A deep learning-based multi-scale assessment and calculation method and system for forest soil carbon sink capacity

Abstract

The present application relates to the technical field of soil carbon sink capacity evaluation, and discloses a forest soil carbon sink capacity multi-scale evaluation and calculation method and system based on deep learning, the method comprising: dividing the forest to be evaluated into multiple forest soil regions by adopting a clustering method based on regional carbon sink potential; performing layer feature extraction on the multi-layer multi-source data of the forest soil region to obtain the multi-layer features of each forest soil region; constructing the multi-scale multi-layer features of the forest soil region and performing correlation fusion to obtain hierarchical change fusion features; and evaluating the carbon sink capacity of the forest soil region by using a soil carbon sink capacity evaluation model.The present application divides the forest soil region based on multi-source multi-layer data, fuses the multi-scale features, dynamically models and evaluates the fusion features by using a deep learning model, accurately evaluates the carbon sink capacity of the forest soil, and realizes efficient and fine evaluation and calculation of the carbon cycle process of a complex ecological system.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of soil carbon sink capacity evaluation, and in particular to a deep learning-based multi-scale forest soil carbon sink capacity evaluation method and system. BACKGROUND

[0002] Forest soil is one of the largest terrestrial carbon pools in the world, with a carbon storage of more than 60% of the total carbon storage of terrestrial ecosystems, and is an important part of global carbon cycle. It plays an irreplaceable role in slowing the increase of carbon dioxide concentration in the atmosphere and maintaining the stability of the ecosystem. The carbon sink capacity of forest soil not only reflects the comprehensive function of soil in long-term carbon sequestration and short-term carbon dynamic exchange, but also directly affects the regional and even global carbon balance pattern and the mitigation effect of climate change.

[0003] However, the forest soil carbon sink process is a complex system driven by multiple biological-physical-chemical processes, and its changes are influenced by multiple factors such as soil physical and chemical properties (such as organic carbon density, active carbon proportion, water content), vegetation growth state and litter input, regional climate conditions (temperature, humidity, precipitation pattern), and human activities (logging, fertilization, land use change). These influencing factors not only have high spatial heterogeneity, but also show obvious periodicity, seasonality and mutagenicity in time scale, resulting in strong nonlinearity, multi-scale and dynamic evolution characteristics of forest soil carbon sink capacity.

[0004] At the same time, the change of soil carbon sink is also closely related to topography, soil microbial community structure, soil respiration rate and other micro ecological processes, and there are significant interactions and coupling feedbacks among these processes. For example, in the same forest area, different slope positions, different soil textures and different vegetation types may show completely different carbon input and release dynamics, thus forming a spatial puzzle effect on the overall carbon balance. In addition, under the background of climate change, extreme weather events occur frequently, making the stability and predictability of forest soil carbon sink capacity face greater challenges.

[0005] In recent years, academia and industry have conducted a large number of explorations on carbon sink calculation methods. For example, patent CN112884339B proposes a rural community carbon emission calculation method and system, the steps of which include: determining the calculation boundary of rural community carbon emission according to different sources of carbon emission; calculating the carbon emission of carbon source and the carbon absorption of carbon sink by using a carbon emission calculation model; obtaining the total carbon emission of the rural community, and comparing it with the execution standard to analyze the carbon emission level, and then implementing corresponding emission reduction and energy saving measures to realize effective management of rural community carbon emission. This method can accurately calculate the carbon emission level of rural community, but its focus is on the macro calculation of rural community carbon source and carbon sink, mainly relying on preset emission coefficients and empirical parameters, and lacks targeted characterization of the multi-scale and dynamic evolution characteristics of forest soil carbon sink.

[0006] To solve this problem, the present patent proposes a multi-scale evaluation and calculation method and system for forest soil carbon sink capacity based on deep learning, which analyzes multi-level and multi-scale data of forest soil to realize accurate quantitative evaluation of forest soil carbon sink capacity and provide a scientific basis for climate change analysis. SUMMARY

[0007] The present application provides a multi-scale evaluation and calculation method and system for forest soil carbon sink capacity based on deep learning. Traditional methods divide regions according to administrative areas or fixed grids, which cannot reflect the natural spatial distribution of soil carbon sink potential. The present application divides forest soil regions based on regional carbon sink potential, so that the spatial unit boundary is consistent with the natural division of carbon sink capacity changes, significantly improving the scientificity of spatial division and the accuracy of evaluation. Through multi-level feature extraction and multi-scale multi-level correlation fusion, combined with deep learning model for joint modeling of spatial and time-dependent features, high-precision fusion of multi-level and multi-source data is realized. Through multi-scale fusion to extract hierarchical change fusion features, time dynamics and spatial structure are cooperatively modeled, so that the spatiotemporal evolution law of forest soil carbon sink capacity can be more accurately reflected, and the evaluation accuracy and generalization ability of carbon sink capacity are significantly improved.

[0008] To achieve the above-mentioned purpose, the present application provides a multi-scale evaluation and calculation method and system for forest soil carbon sink capacity based on deep learning, comprising the following steps:

[0009] S1: dividing the forest to be evaluated into multiple forest soil regions based on regional carbon sink potential clustering, and periodically collecting multi-level and multi-source data of each forest soil region;

[0010] S2: performing level feature extraction on the multi-level and multi-source data of the forest soil region to obtain the multi-level features of each forest soil region;

[0011] S3: based on the multi-level features, constructing the multi-scale multi-level features of the forest soil region in the time scale change and the spatial scale change, and performing multi-level and multi-scale correlation fusion on the multi-scale multi-level features to obtain hierarchical change fusion features of the forest soil region;

[0012] S4: receiving the hierarchical change fusion features of the forest soil region by using a soil carbon sink capacity evaluation model to evaluate the carbon sink capacity of the forest soil region.

