Deep learning-based cultivated land quality grade evaluation method and system

Abstract

The invention provides a cultivated land quality grade evaluation method and system based on deep learning, and relates to the technical field of cultivated land quality evaluation, and the method comprises the steps: obtaining a target cultivated land hyperspectral remote sensing image, screening each nutrient index sensitive wave band, and obtaining time sequence wave band data, constructing a plough layer nutrient inversion model by using a deep learning algorithm, inputting time sequence wave band data, and outputting a spatial distribution inversion result; constructing an evaluation index set according to site conditions and climate factor data, screening evaluation indexes by using a Delphi method, determining weights corresponding to the evaluation indexes to obtain a minimum data set, performing normalization processing on numerical values of the evaluation indexes, calculating membership values of the evaluation indexes, and obtaining a minimum data set; and carrying out weighted summation processing on the membership value and the weight to obtain a soil fertility index so as to determine the cultivated land quality grade. The accuracy of the inversion result can be improved.

Description

Technical Field

[0001] This application relates to the technical field of farmland quality assessment, and in particular to a method and system for farmland quality grading based on deep learning. Background Technology

[0002] As a fundamental resource for agricultural production, arable land's quality directly impacts grain yield, agricultural product quality, and sustainable agricultural development. With the continuous growth of the global population and the acceleration of urbanization, arable land resources face increasingly severe challenges, such as declining quantity and deteriorating quality. Accurate and scientific evaluation of arable land quality is crucial for rationally planning land use, formulating agricultural policies, ensuring national food security, and promoting the improvement of the agricultural ecological environment.

[0003] Chinese invention patent application publication number CN110046415A discloses a spatiotemporally refined remote sensing dynamic inversion method for soil organic matter content. This method integrates Sentinel-2 and Sentinel-3 satellite image data, uses the Gram-Schmidt spectral sharpening method to achieve high spatial resolution enhancement of the images, and combines multiple vegetation indices and multi-band reflectance to construct an inversion model to obtain a spatiotemporally refined map of soil organic matter content changes within a regional scale.

[0004] However, soil organic matter content is affected by various environmental factors and exhibits significant temporal variations. The aforementioned patents, when processing time-series data, primarily rely on static inversion models, failing to fully utilize the dynamic information of the time-series data and thus limiting the accuracy of the inversion results to some extent. Summary of the Invention

[0005] To improve the accuracy of the inversion results, this application provides a method and system for evaluating the quality grade of cultivated land based on deep learning.

[0006] Firstly, this application provides a deep learning-based method for evaluating the quality grade of arable land, employing the following technical solution: A deep learning-based method for evaluating the quality grade of arable land includes the following steps: Hyperspectral remote sensing images of the target farmland are acquired, and the sensitive bands of each nutrient index are obtained through spectral bands. The time-series band data of each nutrient index are obtained, and a topsoil nutrient inversion model is constructed using a deep learning algorithm. The time-series band data of each nutrient index are input into the topsoil nutrient inversion model, and the spatial distribution inversion results of topsoil nutrients are output. Based on the spatial distribution inversion results of topsoil nutrients, site condition data, and climate factor data, an evaluation index set is constructed. The Delphi method is used to screen the evaluation indicators and output the weights of each evaluation indicator to obtain the minimum dataset for farmland quality evaluation. The evaluation indicators in the minimum dataset are normalized, and the membership degree value of each evaluation indicator is calculated through the membership function. The soil fertility index is obtained by weighting the indicators together with the weights. The farmland quality grade is determined based on the soil fertility index.

[0007] This application acquires hyperspectral remote sensing images of target cultivated land and filters sensitive bands for each nutrient index to obtain time-series band data for each nutrient index. Hyperspectral remote sensing images can capture subtle spectral differences among different nutrients in cultivated land soil, while time-series data considers nutrient changes over time, making the acquired nutrient information more comprehensive and accurate. Subsequently, a deep learning algorithm is used to construct a topsoil nutrient inversion model, which can automatically extract features related to nutrient indices from complex time-series band data and establish a precise mapping relationship between input data and nutrient content, thereby outputting more accurate topsoil nutrient spatial distribution inversion results, providing a key basis for cultivated land quality assessment.

[0008] Subsequently, this application constructs an evaluation index set based on the spatial distribution inversion results of topsoil nutrients, site condition data, and climate factor data, comprehensively considering multiple factors affecting arable land quality, thus providing a more comprehensive reflection of arable land quality characteristics. Then, the Delphi method is used to screen the evaluation indicators and output the weights of each indicator, making the selection of evaluation indicators and the determination of their weights more scientific and reasonable.

[0009] This application allows for the regular monitoring and evaluation of arable land quality, enabling timely detection of changing trends. It can also issue timely warnings when arable land quality declines or potential risks emerge.

[0010] Optionally, the sensitive bands for screening nutrient indicators by spectral bands include: The multicollinearity between spectral bands is eliminated by the continuous projection algorithm, and the spectral bands corresponding to the first N projection vectors are integrated into the first set. Calculate the correlation coefficient between each spectral band and nutrient index, and integrate the spectral bands with an absolute value of the correlation coefficient greater than the preset correlation coefficient threshold into a second set; The intersection of the first set and the second set is taken as the sensitive band.

[0011] This application employs a continuous projection algorithm (SPA) to progressively select band combinations with minimal redundancy from the original spectral bands through vector projection. This reduces redundant information between bands, making the bands in the first set more independent and representative, thus improving the accuracy of nutrient index retrieval. Next, the correlation coefficient between each spectral band and the nutrient index is calculated. The larger the absolute value of the correlation coefficient, the stronger the linear relationship between the spectral band and the nutrient index, meaning the band is more sensitive to changes in the nutrient index. This application integrates spectral bands with absolute correlation coefficient values ​​greater than a preset threshold into a second set. This allows for the selection of bands closely related to nutrient indexes from numerous spectral bands, focusing on those most sensitive to nutrient information, further narrowing the range of sensitive bands and improving the targeting and effectiveness of the selection. This application uses the intersection of the first and second sets as the sensitive bands, ensuring that the selected bands are highly sensitive to nutrient information.

[0012] By adopting the above scheme, this application can more accurately capture the spectral characteristics of nutrient indicators, thereby improving the model's accuracy in retrieving nutrient content.

[0013] Optionally, the topsoil nutrient inversion model includes an iTransformer layer, which encodes the time-series spectral data of each nutrient index into a high-dimensional feature vector, and captures the change pattern of the nutrient index through adaptive nonlinear mapping in the high-dimensional feature space to obtain the mapping result. The mapping result is then converted into the predicted value of the nutrient index through a regression head. Based on the spatial geographic coordinates and predicted values ​​of each nutrient index sampling point, geostatistical methods are used to analyze the spatial autocorrelation characteristics of the nutrient index, generating the spatial distribution inversion results of topsoil nutrients.

