental target settings in the plant function database), the response process is immediately initiated. First, based on the current root growth trajectory map, the target rhizosphere region exhibiting environmental anomalies is precisely located (specifically, to three-dimensional spatial coordinates). Then, the regulation command execution system is triggered to implement precise intervention measures in this specific target region. The intervention measures are selected based on the specific type of anomaly detected, mainly including: injecting pH-adjusting solutions (such as lime water to raise the pH of acidic soil, or sulfur powder / acidic solution to lower the pH of alkaline soil) into the target rhizosphere region to optimize the soil acid-base environment; increasing the application of rhizosphere growth-promoting bacteria (PGPR) agents (such as specific strains with the ability to solubilize phosphorus, fix nitrogen, secrete growth hormones, or chelate heavy metals) to promote plant growth or change the form of heavy metals through microbial activity; or adjusting the irrigation amount in this area (increasing or decreasing) to improve rhizosphere moisture conditions. These operations are not a broad-based approach to the entire area, but rather a targeted, precise adjustment of identified problem areas to maximize resource utilization efficiency and repair effectiveness. Following the adjustment, monitoring and feedback continue, forming a continuous optimization loop of "monitoring-analysis-decision-execution-re-monitoring".

[0049] Compared with existing technologies, this method represents a significant leap forward in ecological restoration technology for coal mining subsidence areas. Traditional methods often rely on limited manual sampling points for environmental assessment, with plant selection primarily based on experience and limited literature data. Configuration schemes are static and lack specificity, leading to delayed and inefficient post-restore maintenance and adjustments. This method, however, achieves high-precision, comprehensive, and real-time dynamic perception of the subsidence area environment (especially pollutants and key soil parameters) through the deep integration of UAVs, multi-source sensors, the Internet of Things, and machine learning technologies. The constructed intelligent matching model can deeply explore the complex relationship between plant functional characteristics and environmental stress, outputting highly customized and functionally complementary plant community combinations. Spatial configuration maps combined with topographic zoning ensure precise implementation of the scheme. Most importantly, by establishing a dense rhizosphere microenvironment monitoring network, three-dimensional dynamic root scanning technology, and a dynamic feedback control mechanism based on real-time data, this method completely changes the traditional "one-time planting, passive waiting" model, achieving closed-loop intelligent management of the restoration process that is "perceptible, analyzable, decision-making, executable, and optimizable." This significantly improves the success rate of plant establishment, growth vigor, and restoration efficiency (such as heavy metal accumulation efficiency and soil structure improvement speed) in harsh subsidence areas, greatly reducing the risk of restoration failure and subsequent maintenance costs. It provides strong technical support for the rapid, stable, and precise restoration of complex and damaged mining ecosystems. Its level of automation and intelligence, as well as its ability to dynamically and precisely regulate the rhizosphere microenvironment, are unmatched by existing technologies.

[0050] In another technical solution, the machine learning model in step S2 is trained using an incremental learning mechanism. After planting, it continuously receives rhizosphere microenvironment data collected in step S4 and plant growth status assessment indicators obtained by UAV hyperspectral imaging equipment. The plant growth status assessment indicators include chlorophyll content index, canopy nitrogen content, and biomass accumulation rate. When the actual growth index of any species in the target plant combination scheme is found to be lower than the preset growth threshold three times in a row, the model parameter update module is triggered. The weight allocation of pollution tolerance threshold and heavy metal enrichment coefficient is recalculated using the newly added data, and the corrected target plant combination scheme is output. The model parameter update module generates a version record after each run and uses the corrected target plant combination scheme to dynamically adjust the planting scheme of subsequent batches of plants through spatial configuration maps.

[0051] The machine learning model possesses continuous evolution capabilities. After the initial plant planting, it does not operate statically but dynamically optimizes through an incremental learning mechanism. This mechanism continuously receives two types of key data: first, rhizosphere microenvironment data (such as real-time heavy metal ion concentration, pH fluctuations, and microbial activity) periodically collected by embedded soil multi-parameter sensors; and second, plant growth status assessment indicators acquired by UAV multispectral cameras. These indicators are obtained by analyzing the vegetation reflectance spectral characteristics using multispectral imaging technology, specifically including chlorophyll content indices reflecting photosynthetic efficiency (such as the NDVI index), canopy nitrogen content characterizing nutrient status (analyzed using near-infrared bands), and biomass accumulation rate calculated based on 3D point cloud reconstruction (reflecting plant growth vitality). All data are bound to the coordinates of remediation units through spatiotemporal tags, forming a dynamic database covering the entire area.

[0052] When the system detects that the actual growth index of any species in the target plant combination falls below a preset threshold three times consecutively (e.g., the chlorophyll index of an enriched plant is consistently below the healthy baseline value), the model parameter update module will be automatically activated. The "three consecutive times" threshold here has fault-tolerant design significance: the first anomaly may be due to weather interference, the second indicates a risk, and the third confirms adaptation failure. When the update module runs, it first retrieves newly added rhizosphere microenvironment data and growth indicators, and recalculates the weight allocation of key parameters using a weighted algorithm (e.g., adjusting the priority of pollution tolerance threshold and heavy metal enrichment coefficient). For example, if a sudden migration of heavy metals in a region causes the original plant tolerance threshold to fail, the model will reduce the weight of that threshold and instead increase the assessment weight of root exudate regulation capacity. The updated target plant combination scheme not only corrects species selection but also adjusts functional ratios (e.g., increasing the proportion of associated nitrogen-fixing plants). All update records generate timestamped version records (e.g., Ver2.1.3) and are dynamically mapped to subsequent planting batches through spatial configuration maps (e.g., replacing failed species in the next month's replanting plan).

