Land resource utilization state evaluation method and system based on big data analysis

Abstract

The invention discloses a homeland resource utilization state evaluation method and system based on big data analysis, and the method comprises the steps: strengthening the modeling capability of the shape, boundary and ground feature space context relation of a land parcel through employing a 3D-CNN (convolutional neural network) embedded with a coordinate attention mechanism, obtaining spatial features, and carrying out the estimation of the utilization state of homeland resources. Modeling the time sequence data of the segmented land parcel units by adopting an encoder, and capturing seasonal and interannual change modes of spectral characteristics to obtain time sequence characteristics; fusing the spatial features and the time sequence features based on cross attention to obtain a fused feature vector; and inputting the fusion feature vector into an MLP multilayer perceptron, outputting a land utilization strength index and an ecological bearing capacity index, calculating an SHA-384 hash value of an evaluation result, and storing the hash value in a block chain. And dynamic monitoring and early warning of the land utilization health state are realized.

Description

Technical Field

[0001] This invention relates to the field of land and resources management technology, and in particular to a method and system for assessing the status of land and resources utilization based on big data analysis. Background Technology

[0002] Land resources are the material foundation for national economic and social development. Accurate and efficient assessment of their utilization status is crucial for achieving sustainable development and refined governance. Currently, the assessment of land resource utilization status faces challenges such as data integration difficulties, inconsistent spatiotemporal benchmarks for multi-source data, and insufficient data integrity after preprocessing. Traditional methods focus only on spatial or temporal dimensions, failing to simultaneously capture the spatial context of land features and spectral temporal patterns, thus limiting assessment accuracy. Furthermore, assessment results are often stored in local files, making them susceptible to tampering and lacking traceability, failing to meet the verifiable and traceable requirements of land management. Summary of the Invention

[0003] The purpose of this invention is to solve the above problems by designing a method and system for assessing the status of land resource utilization based on big data analysis.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned land resource utilization status assessment method based on big data analysis: Collect time-series remote sensing data and GIS data, and unify the collected data to the same spatiotemporal reference through spatial interpolation, resampling and normalization to obtain multimodal data; By utilizing a 3D-CNN convolutional neural network with embedded coordinate attention mechanism, multimodal data is stacked according to spatial coordinates and time dimensions to form a three-dimensional data cube, which serves as the input data for the 3D-CNN convolutional neural network. This enhances the modeling ability of terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. The study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot unit and arranged in chronological order to obtain the time series of the plot units. The time series is input into the Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, captures the seasonal and interannual variation patterns of spectral features, and outputs the time series features. The spatial and temporal features are fused based on cross-attention to obtain a fused feature vector; The fused feature vector is input into the MLP multilayer perceptron, which outputs the land use intensity index and the ecological carrying capacity index. The SHA-384 hash value of the evaluation result is calculated and stored on the blockchain.

[0005] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the collection of time-series remote sensing data and GIS data, through spatial interpolation, resampling, and normalization, unifies the collected data to the same spatiotemporal reference, resulting in multimodal data, including: Collect time-series remote sensing data and GIS data. The time-series remote sensing data includes at least optical remote sensing, synthetic aperture radar (SAR) imagery, and time-series frequency data. The GIS data includes at least basic geographic data, land management data, and ecological environment data. The IDW (Inverse Distance Weighted Interpolation) method is used for GIS raster data to fill data gaps, unifying remote sensing and GIS data of different resolutions to the same spatial resolution using bilinear interpolation. The remote sensing DN values ​​were converted to reflectance and then normalized to the [0,1] interval; the GIS data were normalized using min-max to eliminate the influence of dimensions and obtain multimodal data.

[0006] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the 3D-CNN convolutional neural network with embedded coordinate attention mechanism stacks multimodal data into a three-dimensional data cube according to spatial coordinates and time dimensions, which serves as the input data for the 3D-CNN convolutional neural network. This enhances the modeling ability of terrain shape, boundaries, and spatial context relationships of land features, resulting in spatial features, including: By performing convolution operations on 3D tensors using 2-3 layers of 3D convolution kernels, the spatial local features of the land parcels can be captured. A coordinate attention module is inserted after each 3D convolution layer. Global average pooling is performed on the feature map output by the convolution to obtain the channel attention weights. Pooling is performed on the length and width spatial dimensions respectively to obtain the spatial coordinate attention weights. Multiply the channel attention weights by the spatial coordinate attention weights and apply the result to the original feature map to obtain the enhanced spatial features. The spatial features are reduced by using a 3D max pooling layer, and nonlinearity is introduced by using a ReLU activation function to output the spatial features of each plot unit.

[0007] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot unit and arranged chronologically to obtain a time series of the plot units. This time series is input into a Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, capturing the seasonal and interannual variation patterns of spectral features, and outputting time series features, including: A position vector related to the time step is generated by using a sine function, and then added to the original time series data to retain the time information; The time series is split into multiple subsequences. For each subsequence, the similarity between the query, key, and value is calculated to obtain the attention weight. By normalizing the weights using Softmax, we focus on key time points related to the current time step, and use the weights to sum the values ​​to capture temporal dependencies and obtain temporal features.

