A deep learning-based natural resource asset multi-factor intelligent valuation method and system

Abstract

The present application relates to natural resource asset evaluation and artificial intelligence technical field, provide a kind of based on deep learning's natural resource asset multi-element intelligent valuation method and system, comprising: S1: multi-source heterogeneous data acquisition and fusion, construct unified data base;S2: based on unified data base, construct multi-level feature engineering system, including basic feature, interactive feature and depth feature;S3: construct multi-task deep learning model;S4: simultaneously execute the single-element value evaluation of different natural resource assets and the combined total value prediction of multi-element combination assets;S5: based on attention mechanism, calculate the synergistic weight between different elements in the multi-element combination assets, and quantify combination premium;S6: output valuation result.The present application can preferably intelligent valuation.

Description

Technical Field

[0001] This invention relates to the fields of natural resource asset valuation and artificial intelligence technology, specifically to a multi-factor intelligent valuation method and system for natural resource assets based on deep learning. Background Technology

[0002] Under the reform of the overall supply (combined supply) of natural resource assets, it is necessary to integrate various natural resource elements such as land, minerals, forests, and water resources into a unified "asset package" and allocate them as a whole through market-oriented means. This institutional innovation urgently requires a new assessment method and technical system that can simultaneously handle multiple asset elements, accurately quantify the synergistic value between elements, and realize the monetization of ecological value.

[0003] Currently, natural resource asset valuation mainly employs the following three traditional methods: 1) Cost approximation method; The value of resource assets is determined based on the sum of all costs incurred in developing land or acquiring resources, plus a certain amount of profit and taxes. This method is stipulated in both the "Urban Land Valuation Regulations" (GB / T 18508-2014) and the "Application Guidelines for Mineral Rights Valuation". Its main drawbacks are that it is cost-oriented, fails to reflect market supply and demand and the scarcity value of resources, and completely ignores the ecological service value of resources.

[0004] 2) Income capitalization approach; This method involves discounting the expected future returns of resource assets at a certain discount rate to calculate the sum of their present values ​​as of the valuation date. It is widely used in "Land Valuation Practice" and "Guidelines for Mineral Rights Valuation" (CMVS 00001-2008). Its main drawbacks are: strong subjectivity in predicting future returns, a lack of unified standards for selecting discount rates, and difficulty in incorporating non-market-based ecosystem service revenues such as carbon sequestration and water conservation into the calculation framework.

[0005] 3) Market comparison method; The value of the appraised asset is estimated by using the transaction prices of comparable cases with similar conditions, after adjustments. This method is detailed in standards such as the "Real Estate Valuation Standard" (GB / T 50291-2015). Its main drawback is that it is highly dependent on the quantity and quality of comparable cases. However, in natural resource asset transactions, especially in scenarios with a combination of supply factors, comparable cases are extremely scarce, making this method often unapplicable.

[0006] The three traditional methods mentioned above are all designed for single resource elements and face fundamental limitations in multi-element combination supply scenarios: they cannot quantify the synergistic effects and combination premiums between elements, cannot incorporate the value of ecosystem services into a unified valuation framework, and have low evaluation efficiency (a single asset valuation usually takes 15-30 days).

[0007] In recent years, machine learning technology has made some progress in the application of asset valuation: (1) Land valuation based on traditional machine learning; Some scholars have proposed a land price prediction model based on Random Forest, which uses spatial and socioeconomic characteristics to predict land prices. 2 The value reached 0.85. Some scholars have used gradient boosting trees (XGBoost) to predict urban land prices in China, achieving better results than traditional regression models. However, the above studies only focus on land as a single factor, without considering other natural resource types such as minerals, forests, and water resources, nor do they consider the collaborative valuation of multiple factor combinations.

[0008] (2) Real estate valuation based on deep learning; Scholars have pioneered the application of deep neural networks to real estate valuation, achieving automated property value assessment through the fusion of satellite imagery and structured data. Others have proposed a spatiotemporal prediction model for real estate prices combining convolutional neural networks (CNN) and long short-term memory networks (LSTM). While these studies represent innovations in deep learning methods, their applications are limited to real estate (building + land) valuation, and the types of data and feature dimensions they handle are far fewer than those required for multi-factor valuation of natural resource assets.