[0053] The incremental learning mechanism overcomes the limitations of traditional static vegetation configuration. Compared to conventional methods that passively remedy remediation after failure, this approach significantly improves the response speed to sudden environmental changes in the subsidence area by driving model iteration through real-time data. For example, when soil pollution spreads due to rainfall, traditional methods require months of manual investigation to detect plant wilting, while this approach can automatically identify and reconstruct the plan within 3-4 weeks (depending on the monitoring cycle), reducing the risk of remediation failure to a negligible level and minimizing resource waste caused by misjudgment.

[0054] The methods for obtaining the plant growth status assessment indicators specifically include:

[0055] The vegetation canopy reflectance spectrum is collected by the drone multispectral camera deployed in step S4, and the reflectance spectrum is converted into chlorophyll content index based on the normalized vegetation index calculation formula.

[0056] A convolutional neural network model was used to analyze the visible and near-infrared images of the canopy, outputting a pixel-level canopy nitrogen content distribution map and calculating the regional average value.

[0057] A three-dimensional model of the plant was reconstructed by combining lidar point cloud data, and the biomass accumulation rate was inverted based on the volume growth rate.

[0058] The method for setting the preset growth threshold for each plant is as follows: retrieve the historical growth data of the species in the non-polluted area from the plant function database, and take the average decrease of 30% in chlorophyll content index, 25% in canopy nitrogen content, and 40% in biomass accumulation rate as the dynamic judgment benchmark values.

[0059] The time interval between three consecutive monitoring cycles is automatically adjusted based on the plant growth stages recorded in the spatial configuration map, with a 7-day interval for fast-growing herbaceous plants and a 30-day interval for woody plants.

[0060] The chlorophyll content index was obtained using vegetation canopy reflectance spectral data collected by a UAV multispectral camera. Based on the Normalized Difference Vegetation Index (NDVI) calculation principle, the ratio of the difference in reflectance between the red light band (approximately 630-690 nm) and the near-infrared band (approximately 760-900 nm) to their sum was converted into an index value ranging from 0 to 1. A higher index indicates stronger chlorophyll activity. The quantification of canopy nitrogen content combined machine learning and image analysis: first, a convolutional neural network (CNN) model was used to segment leaf pixels in the visible light and near-infrared images of the canopy; then, a pixel-level distribution map was generated based on nitrogen characteristic spectra (such as the absorption valley at 550 nm); finally, the regional average value within the restoration unit was calculated. The biomass accumulation rate was measured using lidar point cloud data: by comparing the reconstructed 3D plant models (such as branch volume and leaf area) from adjacent scan cycles, a time-biomass curve was fitted and its slope (unit: kg / m²) was calculated. 3 / sky).

[0061] The preset thresholds are set using species-differentiated benchmarks. The optimal historical growth data for the species in a non-polluted area is retrieved from the plant function database, and key indicators are adjusted downwards by a specific percentage as dynamic judgment benchmarks. For example, a 25%-35% decrease in the historical average chlorophyll content index (typically 30%), a 20%-30% decrease in the average canopy nitrogen content (typically 25%), and a 35%-45% decrease in the average biomass accumulation rate (typically 40%) are set as the threshold red lines. Plant growth stage information (such as germination stage and rapid growth stage) recorded in the spatial configuration map triggers automatic adjustment of the monitoring interval. For fast-growing herbaceous plants with short growth cycles (such as ryegrass), a dense monitoring strategy is adopted (interval of 5-10 days, typical value 7 days); while for woody plants (such as black locust), due to their slow growth, the interval is extended to 25-35 days (typical value 30 days), avoiding data redundancy while ensuring timely early warning.

[0062] By integrating multi-source sensing technologies and setting intelligent thresholds, the accuracy problem of plant health assessment in mining areas has been solved. Traditional methods, relying on manual visual inspection or single indicators (such as plant height), are easily affected by subjectivity. This solution, however, integrates multi-dimensional data such as spectra, images, and 3D point clouds, significantly enhancing the correlation between assessment results and actual physiological states. For example, when heavy metal stress leads to metabolic disorders in plants, the chlorophyll index may issue an early warning weeks before visible symptoms, providing a critical window for intervention. Dynamic periodic adjustments optimize the allocation of monitoring resources, reducing remediation costs to an economically feasible level.

[0063] In step S2, the model parameter update module performs pollutant migration early warning analysis simultaneously during operation. Based on the spatial distribution data of pollutants obtained in step S1 and the rate of change of specific heavy metal ion concentrations collected in real time in step S4, the abnormal diffusion area of ​​pollutants is identified by spectrum analysis. When the rate of change of specific heavy metal ion concentrations exceeds three standard deviations of the background value of the area, the incremental learning process is immediately interrupted and the emergency model optimization thread is started to prioritize the reconstruction of the target plant combination scheme in the abnormal diffusion area. The corrected scheme output by the emergency model optimization thread is updated to the spatial configuration map of the corresponding area in real time through an independent communication channel, and at the same time, the control command execution system in step S5 is triggered to carry out rhizosphere microenvironment pre-regulation operation in the area.

