Land management analysis method and system based on big data

Abstract

The application discloses a land management analysis method and system based on big data, relates to the technical field of land management, and obtains remote sensing image data and geographic national condition monitoring vector data of a target region, filters a small number of ground object samples, performs geometric transformation and mosaic data enhancement, generates a training data set, extracts features of the training data set by using a multi-resolution backbone network, a pyramid feature extraction structure and a channel space collaborative attention unit, generates a local feature map, converts the training data set into serialized feature input, obtains a long-range feature map based on the serialized feature input, constructs a first land classification model, obtains a full-scale feature map based on the local feature map and the long-range feature map, constructs a composite loss function, iteratively trains and outputs a second land classification model, and predicts remote sensing image data of the target region to obtain a spatial distribution boundary, analyzes various land use types, and obtains a land management analysis result.

Description

Technical Field

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

[0002] Land resources are a vital national fundamental natural resource. Utilizing remote sensing big data for land use classification and dynamic monitoring is of great significance for natural resource surveys, ecological environmental protection, and national spatial planning. With the rapid development of remote sensing Earth observation technology, high-resolution remote sensing imagery has become an important data source for land management analysis.

[0003] Currently, Chinese invention application number 202510929368.4 discloses a land resource management data analysis system. This system acquires spectral reflectance changes under multi-directional illumination on the land surface and extracts reflectance gradient response information. It then constructs a texture vector set by combining regional texture structure features. Furthermore, it identifies potential fission boundaries by comparing the texture vectors of adjacent regions and performs fission segment classification and land type reclassification. However, the existing technology has the following shortcomings: it struggles to handle extreme class imbalances and the identification of small target features. The existing technology mainly relies on illumination reflectance gradients and texture clustering for region division. When facing scenarios with extremely uneven distribution of land features, due to the lack of data augmentation and weight balancing mechanisms for minority class samples, the features of small target features are easily suppressed by the majority class background features, leading to missed classifications. Or misclassification; limited by the local receptive field, it is difficult to capture long-range spatial dependencies. Existing technologies identify boundary breakpoints by comparing the construction texture vectors of adjacent regions. Essentially, they determine boundary continuity through local neighborhood analysis, lacking the ability to model global contextual information of images. They cannot establish long-range dependencies like state-space models in deep learning. When faced with linear features such as farmland and roads that are occluded or have discontinuous textures, it is easy to cause rough and broken segmentation boundaries, failing to maintain the integrity of the spatial structure of features. There is a lack of deep fusion and quantitative evaluation of multi-scale features. Although existing technologies combine reflection gradients and texture vectors, they are mainly based on the calculation and clustering of shallow physical features. They are lacking in the extraction and adaptive fusion of multi-scale semantic features, making it difficult to extract deep semantic information while maintaining high-resolution details. Summary of the Invention

[0004] The technical problem solved by this invention is that existing technologies struggle to achieve both precise extraction of multi-scale features and complete preservation of land cover boundaries in high-resolution remote sensing images of complex scenes in arid oasis transition zones. Furthermore, they cannot effectively overcome the suppression of sparse structures and artificial land excavation features by the vast desert caused by extreme class imbalances, and the difficulty of establishing long-range spatial dependencies due to the limited receptive field of traditional convolutional neural networks. This results in rough and fragmented boundaries between linear features such as cultivated land and roads in land cover classification, severe fragmentation of land cover patches, and low recognition rates for small target features in complex backgrounds, leading to frequent omissions or misclassifications. Consequently, these issues severely affect the accuracy and reliability of precise land resource management and spatiotemporal evolution analysis of ecological value.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a land management analysis method based on big data, comprising the following steps:

[0006] Step S1: Acquire remote sensing image data and geographic national condition monitoring vector data of the target area, screen a few land cover samples and perform geometric transformation and mosaic data augmentation to generate a training dataset;

[0007] Step S2: Use a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to extract features from the training dataset and generate local feature maps.

[0008] Step S3: Convert the training dataset into serialized feature input, and obtain a long-range feature map based on the serialized feature input;

[0009] Step S4: Construct the first land classification model based on the multi-resolution backbone network, pyramid feature extraction structure and visual state space encoder; obtain the full-scale feature map based on local feature map and long-range feature map; construct the composite loss function and iteratively train to output the second land classification model.

[0010] Step S5: Based on the second land classification model, predict the remote sensing image data of the target area to obtain the spatial distribution boundary, construct the land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain the land management analysis results.

[0011] Preferably, step S1 includes the following sub-steps:

[0012] Step S101: Obtain remote sensing image data and geographic national condition monitoring vector data of the target area; perform radiometric correction, geometric correction and image fusion processing on the remote sensing image data; and perform spatial registration and cropping on the processed remote sensing image data and geographic national condition monitoring vector data to generate a land sample dataset with uniform resolution and size.

[0013] Step S102: Calculate the percentage of pixels for each land cover category in the land sample dataset, mark land cover categories with a percentage of pixels less than a preset percentage threshold as minority land cover categories, and traverse the land sample dataset to select image blocks that include the minority land cover categories, and construct a minority sample set based on the image blocks.

[0014] Preferably, step S1 further includes the following sub-steps:

[0015] Step S103: Perform geometric transformation operations of vertical flipping, horizontal mirroring, and 90-degree rotation sequentially on the image blocks in the minority class sample set to generate the first image block. At the same time, apply mosaic data augmentation technology to randomly select four image blocks from the minority class sample set for random scaling, cropping, and stitching to generate the second image block. Merge the first and second image blocks into the land sample dataset to generate the training dataset. The training dataset includes remote sensing image blocks and semantic segmentation ground truth labels corresponding to each remote sensing image block. The semantic segmentation ground truth labels include the true label distribution and true area of ​​each land cover category. Land cover categories other than minority class samples in the training dataset are defined as majority class samples.

[0016] Preferably, step S2 includes the following sub-steps:

[0017] Step S201: Construct a multi-resolution backbone network based on the training dataset. The multi-resolution backbone network includes a high-resolution sub-network and a low-resolution sub-network. The high-resolution sub-network is used as the starting stage. In subsequent stages, the low-resolution sub-network is added in parallel according to a preset stage. During the forward propagation of the multi-resolution backbone network, the multi-scale fusion operation is repeatedly performed to enable information interaction between sub-networks of different resolutions. The output features of each sub-network of resolutions are upsampled and spliced ​​together to output a fused feature map that maintains high spatial resolution.