[0009] (3) Mineral resource assessment based on machine learning; Some scholars have applied the random forest algorithm to assess mineral resource potential, but their goal is spatial prediction of mineral resource distribution, not economic valuation of mineral assets. Other scholars have used artificial neural networks to estimate mineral resource quantities, but these methods similarly fail to address market valuation and multi-factor collaborative pricing.

[0010] Existing patents (such as CN201910XXX, CN113469570A, CN112016754A) have also failed to solve the technical challenges of valuing multi-factor natural resource asset portfolios.

[0011] In summary, the existing technology has the following main drawbacks: 1. The evaluation method is too simplistic and difficult to adapt to scenarios with multiple factors combined; Traditional valuation methods (cost approach, income approach, market comparison approach) are mainly aimed at single resource elements. When multiple elements such as land, minerals, forests, and water resources are combined into an "asset package" for overall supply, they cannot accurately quantify the synergistic effect and combination premium between elements, resulting in a systematic underestimation of the valuation results.

[0012] 2. The assessment efficiency is low and cannot meet the needs of large-scale rapid valuation. Traditional manual appraisal requires on-site inspection, data collection, and expert review, and the appraisal cycle for a single asset usually takes 15-30 days, which is seriously inefficient when facing regional and batch asset allocation needs.

[0013] 3. The influencing factors are not fully considered, resulting in low evaluation accuracy; Traditional methods rely mainly on historical transaction cases and appraisers' experience, making it difficult to fully consider multi-dimensional influencing factors such as location conditions, resource endowment, ecological value, market supply and demand, and policy guidance. As a result, the appraisal results deviate significantly from the market transaction price (usually with a deviation rate of 15%-30%).

[0014] 4. Ecological value is difficult to quantify, and resource value is underestimated; The ecological service functions of plateau lakes such as Fuxian Lake, such as water resources and the carbon sequestration value of the Ailao Mountain forests, lack scientific quantitative methods in traditional assessment systems, resulting in the inability to reflect the true value of ecological resources.

[0015] 5. Lack of a dynamic updating mechanism results in poor timeliness of evaluation results; Market environment, policy conditions, and resource status are constantly changing, but traditional assessment results are static values ​​and lack the ability to be dynamically updated based on real-time data.

[0016] Therefore, there is an urgent need for an intelligent valuation method and system that can uniformly process multiple types of natural resource elements, quantify and combine premiums, assess ecological value, and have dynamic updating capabilities. Summary of the Invention

[0017] The present invention provides a method and system for intelligent valuation of multiple factors of natural resource assets based on deep learning, which can overcome some or all of the defects of the prior art.

[0018] According to the present invention, a multi-factor intelligent valuation method for natural resource assets based on deep learning includes the following steps: Step S1: Collect and fuse multi-source heterogeneous data to build a unified data foundation. The data sources of the unified data foundation include resource baseline data, spatial geographic data, location condition data, market transaction data, remote sensing monitoring data, ecological environment data, policy planning data, and socio-economic data. Step S2: Based on a unified data foundation, construct a multi-level feature engineering system, including basic features, interactive features, and deep features; Step S3: Construct a multi-task deep learning model, which includes a shared feature extraction network and multiple task-specific output heads, wherein the multiple task-specific output heads include at least a value assessment head for different natural resource types and a combined premium estimation head; Step S4: Using the multi-task deep learning model, simultaneously perform single-factor value assessment of different natural resource assets and prediction of the total combined value of multi-factor combined assets; Step S5: Based on the attention mechanism, calculate the synergistic weights between different elements in the multi-element portfolio asset and quantify the portfolio premium, which is the difference between the total portfolio value and the sum of the values ​​of each individual element. Step S6: Output the valuation results, which include at least the value of each individual element, the total value of the combination, the combination premium, and model interpretability information.