[0008] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the step of fusing the spatial and temporal features based on cross-attention to obtain a fused feature vector includes: Spatial feature vectors are used as queries, and temporal feature vectors are used as keys and values. Spatial features guide the weight allocation of temporal features. The similarity between the query and the key is calculated by performing a dot product operation. The similarity is then normalized to an attention weight between 0 and 1 using the Softmax function. The normalized weights are then used to perform a weighted summation of the temporal features to obtain the fused feature vector.

[0009] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the step of inputting the fused feature vector into an MLP (Multilayer Perceptron) and outputting the land use intensity index and ecological carrying capacity index includes: The input layer receives the fused feature vector, and the hidden layer has 2-3 hidden layers. The first layer has 256 neurons, the second layer has 128 neurons, the ReLU activation function is used, and a Dropout layer is added to each layer. The output layer has two neurons. The fused feature vectors are propagated forward through the MLP and transformed nonlinearly through the hidden layer to generate two values ​​in the output layer. The two values ​​are normalized to the [0,100] interval by the Sigmoid function to obtain the land use intensity index and the ecological carrying capacity index.

[0010] Furthermore, in the aforementioned land resource utilization status assessment method based on big data analysis, the calculation of the SHA-384 hash value of the assessment result, and the storage of the hash value on the blockchain, includes: The hash algorithm is used to calculate the land use intensity index and the ecological carrying capacity index, generate a hash value of fixed length, and store the hash value on the blockchain.

[0011] Furthermore, the land and resources utilization status assessment system based on big data analysis includes the following modules: The multimodal data acquisition module is used to collect multi-temporal remote sensing data and GIS data. Through spatial interpolation, resampling and normalization, the collected data is unified to the same spatiotemporal reference to obtain multimodal data. The feature data extraction module utilizes a 3D-CNN convolutional neural network with embedded coordinate attention mechanism to stack multimodal data into a three-dimensional data cube according to spatial coordinates and time dimensions. This cube serves as the input data for the 3D-CNN convolutional neural network, enhancing its ability to model terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. The study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot unit and arranged in chronological order to obtain the time series of the plot units. The time series is input into the Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, captures the seasonal and interannual variation patterns of spectral features, and outputs the time series features. The feature vector fusion module is used to fuse the spatial features and temporal features based on cross-attention to obtain a fused feature vector; The state assessment module is used to input the fused feature vector into the MLP multilayer perceptron, output the land use intensity index and the ecological carrying capacity index, calculate the SHA-384 hash value of the assessment result, and store the hash value on the blockchain.

[0012] Furthermore, in the land and resources utilization status assessment system based on big data analysis, the utilization status assessment module includes the following sub-modules: Set up a submodule for the input layer to receive the fused feature vector, and set up 2-3 hidden layers. The first layer has 256 neurons and the second layer has 128 neurons. The ReLU activation function is used, and a Dropout layer is added to each layer. The evaluation submodule sets up two neurons in the output layer, fuses the feature vectors through forward propagation of the MLP, passes through the nonlinear transformation of the hidden layer, and generates two values ​​in the output layer. The two values ​​are then normalized to the [0,100] interval by the Sigmoid function to obtain the land use intensity index and the ecological carrying capacity index.

[0013] Furthermore, in the land and resources utilization status assessment system based on big data analysis, the utilization status assessment module includes the following sub-modules: The calculation submodule is used to call the hash algorithm to calculate the land use intensity index and the ecological carrying capacity index, generate a fixed-length hash value, and store the hash value on the blockchain.

[0014] Its beneficial effects lie in the fact that, through end-to-end deep learning models, it achieves full automation from data to assessment reports, greatly reducing human intervention and subjective bias. Multimodal deep semantic fusion, with its innovative cross-modal attention fusion module, goes beyond simple data stacking, achieving complementarity and enhancement of data from different sources at a high-level semantic level, improving the richness and accuracy of feature representation. Dynamic and forward-looking, through temporal deep learning networks, the model can not only assess the current situation but also capture changing trends, enabling dynamic monitoring and early warning of land use health. Data credibility is guaranteed; blockchain technology ensures data immutability, enhancing the credibility and legal validity of the assessment results. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the land resource utilization status assessment method based on big data analysis in this invention. Figure 2 This is a schematic diagram of the second embodiment of the land resource utilization status assessment method based on big data analysis in this invention. Figure 3 This is a schematic diagram of the first embodiment of the land resource utilization status assessment system based on big data analysis in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, the land resource utilization status assessment method based on big data analysis includes the following steps: Step 101: Collect time-series remote sensing data and GIS data, and unify the collected data to the same spatiotemporal reference through spatial interpolation, resampling and normalization to obtain multimodal data; Specifically, in this embodiment, time-series remote sensing data and GIS data are collected. The time-series remote sensing data includes at least optical remote sensing, synthetic aperture radar (SAR) imagery, and time-series frequencies. The GIS data includes at least basic geographic data, land management data, and ecological environment data. The IDW inverse distance weighted interpolation method is used for the GIS raster data to fill data gaps and unify remote sensing and GIS data of different resolutions to the same spatial resolution. Bilinear interpolation is used to convert the remote sensing DN values ​​into reflectance and then normalize them to the [0,1] interval. The GIS data is normalized using min-max to eliminate the influence of dimensions and obtain multimodal data.

