An intelligent land use identification method and system

Abstract

The application relates to the technical field of intelligent remote sensing information processing, and discloses a method and system for intelligent land use identification. The method comprises receiving multi-temporal remote sensing data, extracting spectral and spatial features, and fusing terrain and soil data; using a feature selection algorithm to evaluate feature importance, combining environmental data to dynamically adjust feature weights for fusion clustering; calculating the long-term change rate and short-term fluctuation amplitude of the time series features, dividing the stability and changeability regions and extracting the representative features respectively; training and optimizing the integrated learning model based on these features, and finally realizing high-precision land use identification. The method reduces the complex environmental interference through dynamic feature weight optimization, accurately captures the dynamic change law of the land class by using the time series stability partition, and thus improves the distinguishing accuracy and timeliness of the spectral similar land class and the periodic change land class.

Description

Technical Field

[0001] This invention relates to the field of intelligent processing technology of remote sensing information, specifically to an intelligent land use identification method and system. Background Technology

[0002] Currently, land use identification based on remote sensing imagery mainly relies on the analysis of images from a single period or a limited number of temporal phases. Conventional techniques typically extract spectral, textural, and spatial structure features from images and directly input them into classification models for training. When processing multi-source data, the commonly used feature fusion strategies are often simple stitching or superposition based on fixed empirical weights, failing to fully consider the dynamic influence of environmental factors such as topography and soil on the discriminative ability of different features. This rigid feature processing approach limits the accuracy of models in identifying land parcels in complex geographical environments.

[0003] Existing methods for utilizing time-series information often focus on the increase of observation time points or simple change detection. A common approach is to stack image features from different time phases as input, hoping the model can learn temporal patterns on its own. This approach fails to explicitly quantify and analyze the intrinsic patterns of land cover dynamics. The model struggles to effectively distinguish land types with similar spectral characteristics but vastly different temporal behaviors.

[0004] Existing technologies suffer from weak correlations between feature fusion strategies and geoscientific mechanisms, as well as a limited temporal analysis dimension that fails to accurately characterize the dynamic stability of land use. This results in spatial ambiguity and temporal lag in land use identification results, making it difficult to meet the application requirements of high-precision, real-time dynamic monitoring. Achieving adaptive optimization and fusion of multi-source features and effectively analyzing the stability and change information contained in time series data is a key challenge for improving the accuracy and reliability of land use identification. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent land use identification method and system to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides an intelligent land use identification method, the method comprising: It receives multi-temporal image data from remote sensing satellites and UAVs, and performs radiometric and geometric correction on the image data; Spectral reflectance features and spatial structure features are extracted from the corrected image data, while topographic elevation data and soil physicochemical property data are obtained. A feature selection algorithm was used to evaluate the importance scores of spectral reflectance features and spatial structure features, and the feature weights were adjusted by combining topographic elevation data and soil physicochemical property data. Feature fusion is performed based on the adjusted feature weights to generate a multimodal feature set, and cluster analysis is performed on the multimodal feature set to identify potential land use patterns. By introducing time dimension analysis, the multimodal feature sets corresponding to multi-temporal image data are arranged in chronological order, and the long-term rate of change and short-term fluctuation amplitude of the feature values ​​are calculated. Based on the long-term rate of change and the short-term fluctuation range, land use stability areas and change areas are divided, and representative features are extracted for each. An ensemble learning model is trained using representative features, and its hyperparameters are optimized using a grid search method. The feature data of the area to be identified is input into the optimized ensemble learning model, and the land use classification results are output.

[0007] Preferably, the importance scores for evaluating spectral reflectance features and spatial structure features using the feature selection algorithm include: Calculate the mutual information value between each spectral reflectance feature and the land use category to obtain a preliminary score of the importance of the spectral features; Calculate the chi-square statistic between each spatial structure feature and land use category to obtain a preliminary score of spatial feature importance; The initial scores for the importance of spectral features and spatial features are combined, and the final score for the importance of features is determined using the entropy weighting method.