[0019] As a preferred option, in step S1, the resource baseline data includes the third national land survey, mineral reserves, forest resource inventory, and water resource survey. Spatial geographic data includes elevation, slope, aspect, soil type, and climate data; Location data includes distance from the city center, accessibility, and density of surrounding points of interest (POIs). Market transaction data includes historical transaction prices, listing prices, and records of failed auctions; Remote sensing monitoring data includes multispectral imagery, NDVI vegetation index, and land surface temperature; Ecological and environmental data include water quality monitoring, air quality, and ecological protection red lines; Policy planning data includes land spatial planning, industrial policies, and environmental protection policies; Socioeconomic data includes GDP, population density, and industrial structure.

[0020] As a preferred option, the construction of the multi-level feature engineering system in step S2 specifically includes: The first layer: Constructing more than 80 basic features, including resource attribute features, spatial location features, location condition features, ecological environment features, market supply and demand features, policy and planning features, and time features; The second layer generates interactive features of more than 100 dimensions through feature cross-referencing, which are used to capture non-linear relationships between elements; The third layer: automatically learns high-dimensional abstract deep features through deep neural networks.

[0021] Preferably, the multi-task deep learning model in step S3 further includes an ecological value intelligent quantification submodule, which is used to perform at least one of the following evaluations: Carbon sink value assessment: Carbon sink volume is estimated based on remote sensing net primary productivity inversion model, and carbon sink value is calculated in combination with dynamically acquired carbon trading prices; Water conservation value estimation: The water conservation volume is calculated based on the water conservation volume model, and the conservation value is calculated in combination with shadow prices; Biodiversity conservation value assessment: Conservation value is calculated based on habitat quality assessment models, species scarcity, and social willingness to pay.

[0022] As a preferred option, the specific method for quantifying the portfolio premium based on the attention mechanism in step S5 includes: 5.1) Input the feature vectors of each element in the multi-element portfolio asset into the multi-head attention layer and calculate the attention weight of each element; 5.2) The attention weights are used to weight and fuse the feature vectors of each element to generate combined features; 5.3) Input the combined features into a fully connected network to predict the total combined value; 5.4) Calculate the difference between the total value of the combination and the sum of the values ​​of each individual element to obtain the combination premium; 5.5) Calculate the ratio of the combined premium to the sum of the values ​​of each individual factor to obtain the combined premium rate.

[0023] Preferably, the training of multi-task deep learning models employs a multi-task weighted loss function. :

[0024] in, For the first i The loss function for each task is expressed as the sum of the mean squared error and the relative error penalty term: , Assigning weights to each task. This is the relative error penalty coefficient. , For the first i The model predictions and actual values ​​for each task.

[0025] As a preferred option, a dynamic update mechanism is also included, which includes: A) Regular updates: Automatic model retraining is triggered every quarter; B) Data-driven update: Incremental training is triggered when the amount of new transaction data reaches a preset threshold; C) Bias-driven update: Model correction is triggered when the prediction bias exceeds a preset threshold multiple times in a row; D) Policy-driven updates: Manually triggered updates in response to significant policy changes; The update process includes: new data validation, incremental training, A / B testing, model switching, and performance monitoring.

[0026] Preferably, the model interpretability information output in step S6 includes: the ranking of the contribution of each feature to the valuation result or the strength of the synergistic effect between elements generated by SHAP value analysis or attention weight visualization.

[0027] This invention provides a deep learning-based intelligent valuation system for multi-factor natural resource assets, which employs the aforementioned deep learning-based intelligent valuation method for multi-factor natural resource assets and includes: The data acquisition layer is configured to: collect natural resource asset-related data from multiple heterogeneous data sources, and clean, merge, and process the collected data in real time to build a unified data foundation; The feature engineering layer, connected to the data acquisition layer, is configured to: perform feature extraction, feature selection, and feature encoding on the data in the unified data foundation to generate a multi-level feature system including basic features, interactive features, and deep features; The model training layer, connected to the feature engineering layer, is configured to: construct and train a multi-task deep learning model based on the multi-level feature system; the model includes a shared feature extraction network, multiple task-specific output heads, and a combined premium quantization unit based on an attention mechanism; and train, optimize, and evaluate the model. The intelligent valuation application layer, connected to the model training layer, is configured to: utilize a trained multi-task deep learning model to provide single-factor valuation services, multi-factor combined valuation services, ecological value quantification services, and dynamic model update services, and output valuation results.