[0020] During the multi-source data targeted acquisition phase, data types need to be precisely selected based on the assessment area scale and cycle: When acquiring remote sensing images, optical images should be prioritized for large-scale provincial assessments, simultaneously supplemented by Sentinel-1 SAR images; the time series frequency should be designed with a dual dimension of monthly and quarterly data. Monthly acquisition is used to capture short-term dynamics such as wheat greening, grain filling, and seasonal water changes, while quarterly acquisition supports annual trend analysis. If it is necessary to assess the long-term evolution of land use, 3-5 years of continuous data should be prioritized to cover interannual differences. GIS data acquisition needs to focus on three layers of requirements: spatial positioning, management benchmarks, and ecological background. In basic geographic data, the administrative division vector map should use the latest annual update to ensure boundary accuracy, and the digital elevation model (DEM) should use a 30-meter resolution ASTERGDEM; land management data should include the land use status vector map and the cultivated land protection red line vector map for the past 1-2 years; ecological environment data should integrate river / wetland vector boundaries and soil type raster maps corresponding to the Second National Soil Census. In the process of unifying the spatiotemporal reference, the spatial aspect strictly follows the "Basic Technical Regulations for National Geodetic Surveying" and adopts the 2000 National Geodetic Coordinate System, with CGCS2000 as the sole spatial reference. For mid-latitude regions and low-latitude regions, such as South China, the UTM projection is used. In terms of time, all data must be calibrated for the time zone.

[0021] The data preprocessing stage requires targeted elimination of format and accuracy differences: Spatial interpolation mainly addresses two types of data gaps: firstly, missing pixels in optical remote sensing images due to cloud cover or sensor malfunction; and secondly, sparse areas of GIS raster data, such as soil moisture sampling points. In practice, for small-scale assessments, such as at the municipal level, inverse distance weighting and IDW interpolation can be used; for large-scale assessments, such as at the provincial level, Kriging interpolation is preferred. Resampling requires first refining the original resolution of each data set. For example, Sentinel-2 multispectral bands are 10-60 meters, DEMs are 30 meters, and land use vector maps are converted to raster with a default resolution of 10 meters. The final resolution is unified to 30 meters. For optical images, such as Sentinel-2, bilinear interpolation is used; for vector-to-raster data, nearest neighbor interpolation is used for land cover types. Normalization requires two types of processing: First, remote sensing images undergo radiometric calibration to convert DN values ​​and original grayscale values ​​into apparent emissivity. Then, atmospheric correction is applied, such as using the 6S model to remove atmospheric scattering and absorption effects, yielding the true surface reflectance. Finally, reflectance is normalized to the [0,1] interval using reflectance minus minimum / maximum minus maximum, eliminating gain differences between different satellite sensors, such as the different DN value ranges of Landsat and Sentinel-2. In GIS data, DEM elevation values, ranging from -15 meters to 5000 meters, and soil organic matter content, ranging from 0.5 to 15 g / kg, are all compressed to the [0,1] interval using min-max normalization. This ensures that elevation in meters and organic matter in g / kg are on the same comparative dimension during subsequent model calculations, avoiding model bias towards higher-weighted elevation features due to differences in numerical magnitudes, such as elevation values ​​in the thousands and organic matter values ​​in the tens. Multimodal data integration uses 30m×30m grids as the basic unit, and associates data in a two-dimensional spatial location-time relationship: each grid not only contains spatial features such as elevation, slope, aspect, and land cover type, but also includes temporal features such as monthly scale spectral reflectance, NDVI, normalized difference vegetation index, EVI, enhanced vegetation index, NDBI, and building index. For example, a certain farmland grid needs to be associated with complete information such as NDVI value from January to December 2023, elevation of 50 meters, slope of 2°, and soil type of alluvial soil. The resulting multimodal dataset can support subsequent plot unit segmentation.

[0022] Step 102: Using a 3D-CNN convolutional neural network with embedded coordinate attention mechanism, multimodal data is stacked according to spatial coordinates and time dimensions to form a three-dimensional data cube, which is used as input data for the 3D-CNN convolutional neural network to enhance the modeling ability of terrain shape, boundary and spatial context relationship of land features, and obtain spatial features. The study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover and soil moisture index are extracted for each plot unit and arranged in chronological order to obtain the time series of plot units. The time series is input into the Transformer encoder, and the dependency relationship between different time steps is learned through self-attention mechanism to capture the seasonal and interannual variation patterns of spectral features and output the time series features. Specifically, in this embodiment, a 3D-CNN convolutional neural network with an embedded coordinate attention mechanism is used to stack multimodal data after unifying the spatiotemporal reference, including spectral, texture, and vegetation index features of multi-temporal remote sensing images, as well as spatial attributes such as terrain elevation, slope and aspect, land use type, and administrative division boundaries after rasterization of GIS vector data, to form a three-dimensional data cube, x, y, t, c, where x / y is the spatial resolution, t is the time series length, and c is the number of feature channels. This cube serves as the input data for the 3D-CNN, enhancing the ability to model terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. First, the study area is divided into several plot units through spatial segmentation, such as GIS-based administrative boundaries and land use patches, or image-based semantic segmentation models. For each plot unit, corresponding spectral features, such as NDVI, EVI, reflectance bands, vegetation cover, and soil moisture, are extracted from multi-temporal remote sensing data and arranged in chronological order to form the time series data of that plot. Simultaneously, the corresponding GIS attributes of the plot, such as plot area, ownership, and slope, are fused as auxiliary features. The time series data of each plot unit is first standardized, and then organized into sequences according to time steps, such as quarters or years, and input into a Transformer encoder. The encoder learns the dependencies between different time steps through a self-attention mechanism, capturing the seasonality of spectral features, such as crop growth cycles and interannual variation patterns, such as land use type conversion. The hidden state of each time step output by the encoder is used to obtain the temporal feature vector of that plot unit through global average pooling or max pooling, which is used for subsequent fusion.