[0013] As a further improved method of the present application:

[0014] Optionally, the forest to be evaluated is divided into a plurality of forest soil regions in a clustering manner based on the regional carbon sink potential, comprising:

[0015] The forest to be evaluated is divided into rectangular grid regions according to a fixed spatial resolution (for example, 10 meters 10 meters), and the rectangular grid regions are composed of a plurality of rectangular grids;

[0016] The soil of each rectangular grid is sampled, and the soil organic carbon content in the soil sampling result is extracted as the soil organic carbon content of the rectangular grid, and the remote sensing image of the rectangular grid is obtained, the reflectivity of the rectangular grid in different wave bands is calculated, and the annual average precipitation of the rectangular grid is obtained;

[0017] The leaf area index of the rectangular grid is obtained based on the reflectivity of the rectangular grid in different wave bands;

[0018] The carbon sink potential of the rectangular grid is calculated based on the soil organic carbon content, the leaf area index and the annual average precipitation of the rectangular grid:

[0019] ;

[0020] Wherein, C represents the carbon sink potential of the rectangular grid, represents the soil organic carbon content of the rectangular grid, represents the leaf area index of the rectangular grid, represents the annual average precipitation of the rectangular grid; represents the maximum soil organic carbon content in all rectangular grids in the forest to be evaluated, represents the maximum leaf area index in all rectangular grids in the forest to be evaluated, represents the maximum annual average precipitation in all rectangular grids in the forest to be evaluated;

[0021] All are potential weight coefficients, and are set to 0.4, 0.4, 0.2 respectively;

[0022] Merge rectangular grids that have adjacent edges and the same carbon sink potential until any two rectangular grids with adjacent edges have different carbon sink potentials;

[0023] Extract the rectangular grid with the smallest current area as the rectangular grid to be clustered. Calculate the merging coefficient between the rectangular grid to be clustered and its adjacent rectangular grids. Then, cluster and merge the rectangular grid to be clustered with the adjacent rectangular grid with the highest merging coefficient. The adjacent rectangular grid is the rectangular grid that has an adjacent edge to the rectangular grid to be clustered.

[0024] The formula for calculating the merging coefficient is as follows:

[0025] ;

[0026] in, This represents the merging coefficient between the rectangular grid to be clustered and its adjacent rectangular grids. This represents the adjacent side length between the rectangular grid to be clustered and its adjacent rectangular grids. This represents the difference in carbon sink potential between the rectangular grid to be clustered and its adjacent rectangular grids. Indicates the penalty coefficient, set It is 1.2;

[0027] Repeat the above operations of extracting the smallest rectangular grid and clustering and merging until the area of ​​the extracted rectangular grid is higher than the preset area threshold.

[0028] Optionally, multi-level, multi-source data may be collected periodically for each forest soil region, including:

[0029] The multi-level, multi-source data includes multi-source data at the soil level, multi-source data at the vegetation level, multi-source data at the environmental level, and multi-source data at the remote sensing level.

[0030] The multi-source data at the soil level include soil organic carbon content, soil density, soil layer thickness, soil activated carbon content, and soil respiration flux sequence, where soil respiration flux is the rate at which the soil releases carbon dioxide.

[0031] The multi-source data at the vegetation level includes the area of ​​the survey plot, the increase in underground root biomass and aboveground biomass of the survey plot relative to the previous collection of multi-level multi-source data, and the amount of litter at the survey plot.

[0032] The multi-source environmental data includes soil temperature, soil moisture content, air temperature, annual precipitation, and evaporation. Soil temperature and soil moisture sensors are deployed in forest soil areas to collect soil temperature and soil moisture content, respectively. Air temperature, annual precipitation, and evaporation are also collected based on meteorological station observations.

[0033] The multi-source data of the remote sensing layer includes remote sensing images of the forest soil area in different wave bands.

[0034] Optionally, the multi-layer multi-source data of the forest soil area is subjected to layer feature extraction, including:

[0035] The multi-layer features include soil layer features, vegetation layer features, environmental layer features, and remote sensing layer features.

[0036] The multi-source data of the soil layer is subjected to feature extraction to obtain soil layer features, including soil organic carbon density, active carbon ratio, and soil respiration rate variation coefficient.

[0037] The multi-source data of the vegetation layer is subjected to feature extraction to obtain vegetation layer features, including vegetation carbon input rate, wherein the vegetation carbon input rate is the amount of carbon fixed and input into the soil through photosynthesis per unit time, reflecting the source strength of carbon sink.

[0038] The calculation method of the vegetation carbon input rate is:

[0039] ;

[0040] wherein, represents the vegetation carbon input rate of the forest soil area, represents the area of the survey plot in the process of obtaining the multi-source data of the vegetation layer, represents the periodic collection time interval of the multi-layer multi-source data, represents the aboveground biomass growth amount of the survey plot relative to the last collected multi-layer multi-source data, represents the underground root biomass growth amount of the survey plot relative to the last collected multi-layer multi-source data, and R represents the amount of litter at the survey plot, represents the mass proportion of carbon elements in plant biomass, and is set to 50%;

[0041] The multi-source data of the environmental layer is subjected to feature extraction to obtain environmental layer features, including temperature and humidity adjustment coefficient and water-heat stress index.

[0042] The calculation method of the temperature and humidity adjustment coefficient is:

[0043] ;

[0044] wherein, represents the temperature and humidity adjustment coefficient, represents soil temperature, represents soil water content, ​denotes a soil reference temperature, denotes a soil reference water content, denotes a temperature sensitivity coefficient, set = 2, denotes an exponential function with a natural constant as a base, denotes a water regulation width parameter, set = 0.05;

[0045] Feature extraction is performed on multi-source data at a remote sensing level to obtain remote sensing level features, which include an enhanced vegetation index and an enhanced vegetation index change rate.

[0046] Optionally, multi-scale multi-level features of the forest soil region are constructed in terms of temporal scale changes and spatial scale changes, including:

[0047] The multi-level features of the collected forest soil region are standardized to obtain standardized level features;

[0048] A K-nearest neighbor level feature sequence of the forest soil region and a standardized level feature time sequence in a time sequence level are obtained, the K-nearest neighbor level feature sequence of the forest soil region being a sequence composed of neighbor standardized level features of K nearest neighboring forest soil regions closest to the forest soil region, K being set as 9, and the standardized level feature time sequence being a sequence composed of standardized level features of L multi-level multi-source data collected in the past by the forest soil region, L being set as 5;

[0049] The standardized level features of the forest soil region, the K-nearest neighbor level feature sequence, and the standardized level feature time sequence in the time sequence level are taken as multi-scale multi-level features of the forest soil region.