[0028] The beneficial effects of this invention are as follows: This invention supports unified valuation of multiple resource elements such as land, minerals, forests, and water resources. It can accurately quantify the synergistic effects and combination premiums of multi-element combinations, realize the intelligent quantification and monetization of ecological value, and significantly improve the evaluation efficiency (second-level response) and accuracy (deviation rate <5%). It establishes a dynamic update mechanism based on real-time data, as detailed below: I. For the first time, unified intelligent valuation of multi-element natural resource assets has been achieved; It breaks through the limitations of traditional methods that can only target a single element, and achieves collaborative valuation of multiple resources such as land, minerals, forests, and water resources under a unified framework.

[0029] II. Accurately quantify portfolio premiums to improve asset allocation efficiency; By modeling the synergistic effect of elements through attention mechanisms, the premium of the "asset package" combination can be accurately quantified, resulting in a 20%-35% increase in value-added rate compared to traditional methods.

[0030] III. Realizing the intelligent quantification and monetization of ecological value; For the first time, the value of ecological services such as carbon sequestration, water conservation, and biodiversity has been incorporated into the valuation system, allowing the value of ecological resources to be fully reflected.

[0031] IV. Evaluation efficiency is improved by more than 100 times; The valuation of a single asset has been reduced from the traditional 15-30 days to less than 3 seconds, and regional batch valuation is supported.

[0032] V. The accuracy of the assessment is significantly improved; By comprehensively considering multiple influencing factors through deep learning, the evaluation bias rate has been reduced from the traditional 15%-30% to less than 5%.

[0033] VI. Establish a dynamic updating mechanism to maintain the timeliness of valuations; Adaptive updates based on real-time data and transaction feedback ensure that valuation results always reflect the latest market conditions.

[0034] VII. Highly interpretable and supports decision analysis; By analyzing feature importance and visualizing attention weights, the valuation basis is clearly presented, supporting management decisions. Attached Figure Description

[0035] Figure 1 This is a flowchart of a multi-factor intelligent valuation method for natural resource assets based on deep learning, as shown in Example 1. Figure 2 This is a schematic diagram of the multi-task deep learning model in Example 1; Figure 3 This is a schematic diagram of a multi-factor intelligent valuation system for natural resource assets based on deep learning, as shown in Example 1. Detailed Implementation

[0036] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0037] Example 1 like Figure 1 As shown, this embodiment provides a multi-factor intelligent valuation method for natural resource assets based on deep learning, which includes the following steps: Step S1: Collect and integrate multi-source heterogeneous data to build a unified data foundation. The data sources of the unified data foundation include resource baseline data, spatial geographic data, location condition data, market transaction data, remote sensing monitoring data, ecological environment data, policy planning data, and socio-economic data.

[0038] Specifically: 1.1) Baseline resource data: Third National Land Survey, mineral reserves, forest resource inventory, water resource survey, etc.; 1.2) Spatial geographic data: elevation, slope, aspect, soil type, climate data, etc.; 1.3) Location data: distance from the city center, accessibility, density of surrounding POIs, etc.; 1.4) Market transaction data: historical transaction prices, listing prices, and records of failed auctions; 1.5) Remote sensing monitoring data: multispectral imagery, vegetation index (NDVI), land surface temperature, etc. 1.6) Ecological and environmental data: water quality monitoring, air quality, ecological protection red lines, etc.; 1.7) Policy and planning data: territorial spatial planning, industrial policies, environmental protection policies, etc.; 1.8) Socioeconomic data: GDP, population density, industrial structure, etc.

[0039] Data fusion employs a unified spatiotemporal reference processing method: Spatial reference: uniformly converted to the CGCS2000 coordinate system; Time base: Establish timestamp index to support multi-time phase data association; Data format: Convert to a standardized vector / raster format.

[0040] Step S2: Based on a unified data foundation, construct a multi-level feature engineering system, including basic features, interactive features, and deep features.