[0018] Step S202: Input the fused feature map into the pyramid feature extraction structure, perform pyramid pooling operation using pooling kernels of different sizes to capture global and local features respectively, obtain pooling results, and then upsample the pooling results of different scales to restore them to their original size and concatenate them with the fused feature map to generate an intermediate feature map;

[0019] Step S203: Apply a channel spatial collaborative attention unit to the intermediate feature map, use one-dimensional convolution without dimensionality reduction to calculate cross-channel interaction information, generate channel attention weights, perform spatial average pooling along the horizontal and vertical directions respectively to generate horizontal and vertical feature vectors, encode the horizontal and vertical feature vectors using convolution transformation to generate a spatial attention weight map, and perform element-wise multiplication weighting on the intermediate feature map using the channel attention weights and spatial attention weight map to output a local feature map.

[0020] Preferably, step S3 includes the following sub-steps:

[0021] Step S301: Divide the image data in the training dataset into n image blocks, and flatten the image blocks into a one-dimensional feature sequence according to the spatial arrangement order from left to right and from top to bottom. At the same time, embed position encoding information to mark the spatial relative position of each image block and construct a serialized feature input.

[0022] Step S302: The serialized feature input is imported into the visual state space encoder, which includes a scale visual state space block. The continuous-time state space equation parameters are converted into discretized parameters using the zero-order preservation technique. The serialized feature input is traversed according to the spatial arrangement order in the one-dimensional feature sequence. During the traversal, the feature vector extracted from the serialized feature input at the current time is defined as the current input, and the state variable recursively calculated and stored at the previous time is defined as the historical hidden state. Through the interactive calculation between the current input and the historical hidden state, the long-range dependency relationship between image blocks at different spatial locations in the serialized feature input is captured, and the updated feature vector is recombined to generate the output sequence after state update.

[0023] Step S303: The output sequence is dimensionally reshaped, and the output sequence is restored to a two-dimensional spatial structure according to the spatial relative position. The long-range dependency is mapped to a spatial distribution feature spanning the entire image, and a long-range feature map is generated.

[0024] Preferably, step S4 includes the following sub-steps:

[0025] Step S401: Construct an adaptive perception fusion module, which includes a feature stitching unit, a weight generation unit, and a weighted fusion unit. The feature stitching unit aligns the local feature map and the long-range feature map, and then stitches the aligned local and long-range feature maps along the channel dimension to generate a hybrid feature. The hybrid feature is then input into the weight generation unit, which calculates a spatial attention weight map through convolution operations and activation functions. The spatial attention weight map represents the degree of dependence of each pixel position on local and global features. The weighted fusion unit performs a pixel-by-pixel weighted summation of the local and long-range feature maps based on the spatial attention weight map to generate a full-scale feature map.

[0026] Preferably, step S4 further includes the following sub-steps:

[0027] Step S402: Classify and predict the full-scale feature map to obtain the predicted probability distribution and predicted region. Construct a composite loss function based on the true label distribution and true region in the training dataset.

[0028] The composite loss function includes a cross-entropy loss term and a Dice coefficient loss term. The difference between the predicted probability distribution and the true label distribution is calculated using the cross-entropy loss term, and the error between the predicted region and the true region is calculated using the Dice coefficient loss term. The difference and the error are weighted and summed to obtain the total loss value of the current training batch.

[0029] Preferably, step S4 further includes the following sub-steps:

[0030] Step S403: The multi-resolution backbone network, pyramid feature extraction structure, visual state space encoder and adaptive perception fusion module are cascaded to construct the first land classification model, and the training dataset is divided into training subset and validation subset according to a preset ratio;

[0031] Iterative training is performed on the first land classification model based on the training subset, and the parameters of the first land classification model are updated according to the total loss value using the backpropagation algorithm, specifically including:

[0032] At the end of each training round, the validation subset is predicted using the first land classification model to construct a confusion matrix. Based on the confusion matrix, the precision and recall of each land cover category are calculated. The harmonic mean of the precision and recall is defined as the F1 score. The standard deviation and mean of the F1 scores of all land cover categories are calculated. The ratio of the standard deviation and mean is defined as the coefficient of variation.

[0033] When the coefficient of variation is greater than the preset balance threshold, it is determined that the first land classification model is biased towards the majority class. Based on the reciprocal of the F1 score, the weight ratio of the minority class samples in the composite loss function in the next training round is increased. The training is repeated until the total loss value converges, and the second land classification model is output.

[0034] Preferably, step S5 includes the following sub-steps:

[0035] Step S501: Based on the second land classification model, perform pixel-by-pixel classification prediction on the remote sensing image data of the target area in the start year and end year, generate the thematic maps of land use status corresponding to the start year and end year respectively, and perform noise reduction and smoothing on the thematic maps of land use status to obtain the spatial distribution boundaries of each land use type.

[0036] Step S502: Spatial overlay analysis of the land use status thematic maps of the starting year and the ending year is performed to construct a land use type transfer matrix. Based on the land use type transfer matrix, the transfer-out area, transfer-in area and retention area of ​​each land use type are calculated. The ratio of the retention area to the initial area of ​​the starting year is calculated to obtain the land use stability index.

[0037] Step S503: Based on the land use type transfer matrix, extract the net change in area of ​​the target land use type during the study period, calculate the ratio of the absolute value of the net change in area to the initial area, and divide the obtained ratio by the time span of the study period to obtain a single land use dynamic index.

[0038] Step S504, based on the spatial distribution boundary, performs landscape pattern evolution analysis, specifically including:

[0039] A topological analysis is performed on the spatial distribution boundary to extract the geometric perimeter of patches in each land use type, and the number of patches in each land use type is counted. The sum of the reserved area and the transferred area is recorded as the current total area of ​​each land use type.

[0040] Calculate the landscape shape index based on the geometric perimeter and the current total area;

[0041] Calculate the patch density index based on the number of patches and the total area of ​​the target region;

[0042] Step S505: The land use stability index, single land use dynamic index, landscape shape index and patch density index are used as the land management analysis results, and a multi-dimensional land use comprehensive evaluation report is generated based on the land management analysis results.

[0043] A land management and analysis system based on big data includes a data acquisition module, an extraction module, a construction module, a training module, and an analysis module.

[0044] The acquisition module is used to acquire remote sensing image data and geographic national condition monitoring vector data of the target area, screen a few land cover samples and perform geometric transformation and mosaic data enhancement to generate a training dataset.

[0045] The extraction module is used to extract features from the training dataset using a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to generate local feature maps.

[0046] The building module is used to convert the training dataset into a serialized feature input, and uses a visual state space encoder to perform discretization parameter transformation and recursive state update of the state space equation to capture long-range dependencies and generate a long-range feature map.