[0014] Time-series spectral data contains spectral information about nutrient changes over time. Directly analyzing the raw data can lead to problems such as high data dimensionality and information redundancy. This application employs an iTransformer layer to encode the time-series spectral data of each nutrient index into high-dimensional feature vectors. By encoding into high-dimensional feature vectors, the most representative and discriminative features in the data can be extracted, noise and irrelevant information can be removed, and the data can be made more structured and separable in the feature space.

[0015] This application employs adaptive nonlinear mapping in a high-dimensional feature space, enabling better capture of nutrient index variation patterns. Nutrient changes in soil are often influenced by a complex interplay of factors, exhibiting nonlinear dynamic characteristics. Traditional linear models struggle to accurately describe these complex relationships, while adaptive nonlinear mapping automatically adjusts the mapping function based on data distribution and characteristics, more flexibly fitting the changing patterns of nutrient indices. Subsequently, a regression head transforms the mapping results into predicted values ​​of nutrient indices, achieving accurate predictions.

[0016] Finally, based on the spatial geographic coordinates and predicted values ​​of each nutrient index sampling point, this application uses geostatistical methods to analyze the spatial autocorrelation characteristics of the nutrient index. According to the spatial autocorrelation characteristics, geostatistical methods can use information from known sampling points to interpolate and estimate the nutrient content in unknown areas, thereby generating spatial distribution inversion results for topsoil nutrients. The generated spatial distribution inversion results for topsoil nutrients can visually demonstrate the distribution of nutrients in farmland.

[0017] Optionally, in the process of encoding the time-series spectral data of each nutrient index into a high-dimensional feature vector in the iTransformer layer, a self-attention mechanism is used to weight different spectral bands in the time-series spectral data.

[0018] Spectral bands in time-series spectral data do not exist in isolation; they may have complex interactions and dependencies. Self-attention mechanisms can capture these complex nonlinear relationships by calculating attention weights between bands. Furthermore, by calculating the attention weights between each band and other bands, self-attention mechanisms can highlight spectral bands that are more critical for predicting nutrient indicators.

[0019] Optionally, the iTransformer layer integrates an LSTM time-series module that captures the time-series features of time-series spectral data for the same nutrient index. The self-attention module to which the self-attention mechanism belongs is connected in parallel with the LSTM time-series module. The outputs of the self-attention module and the LSTM time-series module are fused through residual connection to output a joint feature vector, which is then updated into a high-dimensional feature vector.

[0020] In time-series spectral data, changes in nutrient indicators are often not short-term random fluctuations, but rather the result of long-term influence from multiple factors. The LSTM time-series module can process each time step in the time-series spectral data one by one, dynamically deciding which information to retain, which to forget, and how to update the current state based on the current input and the hidden states of previous time steps, thereby capturing the long-term trends and patterns of nutrient indicators over time.

[0021] The self-attention mechanism, by calculating the attention weights between each band and all other bands, highlights the spectral bands that are more critical for predicting nutrient indicators, as well as the complex combination patterns between bands. Unlike LSTM, which focuses on the temporal dimension of dependence, the self-attention module focuses more on uncovering the intrinsic connections between different spectral bands at the same time step.

[0022] This application addresses the vanishing gradient problem during deep network training by introducing residual connections, allowing gradients to flow directly through network layers and enabling the network to learn more complex feature representations. In this application, the outputs of the self-attention module and the LSTM temporal module are fused through residual connections, allowing temporal features and inter-band interaction features to be added together positionally. Residual connections enable the model to retain the output information of the original modules during the fusion process, while simultaneously learning the differences and complementary information between them, thereby generating a more comprehensive and richer joint feature vector.

[0023] Optionally, the method further includes: Based on domain knowledge, a knowledge graph is constructed. Based on the correlation between arable land nutrient data, site condition data, and climate factor data in the knowledge graph, the scope of spectral band selection, the scope of site condition data collection, and the scope of climate factor data collection are initially defined.

[0024] Knowledge graphs clearly visualize the relationships between soil nutrient data, site condition data, and climate factor data. Based on these relationships, it is possible to identify which factors have a significant impact on soil nutrients, thereby enabling a targeted determination of the data collection scope and minimizing the need to blindly collect large amounts of irrelevant data.

[0025] The initially defined data collection scope was determined based on the relationship between various factors and arable land nutrients within the domain knowledge framework, making the collected data more likely to contain information closely related to nutrient prediction and analysis. Taking soil phosphorus research as an example, the knowledge graph shows that soil pH and evaporation have a significant impact on soil phosphorus availability. When collecting data, following the guidance of the knowledge graph, focusing on collecting data under different soil pH ranges and evaporation conditions can more accurately reflect the relationship between soil phosphorus and these factors, improving the effectiveness of the data.

[0026] By limiting the data collection scope through knowledge graphs, this application effectively controls the data dimension and scale of the input model, reduces interference from irrelevant data, and eliminates the need for the model to process a large amount of redundant information, thereby reducing the complexity of the model.

[0027] Optionally, when using the Delphi method to screen evaluation indicators and their corresponding weights, the method further includes: The weight relationships in the knowledge graph are transformed into a reference weight matrix, and the initial weight values ​​of each evaluation indicator in the first round of the Delphi method questionnaire are generated based on the reference weight matrix. After using the Delphi method to screen evaluation indicators and their corresponding weights, the weight adjustment strategy for each evaluation indicator is obtained. Based on the weight adjustment strategy, the adjusted weight values ​​of each evaluation indicator are obtained, and the knowledge graph is corrected using the adjusted weight values.

[0028] This application transforms the weight relationships of knowledge graphs into a reference weight matrix, converting implicit and scattered knowledge into an intuitive numerical form. This provides initial weight values ​​based on professional knowledge for the first round of the Delphi method questionnaire. Reasonable initial weights offer experts a clear starting point, enabling them to quickly focus on the parts of the weights that need adjustment during evaluation, rather than starting from a completely blank state to consider the weights of each indicator. This helps experts more efficiently utilize their professional knowledge and experience to evaluate and revise the initial weights.

[0029] This application obtains the weight adjustment strategy for each evaluation indicator and obtains the adjusted weight value based on the strategy, enabling dynamic optimization of the weight. Subsequently, the knowledge graph is modified using the adjusted weight value, which can integrate the new knowledge and experience accumulated by experts in the Delphi method evaluation process into the knowledge graph, enriching and improving the content of the knowledge graph and helping to maintain the timeliness and accuracy of the knowledge graph.

[0030] Optionally, the method further includes: Based on geospatial rules, the knowledge graph is divided into several subgraphs. The subgraph corresponding to the target farmland is queried by the latitude and longitude coordinates of the target farmland, and it is denoted as the target graph. Extract the region weight matrix from the target map and use the region weight matrix as the new reference weight matrix.

[0031] Geospatial rules can rationally divide a knowledge graph into several sub-graphs based on geographical features, administrative divisions, and other factors. Each sub-graph represents a specific geographical region and has relatively independent geographical attributes and characteristics. Using the latitude and longitude coordinates of the target farmland, this application can quickly and accurately locate the target map where the target farmland is located within the segmented sub-graphs. By extracting a regional weight matrix from the target map, this application can fully consider the special circumstances of the region where the target farmland is located, making the weight settings more consistent with the reality of that region.