[0050] Optionally, multi-level and multi-scale correlation fusion is performed on the multi-scale multi-level features to obtain hierarchical change fusion features of the forest soil region, including:

[0051] Multi-scale multi-level features of a forest soil region Q to be correlated and fused are obtained; it should be noted that the forest soil region Q is any forest soil region among the first to Nth forest soil regions;

[0052] The standardized level features of the forest soil region Q in the multi-scale multi-level features are extracted, and the standardized level features are mapped to a high-dimensional space by using an encoder to obtain high-dimensional mapping features of the forest soil region Q;

[0053] K-neighbor layer feature sequence of the forest soil region Q in the multi-scale multi-layer feature is extracted, and the distance from the center of the adjacent forest soil region to the forest soil region Q associated with each neighbor normalized layer feature is obtained, the weight of the neighbor normalized layer feature is calculated based on the distance, the K-neighbor layer feature sequence is graph-convolved by applying a graph convolution network, and a spatial fusion feature of the forest soil region Q is obtained;

[0054] A normalized layer feature time sequence of the forest soil region Q in the multi-scale multi-layer feature is extracted, and the extracted normalized layer feature time sequence is input into a pre-trained bidirectional long short-term memory network to capture the time dimension variation law of the forest soil region Q, and a time sequence dependent feature representing the dynamic change trend of the forest soil region Q is output.

[0055] The high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature of the forest soil region Q are taken as the hierarchical change fusion feature of the forest soil region Q.

[0056] Optionally, the soil carbon sink capacity evaluation model includes an input layer, a soft gating modulation unit, a double-path multi-layer perceptron, and an output layer, comprising:

[0057] The input layer is used to receive the hierarchical change fusion feature of the forest soil region;

[0058] The soft gating modulation unit is used to calculate the gating modulation weight of the high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature in the hierarchical change fusion feature, and to weight and fuse the high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature to obtain a multi-scale hierarchical change fusion feature;

[0059] The double-path multi-layer perceptron includes a carbon sink rate-multi-layer perceptron and a carbon stock-multi-layer perceptron, which respectively receive the multi-scale hierarchical change fusion feature and respectively output a carbon sink rate evaluation result and a carbon stock evaluation result, which are taken as the carbon sink capacity of the forest soil region.

[0060] Optionally, the soil carbon sink capacity evaluation model receives the hierarchical change fusion feature of the forest soil region to evaluate the carbon sink capacity of the forest soil region, comprising:

[0061] The hierarchical change fusion feature of the forest soil region is input into the soil carbon sink capacity evaluation model, and the input layer, the soft gating modulation unit, the double-path multi-layer perceptron and the output layer in the soil carbon sink capacity evaluation model process the hierarchical change fusion feature in turn to obtain the carbon sink capacity of the forest soil region.

[0062] The application further provides a forest soil carbon sink capacity multi-scale evaluation and calculation system, comprising a region division device, a feature extraction module and a carbon sink capacity evaluation device.

[0063] The region division device is used for dividing the forest to be evaluated into a plurality of forest soil regions in a clustering manner based on regional carbon sink potential, and periodically collecting multi-layer multi-source data of each forest soil region.

[0064] The feature extraction module is used for performing layer feature extraction on the multi-layer multi-source data of the forest soil region, obtaining multi-layer features of each forest soil region, constructing multi-scale multi-layer features of the forest soil region in time scale change and spatial scale change based on the multi-layer features, and performing multi-layer and multi-scale correlation fusion on the multi-scale multi-layer features to obtain hierarchical change fusion features of the forest soil region.

[0065] The carbon sink capacity evaluation device is used for receiving the hierarchical change fusion features of the forest soil region by using a soil carbon sink capacity evaluation model, and evaluating the carbon sink capacity of the forest soil region.

[0066] To realize the deep learning-based forest soil carbon sink capacity multi-scale evaluation and calculation method as described above.

[0067] Compared with the prior art, the application provides a deep learning-based forest soil carbon sink capacity multi-scale evaluation and calculation method and system, which has the following beneficial effects:

[0068] Firstly, the soil organic carbon content directly reflects the existing carbon storage and potential stable capacity of the soil, embodies the accommodation space and stability of the soil for new organic carbon, the leaf area index reflects the total green leaf area per unit of land surface, is a key driving factor for carbon fixation and input of organic matter to the soil, and the annual average precipitation determines the water supply and directly affects the plant productivity and the decomposition rate of soil organic matter, the interaction of the three determines the carbon sink growth potential of the rectangular grid under given climate and vegetation conditions, can comprehensively depict the carbon sink capacity from the three dimensions of storage, input and environmental constraints, divides the forest to be evaluated into a plurality of forest soil regions in a clustering manner based on regional carbon sink potential, ensures that the forest soil regions are relatively consistent in carbon storage capacity, vegetation input and water conditions, improves the pertinence of subsequent model prediction and management measures, and avoids the internal heterogeneity caused by the traditional method of dividing only according to single factors such as terrain and vegetation type, thereby reducing the carbon sink estimation error.

[0069] Meanwhile, the multi-scale hierarchical change fusion features extracted by the application have high-dimensional semantic expression ability, spatial neighborhood correlation information and time sequence change law, and can comprehensively depict the multi-dimensional driving factors in the forest soil carbon cycle process. On this basis, the application respectively constructs a carbon sink rate-multilayer perception machine and a carbon storage-multilayer perception machine, realizes double-path interpretation of the fusion features: the former focuses on depicting the carbon absorption and fixation capacity of forest soil per unit time, and the latter focuses on evaluating the carbon storage scale and long-term stability of forest soil. By modeling the carbon sink rate and the carbon storage respectively, the short-term dynamic ability and the long-term storage potential can be depicted at the same time, and the one-sidedness of evaluation caused by a single index is avoided. BRIEF DESCRIPTION OF DRAWINGS

[0070] Figure 1 A flowchart of a multi-scale evaluation and calculation method for forest soil carbon sink capacity based on deep learning is provided for an embodiment of the application.

[0071] Figure 2 A forest soil regional carbon sink rate distribution map is provided for an embodiment of the application.

[0072] Figure 3 A forest soil regional carbon storage distribution map is provided for an embodiment of the application.

[0073] The implementation, functional characteristics and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0074] It should be understood that the specific embodiments described herein are only used to explain the application and not to limit the application.

[0075] The embodiment of the application provides a multi-scale evaluation and calculation method for forest soil carbon sink capacity based on deep learning. The execution subject of the multi-scale evaluation and calculation method for forest soil carbon sink capacity based on deep learning includes but is not limited to at least one of electronic devices such as a server, a terminal and the like which can be configured to execute the method provided by the embodiment of the application. In other words, the multi-scale evaluation and calculation method for forest soil carbon sink capacity based on deep learning can be executed by software or hardware installed in a terminal device or a server device, and the software can be a blockchain platform. The server includes but is not limited to a single server, a server cluster, a cloud server or a cloud server cluster and the like.

[0076] With reference to Figure 1 , Figure 2 and Figure 3 , embodiment 1 of the application is:

[0077] A multi-scale evaluation and calculation method for forest soil carbon sink capacity based on deep learning, the method comprising:

[0078] S1: The forest to be evaluated is divided into multiple forest soil regions using a clustering method based on regional carbon sink potential, and multi-level, multi-source data for each forest soil region is collected periodically.