[0047] The training module is used to build a first land classification model based on a multi-resolution backbone network, a pyramid feature extraction structure and a visual state space encoder, to perform weighted fusion of local feature maps and long-range feature maps to generate a full-scale feature map, to construct a composite loss function, and to iteratively train and output a second land classification model.

[0048] The analysis module is used to predict the spatial distribution boundary of the target area using remote sensing image data in the start and end years, construct a land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain land management analysis results.

[0049] The beneficial effects of this invention are as follows: By performing mosaic data augmentation and geometric transformation on minority class samples, and combining a training strategy that dynamically adjusts the weights of the composite loss function based on the coefficient of variation index, the invention effectively overcomes the problem of underclassification and suppression of small target features caused by extreme class imbalance in the transition zone between arid oases. It utilizes a multi-resolution backbone network to maintain high-resolution feature flow, introduces a visual state space encoder to perform recursive state updates to capture long-range dependencies, breaks through the bottleneck of limited receptive field in traditional convolutional neural networks, solves the problem of rough and broken segmentation boundaries of linear features such as farmland and roads, constructs an adaptive perception fusion module based on spatial attention weight map to achieve accurate complementarity between local texture and global semantics, and combines multi-dimensional evaluation indicators such as landscape shape index to achieve land management analysis that takes into account the fine extraction of multi-scale features, the preservation of feature boundaries, and the quantitative assessment of spatial morphological evolution. Attached Figure Description

[0050] Figure 1 A flowchart illustrating the steps of a land management analysis method based on big data, as provided in one embodiment of the present invention;

[0051] Figure 2 This is a basic flowchart of a land management and analysis system based on big data, provided as an embodiment of the present invention. Detailed Implementation

[0052] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0053] Example 1, referring to Figure 1 This paper presents a land management analysis method based on big data, which includes the following steps:

[0054] Step S1: Acquire remote sensing image data and geographic national condition monitoring vector data of the target area, select a few land cover samples and perform geometric transformation and mosaic data augmentation to generate a training dataset.

[0055] Step S2: Use a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to extract features from the training dataset and generate local feature maps.

[0056] Step S3: Convert the training dataset into a serialized feature input, and use a visual state space encoder to perform discretization parameter transformation and recursive state update of the state space equation to capture long-range dependencies and generate a long-range feature map.

[0057] Step S4: Construct the first land classification model based on the multi-resolution backbone network, pyramid feature extraction structure and visual state space encoder; perform weighted fusion of local feature maps and long-range feature maps to generate full-scale feature maps; construct a composite loss function; and iteratively train and output the second land classification model.

[0058] Step S5: Based on the second land classification model, predict the remote sensing image data of the target area in the start and end years to obtain the spatial distribution boundary, construct the land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain the land management analysis results.

[0059] This invention effectively overcomes the problem of underclassification and suppression of small-target features caused by extreme class imbalance in arid oasis transition zones by performing mosaic data augmentation and geometric transformation on minority class samples and combining a training strategy that dynamically adjusts the weights of the composite loss function based on the coefficient of variation index. It utilizes a multi-resolution backbone network to maintain high-resolution feature flow and introduces a visual state space encoder to perform recursive state updates to capture long-range dependencies, breaking through the bottleneck of limited receptive field in traditional convolutional neural networks. It solves the problem of rough and broken segmentation boundaries of linear features such as farmland and roads, and constructs an adaptive perception fusion module based on spatial attention weight map to achieve accurate complementarity between local texture and global semantics. Combined with multi-dimensional evaluation indicators such as landscape shape index, it realizes land management analysis that takes into account the fine extraction of multi-scale features, the preservation of feature boundaries, and the quantitative assessment of spatial morphological evolution.

[0060] In a specific embodiment, step S1 includes the following sub-steps:

[0061] Step S101: Obtain remote sensing image data and geographic national condition monitoring vector data of the target area; perform radiometric correction, geometric correction and image fusion processing on the remote sensing image data; and perform spatial registration and cropping on the processed remote sensing image data and geographic national condition monitoring vector data to generate a land sample dataset with uniform resolution and size.

[0062] It should be noted that the generation of the land sample dataset in step S101 specifically includes the rasterization and cropping process of the data. Since the geographic national condition monitoring vector data is in vector format, it is converted into raster images with the same resolution as the remote sensing image data. These raster images constitute the initial semantic segmentation ground truth labels. To meet the input requirements of the deep learning model, a sliding window technique is used to simultaneously crop the registered remote sensing image data and the initial semantic segmentation ground truth labels into multiple units of a uniform size of 512×512 pixels. These cropped units constitute the image blocks in step S102 and the corresponding semantic segmentation ground truth labels. The set of image blocks and semantic segmentation ground truth labels constitutes the land sample dataset.

[0063] Step S102: Calculate the percentage of pixels for each land cover category in the land sample dataset, mark land cover categories with a percentage of pixels less than a preset percentage threshold as minority land cover categories, and traverse the land sample dataset to select image blocks that include minority land cover categories, and construct a minority sample set based on the image blocks.

[0064] It should be noted that in step S102, the logic for filtering image blocks that include a minority of land cover categories is as follows:

[0065] Iterate through the semantic segmentation ground truth labels corresponding to each image block in the land sample dataset, count the number of pixels belonging to the minority land cover category in the semantic segmentation ground truth labels, and if the number of pixels is greater than zero, determine that the corresponding image block is an image block that includes the minority land cover category, and include the corresponding image block into the minority sample set for data augmentation processing in the subsequent step S103.

[0066] Step S103: Perform geometric transformation operations of vertical flipping, horizontal mirroring, and 90-degree rotation on the image blocks in the minority class sample set in sequence to generate the first image block. At the same time, apply mosaic data augmentation technology to randomly select four image blocks from the minority class sample set for random scaling, cropping, and stitching to generate the second image block. Merge the first and second image blocks into the land sample dataset to generate the training dataset. The training dataset includes remote sensing image blocks and semantic segmentation ground truth labels corresponding to each remote sensing image block. The semantic segmentation ground truth labels include the true label distribution and true area of ​​each land cover category. Land cover categories other than minority class samples in the training dataset are defined as majority class samples.

[0067] It should be noted that in step S103, the training dataset is used to train the semantic segmentation model; therefore, data augmentation must maintain the consistency between image features and label information. When performing vertical flipping, horizontal mirroring, and 90-degree rotation operations on image patches in the minority class sample set to generate the first image patch, the same geometric transformation operation is performed on the semantic segmentation ground truth label corresponding to the first image patch, so that the first image patch and the transformed semantic segmentation ground truth label maintain a spatial correspondence.