[0023] Specifically, in this embodiment, 2-3 layers of 3D convolutional kernels are used to perform convolution operations on 3D tensors to capture the spatial local features of land parcels. A coordinate attention module is inserted after each 3D convolution layer to perform global average pooling on the feature map output by the convolution, obtaining channel attention weights. Pooling is performed separately on the length and width spatial dimensions to obtain spatial coordinate attention weights. The channel attention weights are multiplied by the spatial coordinate attention weights and applied to the original feature map to obtain enhanced spatial features. The dimensionality of the spatial features is reduced through a 3D max pooling layer, and nonlinearity is introduced through a ReLU activation function to output the spatial features of each land parcel unit. A position vector related to the time step is generated using a sine function and added to the original time-series data to preserve temporal information. The time-series sequence is split into multiple subsequences, and for each subsequence, the similarity of the query, key, and value is calculated to obtain attention weights. Softmax normalization of the weights focuses on key time points related to the current time step, and the weights are used to sum the values ​​to capture temporal dependencies, obtaining temporal features.

[0024] Before spatial feature extraction, land parcel unit segmentation must be completed. This step employs an object-oriented multi-scale segmentation method, requiring optimization of parameter settings based on the land use characteristics of the assessment area: After inputting preprocessed 30-meter resolution multimodal raster data, the segmentation scale is first determined according to land cover type. For large-scale land types such as cultivated land and forest land, a segmentation scale of 20-30 is selected to ensure parcel integrity; for small-scale land types such as residential land and small water bodies, a segmentation scale of 15-20 is selected to avoid excessive merging. Then, the segmentation criteria weights are set. After segmentation, the accuracy needs to be verified through manual verification and data comparison: 10% of the land parcel units are randomly selected and compared with the latest land use status vector map to ensure that the parcel boundary overlap exceeds 90% and the land use misclassification rate is less than 5%. Finally, discrete land parcel units are output, providing accurate spatial units for subsequent feature extraction. Spatial feature extraction is based on 3D-CNN with embedded coordinate attention. The core of this method is to construct a model input and feature enhancement mechanism that is adapted to the land parcel data. When constructing the 3D-CNN input, the tensor size needs to be determined according to the average number of grid cells in the land parcel. For example, most land parcels contain 30×30 grid cells, so the tensor is set to 32×32×12 to avoid data pruning. The 12 feature channels specifically cover terrain, DEM, slope, aspect, optical, blue, green, red, and near-infrared bands, SAR, VV and VH polarization backscattering coefficients, attribute, NDVI, NDBI, and soil type encoding. At the same time, masking technology is used to remove external grid cells of the land parcel, such as road grid cells at the edge of a certain farmland land parcel, reducing the amount of invalid computation by more than 50%. In the feature calculation stage, local features such as boundary contours and internal textures are initially extracted using two layers of 3D convolutional kernels. Then, a coordinate attention module is inserted: channel attention uses global average pooling with a 2×2×2 window to assign a weight of 0.8 or higher to the NDVI channel of cultivated land and a weight of 0.75 to the NDBI channel of construction land, suppressing interference from irrelevant channels. Spatial coordinate attention uses 3×1×1 pooling along the length direction and 1×3×1 pooling along the width direction, making the weight of the boundary between cultivated land and roads 2-3 times higher than that inside the plot, strengthening shape and boundary features. Finally, 2×2×2 3D max pooling reduces the feature map dimension from 32×32×128 to 16×16×64. Non-linearity is introduced through the ReLU activation function, and after flattening, a 256-dimensional spatial feature vector is output, fully preserving the spatial context of the plot.