[0079] The forest to be assessed was divided into multiple forest-soil regions using a clustering approach based on regional carbon sink potential, including:

[0080] According to a fixed spatial resolution (e.g., 10 meters) The forest to be assessed is divided into rectangular grid areas (10 meters apart), each rectangular grid area consisting of multiple rectangular grids.

[0081] Soil samples were taken from each rectangular grid, and the soil organic carbon content was extracted from the soil sampling results as the soil organic carbon content of the rectangular grid. Remote sensing images of the rectangular grids were also acquired, the reflectance of the rectangular grids in different bands was calculated, and the annual average precipitation of the rectangular grids was obtained.

[0082] The leaf area index of a rectangular grid is obtained by inverting the reflectance of the grid in different spectral bands; specifically, the inversion formula is:

[0083] ;

[0084] ;

[0085] in, The enhanced vegetation index represents the rectangular grid. This represents the gain coefficient, and is set accordingly. It is 2. This represents the near-infrared reflectance of a rectangular grid. This represents the red-band reflectivity of the rectangular grid. This represents the blue light reflectance of the rectangular grid. All are atmospheric correction factors, set They are 6 and 7.5 respectively. This indicates the soil background correction item, and the settings are as follows: =1;

[0086] The leaf area index represents the rectangular grid. These are all empirical regression coefficients, obtained by utilizing... Measurement instrument (e.g., LAI-2000) collects data The data were then paired with the enhanced vegetation index to form a data pair, and regression processing was performed to obtain empirical regression coefficients.

[0087] It should be noted that, based on the enhanced vegetation index, Inversion analysis showed that the enhanced vegetation index was less prone to saturation than the normalized vegetation index in areas with high leaf area (dense forest), and could distinguish between high-leaf-area areas and dense forests. The change, and the calculation formula of the enhanced vegetation index contains atmospheric correction and soil background term, reduces the soil brightness and aerosol interference, makes the enhanced vegetation index more sensitive to the change of canopy structure and optical thickness, improves the detection of the change after seasonal disturbance.

[0088] Based on the soil organic carbon content, leaf area index and annual average precipitation of the rectangular grid, the carbon sink potential of the rectangular grid is calculated:

[0089]

[0090] Wherein, C represents the carbon sink potential of the rectangular grid, represents the soil organic carbon content of the rectangular grid, represents the leaf area index of the rectangular grid, represents the annual average precipitation of the rectangular grid; represents the maximum soil organic carbon content of all rectangular grids in the forest to be evaluated, represents the maximum leaf area index of all rectangular grids in the forest to be evaluated, represents the maximum annual average precipitation of all rectangular grids in the forest to be evaluated;

[0091] All are potential weight coefficients, set 0.4, 0.4, 0.2 respectively;

[0092] Merge the rectangular grids with adjacent edges and the same carbon sink potential until the carbon sink potential of any two rectangular grids with adjacent edges is different; as an embodiment of the present application, the merging result of multiple rectangular grids is still a rectangular grid;

[0093] Extract the rectangular grid with the smallest area as the rectangular grid to be clustered, calculate the merging coefficient between the rectangular grid to be clustered and the adjacent rectangular grid, and cluster and merge the rectangular grid to be clustered and the adjacent rectangular grid with the highest merging coefficient, wherein the adjacent rectangular grid is the rectangular grid with adjacent edges with the rectangular grid to be clustered;

[0094] As a preferred embodiment of the present application, the calculation formula of the merging coefficient is:

[0095]

[0096] Wherein, represents the merging coefficient between the rectangular grid to be clustered and the adjacent rectangular grid, represents the length of the adjacent edge between the rectangular grid to be clustered and the adjacent rectangular grid, represents the carbon sink potential difference between the rectangular grid to be clustered and the adjacent rectangular grid,​​​ denotes a penalty coefficient, is set to 1.2, denotes an exponential function with a natural constant as base;

[0097] The above-mentioned extraction of the current smallest rectangular grid and the cluster merging operation are repeated until the area of the extracted rectangular grid is higher than a preset area threshold (for example, 50 square meters).

[0098] Periodically collect multi-layer multi-source data of each forest soil area, including:

[0099] The multi-layer multi-source data includes soil layer multi-source data, vegetation layer multi-source data, environmental layer multi-source data, and remote sensing layer multi-source data;

[0100] The soil layer multi-source data includes soil organic carbon content, soil density, soil layer thickness, soil active carbon content, and soil respiration flux sequence, wherein the soil respiration flux is the rate of carbon dioxide released by the soil;

[0101] Specifically, soil sampling is performed on the forest soil area, and the soil organic carbon content (for example, dry combustion method), soil layer thickness, soil density (for example, bubble method), and soil active carbon content (for example, chemical oxidation method) of the soil sampling result are determined in the laboratory. A gas analyzer is installed at the center of the forest soil area to collect soil respiration flux and form a soil respiration flux sequence within a short period (one day);

[0102] The vegetation layer multi-source data includes the area of the investigation plot, the increase of the belowground root biomass and the aboveground biomass of the investigation plot relative to the last collected multi-layer multi-source data, and the amount of litter at the investigation plot, wherein the investigation plot is a fixed area at the center of the forest soil area, the belowground root biomass, the aboveground biomass, and the amount of litter of the investigation plot are measured, and the vegetation layer multi-source data is calculated based on the last collected multi-layer multi-source data;

[0103] As an embodiment of the present application, the trunk diameter and the tree height of the investigation plot are measured, the aboveground biomass is estimated by using a regional empirical equation; a litter collector (net basket) is installed to collect the litter such as leaves and branches on the same day, and the amount of litter is determined by weighing and drying; the root system is sampled by drilling and layer sampling, and the belowground root biomass is determined by dry weight in the laboratory;

[0104] The environmental layer multi-source data includes soil temperature, soil moisture content, air temperature, annual average precipitation, and evaporation amount, the soil temperature and the soil moisture content are collected by deploying soil temperature sensors and soil moisture sensors in the forest soil area, and the air temperature, the annual average precipitation, and the evaporation amount are collected based on the observation results of the weather station; ​

[0105] The multi-source data of the remote sensing layer includes remote sensing images of the forest soil region in different wave bands.

[0106] S2: Layer feature extraction is performed on the multi-layer multi-source data of the forest soil region to obtain multi-layer features of each forest soil region.