[0064] The pollutant migration early warning system is deeply integrated into the optimization process of the machine learning model. When the model parameter update module in step S2 runs, the dynamic monitoring of pollutants using spectral analysis is activated simultaneously. This method, based on the initial spatial distribution data of pollutants obtained in step S1 (such as the spatial gradient of heavy metal concentrations) and the real-time collection of specific heavy metal ion concentration change rates (such as the hourly migration rate of cadmium ions) in step S4, decomposes the spatiotemporal data into different frequency components using Fourier transform or wavelet transform to identify abnormal diffusion patterns (e.g., a sudden increase in high-frequency components may indicate a pollution leak). The system presets three standard deviations of the regional background value as a safety boundary (e.g., if the background change rate of lead ions in a certain area is ±0.05 mg / L·h, the three standard deviation threshold is ±0.15 mg / L·h), and this threshold can be dynamically calibrated based on historical data.

[0065] Once a pollutant concentration change rate is detected to exceed a threshold (e.g., the arsenic ion change rate at a monitoring point reaches 0.25 mg / L·h, exceeding the background value of 0.08 mg / L·h by three times), a dual response is immediately executed: first, the regular incremental learning process is interrupted to avoid erroneous data contaminating the model; then, an emergency model optimization thread is initiated. This thread prioritizes abnormal diffusion areas (e.g., a fan-shaped area with a radius of 50 meters centered on the pollution source), and calls high-priority computing resources to reconstruct the target plant combination scheme. Reconstruction strategies include: replacing the original species with insufficient tolerance (e.g., replacing *Acer clover* with *Pteris vittata* with a low enrichment coefficient), and increasing the proportion of plants with strong root soil-fixing capabilities (e.g., increasing the proportion of *Vegetariana* from 15% to 30%). The reconstructed scheme is pushed in real time to the spatial configuration map of the corresponding area through an independent communication channel (e.g., a 5G private network), and the control system in step S5 is linked to implement rhizosphere microenvironment pre-regulation (e.g., pre-injection of chelating agents to reduce heavy metal activity).

[0066] This approach upgrades passive pollution monitoring to proactive risk prevention. Traditional methods only take action after pollutants cause vegetation death, while this approach's early warning mechanism can trigger remediation and pre-regulation in the early stages of pollutant migration (such as the initial stage of groundwater infiltration), giving the remediation system a "immune response" capability. For example, when heavy metals are leached from a coal gangue pile due to acid rain, the entire process from early warning to rhizosphere intervention can be completed within hours, suppressing the damage to plants from pollution spread in its early stages.

[0067] When the emergency model optimization thread reconstructs the target plant combination scheme, it temporarily adds rhizosphere microenvironment monitoring points in the abnormal pollutant diffusion area through the control command execution system. The new monitoring points are arranged in a concentric circle array, with the center point located at the maximum abnormal diffusion concentration gradient, and the array radius covering the pollutant migration prediction boundary. The dynamic data of rhizosphere redox potential and the data of plant root exudate components collected by the new points are transmitted back to the emergency model optimization thread in real time to verify the feasibility of the plant rhizosphere interaction effect in the reconstruction scheme. After the verification is successful, the temporary monitoring points are immediately removed, and the collected data are incorporated into the abnormal working condition dataset of the plant function database for subsequent model training.

[0068] A dynamically adjustable temporary monitoring network was designed for areas with abnormal pollutant diffusion. When the emergency model optimization thread runs, the S5 control system, in conjunction with the synchronous instruction step, adds retrievable sensor modules to the target area. These modules are arranged in a concentric array: the center point is located at the location of the maximum pollutant concentration gradient (e.g., the peak arsenic concentration determined by an interpolation algorithm), and the array radius covers the predicted boundary of pollutant migration (typically the diffusion rate multiplied by 72 hours, with a typical radius of 15-50 meters). The array density is adjustable (e.g., one point per 10 square meters) to ensure the integrity of gradient data acquisition. Each module contains two types of sensors: a dynamic potential sensor (accuracy ±10mV) for monitoring soil redox status, and a microdialysis sampler (capturing molecules such as organic acids and phenols) for analyzing root exudate components.

[0069] Dynamic data on rhizosphere redox potential (reflecting the trend of heavy metal valence state transformation) and root exudate composition data (such as the correlation between citric acid concentration and heavy metal solubility) collected from newly added monitoring sites are transmitted back to the emergency model optimization thread in real time. The thread uses machine learning models (such as random forest regressors) to verify the feasibility of plant rhizosphere interaction effects in the reconstruction scheme: for example, if data shows that the oxalic acid concentration secreted by associated plants in the new combination is insufficient to activate cadmium in the soil by the target enriching plant (such as *Sedum aizoon*), then species with stronger oxalic acid secretion capacity (such as *Solanum nigrum*) are automatically added. After successful verification (usually taking 2-4 hours), the temporary monitoring points are immediately deactivated, and their data are automatically categorized into the abnormal operating condition dataset in the plant function database to enhance the future model's predictive ability for extreme scenarios.