[0008] Preferably, the adjustment of feature weights by combining topographic elevation data and soil physicochemical property data includes: The topographic elevation data was divided into multiple elevation zones, and the distribution frequency of land use types within each elevation zone was statistically analyzed. Normalize the soil physicochemical property data to the same dimension and calculate the correlation coefficient between soil parameters and land use type; Topographic adjustment factors and soil adjustment factors are generated based on distribution frequency and correlation coefficient; The adjusted feature weights are obtained by multiplying the topography adjustment factor and the soil adjustment factor by the feature importance score.

[0009] Preferably, the clustering analysis of the multimodal feature set to identify potential land use patterns includes: A density clustering algorithm is used to group the multimodal feature set to obtain multiple feature clusters; Calculate the centroid and contour coefficient of each feature cluster, and select feature clusters with contour coefficients greater than the threshold as potential land use patterns; Each potential land use pattern is assigned a temporary label, and the range of its characteristic distribution is recorded.

[0010] Preferably, the calculation of the long-term rate of change and short-term fluctuation range of the characteristic value includes: Time series modeling is performed on the multimodal feature set corresponding to multi-temporal image data, and the trend curve of the feature values ​​is fitted. Calculate the average slope of the trend curve within a preset time window as the long-term rate of change; The residual standard deviation of the eigenvalues ​​and their trend curves is calculated as the short-term volatility.

[0011] Preferably, the division of land use stability zones and change zones includes: Set long-term rate of change thresholds and short-term volatility thresholds; Regions with long-term rate of change below the threshold and short-term fluctuation amplitude below the threshold are marked as stability regions. Regions with long-term rates of change exceeding a threshold or short-term fluctuations exceeding a threshold are marked as volatile regions.

[0012] Preferably, the extraction of representative features includes: For the stable region, the mean and variance of the multimodal feature set are calculated as representative features; For regions with variability, the time series extrema and gradient of the multimodal feature set are extracted as representative features.

[0013] Preferably, training the ensemble learning model using representative features includes: A hybrid model of random forest and gradient boosting decision tree is constructed, in which random forest is used to handle features of stable regions and gradient boosting decision tree is used to handle features of variable regions. The performance of the hybrid model was evaluated using ten-fold cross-validation, and the optimal model parameters were selected. The optimization of model hyperparameters using the grid search method includes: Define the number of trees and maximum depth search range for a random forest, and the learning rate and number of iterations search range for a gradient boosting decision tree; Within the search range, perform parameter combination traversal and calculate the cross-validation accuracy for each parameter group; Choose the parameter combination that yields the highest cross-validation accuracy as the optimal hyperparameter.

[0014] Preferably, the step of inputting the feature data of the region to be identified into the optimized ensemble learning model includes: The multimodal feature set of the region to be identified is acquired in real time, and the same preprocessing and feature extraction are performed as on the training data. The preprocessed feature data is input into the optimized ensemble learning model, and the model outputs the land use probability distribution. The final land use category is determined based on the maximum value of the probability distribution.

[0015] Preferably, the present invention also includes an intelligent land use identification system, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the intelligent land use identification method described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This technique, which assesses the importance of spectral and spatial features using feature selection algorithms and dynamically adjusts feature weights by integrating topographic elevation and soil physicochemical property data, departs from the traditional method of simply superimposing or fusioning multi-source features with fixed weights. Topographic and soil data, as key environmental drivers, directly participate in feature weight allocation decisions, providing geoscientific support for the feature fusion process. This dynamic weighting mechanism can adaptively optimize feature combinations based on specific geographic scenarios. The resulting multimodal feature set enhances the ability to represent land use in complex environments, reduces classification ambiguities caused by spectral similarities but different origins, and strengthens the model's adaptability to geographic environmental heterogeneity.