[0068] Similarly, when generating the second image patch using mosaic data augmentation, the system randomly selects four image patches and their corresponding semantic segmentation ground truth labels from the minority class sample set. Based on the same random scaling ratio, cropping coordinates, and stitching order, it generates the second image patch and its corresponding stitched semantic segmentation ground truth labels. The semantic segmentation ground truth labels of the first image patch, the second image patch, and the first and second image patches after synchronous transformation are then merged into the land sample dataset.

[0069] Addressing the technical challenge of minority land cover categories being difficult for models to learn effectively due to significant differences in pixel count among different land cover categories in land sample datasets, this invention considers the inability of conventional data augmentation to resolve the coupling between land cover and background. By traversing and counting the number of pixels in the semantic segmentation ground truth labels for minority land cover categories, units with a pixel count greater than zero are precisely selected to construct a minority sample set. Based on this, a mosaic data augmentation technique is specifically introduced. Four image blocks are randomly selected from the minority sample set and randomly scaled, cropped, and stitched to generate a second image block. Synchronous transformation processing is performed on the semantic segmentation ground truth labels to strictly maintain the consistency between image features and label information. This not only increases the number of samples through geometric transformation but also breaks the inherent binding relationship between minority land cover categories and the original spatial background through mosaic data augmentation. By artificially constructing highly complex combined scenes, the spatial distribution diversity of minority samples is increased, effectively solving the problem that minority land cover category features are suppressed by majority samples during training.

[0070] In a specific embodiment, step S2 includes the following sub-steps:

[0071] Step S201: Construct a multi-resolution backbone network based on the training dataset. The multi-resolution backbone network includes a high-resolution sub-network and a low-resolution sub-network. The high-resolution sub-network is used as the starting stage. In subsequent stages, the low-resolution sub-network is added in parallel according to the preset stages. During the forward propagation of the multi-resolution backbone network, the multi-scale fusion operation is repeatedly performed to enable information interaction between sub-networks of different resolutions. The output features of each sub-network of resolutions are upsampled and spliced ​​together to output a fused feature map that maintains high spatial resolution.

[0072] Step S202: Input the fused feature map into the pyramid feature extraction structure, perform pyramid pooling operation using pooling kernels of different sizes to capture global and local features respectively, obtain pooling results, and then upsample the pooling results of different scales to restore them to their original size and concatenate them with the fused feature map to generate an intermediate feature map.

[0073] Step S203: Apply channel spatial collaborative attention units to the intermediate feature map, use one-dimensional convolution without dimensionality reduction to calculate cross-channel interaction information, generate channel attention weights, perform spatial average pooling along the horizontal and vertical directions respectively to generate horizontal and vertical feature vectors, encode the horizontal and vertical feature vectors using convolution transformation to generate a spatial attention weight map, and perform element-wise multiplication weighting on the intermediate feature map using the channel attention weights and the spatial attention weight map to output a local feature map.

[0074] It should be noted that in step S203, the channel-space collaborative attention unit is designed to enhance the model's sensitivity to key features. The channel-space collaborative attention unit includes necessary feature compression and numerical normalization operations at the underlying logic level, specifically including:

[0075] Before performing one-dimensional convolution, perform convolution on the dimension... The intermediate feature maps are subjected to global average pooling, compressing the spatial information of each channel into a scalar, thereby generating a dimensional vector. The global channel descriptor is input into a one-dimensional convolutional layer without dimensionality reduction. The one-dimensional convolutional layer uses a size of... The convolution kernel is used to capture local dependencies between channels. The output of the convolution must be processed by a sigmoid activation function to map the values ​​to... The interval is used to generate effective channel attention weights.

[0076] After performing spatial average pooling along the horizontal (X-axis) and vertical (Y-axis) directions to obtain orientation-aware horizontal and vertical feature vectors, respectively, the process of encoding these vectors using convolutional transformations actually involves ReLU function and dimension adjustment. Before outputting the final spatial weight map, it also needs to be processed by the Sigmoid activation function to ensure that the value of each pixel in the generated weight map is within the specified range. The values ​​between these values ​​represent the importance probability of the spatial location of a pixel.

[0077] Weighted computation specifically refers to element-wise multiplication. In particular, by using a broadcast mechanism, the generated channel attention weights are multiplied with the intermediate feature map, and then multiplied with the generated spatial attention weight map (the order can be interchanged). This suppresses background noise and enhances the feature response of the target ground object without changing the feature map size, and finally outputs a local feature map.

[0078] To address the challenges of subtle textures in a few types of land features in remote sensing imagery, which are easily obscured by complex backgrounds, conventional convolutional neural networks suffer from spatial resolution loss due to downsampling, and traditional attention mechanisms often lose crucial cross-channel interaction information due to dimensionality reduction operations in fully connected layers, this invention designs a channel-spatial collaborative attention unit based on a multi-resolution backbone network to maintain high spatial resolution. Firstly, considering that dimensionality reduction in conventional channel attention can disrupt inter-channel dependencies, global average pooling combined with non-dimensionality-reduced one-dimensional convolution directly captures cross-channel interaction information, ensuring complete preservation of feature channel correlations. Simultaneously, to solve the problem of spatial location information loss caused by ordinary global pooling, spatial average pooling is performed along both the horizontal and vertical directions, encoding the spatial location information of the two-dimensional image into horizontal and vertical feature vectors. Channel attention weights and spatial attention weight maps generated by the Sigmoid activation function, using broadcasting mechanisms and element-wise multiplication, accurately suppress background noise and significantly enhance the feature response of target land features without changing the feature map size, thereby achieving refined identification and localization of small, discrete land features.

[0079] In a specific embodiment, step S3 includes the following sub-steps:

[0080] Step S301: Divide the image data in the training dataset into n image blocks, and flatten the image blocks into a one-dimensional feature sequence according to the spatial arrangement order from left to right and from top to bottom. At the same time, embed position encoding information to mark the spatial relative position of each image block and construct a serialized feature input.

[0081] Step S302: The serialized feature input is imported into the visual state space encoder, which includes a scale visual state space block. The zero-order preservation technique is used to convert the continuous-time state space equation parameters into discretized parameters. The serialized feature input is traversed according to the spatial arrangement order in the one-dimensional feature sequence. During the traversal, the feature vector extracted from the serialized feature input at the current time is defined as the current input, and the state variable recursively calculated and stored at the previous time is defined as the historical hidden state. Through the interactive calculation of the current input and the historical hidden state, the long-range dependency relationship between image blocks at different spatial locations in the serialized feature input is captured, and the updated feature vectors are recombined to generate the output sequence after state update.