[0025] Temporal feature extraction focuses on capturing the dynamic changes in land features. This requires the standardization of temporal data: for each plot, monthly time-series data is collected for three complete years, totaling 36 time steps. Five core spectral indicators are identified: NDVI (reflecting vegetation cover), EVI (removing soil background interference), NDBI (identifying construction land expansion), NDWI (monitoring water changes), and SAR backscattering coefficient (penetrating vegetation to capture land feature structural stability). The data processing stage requires outlier removal and smoothing to ensure quality: for outliers caused by extreme weather events such as heavy rain or sandstorms (e.g., a sharp drop of more than 0.3 in NDVI compared to adjacent months), a three-period moving average is used for replacement; for minor fluctuations in monthly data, such as monthly changes in NDVI of cultivated land less than 0.05, linear interpolation is used for further smoothing, ultimately forming a 36×5 two-dimensional time-series sequence, with each row corresponding to one time step and each column corresponding to one spectral indicator, ensuring that data in the same column represents the same period's indicators for the same plot. Temporal feature modeling is based on a Transformer encoder, which focuses on overcoming the challenges of temporal order perception and dynamic dependency capture: In the position encoding stage, for the 256-dimensional temporal vector, the even-numbered dimensions (0, 2, 4...) adopt the sine function sin(pos / 10000^(2i / 256)), and the odd-numbered dimensions (1, 3, 5...) adopt the cosine function cos(pos / 10000^(2i / 256)), where pos is the time step from 1 to 36 and i is the dimension index, accurately preserving the seasonal order from January to December and the interannual order from the 1st to the 3rd year. The multi-head attention layer has eight heads, each processing 32-dimensional features. It calculates the query, the similarity between the current time step and the key, and all time steps through dot product operations. After softmax normalization, the attention weights for farmland plots in summer (June-August) to other time steps exceed 0.6, while in winter (December-February), the weights are below 0.1, efficiently capturing seasonal dependence. The outputs of the eight heads are concatenated and input into a feedforward neural network to fit the nonlinear changes in time-series data, such as the year-on-year increase in NDBI of construction land. A residual connection is added after each encoding block, adding the module's input and output vectors to avoid gradient vanishing and layer normalization. The feature distribution is standardized with a mean of 0 and a variance of 1. After three encoding blocks, the output of the last observation point at the 36th time step is taken as the 256-dimensional time-series feature vector, fully preserving the seasonal and interannual variation patterns of land features.

[0026] Step 103: Based on cross-attention, spatial features and temporal features are fused to obtain a fused feature vector; Specifically, in this embodiment, the spatial feature vector is used as the query and the temporal feature vector is used as the key and value. The spatial features guide the weight allocation of the temporal features. The similarity between the query and the key is calculated by the dot product operation. The similarity is normalized to an attention weight between 0 and 1 by the Softmax function. The normalized weights are used to perform a weighted summation of the temporal features to obtain the fused feature vector.

[0027] In the input preparation stage, it is necessary to strictly ensure the matching and consistency of the input data. Each plot unit needs to correspond to a set of spatial feature vector-temporal feature vector input pairs, where both spatial and temporal feature vectors are 256-dimensional, derived from the extraction results mentioned above. The core is to establish a one-to-one mapping relationship through the unique ID of the plot to avoid feature confusion between different plots, such as incorrectly pairing the spatial features of cultivated land with the temporal features of construction land. If the two types of feature vectors have different dimensions due to special circumstances such as model parameter adjustments, such as the spatial feature vector being 384-dimensional and the temporal feature vector being 256-dimensional, then a fully connected layer needs to be introduced for dimension alignment. The number of input neurons in this fully connected layer is consistent with the original feature dimension, and the number of output neurons is uniformly set to 256. Through linear transformation, features of different dimensions are converted to the same feature space to eliminate fusion bias caused by dimension mismatch, such as unbalanced weight distribution and loss of feature information. The process then proceeds to the cross-attention module fusion. The first step is to clarify the division of attention roles, defining spatial feature vectors as queries and temporal feature vectors as keys and values. The core logic of this design is to let spatial attributes dominate the selection of temporal features. For example, the spatial features of cultivated land, including attributes such as gentle terrain and fertile soil, will be used as queries. In the temporal features, including monthly NDVI changes, the focus will be on keys related to spring planting-summer growth-autumn harvest-winter withering, such as high NDVI values ​​from June to August and low NDVI values ​​from December to February. On the other hand, the spatial features of construction land, including attributes such as regular boundaries and flat terrain, will focus on keys in the temporal features with stable NDBI without significant fluctuations, such as NDBI values ​​remaining above 0.6 in each month, achieving accurate matching of temporal features and spatial attributes. The first step is matching; the second step is calculating attention weights by quantifying the similarity between the query and the key through the dot product operation (vector inner product). A higher similarity value indicates a stronger correlation between the temporal feature and the spatial feature. For example, the dot product result of farmland query and summer NDVI key is significantly higher than that of winter NDVI key. The similarity results are then input into the Softmax function for normalization, converting them into weight values ​​in the 0-1 range, with the sum of the weights corresponding to all keys being 1 to ensure the rationality of weight allocation. The third step is weighted fusion of temporal features by summing the elements of the value and temporal feature vector using the normalized weights to obtain the spatially guided version of the temporal feature. Taking farmland as an example, the weight corresponding to summer NDVI, such as 0.3, is much higher than the weight corresponding to winter NDVI, such as 0.05. The weighted temporal features highlight seasonal variation patterns and weaken irrelevant temporal noise, such as occasional NDVI anomalies. The fourth step involves feature splicing and re-fusion. The original spatial feature vector (256 dimensions) and the weighted temporal feature vector (256 dimensions) are spliced ​​along the feature dimensions to form a 512-dimensional spliced ​​feature vector. This is then input into a fully connected layer with 512 input neurons and 512 output neurons for linear fusion. A ReLU activation function is introduced to set feature values ​​less than 0 to 0 and retain those greater than 0, breaking the linear correlation between features and strengthening the non-linear expressive power of the fused features, such as simultaneously reflecting the spatial attribute of flat farmland topography and the temporal pattern of NDVI seasonal fluctuations. The fifth step implements residual connections and layer normalization optimization. The residual connections use the sum of the pre-fusion spatial feature vector and the temporal feature vector (256 dimensions) to achieve the desired effect. The vector, through a broadcast mechanism, matches the 512-dimensional fused feature vector dimension with the fused feature vector, effectively mitigating the gradient vanishing problem during deep fusion and preventing the loss of original information due to multi-layer transformations. Layer normalization standardizes the fused feature vector to a mean of 0 and a variance of 1, unifying feature distribution, such as eliminating weight imbalances caused by excessively large spatial feature values ​​and excessively small temporal feature values, ensuring the stability of the fused features. Finally, a 512-dimensional fused feature vector is output for each land parcel unit. This vector fully integrates the parcel's shape, boundaries, spatial context (such as whether there are rivers surrounding the farmland), and temporal information such as seasonal and interannual variations in spectral features (such as the annual expansion trend of construction land), providing high-quality, multi-dimensional feature support for subsequent MLP calculations of land use intensity index and ecological carrying capacity index.