[0070] By employing an "on-demand deployment - precise verification - rapid evacuation" model, the inflexibility of traditional fixed monitoring networks is addressed. Compared to deploying high-density sensors across the entire subsidence area (which is costly), this solution only temporarily activates monitoring resources in high-risk areas, reducing the monitoring cost per unit area to a commercially viable level. Simultaneously, real-time feedback of secretion composition data ensures the scientific validity of the reconstruction plan, avoiding secondary repair failures due to theoretical speculation.

[0071] The temporarily added rhizosphere microenvironment monitoring points are deployed using retrievable sensor modules. These modules include multi-parameter soil sensors with height-adjustable supports and drone hoisting interfaces. When the emergency model optimization thread issues an addition command, the control center automatically dispatches idle fixed monitoring point sensor modules, which are then hoisted to the area of ​​abnormal pollutant diffusion via a multi-rotor drone. The support adjusts the implantation depth according to the coordinates of the concentric circle array to position the sensor probe in the active rhizosphere layer of the target plant. After successful verification, the drone hoists and retrieves the sensor module to its original position for standby. During the support resetting process, soil adhering to the probe surface is removed. The hoisting interface matches the electromagnetic lock of the drone's cargo compartment, and the reset coordinates are determined by the original burial location recorded in the spatial configuration map.

[0072] The main body of the module is a cylindrical shell (approximately 10cm in diameter) with a height-adjustable support. The support can be precisely positioned within a depth range of 15-50cm using an electric push rod (to align the probe with the rhizosphere active layer). The top integrates a drone mounting interface (such as an ISO standard electromagnetic locking slot) that matches the electromagnetic lock of a multi-rotor drone's cargo bay (typical adsorption force ≥200N). The interior of the shell integrates a multi-parameter soil sensor (pH / conductivity / temperature triple probe) and a micro-pump for exudate sampling, with the total weight controlled to within 1.5kg to accommodate drone payloads.

[0073] Upon receiving an addition command, the control center automatically dispatches idle fixed monitoring point modules (such as sensors that are not used during the non-rainy season in slope areas). A multi-rotor drone (with a payload of ≥2kg) flies to the module storage rack, locates the locking position through visual recognition, and lifts the module. After arriving at the target area, it automatically adjusts its hovering height based on GIS coordinates (e.g., 32.5°N, 115.8°E) and terrain data, releasing the module to the ground. The support is then inserted into the soil at a preset depth (e.g., 20cm in herbaceous areas, 40cm in woody areas). During retrieval, the drone flies above the module, the electromagnetic lock is energized and adsorbs it, and during the support retrieval process, a rotating brush automatically removes soil adhering to the probe (rotation speed 200-500rpm), finally returning the module to its original storage rack coordinates (positioning error <5cm).

[0074] This solution enables "drone-based" full lifecycle management of monitoring equipment. Traditional manual sensor deployment requires 2-3 people per day to complete 20 locations, while this solution, through drone swarms, can complete the same task within one hour, reducing labor costs to a negligible level. The automatic soil removal function extends sensor lifespan (preventing salt crystallization corrosion of the probe) and increases module reuse rate to over 95%. This combined operation mode of "flying robots + intelligent modules" provides cost-effective dynamic monitoring support for subsidence area restoration projects.

[0075] In another technical solution, the control command execution system in step S5 includes a micro-injection device carried by a multi-rotor drone. The micro-injection device includes a pH adjustment liquid storage tank, a bacterial agent storage tank, and a retractable injection needle. When the control command is triggered, the multi-rotor drone flies to the target rhizosphere area directly above it according to the three-dimensional coordinates recorded on the root growth trajectory map. It injects the pH adjustment liquid or rhizosphere growth-promoting bacterial agent into the soil at a depth of 20 to 30 centimeters below the soil surface through the retractable injection needle. The horizontal distance error between the injection position and the center point of the dense root zone marked on the root growth trajectory map is controlled to be no more than 15 centimeters. The operation of adjusting the irrigation volume is performed through an independently laid drip irrigation network. The outlet position of the drip irrigation network is based on the planting density distribution set by the spatial configuration map.

[0076] A physical execution system for precise rhizosphere intervention is constructed using a multi-rotor drone as a carrier, equipped with a specially designed micro-injection device. This device employs a modular design: the main body comprises three independent storage tanks (typically 0.5-2 liters in volume), each containing a pH adjustment solution (such as citric acid solution for alkaline areas or calcium hydroxide suspension for acidic areas), rhizosphere growth promoter (PGPR) inoculum (such as active bacterial suspensions of *Azologia azotocinus* or *Pseudomonas fluorescens*), and sterile water for rinsing. The core injection component is a retractable injection needle (30-50 cm in length, 5-8 mm in diameter), with a soil-penetrating coating on the needle tip, and a micro-hydraulic pump (pressure output 0.2-0.5 MPa) connected to the end. The entire device connects to a gimbal on the drone's underside via a quick-release interface, the gimbal possessing ±15° sway compensation capability to counteract hovering sway.