[0082] Step S303: The output sequence is dimensionally reshaped. Based on the spatial relative position, the output sequence is restored to a two-dimensional spatial structure. The long-range dependency is mapped to a spatial distribution feature spanning the entire image, generating a long-range feature map. The long-range feature map includes global semantic information.

[0083] It should be noted that the scale-based visual state space block in step S302 and the computational logic within the scale-based visual state space block involve the mathematical principles of continuous system discretization and recursive state updates. To enable those skilled in the art to understand how to implement it, further detailed explanation is provided below:

[0084] Although step S301 is described as flattening the image patch, before actually inputting it into the visual state space encoder, the system performs a linear projection operation on the flattened one-dimensional feature sequence, mapping the pixel-level image patch into a feature vector with potential dimensions, thereby ensuring that subsequent calculations are performed in a high-dimensional semantic space rather than the original pixel space.

[0085] The zero-order preservation technique mentioned in step S302 is a key step in discretizing continuous state-space equations. Specifically, the zero-order preservation technique targets the continuous state transition matrix in the state-space model. and continuous input projection matrix Introduce a learnable timescale parameter. , This also refers to the technical meaning of the term "scale" in the naming of scale-based visual state space blocks, representing the sampling step size of the model for information across different spans, according to the formula... and This converts continuous parameters into corresponding discretized parameters. and This ensures that the model can adapt to discretely sampled remote sensing image data.

[0086] The mathematical logic of interactive computation and state update involves interactive computation between the current input and historical hidden states. It is not a simple addition or subtraction operation, but a weighted iterative process based on linear recursive equations. The specific computation logic is as follows:

[0087] The historical hidden state at the current moment With discretized state transition parameters Multiplying them yields the historical state evolution component, which is the feature vector extracted from the serialized feature input at the current time step. With the discretized input parameters Multiplying these components yields the current input injection component. Adding the historical state evolution component to the current input injection component, the resulting superposition is the updated hidden state at the current time step. That is, satisfying the formula This recursive mechanism allows the hidden state at each time step to compress and encode historical information from all time steps prior to the start of the sequence, thus effectively capturing long-range dependencies.

[0088] The generation of the updated feature vector specifically includes:

[0089] The updated feature vector in step S302 After calculating the new hidden state, the projection matrix is ​​output. The result is obtained by linear transformation with the new hidden state, i.e. The output sequence is then reshaped into a two-dimensional structure in step S303 to form a long-range feature map.

[0090] To address the dilemma that cross-domain road networks and ecological corridors in high-resolution remote sensing imagery often span the entire scene, traditional convolutional neural networks (CNNs) are unable to capture global topological connections due to limited receptive fields, while the Transformer architecture suffers from quadratic increases in computational complexity when processing ultra-large images, this invention constructs a visual state-space encoder. This encoder transforms two-dimensional images into one-dimensional feature sequences and utilizes zero-order preservation techniques to accurately convert continuous-time state equation parameters into discretized parameters adapted to digital images. Through interactive computation between the current input and historical hidden states, a long-range dependency capture mechanism with linear computational complexity is established during the traversal process. This breaks the spatial limitations of traditional convolutional operations and achieves lossless capture of long-range dependencies without significantly increasing computational burden. It effectively solves the problem of semantic breaks easily generated in local analysis of large-span ground features, ensuring the continuity and integrity of the identification of ground features distributed across the entire domain.

[0091] In a specific embodiment, step S4 includes the following sub-steps:

[0092] Step S401: Construct an adaptive perception fusion module. The adaptive perception fusion module includes a feature stitching unit, a weight generation unit, and a weighted fusion unit. The feature stitching unit aligns the local feature map and the long-range feature map. The aligned local feature map and the long-range feature map are then stitched together along the channel dimension to generate a hybrid feature. The hybrid feature is input into the weight generation unit, which calculates a spatial attention weight map through convolution operations and activation functions. The spatial attention weight map is used to characterize the dependence of each pixel position on local and global features. The weighted fusion unit performs pixel-by-pixel weighted summation of the local feature map and the long-range feature map based on the spatial attention weight map to generate a full-scale feature map. The full-scale feature map includes local texture details and global semantic information.

[0093] Step S402: Classify and predict the full-scale feature map to obtain the predicted probability distribution and predicted region. Construct a composite loss function based on the true label distribution and the true region in the training dataset.

[0094] The composite loss function includes a cross-entropy loss term and a Dice coefficient loss term. The cross-entropy loss term calculates the difference between the predicted probability distribution and the true label distribution, while the Dice coefficient loss term calculates the error between the predicted region and the true region. The difference and the error are weighted and summed to obtain the total loss value of the current training batch.

[0095] It should be noted that the calculation logic of the Dice coefficient loss term is to first calculate the set overlap between the predicted region and the real region, and then subtract the set overlap from 1 to obtain the error.

[0096] Step S403: Concatenate the multi-resolution backbone network, pyramid feature extraction structure, visual state space encoder and adaptive perception fusion module to construct the first land classification model, and divide the training dataset into training subset and validation subset according to a preset ratio.

[0097] Iterative training is performed on the first land classification model based on a training subset. The parameters of the first land classification model are updated according to the total loss value using the backpropagation algorithm. Specifically, this includes:

[0098] At the end of each training round, the validation subset is predicted using the first land classification model to construct a confusion matrix. Based on the confusion matrix, the precision and recall of each land cover category are calculated. The harmonic mean of precision and recall is defined as the F1 score. The standard deviation and mean of the F1 scores of all land cover categories are calculated, and the ratio of the standard deviation to the mean is defined as the coefficient of variation.

[0099] When the coefficient of variation is greater than the preset equilibrium threshold, the first land classification model is determined to be biased towards the majority class. Based on the reciprocal of the F1 score, the weight of the minority class in the composite loss function in the next training round is increased. The training is repeated until the total loss value converges, and the second land classification model is output.

[0100] It should be noted that in step S401, due to the slight difference in spatial resolution or number of channels between the local feature map and the long-range feature map, the feature stitching unit first uses bilinear interpolation to unify the local feature map and the long-range feature map to the same spatial resolution before performing stitching, and then uses a 1×1 convolutional layer to map the local feature map and the long-range feature map to the same channel dimension, ensuring that the local feature map and the long-range feature map can be aligned and calculated at the pixel level.

[0101] Spatial attention weight map After Sigmoid activation, the values ​​are distributed as follows: Between. The weighted fusion unit adopts a complementary weighting strategy, using the formula for each pixel position on the local feature map and the long-range feature map. Calculations are performed, in which, For full-scale feature maps, Represents element-wise multiplication. For local feature maps, As a long-range feature map, it ensures that the model can adaptively decide at each location whether to preserve more local texture details or introduce global contextual information.