[0107] The layer feature extraction on the multi-layer multi-source data of the forest soil region includes:

[0108] The multi-layer features include soil layer features, vegetation layer features, environmental layer features, and remote sensing layer features;

[0109] The feature extraction is performed on the multi-source data of the soil layer to obtain soil layer features, and the soil layer features include soil organic carbon density, active carbon ratio, and soil respiration rate variation coefficient, wherein the soil organic carbon density is the mass of organic carbon per unit area of soil, directly reflects the carbon capacity stored in the soil, the active carbon ratio is the proportion of active carbon to total organic carbon, indicates the proportion of carbon capacity that can be quickly utilized by microorganisms in the soil, reflects the carbon cycle activity, and the soil respiration rate variation coefficient is the relative fluctuation degree of soil respiration rate, reflects the soil microbial activity and environmental response heterogeneity; it should be noted that the calculation method of the soil respiration rate variation coefficient is the ratio between the standard deviation and the mean value of the soil respiration flux sequence;

[0110] The feature extraction is performed on the multi-source data of the vegetation layer to obtain vegetation layer features, and the vegetation layer features include vegetation carbon input rate, wherein the vegetation carbon input rate is the amount of carbon fixed and input into the soil through photosynthesis per unit time, which represents the source strength of carbon sink;

[0111] As an embodiment of the present application, the calculation method of the vegetation carbon input rate is:

[0112] ;

[0113] wherein, represents the vegetation carbon input rate of the forest soil region, represents the area of the survey plot in the process of obtaining the multi-source data of the vegetation layer, represents the periodic collection time interval of the multi-layer multi-source data, represents the aboveground biomass growth amount of the survey plot relative to the last collected multi-layer multi-source data, represents the underground root biomass growth amount of the survey plot relative to the last collected multi-layer multi-source data, and R represents the amount of litter at the survey plot, represents the mass proportion of carbon elements in plant biomass, and is set to 50%;

[0114] The multi-source data at the environment level is subjected to feature extraction to obtain environment level features, including a temperature and humidity adjustment coefficient and a water-heat stress index, wherein the temperature and humidity adjustment coefficient reflects the adjustment of the temperature and humidity on the microbial decomposition rate, affecting the carbon mineralization process, and the water-heat stress index reflects the stress degree of the environment water and temperature on the plants and microorganisms, affecting the carbon input and decomposition balance.

[0115] Specifically, the temperature and humidity adjustment coefficient is calculated in the following manner:

[0116] ;

[0117] wherein, represents the temperature and humidity adjustment coefficient, represents the soil temperature, represents the soil water content, represents the soil reference temperature (for example, 10 degrees Celsius), represents the soil reference water content (for example, 25%), represents a temperature sensitivity coefficient, which is set to 2, represents an exponential function with a natural constant as the base, represents a water adjustment width parameter, which is set to 0.05;

[0118] The water-heat stress index is calculated in the following manner:

[0119] ;

[0120] wherein, represents the water-heat stress index, represents the air temperature, represents the annual average precipitation, represents the evaporation amount, represents a preset minimum air temperature (for example, 5 degrees Celsius), represents a preset optimal air temperature (for example, 20 degrees Celsius).

[0121] The multi-source data at the remote sensing level is subjected to feature extraction to obtain remote sensing level features, including an enhanced vegetation index and an enhanced vegetation index change rate, wherein the enhanced vegetation index is a high-sensitivity vegetation index, accurately reflecting the leaf area and vegetation health status, and the enhanced vegetation index change rate is a time change trend of the vegetation vitality, reflecting the carbon input change.

[0122] It should be noted that the multi-scale multi-layer feature extraction method overcomes the limitations of single data source by fusing soil physical and chemical properties, biological carbon input dynamics, environmental regulation factors and remote sensing spatial information, realizes accurate description and dynamic tracking of forest soil carbon sink capacity, and significantly improves the accuracy and robustness of carbon sink capacity evaluation by high-resolution and multi-dimensional feature integration, and enhances the adaptability of soil carbon sink capacity evaluation model to heterogeneous forest ecosystems.

[0123] S3: Based on the multi-layer features, multi-scale multi-layer features of the forest soil region in the time scale change and the spatial scale change are constructed, and the multi-layer and multi-scale correlation fusion of the multi-scale multi-layer features is performed to obtain the hierarchical change fusion features of the forest soil region.

[0124] The multi-scale multi-layer features of the forest soil region in the time scale change and the spatial scale change are constructed, including:

[0125] The multi-layer features of the collected forest soil region are standardized to obtain standardized layer features;

[0126] The K nearest neighbor layer feature sequence of the forest soil region and the time sequence sequence of the standardized layer features are obtained, the K nearest neighbor layer feature sequence of the forest soil region is a sequence composed of K nearest neighbor standardized layer features of the nearest neighbor forest soil region, K is set to 9, the standardized layer feature time sequence is a sequence composed of standardized layer features of the forest soil region in the past L multi-layer multi-source data, L is set to 5, and the nearest neighbor standardized layer feature is a standardized layer feature of the neighboring forest soil region of the forest soil region;

[0127] The standardized layer features of the forest soil region, the K nearest neighbor layer feature sequence, and the time sequence sequence of the standardized layer features are used as the multi-scale multi-layer features of the forest soil region.

[0128] As an embodiment of the present application, the multi-scale multi-layer features of the nth forest soil region are:

[0129] ;

[0130] ;

[0131] ;

[0132] wherein, indicates the multi-scale multi-layer features of the nth forest soil region, , N indicates the number of forest soil regions, and T indicates transposition. a K-nearest neighbor layer feature sequence of the nth forest soil region, a normalized layer feature time sequence of the nth forest soil region;

[0133] proximity normalized layer features of the 1st-Kth neighboring forest soil region closest to the nth forest soil region, in sequence respectively;

[0134] normalized layer features of the 1st-Lth multi-layer multi-source data of the nth forest soil region collected in the past, in sequence respectively;

[0135] a normalized layer feature of the nth forest soil region in the current collection cycle;

[0136] performing multi-layer and multi-scale correlation fusion on the multi-scale multi-layer features to obtain hierarchical change fusion features of the forest soil region, including:

[0137] obtaining multi-scale multi-layer features of a forest soil region Q to be correlated and fused currently; it should be noted that the forest soil region Q is any forest soil region among the 1st to Nth forest soil regions;

[0138] extracting normalized layer features of the forest soil region Q from the multi-scale multi-layer features, and mapping the normalized layer features to a high-dimensional space by using an encoder to obtain high-dimensional mapping features of the forest soil region Q;