[0028] Step 104: Input the fused feature vector into the MLP multilayer perceptron, output the land use intensity index and ecological carrying capacity index, calculate the SHA-384 hash value of the evaluation result, and store the hash value on the blockchain.

[0029] Specifically, in this embodiment, the input layer receives the fused feature vector, and the hidden layer has 2-3 layers. The first layer has 256 neurons, and the second layer has 128 neurons, using the ReLU activation function. A Dropout layer is added to each layer. The output layer has 2 neurons. The fused feature vector is propagated forward through an MLP and undergoes a nonlinear transformation in the hidden layers, generating two values ​​at the output layer. These two values ​​are normalized to the [0, 100] interval using the Sigmoid function to obtain the land use intensity index and the ecological carrying capacity index. A hash algorithm is then used to calculate the land use intensity index and the ecological carrying capacity index, generating a fixed-length hash value, which is stored on the blockchain.

[0030] The MLP model transforms fusion features into quantifiable evaluation indicators, and relies on hash algorithms and blockchain technology to ensure the credibility of the results. The specific process revolves around two major stages: index generation and result storage. First, it focuses on the training of the MLP model and the evaluation of index output. Its core objective is to achieve accurate mapping of fusion features to land use intensity index and ecological carrying capacity index. In the MLP model structure design, the input layer strictly matches the dimension of the fused feature vector, setting 512 neurons to fully receive the 512-dimensional fused features; the hidden layer adopts a 2-3 layer progressive design, with 256 neurons in the first layer and 128 neurons in the second layer. Each layer introduces the ReLU activation function to avoid gradient vanishing during model training and enhance feature expression ability. At the same time, it is paired with a Dropout layer with dropout rate=0.3, randomly blocking 30% of neurons to reduce the model's over-reliance on local data and reduce the risk of overfitting; the output layer sets 2 neurons for the two core evaluation indices, and normalizes the output value to the range [0,100] through the Sigmoid activation function. 0 represents the lowest land use intensity / ecological carrying capacity of the plot, and 100 represents the highest, meeting the intuitive requirements of quantitative evaluation. The model training phase requires first constructing a high-quality training set, collecting land parcel samples with real labels within the evaluation area. The training process uses mean squared error (MSE) as the loss function to accurately measure the deviation between the model's predicted values ​​and the actual values. An Adam optimizer with a learning rate of 0.001 is used to adaptively adjust the learning step size, improving convergence efficiency. Iterative training is conducted for 100-200 rounds until the loss function stably converges, for example, if the loss value fluctuation is less than 0.001 for 10 consecutive rounds. After training, the 512-dimensional fused feature vectors of all land parcel units are input into the model, outputting two indices for each parcel. These indices are then organized into a structured evaluation result table in the format of unique parcel ID - land use intensity index - ecological carrying capacity index. The next step involves hashing and storing the evaluation results on the blockchain to ensure their integrity and traceability. The first step converts the evaluation result table into a standard JSON format, including a unique land parcel ID (consistent with the ID from the feature extraction step), an evaluation timestamp accurate to the second to record the index generation time (avoiding time dimension confusion), the specific values ​​of the two indices, and the data source (e.g., remote sensing image number, GIS data version) to facilitate tracing the quality of the original data. The second step uses the SHA-384 hash algorithm to calculate a 384-bit fixed-length hash value. This hash value is tamper-sensitive; if any information in the JSON data, such as an index value modification or land parcel ID replacement, changes occur, the hash value will show a significant difference, serving as the core verification basis for the integrity of the results.The blockchain storage process prioritizes the land and resources department's consortium blockchain, which differs from public blockchains. Only authorized provincial land and resources departments and municipal natural resources bureaus have read and write permissions, balancing data privacy and management standardization. In practice, the evaluation system, as a node in the consortium blockchain, generates transaction information containing hash values, land parcel IDs, and timestamps, and sends it to multiple verification nodes. Verification nodes perform double verification of the transaction's legality, confirming that the timestamp is within a reasonable range and without any abnormalities such as being ahead or behind. They recalculate the hash value of the JSON data and compare it with the hash value in the transaction to ensure consistency. When more than half of the verification nodes pass the verification, the transaction is packaged into a new blockchain block, and the new block forms a chain link with historical blocks through hash values. Each block header contains the hash value of the previous block, ensuring that the block sequence is immutable. Finally, the block is synchronized to all nodes in the consortium blockchain, completing the permanent storage of the evaluation result hash.