[0077] When the control command is triggered, the UAV automatically plans its flight path based on the three-dimensional coordinates (e.g., spatial point X=125.3m, Y=87.6m, Z=-0.25m) recorded on the root growth trajectory map. After flying directly above the target rhizosphere region, the UAV confirms its altitude above the ground (typical hovering height 2-3 meters) using a laser rangefinder, and then performs the following actions: First, the retractable syringe, driven by a servo motor, descends vertically, penetrating the soil surface layer with a constant pressure of 30-50N; second, the syringe stops at a depth of 20-30 cm (corresponding to the rhizosphere active layer of most herbs and shrubs), and a micro hydraulic pump injects the substance at a preset dosage (e.g., 50-200ml of pH adjustment solution per injection), with slight reciprocating pumping during the injection process to promote diffusion; finally, the syringe retracts and the sterile water flushing channel is activated. The key positioning requirement is: the horizontal distance error between the injection point and the center point of the dense root area marked on the root trajectory map is ≤15 cm (achieved by superimposing the UAV RTK positioning accuracy of ±2 cm and the root coordinate error of ±10 cm). For irrigation volume adjustment, it is carried out through an independently laid drip irrigation network. The outlets are distributed according to the planting density of the spatial configuration map (e.g., 4 drippers per square meter in high-density areas), and the flow rate is adjusted by a remotely controlled solenoid valve (adjustable from 0.5-5L / h).

[0078] Compared to traditional manual spraying or trenching (where the depth error often exceeds 50 cm), drone injection delivers substances directly to the core active layer of the root system, maximizing the utilization rate of the conditioning solution / bacterial agent to near its theoretical limit. For example, in the remediation of a lead-contaminated slope, traditional methods require applying 3 tons of lime powder to cover the entire slope, while this solution only requires injecting 200 liters of targeted conditioning solution into the high-risk rhizosphere zone (0.5%) to achieve the same pH neutralization effect, significantly reducing agent costs and soil disturbance.

[0079] In another technical solution, the specific method for dynamically scanning the three-dimensional coordinates of the plant root system in step S4 includes:

[0080] A ground-based mobile robot equipped with a penetrating radar cruises along a preset path. When a near-ground drone encounters a scanning blind spot due to terrain obstacles, the ground-based mobile robot automatically plans a path to enter the blind spot area. Within a 1-meter distance from the plant trunk, it collects root depth distribution data in a fan-shaped scanning mode, generating supplementary trajectory image segments. The penetrating radar operates at a frequency of 200MHz to 800MHz. The scanning data and the root growth trajectory map collected by the near-ground drone are merged using a spatial coordinate transformation algorithm to form a complete three-dimensional dynamic model of the root system in the collapsed area. The cruise path of the ground-based mobile robot is generated based on the coordinates of the plant planting sites recorded in the spatial configuration map.

[0081] To address the issue of blind spots in root system monitoring under complex terrain, an air-ground collaborative scanning solution is adopted. When a near-ground drone (referring to a rotary-wing drone with a flight altitude of <10 meters) cannot acquire complete data due to terrain obstacles such as steep slopes or dense shrubs, a ground-based mobile robot platform is automatically activated. This robot uses a four-wheel independent suspension chassis (ground clearance ≥20cm) and is equipped with penetrating radar (such as GPR ground-penetrating radar). The radar's operating frequency is adjustable within the range of 200-800MHz (low frequency 200-400MHz for deep root detection, high frequency 600-800MHz for shallow, fine imaging). The robot operates based on the plant planting coordinates recorded in the spatial configuration map (e.g., locust planting point P). xy =33.5,44.2), generating a preset cruising path around the target plant (typically a concentric circle with a radius of 1-3 meters).

[0082] After the robot enters the blind spot of the drone scan (such as behind a coal gangue pile), it initiates a fan-shaped scanning mode within a range of 0.8-1.2 meters from the plant trunk: the radar antenna rotates at a step angle of 10-15°, collecting a set of data every 60° (a single scan covers a 120° fan area). The scanning depth is set to 0-1.5 meters, focusing on collecting root depth distribution data (such as taproot depth and fibrous root density). The collected raw radar signals are processed by temporal domain filtering and then stitched together with the root growth trajectory map obtained by the drone using a spatial coordinate transformation algorithm (such as ICP point cloud registration): first, the local coordinate system established by the robot's laser SLAM (origin at the base of the plant trunk) is transformed to the global GIS coordinate system; second, feature point matching (such as special root bifurcation points) is used to align the data; finally, a complete three-dimensional dynamic model of the root system in the collapsed area is generated (resolution up to 5cm voxels).

[0083] This solution addresses the challenge of full-space root system monitoring in complex subsidence areas. Traditional single-UAV platforms often suffer from data loss rates exceeding 40% on steep slopes or in areas with dense obstacles, while the air-ground collaborative mechanism improves blind spot coverage to nearly 100%. For example, in the restoration of a coal gangue hill, a ground robot successfully scanned the 1.2-meter-deep taproot system of Artemisia annua, which was missed by the UAV. Based on this, adjustments to the irrigation plan significantly improved plant survival rates. The completeness of the three-dimensional dynamic model provides a reliable spatial benchmark for rhizosphere regulation.

[0084] In another technical solution, 72 hours after the completion of the regulation operation in step S5, the effect verification program is launched. Chlorophyll fluorescence imaging data of vegetation in the target rhizosphere region is collected by a multispectral camera carried by a near-ground UAV, and the stem flow rate data of the corresponding region in step S4 is retrieved at the same time. When the chlorophyll fluorescence parameter Fv / Fm value is lower than 0.7 and the stem flow rate does not rise to 120% of the pre-regulation level for 24 consecutive hours, it is determined that the regulation has failed and the target rhizosphere region is automatically marked as a high-risk site. The coordinates of the high-risk site are transmitted in real time to the machine learning model in step S2, triggering secondary analysis of pollutant concentration and reconstruction of plant functional ratio for the site, generating a supplementary remediation plan to be embedded in the next round of spatial configuration map update cycle.