[0032] Optionally, the adjusted weight values ​​of each evaluation indicator are obtained based on the weight adjustment strategy, including: Calculate the KL divergence between the weights corresponding to each evaluation indicator and the initial weights of each evaluation indicator. If the KL divergence is greater than a preset divergence threshold, the weights corresponding to each evaluation indicator are used as the initial strategy distribution in the game, and the adjustment direction of the weights of each evaluation indicator is used as the movement direction in the strategy space. Using a pre-constructed multi-armed slot machine model, the weight adjustment range of each questionnaire dimension is used as an independent arm. A reward function is constructed based on the change in KL divergence. The optimal adjustment strategy is selected using the UCB algorithm. A voting mechanism is used to confirm whether to adjust the weights corresponding to each evaluation indicator according to the optimal adjustment strategy. If so, the adjusted weights of each evaluation indicator are output.

[0033] This application calculates the KL divergence between the weights of each evaluation index and the initial weights, thus quantitatively reflecting the degree of deviation between the current and initial weights. This application treats the weight adjustment magnitude of each evaluation index as an independent arm, simulating the weight adjustment process as an exploration and utilization of multiple optional strategies. The reward function is crucial in reinforcement learning, guiding the agent to make correct decisions. Constructing a reward function based on the change in KL divergence enables the model to optimize in the direction of reducing KL divergence (i.e., making the current weights closer to a reasonable value) when adjusting weights. Positive rewards are given when KL divergence decreases, and negative rewards are given when it increases, thereby incentivizing the model to find the optimal adjustment strategy.

[0034] In this application, the UCB algorithm calculates an upper bound for each adjustment strategy based on its historical performance and uncertainty, and selects the strategy with the largest upper bound as the optimal strategy. The multi-armed slot machine model can gradually adapt to this uncertainty by continuously trying different adjustment ranges (independent arms) and based on feedback information, thus finding the optimal adjustment strategy in an uncertain environment.

[0035] In this application, a voting mechanism is used to confirm whether to adjust the weights of each evaluation indicator according to the optimal adjustment strategy. This can minimize unreasonable weight adjustments caused by individual factors or erroneous judgments and increase the stability and reliability of the decision.

[0036] Secondly, this application provides a deep learning-based system for evaluating the quality of cultivated land, employing the following technical solution: A deep learning-based arable land quality grading system includes: a processor and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes a computer program stored on the computer-readable storage medium, it implements the method as described in the first aspect.

[0037] In summary, this application includes at least one of the following beneficial technical effects: 1. This application acquires hyperspectral remote sensing images of target cultivated land and filters sensitive bands for each nutrient index to obtain time-series band data for each nutrient index. Hyperspectral remote sensing images can capture subtle spectral differences among different nutrients in cultivated land soil, while time-series data considers nutrient changes over time, making the acquired nutrient information more comprehensive and accurate. Subsequently, a deep learning algorithm is used to construct a topsoil nutrient inversion model, which can automatically extract features related to nutrient indices from complex time-series band data and establish a precise mapping relationship between input data and nutrient content, thereby outputting more accurate topsoil nutrient spatial distribution inversion results, providing a key basis for cultivated land quality assessment.

[0038] 2. This application constructs an evaluation index set based on the spatial distribution inversion results of topsoil nutrients, site condition data, and climate factor data, comprehensively considering multiple factors affecting arable land quality, thus providing a more comprehensive reflection of arable land quality characteristics. Subsequently, the Delphi method is used to screen the evaluation indicators and output the weights of each indicator, making the selection of evaluation indicators and the determination of their weights more scientific and reasonable.

[0039] 3. This application allows for the regular monitoring and evaluation of arable land quality, enabling timely detection of changing trends. It can also issue timely warnings when arable land quality declines or potential risks emerge. Attached Figure Description

[0040] Figure 1 This is a flowchart of Embodiment 1 of this application. Detailed Implementation

[0041] The present application will be further described in detail below with reference to the accompanying drawings.

[0042] Example 1: This example discloses a method for evaluating the quality grade of cultivated land based on deep learning, referring to... Figure 1 The method includes: S11 acquiring inversion results, S12 evaluating the target arable land grade. First, hyperspectral remote sensing images of the target arable land are acquired, and time-series band data are obtained by screening sensitive bands for each nutrient index. A deep learning algorithm is used to construct a nutrient inversion model of the topsoil layer and input the time-series band data to output spatial distribution inversion results. Then, an evaluation index set is constructed based on the inversion results combined with site conditions and climate factor data. The evaluation indexes are screened using the Delphi method, and the weights corresponding to each evaluation index are determined to obtain the minimum dataset. The values ​​of the evaluation indexes are normalized, and the membership values ​​of the evaluation indexes are calculated. The soil fertility index is obtained by weighted summation of the membership values ​​and weights to determine the arable land quality grade. The execution process of each step in this embodiment is as follows: S11 acquires inversion results by obtaining hyperspectral remote sensing images of the target farmland through satellite or aerial remote sensing.

[0043] Different nutrient indicators in soil, such as nitrogen, phosphorus, and potassium, exhibit different response characteristics to reflectance in different spectral bands. This embodiment obtains the sensitive bands of each nutrient indicator by acquiring spectral band data for each nutrient indicator.

[0044] In this embodiment, the sensitive bands for obtaining nutrient indicators through spectral bands include: Let the number of soil samples be n and the total number of spectral bands be m. Construct a reflectance matrix. The reflectivity matrix X is centered to obtain .

[0045] Initialize the first set Randomly select an initial band from all spectral bands and add it to the first set S.

[0046] Calculate the projection vectors of the remaining spectral bands relative to the spectral bands in the current first set S, using the following projection criteria:

[0047] in, Let be the projection value of the j-th spectral band. The larger the projection value, the lower the information redundancy between the spectral band and the selected spectral bands, and the stronger the complementarity. This is the transpose of the reflectance matrix after centering. Let be the column vector corresponding to the j-th spectral band in the reflectivity matrix.

[0048] Select the spectral band with the largest projection value and add it to the first set S. Then recalculate the projection values ​​of the remaining spectral bands until N spectral bands are selected. Finally, output the first set S containing the selected N spectral bands.

[0049] For a single nutrient indicator (such as total nitrogen content), construct a nutrient indicator vector. ,in This represents the measured value of this nutrient index for the i-th soil sample.

[0050] Calculate the correlation coefficient between the j-th spectral band and the measured value of this nutrient index of the i-th soil sample, wherein the correlation coefficient... The calculation model is as follows:

[0051] in, This represents the reflectance of the i-th soil sample in the j-th spectral band. Let j be the average reflectance of the j-th spectral band. This represents the measured value of this nutrient index in the i-th soil sample. denoted as the average value of this nutrient index, and n is the number of soil samples.