[0102] To maintain a consistent minimization objective, the Dice coefficient loss term does not directly use set overlap, but is defined as 1 minus set overlap, i.e. ,in, The error calculated for the Dice coefficient loss term. The set overlap is expressed mathematically as follows:

[0103] ;

[0104] in, This refers to the set of pixels corresponding to the predicted region output by the model. This is the set of pixels corresponding to the true regions in the semantic segmentation ground truth labels. This represents the number of pixels in the intersection of the predicted region's pixel set and the actual region's pixel set. and These represent the total number of pixels contained in the predicted region and the actual region, respectively.

[0105] Therefore, the total loss value The mathematical expression is:

[0106] ;

[0107] in, To balance the factors, ensuring that the higher the overlap between the predicted and actual regions, the smaller the total loss value is included. The cross-entropy loss term measures the difference between the predicted probability distribution and the true label distribution. In complex land cover scenarios, assigning the cross-entropy loss term and the Dice coefficient loss term the same initial optimization weights, i.e., 0.5 and (1-0.5), forces the first land classification model to simultaneously consider the accurate classification of local texture details and the structural integrity of macroscopic land cover boundaries in the feature space. This avoids the model getting trapped in local optima that overfit the majority class background or leads to patch fragmentation. Therefore, the balance coefficient... The value is 0.5.

[0108] In step S403, the weight ratio of minority class samples in the composite loss function in the next training round is increased based on the reciprocal of the F1 score. Specifically, this refers to adjusting the class weight parameter in the cross-entropy loss term. When the first land classification model is detected to be biased towards the majority class, a weighting coefficient greater than 1 is generated based on the reciprocal of the minority class F1 score. and will When applied to the corresponding category in the weighted cross-entropy formula, that is... This forces the first land classification model to pay more attention to classification errors of minority class samples in the next round. The class weighting factor is dynamically generated based on the F1 score of the minority class samples. For true labels (0 or 1). The model predicts the probability that a corresponding sample belongs to the minority class.

[0109] To address the technical challenges in land cover classification, such as the tendency of models to predict the majority class due to sparse illegal construction points against a backdrop of large arable land, and the difficulty of conventional loss functions in balancing pixel-level classification accuracy with region-level shape similarity, this invention constructs a dynamic imbalanced optimization composite loss function training mechanism. First, in terms of model architecture, an adaptive perception fusion module utilizes a spatial attention map with complementary weights to achieve pixel-level dynamic optimization of local details and long-range semantics. Second, in terms of training strategy, it breaks away from the traditional fixed-weight training mode, introducing a coefficient of variation index to monitor the dispersion of F1 scores for each class in real time. Once a model bias is detected, the penalty weight for minority class samples in the cross-entropy loss is dynamically increased using the reciprocal of the F1 score. Combined with the constraint of region overlap by the Dice coefficient loss term, the model's majority class bias is dynamically corrected during training. This significantly improves the recognition accuracy and boundary integrity of sparse small target features in a long-tailed distribution without artificially discarding majority class sample information.

[0110] In a specific embodiment, step S5 includes the following sub-steps:

[0111] Step S501: Based on the second land classification model, perform pixel-by-pixel classification prediction on the remote sensing image data of the target area in the start year and end year, generate thematic maps of land use status corresponding to the start year and end year respectively, and perform denoising and smoothing processing on the thematic maps of land use status to obtain the spatial distribution boundaries of each land use type.

[0112] It should be noted that, in order to avoid artificially increasing the number of patches and perimeters due to isolated noise points generated during the classification process, which would lead to an inflated index, a morphological opening operation with a 3×3 window is used to process the current land use status before extracting the boundary. The operation first erodes and then expands, smoothing the jagged boundaries while preserving the shape of the main land features, thus ensuring the robustness of the geometric perimeter calculation.

[0113] Step S502: Spatial overlay analysis of the land use status thematic maps of the starting year and the ending year is performed to construct a land use type transfer matrix. Based on the land use type transfer matrix, the transfer-out area, transfer-in area and retention area of ​​each land use type are calculated. The ratio of the retention area to the initial area of ​​the starting year is calculated to obtain the land use stability index. The land use stability index characterizes the survival capacity of each land use type.

[0114] Step S503: Based on the land use type transition matrix, extract the net change in area of ​​the target land use type during the study period, calculate the ratio of the absolute value of the net change in area to the initial area, and divide the obtained ratio by the time span of the study period to obtain a single land use dynamic index. The single land use dynamic index characterizes the evolution rate of the target land use type.

[0115] It should be noted that net area change refers to the difference in area between the target land use type at the end and the beginning of the study period. The mathematical expression for net area change is:

[0116] ;

[0117] in, This represents the net change in area. The total area of ​​the target land use type in the year of termination. The initial area of ​​the target land use type in the starting year.

[0118] Step S504 involves analyzing the evolution of landscape patterns based on spatial distribution boundaries, specifically including:

[0119] A topological analysis was performed on the spatial distribution boundary to extract the geometric perimeter of patches in each land use type. The number of patches in each land use type was counted, and the sum of the retained area and the area transferred in was recorded as the current total area of ​​each land use type.

[0120] The landscape shape index is calculated based on the geometric perimeter and the current total area. The landscape shape index characterizes the shape complexity of land use type boundaries.

[0121] Based on the number of patches and the total area of ​​the target region, the patch density index is calculated, which characterizes the degree of landscape fragmentation in the target region.

[0122] Step S505: The land use stability index, single land use dynamic index, landscape shape index and patch density index are used as the land management analysis results. Based on the land management analysis results, a multi-dimensional land use comprehensive evaluation report is generated. The report is used to characterize the evolution rate of land use change, spatial fragmentation trend and land type survival stability of the target area during the study period.

[0123] It should be noted that the single land use dynamic index is used to characterize the rate of change of a certain land use type within the study period. The mathematical expression is:

[0124] ;

[0125] in, For the initial area, The total area at the end of the year. The time span of the study period is measured in years.

[0126] Landscape Shape Index The complexity of a patch's shape is characterized by the relationship between its geometric perimeter and area; the greater the deviation from a square or circle, the higher the landscape shape index. Considering the raster characteristics of remote sensing data, this invention employs the following correction formula:

[0127] ;

[0128] in, This represents the sum of the geometric perimeters, which is obtained by statistically analyzing the grid boundary lengths. This represents the current total area for each land use type.

[0129] Patch density index Patch density index, which characterizes the degree of landscape fragmentation. The mathematical expression is:

[0130] ;

[0131] in, The number of patches for each land use type. This refers to the total area of ​​the target region, not the area of ​​a single land type.