[0139] extracting a K-nearest neighbor layer feature sequence of the forest soil region Q from the multi-scale multi-layer features, and obtaining distances from a center of a neighboring forest soil region to the forest soil region Q associated with each nearest normalized layer feature, calculating weights of the nearest normalized layer features based on the distances, and applying a graph convolution network to perform graph convolution on the K-nearest neighbor layer feature sequence to obtain a spatial fusion feature of the forest soil region Q;

[0140] Specifically, a graph convolution process of the graph convolution network is as follows:

[0141] convert the K-nearest neighbor layer feature sequence of the forest soil region Q into a matrix form q, where K is 9, q is a 3x3 matrix, construct an adjacency matrix H of the matrix q and a degree matrix of the adjacency matrix H, where the adjacency matrix H is composed of the weights of the nearest normalized layer features, perform E-layer graph convolution on the matrix q in combination with the degree matrix and the adjacency matrix, take a mean value of all graph convolution results as the spatial fusion feature of the forest soil region Q, and the weight calculation formula of the nearest normalized layer features is as follows:

[0142] ;

[0143] wherein, denotes the cosine similarity between the normalized layer feature of the -th neighbor of the K-neighbor layer feature sequence and the normalized layer feature of the forest soil region Q, denotes the distance between the -th neighbor of the K-neighbor layer feature sequence and the forest soil region Q, denotes the distance control coefficient, which is set to 3;

[0144] denotes the weight of the normalized layer feature of the -th neighbor of the K-neighbor layer feature sequence;

[0145] As an embodiment of the present application, by introducing a feature correlation guided adjacency matrix dynamic adjustment mechanism, the cosine similarity between node features is used to form a modulation factor through nonlinear mapping, the distance between forest soil regions is weighted and updated, and adaptive adjustment of multi-scale adjacency fusion weight is realized. This mechanism can dynamically optimize the spatial relationship strength according to the feature content, strengthen the information interaction between the centers of highly correlated adjacent forest soil regions, and suppress the invalid transmission of the centers of lowly correlated adjacent forest soil regions, thereby significantly improving the accuracy and robustness of spatial information fusion in forest soil carbon sink capacity evaluation;

[0146] The normalized layer feature time sequence of the forest soil region Q in the multi-scale multi-layer feature is extracted, and the extracted normalized layer feature time sequence is input into the pre-trained bidirectional long short-term memory network to capture the time dimension variation law of the forest soil region Q, and output the time-dependent feature representing the dynamic change trend of the forest soil region Q;

[0147] The high-dimensional mapping feature, the spatial fusion feature and the time-dependent feature of the forest soil region Q are used as the hierarchical change fusion feature of the forest soil region Q.

[0148] S4: Using the soil carbon sink capacity evaluation model to receive the hierarchical change fusion feature of the forest soil region, and evaluating the carbon sink capacity of the forest soil region.

[0149] The soil carbon sink capacity evaluation model includes an input layer, a soft gate modulation unit, a double-path multi-layer perceptron and an output layer, including:

[0150] The input layer is used to receive the hierarchical change fusion feature of the forest soil region;

[0151] ​The soft gating modulation unit is used for calculating gating modulation weights of the high-dimensional mapping features, the spatial fusion features and the time sequence dependent features in the hierarchical change fusion features, and performing weighted fusion on the high-dimensional mapping features, the spatial fusion features and the time sequence dependent features to obtain multi-scale hierarchical change fusion features.

[0152] As an embodiment of the present application, the gating modulation weight calculation mode of the forest soil region Q is:

[0153]

[0154]

[0155] wherein, are respectively gating modulation weights of the high-dimensional mapping features , the spatial fusion features and the time sequence dependent features of the forest soil region Q in sequence;

[0156] are respectively similarity modulation values of the high-dimensional mapping features , the spatial fusion features and the time sequence dependent features of the forest soil region Q in sequence, represents a trainable weight vector, represents transposition, represents a small coding network (for example, a small multilayer perception), is used for calculating cosine similarity, represents an activation function, and the selected activation function is a Sigmoid function;

[0157] It should be noted that the soft gating mechanism of feature similarity modulation is introduced in the present application to perform adaptive weighted fusion on the high-dimensional mapping features, the spatial fusion features and the time sequence dependent features, which significantly improves the precision and robustness of the forest soil carbon sink capacity evaluation. Specifically, the small coding network is used to perform latent space mapping on the three features respectively to obtain more compact and discriminative representations. Then, the trainable weight vector is used for linear scoring, the cosine similarity of the features and the global fusion representation is combined, and the similarity modulation factor is formed through the activation function to realize the adaptive adjustment of the fusion weight to the feature content correlation. This mechanism can weaken the feature channels that are inconsistent with the overall semantic, and enhance the contribution of the features that are highly consistent with the global trend, thereby effectively suppressing the negative impact of noise interference and local abnormal data on the carbon sink capacity estimation result.

[0158] ​​The double-path multilayer perceptron includes a carbon sink rate-multilayer perceptron and a carbon stock-multilayer perceptron, the carbon sink rate-multilayer perceptron and the carbon stock-multilayer perceptron respectively receive multi-scale hierarchical change fusion features, and respectively output carbon sink rate evaluation results and carbon stock evaluation results, and the carbon sink rate evaluation results and the carbon stock evaluation results are taken as the carbon sink capacity of the forest soil region.

[0159] The soil carbon sink capacity evaluation model receives the hierarchical change fusion features of the forest soil region, and evaluates the carbon sink capacity of the forest soil region, including:

[0160] The hierarchical change fusion features of the forest soil region are input into the soil carbon sink capacity evaluation model, and the input layer, the soft gate modulation unit, the double-path multilayer perceptron and the output layer in the soil carbon sink capacity evaluation model process the hierarchical change fusion features in sequence to obtain the carbon sink capacity of the forest soil region.