[0052] correlation coefficient Spectral bands with a correlation coefficient greater than a preset threshold (e.g., 0.6) are integrated into a second set.

[0053] The spectral bands common to both the first and second sets are used as sensitive bands.

[0054] In other embodiments, the sensitive bands for different nutrient indicators can also be obtained using stepwise regression analysis. By gradually introducing or removing band variables, the combination of sensitive bands is determined based on the significance and goodness of fit of the regression model. The screening process for the sensitive bands of each nutrient indicator is as follows: Let the number of soil samples be n and the total number of spectral bands be m. Construct a reflectance matrix. ,in, This represents the reflectance of the i-th soil sample in the j-th spectral band.

[0055] For a single nutrient indicator (such as total nitrogen content), construct a nutrient indicator vector. ,in This represents the measured value of this nutrient index for the i-th soil sample.

[0056] Noise was removed using Savitzky-Golay smoothing (window size 5-11 points), and interference from soil particle size and moisture was eliminated through Standard Normal Variable Transform (SNV). The calculation formula is as follows:

[0057] in, Let be the reflectance of the i-th soil sample after normalization in the j-th spectral band. Let be the mean reflectance of the j-th spectral band. Let be the standard deviation of the reflectance of the j-th spectral band.

[0058] The data is divided into a training set (used to construct the significance test model) and a validation set (used to verify the effectiveness of sensitive bands) in a 7:3 ratio. Stratified sampling is used to ensure that the nutrient content distribution of the training set and the validation set is consistent. In this embodiment, the significance test model adopts a linear regression model, which fits the theoretical curve of the data points in the training set.

[0059] The F-statistic is used to test the significance of introducing or removing band variables. The calculation model is as follows:

[0060] in, , , where k is the sum of squared residuals before and after the adjustment of the significance test model, n is the number of bands introduced or removed, p is the total number of bands in the current significance test model, and the significance level is set to 0.05. .

[0061] For example, when n is 60, p is 4, and k is 1. It is 12.8. It is 10.3.

[0062] F=(12.8-10.3)÷1÷10.3÷(60-4-1)≈13.37 Query the F-distribution table, Since 13.37 > 4, the introduction of the new spectral band is statistically significant and can be included in the significance test model.

[0063] This embodiment uses the coefficient of determination. The root mean square error (RMSE) is used to evaluate the performance of the regression model.

[0064]

[0065] Where n is the number of soil samples. Let be the measured value of this nutrient index for the i-th soil sample. Let be the predicted value of this nutrient index for the i-th soil sample. The nutrient index is the average content calculated based on all historical data of the i-th soil sample.

[0066] The variance inflation factor (VIF) is used to test the correlation between spectral bands. In this embodiment, the VIF value needs to be less than 10. The calculation model for the variance inflation factor of the j-th spectral band is as follows:

[0067]

[0068]

[0069] in, This is the square of the multiple correlation coefficient between the j-th spectral band and other spectral bands. The multiple correlation coefficient is the coefficient of determination (i.e., goodness of fit) obtained by constructing a multiple linear regression model with the reflectance of the j-th spectral band as the dependent variable and the reflectance of all other candidate spectral bands as the independent variables. Its core function is to quantify the linear correlation strength between the j-th spectral band and other spectral bands. The closer it is to 1, the more the information in that spectral band can be replaced by other spectral bands, and the stronger the multicollinearity. This represents the adjusted sum of squared residuals from the regression model. It represents the sum of squares of the j-th spectral band. This represents the reflectance of the i-th soil sample in the j-th spectral band; Let be the average reflectance of the j-th spectral band.

[0070] This embodiment uses a stepwise regression algorithm to screen the sensitive bands of each nutrient index. The process includes stages 1 to 4.

[0071] Phase 1: Set the initial model as an intercept term model:

[0072] in, Let be the predicted value of this nutrient index for the i-th soil sample. This is the average value of the nutrient index.

[0073] at this time,

[0074] in, For the initial sum of squared residuals, This represents the measured value of this nutrient index in the i-th soil sample, where n is the number of soil samples. Let be the predicted value of this nutrient index for the i-th soil sample.

[0075]

[0076] Where p is the total number of bands in the current significance test model, with an initial value of 0.

[0077] Set a filtering threshold: The filtering threshold includes an introduction threshold. and removal threshold .

[0078] Phase 2: Gradually introduce sensitive bands (forward selection) By iterating through all unintroduced spectral bands, single-band regression models are constructed for each band:

[0079] in, Let be the predicted value of the nutrient index corresponding to the j-th spectral band. Let be the reflectance of the j-th spectral band. Let be the regression coefficient for the j-th spectral band; Let be the intercept of the j-th spectral band.

[0080] Calculate the F-statistic for each spectral band and select... Furthermore, the spectral band s with the largest F-statistic is introduced into the model, and the model is updated as follows:

[0081] in, Let be the predicted value of the nutrient index after introducing the spectral band s, and p be the number of bands in the current model. Let be the reflectance of the s-th spectral band. Let be the regression coefficient for the s-th spectral band; Let be the intercept of the s-th spectral band.

[0082] Then, recalculate. and R 2 .

[0083] Repeat the two steps of phase 2 until no condition is met. After the spectral bands can be introduced, we enter stage 3.

[0084] Phase 3: Eliminating redundant bands (removing them backward) For each spectral band s in the current model, calculate its F-statistic after removal (i.e., the significance of the band's contribution to the model).

[0085] If there exists a spectral band that satisfies Remove the spectral bands with the smallest F-values, update the model, and recalculate RMSE and R. 2 The regression coefficients for each remaining spectral band.

[0086] Repeat the two steps of stage 3 until all spectral bands in the model are... Entering Phase 4.

[0087] Phase 4: Return to Phase 2 and check if there are any new spectral bands that meet the requirements. It can be introduced that (because after removing redundant spectral bands, the change in the model's degrees of freedom may increase the F-statistic for some spectral bands).

[0088] Phase 2 and Phase 3 are executed alternately until any of the following termination conditions are met, at which point the final set of sensitive bands is output.

[0089] Condition 1: No new spectral bands can be introduced and no redundant spectral bands can be eliminated; Condition 2: Model R 2 The increase is less than 0.01; Condition 3: The number of model bands p is less than or equal to 10; Condition 4: Validation set RMSE val It reaches the minimum value.

[0090] A topsoil nutrient inversion model is constructed using a deep learning algorithm. The topsoil nutrient inversion model aims to use deep learning technology to accurately predict topsoil nutrient indicators from time-series spectral data and further generate a spatial distribution map of nutrient content.

[0091] The topsoil nutrient inversion model includes an iTransformer layer, which is connected in parallel with an LSTM time series module and a self-attention module, and adopts a residual connection structure to capture the temporal dynamic features and band-weighted features of the time series spectrum.