[0085] After step S5 performs the regulatory operation (such as injecting pH adjustment solution or bacterial agent), the process does not end immediately. Instead, it automatically starts the effect verification program after a preset delay period (typically 72±12 hours, used to wait for plant physiological response). This program performs a comprehensive evaluation through dual-channel data acquisition: first, it controls a near-ground UAV equipped with a multispectral camera to fly over the target rhizosphere region and collect chlorophyll fluorescence imaging data of the vegetation (especially the maximum photochemical efficiency Fv / Fm value of photosynthetic system II, which is obtained through excitation light-induced fluorescence kinetics measurement); second, it retrieves the plant stem flow rate sensor data (based on the principle of heat dissipation, unit: g / h) buried in the same area in step S4. These two types of data reflect the photosynthetic function recovery status and water transport recovery efficiency, respectively.

[0086] A strict failure judgment logic is established: when the chlorophyll fluorescence parameter Fv / Fm value remains below 0.65-0.75 (typical threshold 0.7, indicating severely impaired light energy conversion efficiency) and the stem flow rate fails to recover to 115%-125% of the pre-regulation level (typical threshold 120%, indicating no improvement in water transport) for 24 consecutive hours, regulation is deemed to have failed. At this point, the target rhizosphere region is automatically marked as a high-risk site (displayed as a red warning icon in the spatial configuration map). Marking triggers a three-level response: first, the site coordinates are transmitted in real-time to the machine learning model in step S2; second, the model initiates secondary analysis of pollutant concentrations (e.g., re-detecting the distribution of heavy metal speciation at the site); finally, the plant functional composition is reconstructed based on the new data (e.g., replacing the original enriching plants with hyperaccumulating species). The generated supplementary remediation plan will be embedded in the next round of spatial configuration map updates (usually executed within 7 days).

[0087] This solution upgrades the traditional "open-loop" remediation method to an intelligent closed-loop system of "self-diagnosis and self-optimization." Compared to manual acceptance, which requires observing plant appearance changes for several months, this solution can scientifically determine the effect 3-4 days after intervention, avoiding ineffective intervention and continuous waste of resources. For example, when the inoculant is deactivated due to low soil temperature, traditional methods may misjudge it as insufficient dosage and repeat the application. This solution, however, uses stem flow data to pinpoint the root cause of failure and initiates ratio reconstruction, significantly improving the success rate of secondary remediation.

[0088] In another technical solution, when constructing the plant function matching model in step S2, the plant root exudate interaction effect analysis module is set to perform the following operations:

[0089] Data on root exudate composition of all species in the target plant combination scheme were collected, and the concentration ratios of organic acids, phenolic acids, and amino acids in the exudates were determined by gas chromatography-mass spectrometry.

[0090] The rhizosphere chemical compatibility index of plant combinations was calculated based on exudate composition data. The rhizosphere chemical compatibility index was derived by weighting the promoting or inhibiting effects of each species' exudates on the heavy metal accumulation capacity of coexisting plants.

[0091] When the rhizosphere chemical compatibility index is lower than the preset index threshold, species with significant inhibitory effects are removed and companion plants that secrete mutually beneficial substances are added.

[0092] When the updated target plant combination scheme is output to the spatial configuration map, the planting distance between the companion plants that secrete mutually beneficial substances and the target enrichment plants is adjusted to be less than the sum of the average root radius of the two plants.

[0093] A chemical-ecological interaction optimization layer was added to the plant matching model. Before outputting the target plant combination scheme, root exudate composition data of all candidate species were first collected: exudates were collected by in-situ root bag method or hydroponics, and the concentration ratios (unit: μg / g·h) of three key substances (organic acids such as citric acid and oxalic acid), phenolic acids such as vanillic acid and ferulic acid, and amino acids such as glutamic acid and proline) were quantitatively analyzed using gas chromatography-mass spectrometry (GC-MS). The concentration data of each substance was converted into a molar concentration matrix and input into the analysis module.

[0094] Based on exudate data, the rhizosphere chemical compatibility index (RCCI) is calculated. This index quantifies the comprehensive impact of exudates from each species on the heavy metal accumulation capacity of coexisting plants using machine learning models (such as support vector machines). For example, citric acid promotes cadmium dissolution and enhances accumulation efficiency, scoring +0.7, while certain phenolic acids inhibit zinc absorption, scoring -0.4. When the RCCI is lower than a preset threshold (usually 0.6-0.8, with a typical value of 0.75), species optimization is automatically performed: first, species with significant inhibitory effects are removed (such as ryegrass that secretes cinnamic acid), and then companion plants that secrete mutually beneficial substances are screened from the database (such as black nightshade that secretes oxalic acid and promotes lead activation). When updating the scheme and outputting it to the spatial configuration map, the planting distance between mutually beneficial companion plants and target enrichment plants is forcibly constrained (such as alfalfa and sedum 'Southeast' spacing ≤ 40cm), which is less than the average root radius of both (typical herbaceous root radius 50-60cm), ensuring effective interaction of exudates in the rhizosphere.