[0041] The first layer consists of over 80 basic features, including resource attribute features, spatial location features, locational conditions features, ecological environment features, market supply and demand features, policy and planning features, and temporal features, as shown in Table 1. Table 1 Basic Features ; The second layer generates interactive features of more than 100 dimensions through feature cross-referencing, which are used to capture non-linear relationships between elements, including: resource endowment × location conditions (such as "high-quality minerals × convenient transportation"), ecological value × market demand (such as "high carbon sink × carbon trading price"), and planned use × surrounding environment (such as "industrial land × pollution carrying capacity").

[0042] The third layer: automatically learns high-dimensional abstract deep features (128 dimensions) through a deep neural network.

[0043] Step S3: Construct a multi-task deep learning model, such as Figure 2As shown, the model includes a shared feature extraction network and multiple task-specific output heads, which include at least a value assessment head for different natural resource types and a combined premium estimation head.

[0044] The multi-task deep learning model in step S3 also includes an ecological value intelligent quantification submodule, which is used to perform at least one of the following evaluations: Carbon sink value assessment: The amount of carbon sink is estimated based on the remote sensing net primary productivity inversion model, and the value of carbon sink is calculated in combination with dynamically acquired carbon trading prices; Carbon sink value = Carbon sink amount × Carbon trading price (dynamically acquired).

[0045] Water conservation value estimation: The water conservation volume is calculated based on the water conservation volume model, and the conservation value is calculated in combination with the shadow price; Conservation value = Water conservation volume × Shadow price.

[0046] Biodiversity conservation value assessment: Conservation value is calculated based on habitat quality assessment models, species scarcity, and social willingness to pay; Conservation value = f(Habitat quality, Species scarcity, Social willingness to pay).

[0047] Step S4: Using the multi-task deep learning model, simultaneously perform single-factor value assessment of different natural resource assets and prediction of the total combined value of multi-factor combined assets.

[0048] The training of multi-task deep learning models uses a multi-task weighted loss function. :

[0049] in, For the first i The loss function for each task is expressed as the sum of the mean squared error and the relative error penalty term: , Assigning weights to each task. This is the relative error penalty coefficient. , For the first i The model predictions and actual values ​​for each task.

[0050] The training strategy is as follows: Data augmentation: SMOTE oversampling is used for resource types with few samples (such as water resources); Transfer learning: Pre-training with data from developed regions and then fine-tuning on local data; Ensemble learning: Training multiple models (XGBoost, LightGBM, deep neural networks) and then integrating them; Online learning: Each new transaction data point automatically triggers an incremental update of the model.

[0051] It also includes a dynamic update mechanism, which includes: A) Regular updates: Automatic model retraining is triggered every quarter; B) Data-driven update: Incremental training is triggered when the amount of new transaction data reaches a preset threshold (e.g., 100 records); C) Bias-driven update: When the prediction bias exceeds a preset threshold multiple times (10 times) consecutively, model correction is triggered; D) Policy-driven updates: Manually triggered updates in response to significant policy changes.

[0052] The update process includes: new data validation, incremental training, A / B testing, model switching, and performance monitoring.

[0053] Step S5: Based on the attention mechanism, calculate the synergistic weights between different elements in the multi-element portfolio asset and quantify the portfolio premium, which is the difference between the total portfolio value and the sum of the values ​​of each individual element.

[0054] The specific methods for quantifying the portfolio premium based on the attention mechanism in step S5 include: 5.1) Input the feature vectors of each element in the multi-element portfolio asset into the multi-head attention layer and calculate the attention weight of each element; 5.2) The attention weights are used to weight and fuse the feature vectors of each element to generate combined features; 5.3) Input the combined features into a fully connected network to predict the total combined value; 5.4) Calculate the difference between the total value of the combination and the sum of the values ​​of each individual element to obtain the combination premium; 5.5) Calculate the ratio of the combined premium to the sum of the values ​​of each individual factor to obtain the combined premium rate.