[0092] The LSTM timing module consists of an input layer, a bidirectional LSTM layer, and an output layer, with the following specific settings: Input layer: Dimension equals the product of time step and the type of time series band data; Bidirectional LSTM layer: includes 2 hidden layers with a hidden dimension of 64, and bidirectional output fusion (capturing forward / reverse time series features). Output layer: The output of the bidirectional LSTM layer (output dimension is 128) is mapped to 64 dimensions through a fully connected layer, and the output is a temporal feature vector.

[0093] Forward computation: Input time-series spectral data → Output features of each time step from the bidirectional LSTM layer → Extract the features of the last time step → Fuse with a fully connected layer → Output time-series feature vector (64 dimensions).

[0094] The self-attention module consists of an input layer, a QKV mapping layer, a multi-head self-attention layer, and an output layer, with the following specific settings: Input layer: Dimension equals the product of time step and the type of time series band data; QKV mapping layer: Maps input to query (Q), key (K), and value (V), each with a dimension of 64; Multi-head self-attention layer: Employs 4-head attention, with attention weights calculated using the following formula:

[0095] Where d is the attention head dimension, and in this embodiment, the value of d is 16, and the spectral values ​​of each band are weighted by a weight matrix; Output layer: The attention output is mapped to 64 dimensions through a fully connected layer, and the output band-weighted feature vector is output.

[0096] Forward computation: Input time-series spectral data → QKV mapping → Multi-head self-attention weighting → Fully connected layer fusion → Output band-weighted feature vector (64 dimensions).

[0097] The residual connection is used to add the time-series feature vector output by the LSTM time-series module to the band-weighted feature vector output by the self-attention module element-wise, and output a 64-dimensional joint feature vector, i.e., a high-dimensional feature vector.

[0098] This embodiment captures the changing patterns of nutrient indicators through an adaptive nonlinear mapping in a high-dimensional feature space. This adaptive nonlinear mapping is implemented using a multilayer perceptron, which contains multiple hidden layers, each employing a different activation function. The network structure and parameter design of the multilayer perceptron are shown below: The first hidden layer contains 128 neurons and uses the ReLU activation function for initial nonlinear transformation. The second hidden layer contains 64 neurons and uses the LeakyReLU activation function to alleviate gradient vanishing and improve fitting. The third hidden layer contains 32 neurons and uses the Tanh activation function to compress features to [-1, 1]. Output layer: The number of neurons in the output layer is the same as the number of nutrient indicators, and it outputs a nutrient feature vector, i.e., the mapping result.

[0099] The mapping results are converted into predicted values ​​of nutrient indicators using a regression head. The mapping model of the regression head is as follows:

[0100] Among them, H Let be the predicted value of nutrient index x, c and e be the parameters of the regression head, which are obtained by least squares fitting, and x be the numerical value of the nutrient index in the mapping result.

[0101] This embodiment optimizes the parameters of the regression head by fitting the nutrient feature vector in historical data with the true values ​​of nutrient indicators in historical data and minimizing the mean square error between the predicted and true values.

[0102] Based on the spatial geographic coordinates and predicted values ​​of each nutrient index sampling point, geostatistical methods are used to analyze the spatial autocorrelation characteristics of the nutrient index, generating a spatial distribution map of nutrient content in the topsoil layer. The process is as follows: The global Moran index I is calculated. The global Moran index is used to assess the spatial autocorrelation characteristics of nutrient indicators. The calculation model for the global Moran index is as follows:

[0103]

[0104] in, Let be the spatial association weight between the i-th soil sample point and the j-th soil sample point. Let be the sampling point distance between the i-th soil sample and the j-th soil sample. Let be the predicted nutrient value for the i-th soil sample. Let M be the predicted nutrient value for the j-th soil sample, and M be the number of soil samples. This is the global mean of the predicted nutrient values.

[0105] A global Moran's index greater than 0 indicates a positive spatial correlation, meaning the sampling points are clustered. A global Moran's index less than 0 indicates a negative spatial correlation, meaning the sampling points are discretely distributed. A global Moran's index equal to 0 indicates a random distribution of sampling points.

[0106] With fixed spatial correlation weights, the predicted nutrient values ​​of each soil sample point are randomly rearranged to obtain n randomly rearranged samples, the same number as the soil samples. For each randomly rearranged sample, the random Moran index is calculated using the global Moran index calculation model, resulting in a random index distribution.

[0107] The percentage of random samples with a random Moran index greater than or equal to the global Moran index is counted. If the percentage is less than 0.05, it indicates a significant spatial correlation; otherwise, it indicates an insignificant spatial correlation, and the number of soil samples should be increased.

[0108] If the spatial correlation is significant, the distance h between the sampling points is divided into several intervals.

[0109] Calculate the variogram for each interval. The variogram is used to describe the spatial variability of nutrient content, and its calculation formula is as follows:

[0110] in, This represents the average of the squared differences in nutrient values ​​among all pairs of sampling points with a sampling point spacing of h. Let be the predicted nutrient value for the i-th soil sample. The predicted nutrient value is located at a distance h from the sampling point of the i-th soil sample. This represents the number of sample pairs with a sampling point spacing of h.

[0111] Using the midpoint of the interval as the x-axis, the variogram Using the values ​​on the vertical axis, a scatter plot is drawn, which is the experimental variogram cloud. Based on the scatter plot, a spherical model is used to fit the theoretical variogram curve. Based on the theoretical variogram curve, the parameters of the Kriging interpolation algorithm are extracted, including nugget value, sill value, range, etc.

[0112] The nugget value is equal to the variogram value as the horizontal axis approaches 0. The sill value is the variogram value. Once the values ​​stabilize, when the sampling point spacing h increases to a certain threshold, the spatial correlation between sampling points completely disappears (i.e., exceeding the range). At this point, the nutrient differences between sampling points with different spacings tend to stabilize, and the variogram... The variability function no longer changes as the midpoint of the interval increases. The stable value, the range refers to the value of the interval that approaches 100% of the base value.

[0113] Based on the calculation results of the variogram, the spatial correlation is calculated, which is equal to the ratio of the nugget value to the sill value.

[0114] If the spatial correlation is greater than 0.75, the ordinary kriging algorithm is used for interpolation calculation; if the spatial correlation is not greater than 0.75, the co-kriging algorithm is used for interpolation calculation.

[0115] The interpolation results, together with the predicted values ​​of each nutrient index sampling point, constitute the spatial distribution inversion results of topsoil nutrients.

[0116] In the evaluation of S12 target arable land grade, this embodiment follows the following six principles in the selection of evaluation indicators.

[0117] ① The principle of importance: There are many factors that affect the fertility of arable land. The evaluation indicators are for the whole country. For a specific county administrative region, the factors that have the most significant impact on the local arable land fertility must be selected, rather than all of them.

[0118] ② The principle of stability: the selected evaluation factors must have relative stability over time. Selecting time-varying indicators will lead to instability in the evaluation results over time, resulting in poor guidance and practicality. However, arable land fertility is stable over a certain period unless there are significant external factors such as human activities.