[0031] Its beneficial effects lie in the fact that, through end-to-end deep learning models, it achieves full automation from data to assessment reports, greatly reducing human intervention and subjective bias. Multimodal deep semantic fusion, with its innovative cross-modal attention fusion module, goes beyond simple data stacking, achieving complementarity and enhancement of data from different sources at a high-level semantic level, improving the richness and accuracy of feature representation. Dynamic and forward-looking, through temporal deep learning networks, the model can not only assess the current situation but also capture changing trends, enabling dynamic monitoring and early warning of land use health. Data credibility is guaranteed; blockchain technology ensures data immutability, enhancing the credibility and legal validity of the assessment results.

[0032] Please see Figure 2 In the land resource utilization status assessment method based on big data analysis, the Transformer encoder is used to model the time series data of the segmented land parcel units, capture the seasonal and interannual variation patterns of spectral features, and obtain the time series features, including the following steps: Step 201: Generate a position vector related to the time step using a sine function, add it to the original time series data, and retain the time information; Step 202: Divide the time series into multiple subsequences. For each subsequence, calculate the similarity of query, key, and value to obtain the attention weight. Step 203: By normalizing the weights using Softmax, focus on key time points related to the current time step, and sum the weights on the Value to capture temporal dependencies and obtain temporal features.

[0033] The above describes an example of the land and resources utilization status assessment method. Please refer to [link / reference]. Figure 3 In the land and resources utilization status assessment system based on big data analysis, the system includes the following modules: The multimodal data acquisition module is used to collect multi-temporal remote sensing data and GIS data. Through spatial interpolation, resampling and normalization, the collected data is unified to the same spatiotemporal reference to obtain multimodal data. The feature data extraction module utilizes a 3D-CNN convolutional neural network with embedded coordinate attention mechanism to stack multimodal data into a three-dimensional data cube according to spatial coordinates and time dimensions. This cube serves as the input data for the 3D-CNN convolutional neural network, enhancing its ability to model terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. The study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot unit and arranged in chronological order to obtain the time series of the plot units. The time series is input into the Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, captures the seasonal and interannual variation patterns of spectral features, and outputs the time series features. The feature vector fusion module is used to fuse spatial features and temporal features based on cross-attention to obtain a fused feature vector; The state assessment module is used to input the fused feature vector into the MLP multilayer perceptron, output the land use intensity index and the ecological carrying capacity index, calculate the SHA-384 hash value of the assessment result, and store the hash value on the blockchain.

[0034] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for assessing the state of land resource utilization based on big data analysis, characterized in that, The method for assessing the status of land and resources utilization includes the following steps: Collect time-series remote sensing data and GIS data, and unify the collected data to the same spatiotemporal reference through spatial interpolation, resampling and normalization to obtain multimodal data; By utilizing a 3D-CNN convolutional neural network with embedded coordinate attention mechanism, multimodal data is stacked according to spatial coordinates and time dimensions to form a three-dimensional data cube, which serves as the input data for the 3D-CNN convolutional neural network. This enhances the modeling ability of terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. The study area is divided into multiple plot units through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot unit and arranged in chronological order to obtain the time series of the plot units. The time series is input into the Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, captures the seasonal and interannual variation patterns of spectral features, and outputs the time series features. The spatial and temporal features are fused based on cross-attention to obtain a fused feature vector; The fused feature vector is input into the MLP multilayer perceptron, which outputs the land use intensity index and the ecological carrying capacity index. The SHA-384 hash value of the evaluation result is calculated and stored on the blockchain.

2. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The collected time-series remote sensing data and GIS data are unified onto the same spatiotemporal reference through spatial interpolation, resampling, and normalization to obtain multimodal data, including: Collect time-series remote sensing data and GIS data. The time-series remote sensing data includes at least optical remote sensing, synthetic aperture radar (SAR) imagery, and time-series frequency data. The GIS data includes at least basic geographic data, land management data, and ecological environment data. The IDW inverse distance weighted interpolation method is used for GIS raster data to fill data gaps and unify remote sensing and GIS data of different resolutions to the same spatial resolution. Bilinear interpolation was used to convert remote sensing DN values ​​into reflectance and then normalize them to the [0,1] interval; GIS data were normalized using min-max to eliminate the influence of dimensions and obtain multimodal data.

3. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The 3D-CNN convolutional neural network, which utilizes an embedded coordinate attention mechanism, stacks multimodal data into a three-dimensional data cube based on spatial coordinates and time dimensions. This cube serves as the input data for the 3D-CNN convolutional neural network, enhancing its ability to model terrain shapes, boundaries, and spatial context relationships of ground features, thereby obtaining spatial features, including: By performing convolution operations on 3D tensors using 2-3 layers of 3D convolution kernels, the spatial local features of the land parcels can be captured. A coordinate attention module is inserted after each 3D convolution layer. Global average pooling is performed on the feature map output by the convolution to obtain the channel attention weights. Pooling is performed on the length and width spatial dimensions respectively to obtain the spatial coordinate attention weights. Multiply the channel attention weights by the spatial coordinate attention weights and apply the result to the original feature map to obtain the enhanced spatial features. The spatial features are reduced by using a 3D max pooling layer, and nonlinearity is introduced by using a ReLU activation function to output the spatial features of each plot unit.

4. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The study area is divided into multiple plots through spatial segmentation. Spectral features, vegetation cover, and soil moisture indices are extracted from each plot, and these are arranged chronologically to obtain a time series of the plots. The time series is input into a Transformer encoder, which learns the dependencies between different time steps through a self-attention mechanism, capturing the seasonal and interannual variation patterns of spectral features, and outputting time series features, including: A position vector related to the time step is generated by using a sine function, and then added to the original time series data to retain the time information; The time series is divided into multiple subsequences. For each subsequence, the similarity between the query, key, and value is calculated to obtain the attention weight. By normalizing the weights using Softmax, we focus on key time points related to the current time step, and use the weights to sum the values ​​to capture temporal dependencies and obtain temporal features.

5. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The process of fusing the spatial and temporal features based on cross-attention to obtain a fused feature vector includes: Spatial feature vectors are used as queries, and temporal feature vectors are used as keys and values. Spatial features guide the weight allocation of temporal features. The similarity between the query and the key is calculated by performing a dot product operation. The similarity is then normalized to an attention weight between 0 and 1 using the Softmax function. The normalized weights are then used to perform a weighted summation of the temporal features to obtain the fused feature vector.

6. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The process of inputting the fused feature vector into an MLP (Multilayer Perceptron) and outputting the land use intensity index and ecological carrying capacity index includes: The input layer receives the fused feature vector, and the hidden layer has 2-3 hidden layers. The first layer has 256 neurons, the second layer has 128 neurons, the ReLU activation function is used, and a Dropout layer is added to each layer. The output layer has two neurons. The fused feature vectors are propagated forward through the MLP and transformed nonlinearly through the hidden layer to generate two values ​​in the output layer. The two values ​​are normalized to the [0,100] interval by the Sigmoid function to obtain the land use intensity index and the ecological carrying capacity index.

7. The land resource utilization status assessment method based on big data analysis as described in claim 1, characterized in that, The SHA-384 hash value of the calculated evaluation result is stored on the blockchain, including: The hash algorithm is used to calculate the land use intensity index and the ecological carrying capacity index, generate a hash value of fixed length, and store the hash value on the blockchain.

8. A land resource utilization status assessment system based on big data analysis, characterized in that, The land and resources utilization status assessment system includes the following modules: The multimodal data acquisition module is used to collect multi-temporal remote sensing data and GIS data. Through spatial interpolation, resampling and normalization, the collected data is unified to the same spatiotemporal reference to obtain multimodal data. The feature data extraction module utilizes a 3D-CNN convolutional neural network with embedded coordinate attention mechanism to stack multimodal data into a three-dimensional data cube based on spatial coordinates and time dimensions. This cube serves as input data for the 3D-CNN convolutional neural network, enhancing its ability to model terrain shape, boundaries, and spatial context relationships of land features, thereby obtaining spatial features. The study area is divided into multiple plot units through spatial segmentation. For each plot unit, spectral features, vegetation cover, and soil moisture indices are extracted from multi-temporal remote sensing data and arranged chronologically to obtain the time series of the plot unit. The time series of each plot unit is standardized and then input into the Transformer encoder. Through a self-attention mechanism, the module learns the dependencies between different time steps, captures the seasonal and interannual variation patterns of spectral features, and outputs temporal features. The feature vector fusion module is used to fuse the spatial features and temporal features based on cross-attention to obtain a fused feature vector; The state assessment module is used to input the fused feature vector into the MLP multilayer perceptron, output the land use intensity index and the ecological carrying capacity index, calculate the SHA-384 hash value of the assessment result, and store the hash value on the blockchain.

9. The land resource utilization status assessment system based on big data analysis as described in claim 8, characterized in that, The utilization status assessment module includes the following sub-modules: Set up a submodule for the input layer to receive the fused feature vector, and set up 2-3 hidden layers. The first layer has 256 neurons and the second layer has 128 neurons. The ReLU activation function is used, and a Dropout layer is added to each layer. The evaluation submodule sets two neurons in the output layer, fuses the feature vector through the forward propagation of the MLP, passes through the nonlinear transformation of the hidden layer, and generates two values ​​in the output layer. The two values ​​are then normalized to the [0,100] interval by the Sigmoid function to obtain the land use intensity index and the ecological carrying capacity index.

10. The land resource utilization status assessment system based on big data analysis as described in claim 8, characterized in that, The utilization status assessment module includes the following sub-modules: The calculation submodule is used to call the hash algorithm to calculate the land use intensity index and the ecological carrying capacity index, generate a fixed-length hash value, and store the hash value on the blockchain.

Images