[0055] The code is as follows: class CombinationPremiumModel: def __init__(self): self.attention = MultiHeadAttention(heads=4) self.fc = FullyConnected(128 → 1) def forward(self, asset_features): """ asset_features: [Land features, mineral features, forest features, water resource features] """ # 1. Calculate the independent value of each element. individual_values ​​= [ land_model(asset_features[0]), mineral_model(asset_features[1]), forest_model(asset_features[2]), water_model(asset_features[3]) ] # 2. Attention mechanism for calculating collaborative weights attention_weights = self.attention(asset_features) # 3. Weighted fusion features combined_features = weighted_sum(asset_features, attention_weights) # 4. Predict the total value of the portfolio combined_value = self.fc(combined_features) # 5. Calculate the portfolio premium premium = combined_value - sum(individual_values) premium_rate = premium / sum(individual_values) return { 'individual_values': individual_values, 'combined_value': combined_value, 'premium': premium 'premium_rate': premium_rate } Step S6: Output the valuation results, which include at least the value of each individual element, the total value of the combination, the combination premium, and model interpretability information.

[0056] The model interpretability information output in step S6 includes: the ranking of the contribution of each feature to the valuation result or the strength of the synergistic effect between elements, generated through SHAP value analysis or attention weight visualization.

[0057] like Figure 3 As shown, this embodiment provides a deep learning-based intelligent valuation system for multi-factor natural resource assets, which employs the aforementioned deep learning-based intelligent valuation method for multi-factor natural resource assets and includes: The data acquisition layer is configured to: collect natural resource asset-related data from multiple heterogeneous data sources, and clean, merge, and process the collected data in real time to build a unified data foundation; The feature engineering layer, connected to the data acquisition layer, is configured to: perform feature extraction, feature selection, and feature encoding on the data in the unified data foundation to generate a multi-level feature system including basic features, interactive features, and deep features; The model training layer, connected to the feature engineering layer, is configured to: construct and train a multi-task deep learning model based on the multi-level feature system; the model includes a shared feature extraction network, multiple task-specific output heads, and a combined premium quantization unit based on an attention mechanism; and train, optimize, and evaluate the model. The intelligent valuation application layer, connected to the model training layer, is configured to: utilize a trained multi-task deep learning model to provide single-factor valuation services, multi-factor combined valuation services, ecological value quantification services, and dynamic model update services, and output valuation results.

[0058] Example 2 The background of this embodiment is: a phosphate mine development project involving mining rights (20 million tons of phosphate mine), industrial land (80 mu), temporary land (70 mu), and forest land use permit (50 mu).

[0059] Implementation steps: Step 1: Data Collection; Collect the following data: Mineral reserves: 20 million tons, grade P2O5 28%; Land area: 80 mu of industrial land, 70 mu of temporary land, and 50 mu of forest land; Spatial location: Longitude 102.5°E, Latitude 24.3°N, Altitude 1850m; Location: 35km from the city center and 8km from the expressway; Ecological environment: NDVI=0.65, not an ecological red line area; Market data: The average transaction price of phosphate rock in the region is 55 yuan / ton, and the price of industrial land is 1.2 million yuan / mu; Policy data: Compliant with mineral resource planning, but requires supporting ecological restoration; Step 2: Feature Engineering; Extracting 80-dimensional basic features: Resource attributes: reserves of 20 million tons, grade of 28%, and exploitable lifespan of 15 years, etc.; Spatial characteristics: latitude and longitude, altitude, slope of 12°, etc. Location characteristics: 35km from the city center, transportation convenience score 0.75, etc. Ecological characteristics: NDVI=0.65, area requiring restoration of 50 acres, etc. Market characteristics: Phosphate rock price is 55 yuan / ton, land price is 1.2 million yuan / mu, etc. Generate 100-dimensional interactive features: Reserves × Grade = 560 million tons; Distance × Transportation = 26.25 (Comprehensive Accessibility Index); Repair area × repair cost = 2.5 million yuan (estimated); Step 3: Model Inference; 3.1) Single-factor valuation; Mining rights value = mineral model (mineral characteristics) = 115 million yuan; Industrial land value = land model (land characteristics) = 96 million yuan; Temporary land use value = Land model (temporary land use characteristics) = 4.2 million yuan; Forest land use value = forest model (forest land characteristics) = 1.5 million yuan; Total value of individual factors = 1.15 + 0.96 + 0.042 + 0.015 = 216.7 million yuan; 3.2) Estimate the portfolio premium; Input: [Mineral characteristics, Industrial land characteristics, Temporary land characteristics, Forest land characteristics]; Attention weight calculation: Coordination weight between mining and industrial land use: 0.85 (strong synergy, mining requires supporting land use); Mineral resources - Temporary land use coordination weight: 0.72 (medium coordination, temporary land use is required for construction); Industrial land-forest land synergy weight: -0.35 (negative synergy, compensation is required for forest land occupation); Total portfolio value forecast: RMB 287.5 million; The combined premium is 2.875 - 2.167 = 0.708 billion yuan. The portfolio premium rate = 0.708 / 2.167 = 32.7%; 3.3) Quantification of ecological value; Forest land carbon sequestration value: Carbon sequestration amount = 50 mu × 0.8 tCO2 / mu / year × 15 years = 600 tCO2; Carbon sequestration value = 600 × 80 yuan / tCO2 = 48,000 yuan; Internalizing the costs of ecological restoration: Repair cost = 50 mu × 50,000 yuan / mu = 2.5 million yuan (deducted from the combined premium); The adjusted total portfolio value is 2.875 + 0.0048 - 0.25 = 262.98 million yuan.