[0119] ③ The principle of differentiation: This principle is divided into spatial differentiation and differentiation of indicator factors. One of the purposes of farmland fertility evaluation is to identify the dominant factors affecting farmland fertility and guide the optimal allocation of farmland resources. If the evaluation indicators do not differ spatially and in terms of attributes, they cannot reflect the differences in farmland fertility.

[0120] ④ The principle of accessibility: Indicators that can be obtained through conventional methods, such as soil nutrient content and irrigation / drainage conditions. While some indicators significantly impact arable land productivity, they are difficult or expensive to obtain, and currently not feasible. Examples include biological indicators such as the types and quantities of soil organisms and the quantity of certain enzymes in the soil.

[0121] ⑤ The principle of simplicity applies: more indicators are not necessarily better. Too many indicators will increase the workload and cost, and may not reveal the main factors affecting arable land fertility. Generally, 8-15 indicators can meet the evaluation needs.

[0122] ⑥ The principles of comprehensiveness and wholeness: Comprehensiveness means considering all types of arable land in the county, not just large areas. The characteristics of any arable land plot that can form a map feature on a 1:50,000 scale must be considered, rather than simply accepting the majority. The principle of wholeness refers to considering indicators that will have a significant impact on arable land fertility over time.

[0123] In this embodiment, an evaluation index set is constructed based on the spatial distribution inversion results of topsoil nutrients, site condition data, and climate factor data. The evaluation index set is shown in Table 1.

[0124] Table 1 Primary evaluation indicators Secondary evaluation indicators How to obtain Nutrient indicators in the topsoil Soil total nitrogen, available phosphorus, available potassium, organic matter content, etc. S11 Obtain the topsoil nutrient inversion results from the inversion results. Site condition indicators Topographic slope, altitude, soil texture (ratio of sand / silt / clay), soil layer thickness, and soil pH value. Topographic remote sensing interpretation, soil profile survey Climate factor indicators Average annual precipitation, average annual temperature, number of frost-free days, and accumulated temperature ≥10℃ Regional meteorological station observation data or meteorological remote sensing products This embodiment uses the Delphi method to screen evaluation indicators and output the weights of each evaluation indicator to obtain the minimum dataset for farmland quality evaluation, including steps 1 to 4: Step 1: Select 5 to 15 experts covering fields such as soil science, agricultural resources and environment, and geographic information science.

[0125] Step 2: Submit the initial evaluation index set, along with the definition of each index, data sources, and an explanation of its impact mechanism on arable land quality to the experts.

[0126] Experts are requested to score the indicators based on three dimensions: necessity, scientific validity, and operability (using a 1-5 point scale, with 5 being the best), and to provide suggestions for adding, deleting, or modifying the indicators.

[0127] After collecting the inquiry forms, the average score of each indicator is calculated, and evaluation indicators with an average score of less than 3 are removed.

[0128] Step 3: Feed back the statistical results of the first round (including indicator scores, reasons for exclusion, and suggestions for addition) to the experts and conduct the second round of inquiries.

[0129] Experts, taking into account the statistical results, re-scored and predicted the weights of the adjusted indicator set (the weights were based on a 100-point scale, with the sum of all indicator weights being 100).

[0130] Repeat steps 2-3 times until the coordination coefficient of the expert opinions is not less than 0.7, indicating that the expert opinions tend to be consistent, and then terminate the inquiry.

[0131] Step 4: Retain the evaluation indicators with an average score of not less than 4 after multiple rounds of inquiries to form the minimum dataset for farmland quality evaluation. Calculate the mean of the expert weights of each evaluation indicator in the minimum dataset, normalize the mean, and obtain the final weights of each evaluation indicator in the minimum dataset.

[0132] In other embodiments, the Analytic Hierarchy Process (AHP) is also used to verify the weights determined by the Delphi method. Specifically, a pairwise comparison judgment matrix is ​​constructed between the indicators (the judgment matrix adopts a 1-9 scale, where 1 indicates that the two indicators are equally important and 9 indicates that one indicator is much more important than the other). The consistency ratio CR of the judgment matrix is ​​calculated. If CR < 0.1, the weight allocation is considered reasonable; if CR ≥ 0.1, experts need to be invited to readjust the weights.

[0133] The evaluation indicators in the minimum dataset are normalized, and the membership values ​​of each evaluation indicator are calculated using membership functions. These are then weighted to obtain the soil fertility index. The process is as follows: Based on the relationship between evaluation indicators and arable land quality, the evaluation indicators are divided into three categories: Positive effect indicators: The higher the value of the evaluation indicator, the higher the quality of cultivated land (such as organic matter content, soil layer thickness, and annual precipitation).

[0134] Negative effect indicators: The higher the value of the evaluation indicator, the lower the quality of cultivated land (such as slope).

[0135] Interval effect index: When the evaluation index value is within a specific interval, the quality of cultivated land is relatively high; when it exceeds the interval, the quality of cultivated land declines (such as soil pH value).

[0136] For the positive effect index, the linear normalization method is used, and the calculation model is as follows:

[0137] in, This represents the linearly normalized value of the i-th positive effect index in the smallest dataset. This is the original value of the indicator. It represents the maximum value of the i-th positive effect index in the smallest dataset. It is the minimum value of the i-th positive effect index in the smallest dataset.

[0138] The negative effect index is calculated using the inverse linear normalization method, and the calculation model is as follows:

[0139] in, This is the normalized value of the i-th negative effect index in the smallest dataset. This is the original value of the indicator. It represents the maximum value of the i-th negative effect index in the smallest dataset. It is the minimum value of the i-th negative effect index in the minimum dataset.

[0140] For the interval effect index, a piecewise normalization method is used, and the calculation model is as follows:

[0141] in, This represents the normalized value of the i-th interval effect index in the smallest dataset. , For the interval endpoints, This represents the maximum value of the i-th interval effect index in the smallest dataset. It represents the minimum value of the i-th interval effect index in the smallest dataset.

[0142] For the positive effect index, an S-shaped membership function is chosen, and the calculation model is as follows:

[0143] in, It is an S-shaped membership function. · It is an exponential function. This is the slope parameter, and its value is a number greater than zero. The inflection point parameter corresponds to the membership degree. The index value when it equals 0.5 It is the normalized value of the i-th positive evaluation index in the minimum dataset.

[0144] For the negative effect index, we choose the inverse S-shaped membership function, and the calculation model is as follows:

[0145] in, It is an inverse S-shaped membership function. This represents the normalized value of the i-th negative effect index in the smallest dataset. · It is an exponential function. This is the slope parameter, and its value is a number greater than zero. These are the parameters for the inflection point.

[0146] For the interval effect index, a parabolic membership function is chosen, and the calculation model is as follows:

[0147] in, For parabolic membership functions, This represents the normalized value of the i-th interval effect index in the smallest dataset. These are the endpoint values ​​after normalization.

[0148] In other embodiments, the normalization method and membership function of the evaluation index can be set according to requirements.