[0095] This approach addresses the critical deficiency of traditional plant configurations that neglect allelopathic effects. For example, in one subsidence area, remediation failed because the salicylic acid secreted by interplanted willows inhibited the cadmium accumulation capacity of associated castor beans. This new scheme, by predicting secretion interactions, upgrades the plant community from "physical coexistence" to "chemical synergy," increasing heavy metal accumulation efficiency to near the optimal level for a single species, while reducing replanting costs caused by interspecific inhibition.

[0096] It should be noted that although the steps are described in a specific order above, this does not mean that they must be performed in that order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required functionality is achieved. The number of devices and processing scale described herein are for simplification of the invention; applications, modifications, and variations of this invention will be readily apparent to those skilled in the art.

[0097] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. They can be applied to various fields suitable for the present invention. For those skilled in the art, other modifications can be easily made. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and illustrations shown and described herein.

Claims

1. A method for selecting and configuring greening plants in coal mining subsidence areas, characterized in that, Includes the following steps: S1: Using a drone equipped with a hyperspectral imaging device and an X-ray fluorescence sensor, spatial distribution data of pollutants and soil physicochemical parameters of the entire subsidence area are obtained. The spatial distribution data of pollutants includes information on heavy metal types and concentration gradients, and the soil physicochemical parameters include pH value, organic matter content and soil moisture content. S2: Construct a plant function matching model. Input the spatial distribution data of pollutants and soil physicochemical parameter data into a pre-trained machine learning model. The machine learning model outputs a target plant combination scheme based on a plant function database. The plant function database includes the pollution tolerance threshold, heavy metal enrichment coefficient, rhizosphere exudate type and ecological niche characteristics of candidate plants. The target plant combination scheme includes specific plant species and their functional ratios for different pollution zones. S3: Divide the subsidence area into restoration units based on the topographic features, and generate a spatial configuration map in combination with the target plant combination scheme. The restoration unit includes the slope area, water accumulation area and coal gangue accumulation area. The spatial configuration map records the plant species combination, planting density and hierarchical structure in each restoration unit. S4: Deploy an IoT monitoring network to periodically collect plant rhizosphere microenvironment data through buried soil multi-parameter sensors, and simultaneously control a near-ground UAV equipped with a micro rhizosphere probe to dynamically scan the three-dimensional coordinates of plant roots, generate a root growth trajectory map, and spatially correlate it with the rhizosphere microenvironment data. S5: Establish a dynamic feedback mechanism. When the associated rhizosphere microenvironment data deviates from the preset threshold range, locate the target rhizosphere region based on the current root growth trajectory map, trigger the control command execution system, and perform at least one of the following operations: inject pH adjustment solution into the target rhizosphere region, apply more rhizosphere growth-promoting bacteria agent, and adjust the irrigation amount. In step S2, the machine learning model is trained using an incremental learning mechanism. After the plants are planted, it continuously receives rhizosphere microenvironment data collected in step S4 and plant growth status assessment indicators obtained by the UAV hyperspectral imaging device. The plant growth status assessment indicators include chlorophyll content index, canopy nitrogen content, and biomass accumulation rate. When the actual growth index of any species in the target plant combination scheme is found to be lower than the preset growth threshold for three consecutive times, the model parameter update module is triggered. The weight allocation of pollution tolerance threshold and heavy metal enrichment coefficient is recalculated using the newly added data, and the corrected target plant combination scheme is output. The model parameter update module generates a version record after each run, and uses the corrected target plant combination scheme to dynamically adjust the plant planting scheme of subsequent batches through spatial configuration maps.

2. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, The methods for obtaining the plant growth status assessment indicators specifically include: The vegetation canopy reflectance spectrum is collected by the drone multispectral camera deployed in step S4, and the reflectance spectrum is converted into chlorophyll content index based on the normalized vegetation index calculation formula. A convolutional neural network model was used to analyze the visible and near-infrared images of the canopy, outputting a pixel-level canopy nitrogen content distribution map and calculating the regional average value. A three-dimensional model of the plant was reconstructed by combining lidar point cloud data, and the biomass accumulation rate was inverted based on the volume growth rate. The method for setting the preset growth threshold for each plant is as follows: retrieve the historical growth data of the species in the non-polluted area from the plant function database, and take the average decrease of 30% in chlorophyll content index, 25% in canopy nitrogen content, and 40% in biomass accumulation rate as the dynamic judgment benchmark values. The time interval between three consecutive monitoring cycles is automatically adjusted based on the plant growth stages recorded in the spatial configuration map, with a 7-day interval for fast-growing herbaceous plants and a 30-day interval for woody plants.

3. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, In step S2, the model parameter update module performs pollutant migration early warning analysis simultaneously during operation. Based on the spatial distribution data of pollutants obtained in step S1 and the rate of change of specific heavy metal ion concentrations collected in real time in step S4, the abnormal diffusion area of ​​pollutants is identified by the spectrum analysis method. When the rate of change of a specific heavy metal ion concentration exceeds three standard deviations of the regional background value, the incremental learning process is immediately interrupted and the emergency model optimization thread is started. The target plant combination scheme in the abnormal diffusion area is reconstructed first. The corrected scheme output by the emergency model optimization thread is updated to the spatial configuration map of the corresponding area in real time through an independent communication channel. At the same time, the step S5 regulation command execution system is triggered to carry out rhizosphere microenvironment pre-regulation operation in the area.

4. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 3, characterized in that, When the emergency model optimization thread reconstructs the target plant combination scheme, it temporarily adds rhizosphere microenvironment monitoring points in the abnormal pollutant diffusion area through the control command execution system. The new monitoring points are arranged in a concentric circle array, with the center point located at the maximum abnormal diffusion concentration gradient, and the array radius covering the pollutant migration prediction boundary. The dynamic data of rhizosphere redox potential and the data of plant root exudate components collected by the new points are transmitted back to the emergency model optimization thread in real time to verify the feasibility of the plant rhizosphere interaction effect in the reconstruction scheme. After the verification is successful, the temporary monitoring points are immediately removed, and the collected data are incorporated into the abnormal working condition dataset of the plant function database for subsequent model training.

5. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 4, characterized in that, The temporarily added rhizosphere microenvironment monitoring points are deployed using retrievable sensor modules. These modules include a multi-parameter soil sensor with a height-adjustable support and a drone hoisting interface. When the emergency model optimization thread issues an addition command, the control center automatically dispatches idle fixed monitoring point sensor modules and hoists them to the area of ​​abnormal pollutant diffusion via a multi-rotor drone. The support adjusts the implantation depth according to the coordinates of the concentric circle array, positioning the sensor probe in the active rhizosphere layer of the target plant. After successful verification, the drone hoists and retrieves the sensor module to its original position for standby, and the soil adhering to the probe surface is removed during the support resetting process. The hoisting interface is matched with the electromagnetic latch of the UAV cargo compartment, and the reset coordinates are determined by the original buried position recorded in the spatial configuration map.

6. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, In step S5, the control command execution system includes a micro-injection device carried by a multi-rotor drone. This micro-injection device includes a pH adjustment liquid storage tank, a bacterial agent storage tank, and a retractable injection needle. When the control command is triggered, the multi-rotor drone flies to the target rhizosphere region directly above it according to the three-dimensional coordinates recorded on the root growth trajectory map. It then injects the pH adjustment liquid or rhizosphere growth-promoting bacterial agent into the soil at a depth of 20 to 30 centimeters below the soil surface through the retractable injection needle. The horizontal distance error between the injection position and the center point of the dense root zone marked on the root growth trajectory map is controlled to be no more than 15 centimeters. The operation of adjusting the irrigation volume is performed through an independently laid drip irrigation network. The outlet position of the drip irrigation network is based on the planting density distribution set by the spatial configuration map.

7. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, The specific method for dynamically scanning the three-dimensional coordinates of plant roots in step S4 includes: A ground-based mobile robot equipped with a penetrating radar cruises along a preset path. When a near-ground drone encounters a scanning blind spot due to terrain obstacles, the ground-based mobile robot automatically plans a path to enter the blind spot area. Within a 1-meter distance from the plant trunk, it collects root depth distribution data in a fan-shaped scanning mode, generating supplementary trajectory image segments. The penetrating radar operates at a frequency of 200MHz to 800MHz. The scanning data and the root growth trajectory map collected by the near-ground drone are merged using a spatial coordinate transformation algorithm to form a complete three-dimensional dynamic model of the root system in the collapsed area. The cruise path of the ground-based mobile robot is generated based on the coordinates of the plant planting sites recorded in the spatial configuration map.

8. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, 72 hours after the completion of the regulation operation in step S5, the effect verification program is launched. Chlorophyll fluorescence imaging data of vegetation in the target rhizosphere region is collected by a multispectral camera carried by a near-ground UAV, and the stem flow rate data of the corresponding region in step S4 is retrieved at the same time. When the chlorophyll fluorescence parameter Fv / Fm value is lower than 0.7 and the stem flow rate does not rise back to 120% of the pre-regulation level for 24 consecutive hours, it is determined that the regulation has failed and the target rhizosphere region is automatically marked as a high-risk site. The coordinates of high-risk sites are transmitted in real time to the machine learning model in step S2, triggering a secondary analysis of pollutant concentration and reconstruction of plant functional ratios for that site, generating a supplementary remediation plan to be embedded in the next round of spatial configuration map update cycle.

9. The method for screening and configuring greening plants in coal mining subsidence areas according to claim 1, characterized in that, When constructing the plant function matching model in step S2, the plant root exudate interaction effect analysis module is set to perform the following operations: Data on root exudate composition of all species in the target plant combination scheme were collected, and the concentration ratios of organic acids, phenolic acids, and amino acids in the exudates were determined by gas chromatography-mass spectrometry. The rhizosphere chemical compatibility index of plant combinations was calculated based on exudate composition data. The rhizosphere chemical compatibility index was derived by weighting the promoting or inhibiting effects of each species' exudates on the heavy metal accumulation capacity of coexisting plants. When the rhizosphere chemical compatibility index is lower than the preset index threshold, species with significant inhibitory effects are removed and companion plants that secrete mutually beneficial substances are added. When the updated target plant combination scheme is output to the spatial configuration map, the planting distance between the companion plants that secrete mutually beneficial substances and the target enrichment plants is adjusted to be less than the sum of the average root radius of the two plants.

Images