[0161] Referring to Figure 2 and Figure 3 , a distribution map of the carbon sink capacity of the forest soil region of the forest to be evaluated is shown, specifically, Figure 2 a carbon sink rate distribution map of the forest soil region, wherein the deeper the color of the forest soil region, the higher the carbon sink rate, Figure 3 a carbon stock distribution map of the forest soil region, wherein the deeper the color of the forest soil region, the higher the carbon stock. Embodiment

[0162] A forest soil carbon sink capacity multi-scale evaluation and calculation system includes a region division device, a feature extraction module and a carbon sink capacity evaluation device:

[0163] The region division device is used to divide the forest to be evaluated into a plurality of forest soil regions by using a clustering method based on regional carbon sink potential, and periodically collects multi-layer multi-source data of each forest soil region;

[0164] The feature extraction module is used to extract layer features from the multi-layer multi-source data of the forest soil region to obtain multi-layer features of each forest soil region, construct multi-scale multi-layer features of the forest soil region in time scale change and spatial scale change based on the multi-layer features, and perform multi-layer, multi-scale correlation fusion on the multi-scale multi-layer features to obtain hierarchical change fusion features of the forest soil region;

[0165] The carbon sink capacity evaluation device is used to receive the hierarchical change fusion features of the forest soil region by using a soil carbon sink capacity evaluation model, and evaluate the carbon sink capacity of the forest soil region;

[0166] To realize a deep learning-based forest soil carbon sink capacity multi-scale evaluation and calculation method as described in Embodiment 1.

[0167] It should be understood that the above-described embodiments are merely intended for description and are not limited to the structure of the patent application.

[0168] It should be noted that the above-mentioned embodiment numbers of the present application are only for description, and do not represent the advantages and disadvantages of the embodiments. And the terms "include", "contain" or any other variants in this paper are intended to cover non-exclusive inclusion, so that the process, device, article or method including a series of elements not only includes those elements, but also includes other elements not explicitly listed, or includes elements inherent to such process, device, article or method. Without more limitations, the element defined by the sentence "including a" does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0169] From the above description of the embodiments, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be realized by software and the necessary general hardware platform, of course, they can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disc, optical disc) as described above, and includes a number of instructions for making a terminal device (which can be a mobile phone, computer, server, or network device, etc.) execute the methods described in various embodiments of the present application.

[0170] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the present application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. A deep learning-based multi-scale assessment and calculation method for forest soil carbon sink capacity, characterized in that, The method comprises: S1: using a clustering method based on regional carbon sink potential to divide the forest to be evaluated into a plurality of forest soil regions, periodically collecting multi-layer multi-source data of each forest soil region; S2: extracting layer features of the multi-layer multi-source data of the forest soil region to obtain multi-layer features of each forest soil region; S3: based on the multi-layer features, constructing multi-scale multi-layer features of the forest soil region in the time scale change and the spatial scale change, and performing multi-layer and multi-scale correlation fusion on the multi-scale multi-layer features to obtain hierarchical change fusion features of the forest soil region; S4: using a soil carbon sink capacity evaluation model to receive the hierarchical change fusion features of the forest soil region, and evaluating to obtain the carbon sink capacity of the forest soil region; The method comprises: According to a fixed spatial resolution, the forest to be evaluated is divided into rectangular grid regions, and the rectangular grid regions are composed of a plurality of rectangular grids; The soil of each rectangular grid is sampled, and the soil organic carbon content in the soil sampling result is extracted as the soil organic carbon content of the rectangular grid, and the remote sensing image of the rectangular grid is obtained, the reflectivity of the rectangular grid in different wave bands is calculated, and the annual average precipitation of the rectangular grid is obtained; Based on the reflectivity of the rectangular grid in different wave bands, the leaf area index of the rectangular grid is obtained; Based on the soil organic carbon content, the leaf area index and the annual average precipitation of the rectangular grid, the carbon sink potential of the rectangular grid is calculated; Rectangular grids with adjacent edges and consistent carbon sink potential are merged until the carbon sink potential of any two rectangular grids with adjacent edges is different; The smallest rectangular grid is extracted as the rectangular grid to be clustered, the merging coefficient between the rectangular grid to be clustered and the adjacent rectangular grid is calculated, and the rectangular grid to be clustered and the adjacent rectangular grid with the highest merging coefficient are clustered and merged, wherein the adjacent rectangular grid is a rectangular grid with adjacent edges to the rectangular grid to be clustered; The above extraction of the smallest rectangular grid and the clustering and merging operation are repeated until the area of the extracted rectangular grid is higher than a preset area threshold.

2. The method of claim 1, wherein the method comprises: obtaining a plurality of soil carbon data sets; and training a deep learning model using the plurality of soil carbon data sets. The formula for calculating the carbon sink potential of the rectangular grid is: ； C represents the soil organic carbon content of the rectangular grid, C represents the soil organic carbon content of the rectangular grid, C represents the leaf area index of the rectangular grid, C represents the annual average precipitation of the rectangular grid; C represents the maximum soil organic carbon content of all rectangular grids in the forest to be evaluated, C represents the maximum leaf area index of all rectangular grids in the forest to be evaluated, C represents the maximum annual average precipitation of all rectangular grids in the forest to be evaluated, are potential weight coefficients, set are 0.4, 0.4, 0.2, respectively. The formula for calculating the merging coefficient is: ； wherein, represents a merging coefficient between the rectangular grid to be clustered and the adjacent rectangular grid, represents an adjacent edge length between the rectangular grid to be clustered and the adjacent rectangular grid, represents a sink potential difference between the rectangular grid to be clustered and the adjacent rectangular grid, represents a penalty coefficient, set to 1.

2.

3. The method of claim 2, wherein the method comprises: determining the soil carbon sink capacity of the forest soil based on the deep learning model. Periodically collecting multi-layer multi-source data of each forest soil region comprises: The multi-layer multi-source data comprises soil layer multi-source data, vegetation layer multi-source data, environmental layer multi-source data and remote sensing layer multi-source data; The soil layer multi-source data comprises soil organic carbon content, soil density, soil layer thickness, soil active carbon content and soil respiration flux sequence, wherein the soil respiration flux is the rate of soil carbon dioxide release; The vegetation layer multi-source data comprises the area of the investigation sample plot, the increase of the underground root biomass and the increase of the aboveground biomass of the investigation sample plot relative to the last collected multi-layer multi-source data, and the amount of litter at the investigation sample plot; The multi-source data of the environment level includes soil temperature, soil moisture content, air temperature, annual average precipitation, and evaporation, and the soil temperature and soil moisture content are collected by deploying soil temperature sensors and soil moisture sensors in the forest soil area, and the air temperature, annual average precipitation, and evaporation are collected based on observation results of a weather station; The multi-source data of the remote sensing level includes remote sensing images of the forest soil area at different wave bands.