[0149] The soil fertility index is obtained by weighting the membership values ​​of each evaluation indicator with their respective weights. The calculation model for the soil fertility index is as follows:

[0150] in, The soil fertility index of the target arable land. , Let be the final weight of the i-th evaluation metric in the minimum dataset. is the membership value corresponding to the i-th evaluation index in the minimum dataset.

[0151] The quality grade of arable land is determined based on the soil fertility index. In this embodiment, the correspondence between the arable land quality grade and the soil fertility index is shown in Table 2.

[0152] Table 2 Farmland quality grade soil fertility index Farmland quality grade soil fertility index First Class ≥0.9040 Sixth Class [0.7140,0.7520) Second Class [0.8660,0.9040) seventh class [0.6760,0.7140) Third Class [0.8280,0.8660) Eighth class [0.6380,0.6760) Fourth Class [0.7900,0.8280) Ninth class [0.6000,0.6380) Fifth Class [0.7520,0.7900) Tenth Rank <0.6000 Example 2: This example differs from Example 1 in that the method further includes: A knowledge graph is constructed based on domain knowledge. This knowledge graph uses soil nutrient data, site condition data, and climate factor data as entities. The latitude and longitude coordinates of the sampling points used to collect data for each entity are used as attributes. The directed edges of the knowledge graph include two types: site conditions and soil nutrients, and climate factors and soil nutrients. The weight calculation process for the directed edges is as follows:

[0153] in, Here, B is the correlation coefficient, and B is the sample size. This refers to the observed values ​​of site nutrients or climatic factors in a specific soil sample. It is the average value calculated from all historical site nutrient data or climate factor data for a given soil sample. This represents the observed values ​​of arable land nutrients in a specific soil sample. It is the average value of nutrient content in a soil sample, calculated based on all historical data of cultivated land.

[0154] Significantly related entity pairs are selected based on the correlation coefficient, i.e., entity pairs with a correlation coefficient greater than 0.6, and are denoted as target entity pairs. Entities in the target entity pairs are denoted as target entities. Directed edges between target entity pairs are retained, and directed edges between non-target entity pairs in the knowledge graph are deleted.

[0155] For each arable land nutrient index (such as organic matter), a geographically weighted regression model is constructed with it as the dependent variable and all target entities that are significantly related to it as independent variables. This model can estimate the regression coefficients for each sampling point. Then, all regression coefficients estimated by the geographically weighted regression model are standardized. The directed edge weights of target entity pairs in the knowledge graph are defined as the median of the standardized regression coefficients of the target entity pairs at all sampling points.

[0156] Based on the relationships between arable land nutrient data, site condition data, and climate factor data in the knowledge graph, the selection range of spectral bands, the collection range of site condition data, and the collection range of climate factor data are initially defined. The limitations in this embodiment are as follows: The spectral band screening range is limited to 760-900nm (organic matter sensitive band), 1550-1750nm (total nitrogen sensitive band), and 2100-2300nm (phosphorus and potassium sensitive band). Based on the above spectral band screening range, a total of 45 target bands were screened.

[0157] The scope of site condition data collection is limited to terrain slope ≤15°, because areas with slope >15° account for less than 2% of the target area, and the quality grade of cultivated land is generally lower than 8, so it has no key evaluation value. Soil texture is limited to alluvial soil and brown soil, which account for 92% of the soil types in the target area.

[0158] The climate factor data collection scope is limited to annual precipitation of 350-650mm and accumulated temperature of ≥10℃ of 3200-4800℃, covering 98% of the climate scenarios in the target area.

[0159] This embodiment, when using the Delphi method to screen evaluation indicators and their corresponding weights, also includes: The weights between entities in the knowledge graph are structured according to indicators to form a reference weight matrix. The specific process is as follows: Define the dimensions of the reference weight matrix, including row and column dimensions. The row dimension equals the sum of the types of site condition data and the types of climate factor data. The column dimension equals the total number of arable land nutrient indicators. The value of the element in the i-th row and j-th column of the reference weight matrix equals the correlation coefficient between the i-th influencing factor (i.e., site condition or climate factor) and the j-th arable land nutrient indicator.

[0160] Based on geospatial rules, in this embodiment, geospatial rules refer to administrative region division rules, which divide the knowledge graph into several subgraphs. Each subgraph contains all farmland nutrient entities, site condition entities, and climate factor entities of the corresponding administrative region. Using the query capability of the subgraphs, the subgraph corresponding to the target farmland can be queried by the latitude and longitude coordinates of the target farmland, and it is denoted as the target graph.

[0161] A regional weight matrix is ​​extracted from the target map. The construction method of the regional weight matrix is ​​the same as that of the reference weight matrix, and will not be repeated in this embodiment. Initial values ​​of each evaluation indicator in the first round of the Delphi method questionnaire are generated based on the regional weight matrix. When using the Delphi method to screen evaluation indicators, these initial values ​​are displayed to experts in the first round of the questionnaire as a scoring reference.

[0162] After using the Delphi method to screen evaluation indicators and their corresponding weights, a weight adjustment strategy for each evaluation indicator is obtained based on the weights of each evaluation indicator and the initial values ​​of each evaluation indicator. The adjusted weight values ​​of each evaluation indicator are then calculated based on the weight adjustment strategies, and the knowledge graph is corrected using the adjusted weight values.

[0163] The calculation of the adjusted weight values ​​for each evaluation indicator based on the weight adjustment strategy includes the following steps: Calculate the KL divergence between the weights corresponding to each evaluation index and the initial weights of each evaluation index. The calculation model for the KL divergence is as follows:

[0164] in, The KL divergence between the weights corresponding to each evaluation indicator and the initial weights of each evaluation indicator; Let be the weight corresponding to the i-th evaluation indicator; is the initial weight value of the i-th evaluation index; log(·) is the logarithm operation.

[0165] If the KL divergence is greater than a preset divergence threshold, the weights corresponding to each evaluation index are used as the initial strategy distribution in the game, and the adjustment direction of the weights of each evaluation index is used as the movement direction in the strategy space. Using a pre-constructed multi-armed slot machine model, the adjustment range of the weights of each evaluation index is used as the independent arm. A reward function is constructed based on the change in KL divergence. In this embodiment, when the KL divergence increases, a penalty of 0.1 is added; conversely, a reward of 0.05 is added. In other embodiments, the penalty and reward can be set according to specific needs.

[0166] The optimal adjustment direction is selected using the UCB algorithm, whose computational model is as follows:

[0167] in, Indicates action The upper confidence bound, i.e., the evaluation value for which the UCB algorithm selects the action, It is an action The average reward is equal to the action. Total historical rewards divided by action Number of selections, For action Number of selections, The total number of attempts is the sum of the number of times each action is selected, and ln(·) is the logarithm operation.

[0168] In the UCB algorithm, the action is the optional operation or strategy that needs to be selected at each decision step; that is, the weight adjustment strategy. The weight adjustment strategy corresponding to the maximum value is taken as the optimal adjustment strategy.