[0060] Step 4: Output the results; The asset portfolio valuation report is as follows: I. Value of a single factor; Mining rights: RMB 115 million; Industrial land: RMB 96 million; Temporary land use: RMB 4.2 million; Forest land use rights: RMB 1.5 million; Subtotal: RMB 216.7 million; II. Combination premium; Synergistic value-added: RMB 70.8 million; Premium rate: 32.7%; III. Adjustment of Ecological Value; Carbon sequestration value: +480,000 yuan; Remediation cost: -25 million yuan; IV. Final valuation; Total portfolio value: RMB 262.98 million; Suggested listing price: RMB 265 million; V. Basis for Valuation; Model version: v2.3.1; Training samples: 1,250; Model accuracy: R 2 =0.92, MAPE=4.3%; Confidence interval: RMB 251 million - RMB 275 million; Step 5: Interpretability analysis; The contribution of each feature to the valuation was analyzed using SHAP values, as shown in Table 2: Table 2. Contribution of each feature to the valuation ; Implementation results: The asset package was ultimately sold for 268 million yuan, with a deviation of only 1.9% from the model's prediction, validating the model's accuracy. Compared to the traditional itemized valuation result (205 million yuan), the value increased by 63 million yuan, representing a 30.7% increase.

[0061] The present invention and its embodiments have been described above illustratively. This description is not restrictive, and the figures shown are only one embodiment of the present invention; the actual structure is not limited thereto. Therefore, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the present invention, such designs should fall within the protection scope of the present invention.

Claims

1. A multi-factor intelligent valuation method for natural resource assets based on deep learning, characterized in that, Includes the following steps: Step S1: Collect and fuse multi-source heterogeneous data to build a unified data foundation. The data sources of the unified data foundation include resource baseline data, spatial geographic data, location condition data, market transaction data, remote sensing monitoring data, ecological environment data, policy planning data, and socio-economic data. Step S2: Based on a unified data foundation, construct a multi-level feature engineering system, including basic features, interactive features, and deep features; Step S3: Construct a multi-task deep learning model, which includes a shared feature extraction network and multiple task-specific output heads, wherein the multiple task-specific output heads include at least a value assessment head for different natural resource types and a combined premium estimation head; Step S4: Using the multi-task deep learning model, simultaneously perform single-factor value assessment of different natural resource assets and prediction of the total combined value of multi-factor combined assets; Step S5: Based on the attention mechanism, calculate the synergistic weights between different elements in the multi-element portfolio asset and quantify the portfolio premium, which is the difference between the total portfolio value and the sum of the values ​​of each individual element. Step S6: Output the valuation results, which include at least the value of each individual element, the total value of the combination, the combination premium, and model interpretability information.

2. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 1, characterized in that, In step S1, the resource baseline data includes the third national land survey, mineral reserves, forest resource inventory, and water resource survey. Spatial geographic data includes elevation, slope, aspect, soil type, and climate data; Location data includes distance from the city center, accessibility, and density of surrounding points of interest (POIs). Market transaction data includes historical transaction prices, listing prices, and records of failed auctions; Remote sensing monitoring data includes multispectral imagery, NDVI vegetation index, and land surface temperature; Ecological and environmental data include water quality monitoring, air quality, and ecological protection red lines; Policy planning data includes land spatial planning, industrial policies, and environmental protection policies; Socioeconomic data includes GDP, population density, and industrial structure.

3. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 1, characterized in that, Step S2, which involves constructing a multi-level feature engineering system, specifically includes: The first layer: Constructing more than 80 basic features, including resource attribute features, spatial location features, location condition features, ecological environment features, market supply and demand features, policy and planning features, and time features; The second layer generates interactive features of more than 100 dimensions through feature cross-referencing, which are used to capture non-linear relationships between elements; The third layer: automatically learns high-dimensional abstract deep features through deep neural networks.

4. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 3, characterized in that, The multi-task deep learning model in step S3 also includes an ecological value intelligent quantification submodule, which is used to perform at least one of the following evaluations: Carbon sink value assessment: Carbon sink volume is estimated based on remote sensing net primary productivity inversion model, and carbon sink value is calculated in combination with dynamically acquired carbon trading prices; Water conservation value estimation: The water conservation volume is calculated based on the water conservation volume model, and the conservation value is calculated in combination with shadow prices; Biodiversity conservation value assessment: Conservation value is calculated based on habitat quality assessment models, species scarcity, and social willingness to pay.

5. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 4, characterized in that, The specific methods for quantifying the portfolio premium based on the attention mechanism in step S5 include: 5.1) Input the feature vectors of each element in the multi-element portfolio asset into the multi-head attention layer and calculate the attention weight of each element; 5.2) The attention weights are used to weight and fuse the feature vectors of each element to generate combined features; 5.3) Input the combined features into a fully connected network to predict the total combined value; 5.4) Calculate the difference between the total value of the combination and the sum of the values ​​of each individual element to obtain the combination premium; 5.5) Calculate the ratio of the combined premium to the sum of the values ​​of each individual factor to obtain the combined premium rate.

6. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 5, characterized in that, The training of multi-task deep learning models uses a multi-task weighted loss function. : ； in, For the first i The loss function for each task is expressed as the sum of the mean squared error and the relative error penalty term: , Assigning weights to each task. This is the relative error penalty coefficient. , For the first i The model predictions and actual values ​​for each task.

7. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 6, characterized in that, It also includes a dynamic update mechanism, which includes: A) Regular updates: Automatic model retraining is triggered every quarter; B) Data-driven update: Incremental training is triggered when the amount of new transaction data reaches a preset threshold; C) Bias-driven update: Model correction is triggered when the prediction bias exceeds a preset threshold multiple times in a row; D) Policy-driven updates: Manually triggered updates in response to significant policy changes; The update process includes: new data validation, incremental training, A / B testing, model switching, and performance monitoring.

8. The method for intelligent multi-factor valuation of natural resource assets based on deep learning according to claim 7, characterized in that, The model interpretability information output in step S6 includes: the ranking of the contribution of each feature to the valuation result or the strength of the synergistic effect between elements, generated through SHAP value analysis or attention weight visualization.

9. A multi-factor intelligent valuation system for natural resource assets based on deep learning, characterized in that: It employs a deep learning-based multi-factor intelligent valuation method for natural resource assets as described in any one of claims 1-8, and includes: The data acquisition layer is configured to: collect natural resource asset-related data from multiple heterogeneous data sources, and clean, merge, and process the collected data in real time to build a unified data foundation; The feature engineering layer, connected to the data acquisition layer, is configured to: perform feature extraction, feature selection, and feature encoding on the data in the unified data foundation to generate a multi-level feature system including basic features, interactive features, and deep features; The model training layer, connected to the feature engineering layer, is configured to: construct and train a multi-task deep learning model based on the multi-level feature system; the model includes a shared feature extraction network, multiple task-specific output heads, and a combined premium quantization unit based on an attention mechanism; and train, optimize, and evaluate the model. The intelligent valuation application layer, connected to the model training layer, is configured to: utilize a trained multi-task deep learning model to provide single-factor valuation services, multi-factor combined valuation services, ecological value quantification services, and dynamic model update services, and output valuation results.

Images