4. The method of claim 3, wherein the method comprises: determining the soil carbon sink capacity of the forest soil based on the deep learning model. The multi-layer multi-source data of the forest soil area is subjected to layer feature extraction, including: The multi-layer features include soil layer features, vegetation layer features, environment layer features, and remote sensing layer features; The multi-source data of the soil layer is subjected to feature extraction to obtain soil layer features, and the soil layer features include soil organic carbon density, active carbon ratio, and soil respiration rate variation coefficient; The multi-source data of the vegetation layer is subjected to feature extraction to obtain vegetation layer features, and the vegetation layer features include vegetation carbon input rate, wherein the vegetation carbon input rate is the amount of carbon fixed and input into the soil through photosynthesis per unit time, reflecting the source strength of carbon sink; The calculation method of the vegetation carbon input rate is: ； in, This indicates the rate of vegetation carbon input in forest soil areas. This indicates the area of ​​the surveyed plots during the multi-source data acquisition process at the vegetation level. This indicates the periodic collection time interval for multi-level, multi-source data. This indicates the increase in aboveground biomass in the surveyed plots relative to the previous collection of multi-level, multi-source data. R represents the increase in underground root biomass in the surveyed plot relative to the previous multi-level, multi-source data collection, while R represents the amount of litter at the surveyed plot. This indicates the mass proportion of carbon in plant biomass, and is set as follows: It is 50%; The multi-source data of the environment layer is subjected to feature extraction to obtain environment layer features, and the environment layer features include temperature and humidity adjustment coefficient and water-heat stress index; The calculation method of the temperature and humidity adjustment coefficient is: ； wherein, represents a temperature and humidity adjustment coefficient, represents a soil temperature, represents a soil water content, represents a soil reference temperature, represents a soil reference water content, represents a temperature sensitivity coefficient, set to 2, represents an exponential function with a natural constant as a base, represents a water adjustment width parameter, set to 0.05; The multi-source data of the remote sensing layer is subjected to feature extraction to obtain remote sensing layer features, and the remote sensing layer features include enhanced vegetation index and enhanced vegetation index change rate.

5. The method for multi-scale assessment and calculation of forest soil carbon sequestration capacity based on deep learning as described in claim 4, characterized in that, Multi-scale multi-layer features of the forest soil area in time series scale change and spatial scale change are constructed, including: The multi-layer features of the collected forest soil area are subjected to standardization processing to obtain standardized layer features; K-neighbor layer feature sequences of the forest soil area and time series sequences of the standardized layer features are obtained, the K-neighbor layer feature sequences of the forest soil area are sequences composed of neighbor standardized layer features of K nearest neighboring forest soil areas of the forest soil area, K is set to 9, and the time series sequences of the standardized layer features are sequences composed of standardized layer features of the collected L multi-layer multi-source data of the forest soil area in the past, L is set to 5; The standardized layer features of the forest soil area, the K-neighbor layer feature sequences, and the time series sequences of the standardized layer features of the forest soil area are used as multi-scale multi-layer features of the forest soil area.

6. The method of claim 5, wherein the method comprises: Multi-layer and multi-scale correlation fusion is performed on the multi-scale multi-layer features to obtain hierarchical change fusion features of the forest soil area, including: Multi-scale multi-layer features of a forest soil area Q to be correlated and fused are obtained, the forest soil area Q is any forest soil area in the first to Nth forest soil areas; Standardized layer features of the forest soil area Q in the multi-scale multi-layer features are extracted, and the standardized layer features are mapped to a high-dimensional space by using an encoder to obtain high-dimensional mapping features of the forest soil area Q; The K-neighbor layer feature sequence of the forest soil region Q in the multi-scale multi-layer feature is extracted, and the distance from the center of the adjacent forest soil region associated with each neighbor normalized layer feature to the forest soil region Q is obtained. The weight of the neighbor normalized layer feature is calculated based on the distance. The graph convolution network is applied to the graph convolution of the K-neighbor layer feature sequence, so as to obtain the spatial fusion feature of the forest soil region Q. The normalized layer feature time sequence of the forest soil region Q in the multi-scale multi-layer feature is extracted, and the extracted normalized layer feature time sequence is input into the pre-trained bidirectional long short-term memory network, so as to capture the time dimension change rule of the forest soil region Q, and output the time sequence dependent feature representing the dynamic change trend of the forest soil region Q. The high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature of the forest soil region Q are taken as the hierarchical change fusion feature of the forest soil region Q.

7. The method of claim 1, wherein the method comprises: determining the soil carbon sink capacity of the forest soil based on the deep learning model. The soil carbon sink capacity evaluation model includes an input layer, a soft gating modulation unit, a double-path multi-layer perceptron and an output layer, including: The input layer is used to receive the hierarchical change fusion feature of the forest soil region; The soft gating modulation unit is used to calculate the gating modulation weight of the high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature in the hierarchical change fusion feature, and to weight fuse the high-dimensional mapping feature, the spatial fusion feature and the time sequence dependent feature, so as to obtain the multi-scale hierarchical change fusion feature; The double-path multi-layer perceptron includes a carbon sink rate-multi-layer perceptron and a carbon stock-multi-layer perceptron, which respectively receive the multi-scale hierarchical change fusion feature and respectively output the carbon sink rate evaluation result and the carbon stock evaluation result, which are taken as the carbon sink capacity of the forest soil region.

8. The method of claim 7, wherein the method comprises: The soil carbon sink capacity evaluation model receives the hierarchical change fusion feature of the forest soil region, and evaluates the carbon sink capacity of the forest soil region, including: The hierarchical change fusion feature of the forest soil region is input into the soil carbon sink capacity evaluation model, and the input layer, the soft gating modulation unit, the double-path multi-layer perceptron and the output layer in the soil carbon sink capacity evaluation model process the hierarchical change fusion feature in turn, so as to obtain the carbon sink capacity of the forest soil region.

9. A forest soil carbon sink capacity multi-scale assessment estimation system, characterized in that, The forest soil carbon sink capacity multi-scale evaluation system includes a region division device, a feature extraction module and a carbon sink capacity evaluation device: The region division device is used to divide the forest to be evaluated into a plurality of forest soil regions by using a clustering method based on regional carbon sink potential, and periodically collect multi-layer multi-source data of each forest soil region; The feature extraction module is used to extract layer features from the multi-layer multi-source data of the forest soil region, to obtain multi-layer features of each forest soil region, to construct multi-scale multi-layer features of the forest soil region in time scale change and spatial scale change based on the multi-layer features, and to correlate and fuse the multi-scale multi-layer features in multi-layer and multi-scale, so as to obtain the hierarchical change fusion feature of the forest soil region; The carbon sink capacity evaluation device is used for receiving hierarchical change fusion features of a forest soil area by using a soil carbon sink capacity evaluation model, and evaluating the carbon sink capacity of the forest soil area. To achieve a deep learning-based multi-scale evaluation and calculation method for forest soil carbon sink capacity according to any one of claims 1-8.

Images