[0169] A voting mechanism is used to confirm whether the weights of each evaluation indicator should be adjusted according to the optimal adjustment strategy. That is, the optimal adjustment strategy is sent to each expert, and each expert votes on the optimal adjustment strategy. If the number of votes in favor accounts for more than 50% of the total votes, the weights of each evaluation indicator can be adjusted according to the optimal adjustment strategy. Otherwise, the weights of each evaluation indicator cannot be adjusted according to the optimal adjustment strategy.

[0170] If the weights of each evaluation indicator can be adjusted according to the optimal adjustment strategy, then the KL divergence between the adjusted weight values ​​and the initial weight values ​​of each evaluation indicator is recalculated and recorded as the first data.

[0171] If the weights of each evaluation indicator cannot be adjusted according to the optimal adjustment strategy, an alarm signal will be sent to each expert.

[0172] Determine whether the first data is greater than the preset divergence threshold. If so, readjust the weights of each evaluation indicator based on the adjusted weight values ​​as the initial strategy distribution in the game; otherwise, output the adjusted weight values ​​of each evaluation indicator.

[0173] Example 3: This example provides a deep learning-based farmland quality grading system, including: a processor, and a memory communicatively connected to the processor; The memory is provided with a computer-readable storage medium, and a computer program is stored on the computer-readable storage medium. When the processor processes the computer program stored on the computer-readable storage medium, it implements the deep learning-based method for evaluating the quality grade of cultivated land.

[0174] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A deep learning-based cultivated land quality grade evaluation method, characterized by, The method comprises the following steps: acquiring hyperspectral remote sensing images of target arable land, obtaining sensitive bands of each nutrient index through spectral bands, obtaining time-series band data of each nutrient index, constructing a plough layer nutrient inversion model by using a deep learning algorithm, inputting the time-series band data of each nutrient index into the plough layer nutrient inversion model, and outputting a spatial distribution inversion result of the plough layer nutrient; constructing an evaluation index set according to the spatial distribution inversion result of the plough layer nutrient, the site condition data and the climate factor data, screening the evaluation indexes by using the Delphi method and outputting the weight of each evaluation index, obtaining a minimum data set of arable land quality evaluation, normalizing the evaluation indexes in the minimum data set, calculating the membership values of the evaluation indexes by using a membership function, combining the weights to obtain an index of land capacity, and determining the quality grade of the arable land according to the index of land capacity.

2. The deep learning-based cultivated land quality grade evaluation method according to claim 1, characterized by, The method for screening the sensitive bands of the nutrient indexes through the spectral bands comprises the following steps: eliminating the multicollinearity among the spectral bands by using a continuous projection algorithm, integrating the spectral bands corresponding to the first N projection vectors into a first set; calculating the correlation coefficients of each spectral band and the nutrient index, integrating the spectral bands with the absolute values of the correlation coefficients greater than a preset correlation coefficient threshold into a second set; taking the intersection of the first set and the second set as the sensitive bands.

3. The deep learning-based cultivated land quality grade evaluation method according to claim 2, characterized in that, The plough layer nutrient inversion model comprises an iTransformer layer, the iTransformer layer encodes the time-series spectral data of each nutrient index into a high-dimensional feature vector, captures the change mode of the nutrient index through an adaptive nonlinear mapping in a high-dimensional feature space, obtains a mapping result, and converts the mapping result into a predicted value of the nutrient index through a regression head; based on the spatial geographic coordinates and the predicted value of each sampling point of the nutrient index, analyzing the spatial autocorrelation characteristics of the nutrient index by using a geostatistical method, and generating a spatial distribution inversion result of the plough layer nutrient.

4. The deep learning-based cultivated land quality grade evaluation method according to claim 3, characterized by, In the process of encoding the time-series spectral data of each nutrient index into a high-dimensional feature vector in the iTransformer layer, the self-attention mechanism is used to weight process different spectral bands in the time-series spectral data.

5. The deep learning-based cultivated land quality grade evaluation method according to claim 4, characterized by, The iTransformer layer integrates an LSTM time module that captures the time-series characteristics of the time-series spectral data of the same nutrient index, the self-attention module to which the self-attention mechanism belongs is connected in parallel with the LSTM time module, the output results of the self-attention module and the LSTM time module are fused through a residual connection, an output joint feature vector is output, and the joint feature vector is updated into a high-dimensional feature vector.

6. The deep learning-based cultivated land quality grade evaluation method according to any one of claims 1-5, characterized in that, The method further comprises the following steps: constructing a knowledge graph according to domain knowledge, and preliminarily limiting the screening range of the spectral bands, the collection range of the site condition data and the collection range of the climate factor data according to the association relationships among the arable land nutrients, the site condition data and the climate factor data in the knowledge graph.

7. The deep learning-based cultivated land quality grade evaluation method according to claim 6, characterized by, When the evaluation indexes and the corresponding weights are screened by using the Delphi method, the method further comprises the following steps: transforming the weight relationships in the knowledge graph into a reference weight matrix, and generating the initial weight values of each evaluation index in the first round of questionnaires of the Delphi method based on the reference weight matrix; After the evaluation indexes and corresponding weights are screened and evaluated by using the Delphi method, the weight adjustment strategy of each evaluation index is obtained, the adjusted weight value of each evaluation index is obtained based on the weight adjustment strategy, and the knowledge graph is corrected by using the adjusted weight value.

8. The deep learning-based cultivated land quality grade evaluation method according to claim 7, characterized in that, The method further comprises: Based on the geospatial rules, the knowledge graph is divided into several subgraphs, the target cultivated land corresponding subgraph is queried by the latitude and longitude coordinates of the target cultivated land, and is recorded as a target graph; The regional weight matrix is extracted in the target graph, and the regional weight matrix is taken as a new reference weight matrix.

9. The deep learning-based cultivated land quality grade evaluation method according to claim 8, characterized by, The adjusted weight value of each evaluation index based on the weight adjustment strategy comprises: The KL divergence of the weight corresponding to each evaluation index and the initial weight value of each evaluation index is calculated, if the KL divergence is greater than a preset divergence threshold, the weight corresponding to each evaluation index is taken as the initial strategy distribution in the game, the weight adjustment direction of each evaluation index is taken as the moving direction in the strategy space, the weight adjustment range of each questionnaire dimension is taken as an independent arm by using the pre-constructed multi-arm tiger machine model, the reward function is constructed according to the change amount of the KL divergence, the UCB algorithm is used to select the optimal adjustment strategy, whether to adjust the weight corresponding to each evaluation index according to the optimal adjustment strategy is confirmed by using the voting mechanism, if yes, the adjusted weight value of each evaluation index is output. 10.A deep learning-based cultivated land quality grade evaluation system, characterized by It comprises: A processor and a memory connected with the processor in communication; A computer readable storage medium is arranged in the memory, and the computer readable storage medium stores a computer program; When the processor processes the computer program stored in the computer readable storage medium, the method in any one of claims 1-9 is realized.

Images