[0132] Traditional land management analysis often focuses on area statistics or simple classification mapping in a single time period, lacking in-depth quantitative methods for land use change processes and micro-characteristics of landscape morphology. This leads to the technical pain point that managers struggle to accurately grasp the activity level of land use changes and spatial fragmentation trends. This invention, based on obtaining spatial distribution boundaries and land use type transfer matrices using a high-precision classification model, constructs a joint analysis system of dynamic change and spatial morphology. This system goes beyond simply acquiring basic data; it further deconstructs land use status by incorporating spatial distribution boundaries. By calculating the land use stability index and a single land use dynamic index, it accurately quantifies the rate and intensity of land use change. Simultaneously, it introduces a landscape shape index to reveal the complex spatial morphological characteristics of land use from a geometric and topological perspective. This progressive processing logic, from quantitative transfer quantification to micro-morphological analysis, generates land management analysis results containing multi-dimensional indicators, providing a quantifiable scientific basis for identifying ecological fragmentation risks and monitoring the disorderly expansion of construction land in national spatial planning.

[0133] Example 2, refer to Figure 2 This paper presents a land management and analysis system based on big data, which includes a data acquisition module, an extraction module, a construction module, a training module, and an analysis module.

[0134] The acquisition module is used to acquire remote sensing image data and geographic national condition monitoring vector data of the target area, screen a few land cover samples and perform geometric transformation and mosaic data augmentation to generate a training dataset.

[0135] The extraction module is used to extract features from the training dataset using a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to generate local feature maps.

[0136] The building block is used to convert the training dataset into a serialized feature input, and uses a visual state-space encoder to perform discretization parameter transformation and recursive state update of the state-space equation to capture long-range dependencies and generate long-range feature maps.

[0137] The training module is used to build a first land classification model based on a multi-resolution backbone network, a pyramid feature extraction structure, and a visual state space encoder. It performs weighted fusion of local feature maps and long-range feature maps to generate a full-scale feature map, constructs a composite loss function, and iteratively trains to output a second land classification model.

[0138] The analysis module is used to predict the spatial distribution boundary of the target area based on remote sensing image data in the start and end years, construct a land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain land management analysis results.

[0139] To address the technical challenge of existing land management analysis systems failing to simultaneously capture the subtle textures of land features in remote sensing imagery and their cross-domain global spatial distribution, leading to missed detections of sparse minority land feature samples and an inability to accurately quantify spatiotemporal evolution, this system innovatively constructs a modular architecture with clearly defined functions. First, the acquisition module utilizes mosaic data enhancement to resolve sample imbalance. Then, an extraction module is designed to focus on mining local high-resolution features using channel space collaborative attention units. Simultaneously, an independently set construction module utilizes a visual state space encoder to specifically capture long-range dependencies through discretized parameter transformation and recursive state updates. This two-pronged design enables the training module to train a second land classification model based on full-scale feature maps, possessing both local sensitivity and global perspective. Finally, the support analysis module generates multi-dimensional land management analysis results based on accurate spatial distribution boundaries and land use type transition matrices, achieving refined panoramic control over regional land use changes.

[0140] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0141] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. A land management analysis method based on big data, characterized in that, Includes the following steps: Step S1: Acquire remote sensing image data and geographic national condition monitoring vector data of the target area, screen a few land cover samples and perform geometric transformation and mosaic data augmentation to generate a training dataset; Step S2: Use a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to extract features from the training dataset and generate local feature maps. Step S3: Convert the training dataset into serialized feature input, and obtain a long-range feature map based on the serialized feature input; Step S3 includes the following sub-steps: Step S301: Divide the image data in the training dataset into n image blocks, and flatten the image blocks into a one-dimensional feature sequence according to the spatial arrangement order from left to right and from top to bottom. At the same time, embed position encoding information to mark the spatial relative position of each image block and construct a serialized feature input. Step S302: The serialized feature input is imported into the visual state space encoder, which includes a scale visual state space block. The continuous-time state space equation parameters are converted into discretized parameters using the zero-order preservation technique. The serialized feature input is traversed according to the spatial arrangement order in the one-dimensional feature sequence. During the traversal, the feature vector extracted from the serialized feature input at the current time is defined as the current input, and the state variable recursively calculated and stored at the previous time is defined as the historical hidden state. Through the interactive calculation between the current input and the historical hidden state, the long-range dependency relationship between image blocks at different spatial locations in the serialized feature input is captured, and the updated feature vector is recombined to generate the output sequence after state update. Step S303: The output sequence is dimensionally reshaped, and the output sequence is restored to a two-dimensional spatial structure according to the spatial relative position. The long-range dependency is mapped to a spatial distribution feature spanning the entire image, and a long-range feature map is generated. Step S4: Construct the first land classification model based on the multi-resolution backbone network, pyramid feature extraction structure and visual state space encoder; obtain the full-scale feature map based on local feature map and long-range feature map; construct the composite loss function and iteratively train to output the second land classification model. Step S4 includes the following sub-steps: Step S401: Construct an adaptive perception fusion module, which includes a feature stitching unit, a weight generation unit, and a weighted fusion unit. The feature stitching unit aligns the local feature map and the long-range feature map, and then stitches the aligned local and long-range feature maps along the channel dimension to generate a hybrid feature. The hybrid feature is then input into the weight generation unit, which calculates a spatial attention weight map through convolution operations and activation functions. The spatial attention weight map is used to characterize the degree of dependence of each pixel position on local and global features. The weighted fusion unit performs a pixel-by-pixel weighted summation of the local feature map and the long-range feature map based on the spatial attention weight map to generate a full-scale feature map. Step S5: Based on the second land classification model, predict the remote sensing image data of the target area to obtain the spatial distribution boundary, construct the land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain the land management analysis results.

2. The land management analysis method based on big data as described in claim 1, characterized in that, Step S1 includes the following sub-steps: Step S101: Obtain remote sensing image data and geographic national condition monitoring vector data of the target area; perform radiometric correction, geometric correction and image fusion processing on the remote sensing image data; and perform spatial registration and cropping on the processed remote sensing image data and geographic national condition monitoring vector data to generate a land sample dataset with uniform resolution and size. Step S102: Calculate the percentage of pixels for each land cover category in the land sample dataset, mark land cover categories with a percentage of pixels less than a preset percentage threshold as minority land cover categories, and traverse the land sample dataset to select image blocks that include the minority land cover categories, and construct a minority sample set based on the image blocks.

3. The land management analysis method based on big data as described in claim 2, characterized in that, Step S1 further includes the following sub-steps: Step S103: Perform geometric transformation operations of vertical flipping, horizontal mirroring, and 90-degree rotation sequentially on the image blocks in the minority class sample set to generate the first image block. At the same time, apply mosaic data augmentation technology to randomly select four image blocks from the minority class sample set for random scaling, cropping, and stitching to generate the second image block. Merge the first and second image blocks into the land sample dataset to generate the training dataset. The training dataset includes remote sensing image blocks and semantic segmentation ground truth labels corresponding to each remote sensing image block. The semantic segmentation ground truth labels include the true label distribution and true area of ​​each land cover category. Land cover categories other than minority class samples in the training dataset are defined as majority class samples.

4. The land management analysis method based on big data as described in claim 3, characterized in that, Step S2 includes the following sub-steps: Step S201: Construct a multi-resolution backbone network based on the training dataset. The multi-resolution backbone network includes a high-resolution sub-network and a low-resolution sub-network. The high-resolution sub-network is used as the starting stage. In subsequent stages, the low-resolution sub-network is added in parallel according to a preset stage. During the forward propagation of the multi-resolution backbone network, the multi-scale fusion operation is repeatedly performed to enable information interaction between sub-networks of different resolutions. The output features of each sub-network of resolutions are upsampled and spliced ​​together to output a fused feature map that maintains high spatial resolution. Step S202: Input the fused feature map into the pyramid feature extraction structure, perform pyramid pooling operation using pooling kernels of different sizes to capture global and local features respectively, obtain pooling results, and then upsample the pooling results of different scales to restore them to their original size and concatenate them with the fused feature map to generate an intermediate feature map; Step S203: Apply a channel spatial collaborative attention unit to the intermediate feature map, use one-dimensional convolution without dimensionality reduction to calculate cross-channel interaction information, generate channel attention weights, perform spatial average pooling along the horizontal and vertical directions respectively to generate horizontal and vertical feature vectors, encode the horizontal and vertical feature vectors using convolution transformation to generate a spatial attention weight map, and perform element-wise multiplication weighting on the intermediate feature map using the channel attention weights and spatial attention weight map to output a local feature map.

5. The land management analysis method based on big data as described in claim 4, characterized in that, Step S4 further includes the following sub-steps: Step S402: Classify and predict the full-scale feature map to obtain the predicted probability distribution and predicted region. Construct a composite loss function based on the true label distribution and true region in the training dataset. The composite loss function includes a cross-entropy loss term and a Dice coefficient loss term. The difference between the predicted probability distribution and the true label distribution is calculated using the cross-entropy loss term, and the error between the predicted region and the true region is calculated using the Dice coefficient loss term. The difference and the error are weighted and summed to obtain the total loss value of the current training batch.

6. The land management analysis method based on big data as described in claim 5, characterized in that, Step S4 further includes the following sub-steps: Step S403: The multi-resolution backbone network, pyramid feature extraction structure, visual state space encoder and adaptive perception fusion module are cascaded to construct the first land classification model, and the training dataset is divided into training subset and validation subset according to a preset ratio; Iterative training is performed on the first land classification model based on the training subset, and the parameters of the first land classification model are updated according to the total loss value using the backpropagation algorithm, specifically including: At the end of each training round, the validation subset is predicted using the first land classification model to construct a confusion matrix. Based on the confusion matrix, the precision and recall of each land cover category are calculated. The harmonic mean of the precision and recall is defined as the F1 score. The standard deviation and mean of the F1 scores of all land cover categories are calculated. The ratio of the standard deviation and mean is defined as the coefficient of variation. When the coefficient of variation is greater than the preset balance threshold, it is determined that the first land classification model is biased towards the majority class. Based on the reciprocal of the F1 score, the weight ratio of the minority class samples in the composite loss function in the next training round is increased. The training is repeated until the total loss value converges, and the second land classification model is output.

7. The land management analysis method based on big data as described in claim 6, characterized in that, Step S5 includes the following sub-steps: Step S501: Based on the second land classification model, perform pixel-by-pixel classification prediction on the remote sensing image data of the target area in the start year and end year, generate the thematic maps of land use status corresponding to the start year and end year respectively, and perform noise reduction and smoothing on the thematic maps of land use status to obtain the spatial distribution boundaries of each land use type. Step S502: Spatial overlay analysis of the land use status thematic maps of the starting year and the ending year is performed to construct a land use type transfer matrix. Based on the land use type transfer matrix, the transfer-out area, transfer-in area and retention area of ​​each land use type are calculated. The ratio of the retention area to the initial area of ​​the starting year is calculated to obtain the land use stability index. Step S503: Based on the land use type transfer matrix, extract the net change in area of ​​the target land use type during the study period, calculate the ratio of the absolute value of the net change in area to the initial area, and divide the obtained ratio by the time span of the study period to obtain a single land use dynamic index. Step S504, based on the spatial distribution boundary, performs landscape pattern evolution analysis, specifically including: A topological analysis is performed on the spatial distribution boundary to extract the geometric perimeter of patches in each land use type, and the number of patches in each land use type is counted. The sum of the reserved area and the transferred area is recorded as the current total area of ​​each land use type. Calculate the landscape shape index based on the geometric perimeter and the current total area; Calculate the patch density index based on the number of patches and the total area of ​​the target region; Step S505: The land use stability index, single land use dynamic index, landscape shape index and patch density index are used as the land management analysis results, and a multi-dimensional land use comprehensive evaluation report is generated based on the land management analysis results.

8. A land management analysis system based on big data, applied in a land management analysis method based on big data as described in any one of claims 1-7, characterized in that, It includes a data acquisition module, an extraction module, a construction module, a training module, and an analysis module; The acquisition module is used to acquire remote sensing image data and geographic national condition monitoring vector data of the target area, screen a few land cover samples and perform geometric transformation and mosaic data enhancement to generate a training dataset. The extraction module is used to extract features from the training dataset using a multi-resolution backbone network, a pyramid feature extraction structure, and a channel space collaborative attention unit to generate local feature maps. The building module is used to convert the training dataset into a serialized feature input, and uses a visual state space encoder to perform discretization parameter transformation and recursive state update of the state space equation to capture long-range dependencies and generate a long-range feature map. The training module is used to build a first land classification model based on a multi-resolution backbone network, a pyramid feature extraction structure and a visual state space encoder, to perform weighted fusion of local feature maps and long-range feature maps to generate a full-scale feature map, to construct a composite loss function, and to iteratively train and output a second land classification model. The analysis module is used to predict the spatial distribution boundary of the target area using remote sensing image data in the start and end years, construct a land use type transition matrix, and analyze each land use type in combination with the spatial distribution boundary to obtain land management analysis results.

Images