[0017] This approach, which analyzes time series data using two quantitative indicators—long-term rate of change and short-term fluctuation amplitude—and then divides the data into stable and variable regions before extracting representative features from each, achieves a refined diagnosis of land dynamic patterns. The long-term rate of change accurately captures the trend of land use type evolution, while the short-term fluctuation amplitude effectively characterizes seasonal, periodic, or sudden changes. This zoning strategy makes feature extraction highly targeted: stable regions focus on their spatiotemporal consistency characteristics, while variable regions highlight their dynamic patterns. This processing method provides the classification model with highly discriminative and clearly patterned training samples, enabling the model to accurately identify land types with vastly different temporal behaviors simultaneously, thereby improving the accuracy and robustness of identification in complex dynamic land use scenarios. Attached Figure Description

[0018] Figure 1 This is a schematic diagram illustrating the working principle of the intelligent land use identification method described in this invention. Figure 2 A flowchart for evaluating feature importance scores; Figure 3 A flowchart for cluster analysis to identify potential land use patterns; Figure 4 A chart analyzing the temporal changes in land use indicators; Figure 5 This is a graph showing the relationship between the hyperparameters of a random forest and its classification accuracy. Figure 6 This is a time series characteristic diagram of the long-term change rate and short-term fluctuation range of land use types. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figure 1 This invention provides an intelligent land use identification method, comprising: receiving multi-temporal image data from remote sensing satellites and UAVs, and performing radiometric and geometric correction on the image data to eliminate atmospheric interference and geometric distortion. Spectral reflectance features and spatial structure features are extracted from the corrected image data, while topographic elevation data and soil physicochemical property data are acquired simultaneously. The spectral reflectance features include multi-band reflectance values, and the spatial structure features involve texture and shape indices. A feature selection algorithm is used to evaluate the importance scores of the spectral reflectance features and spatial structure features, and the feature weights are adjusted in conjunction with the topographic elevation data and soil physicochemical property data. The feature selection algorithm is based on statistical metrics. Feature fusion is performed according to the adjusted feature weights to generate a multimodal feature set, and cluster analysis is performed on the multimodal feature set to identify potential land use patterns. The cluster analysis employs unsupervised learning. A time dimension analysis is introduced, arranging the multimodal feature set corresponding to the multi-temporal image data in chronological order, and calculating the long-term rate of change and short-term fluctuation amplitude of the feature values. The long-term rate of change characterizes trend changes, and the short-term fluctuation amplitude reflects random variation. Land use is divided into stable and variable zones based on long-term change rates and short-term fluctuation amplitudes, and representative features are extracted for each zone. Stable zones exhibit relatively constant features, while variable zones show significant dynamics. An ensemble learning model is trained using these representative features, and its hyperparameters are optimized using a grid search method. The ensemble learning model incorporates multiple algorithms. The feature data of the area to be identified is input into the optimized ensemble learning model, which outputs land use classification results based on the probability maximization principle.

[0021] Example 1: See Figure 2In its implementation, the feature selection algorithm evaluates the importance scores of spectral reflectance features and spatial structure features. It calculates the mutual information value between each spectral reflectance feature and land use category. Spectral reflectance features are the spectral values ​​of pixels in multi-band remote sensing images, and land use categories are predefined classification systems such as cultivated land, forest land, and construction land. The mutual information value measures the statistical dependence between the spectral reflectance feature and the land use category variable; a higher mutual information value indicates a greater contribution of the spectral reflectance feature to distinguishing different land use categories, thus obtaining a preliminary score for the importance of the spectral feature. The algorithm then calculates the chi-square statistic between each spatial structure feature and the land use category. Spatial structure features include texture features (such as entropy and contrast) and shape features (such as area and aspect ratio). The chi-square test is used to assess the independence between different values ​​of the spatial structure feature and the distribution of land use categories; a higher chi-square statistic indicates a more significant association between the spatial structure feature and the land use category, thus obtaining a preliminary score for the importance of the spatial feature.

[0022] In some embodiments, calculating the mutual information value requires estimating the joint probability distribution and marginal probability distribution of spectral reflectance features and land use categories. Histogram methods or kernel density estimation methods are typically used for probability density estimation. The chi-square statistic is calculated based on a contingency table constructed from discretized spatial structure features and land use categories, obtained by comparing the difference between observed frequencies and expected frequencies. The preliminary scores for spectral feature importance and spatial feature importance are fused, and the entropy weighting method is used to determine the final feature importance score. The entropy weighting method is an objective weighting method based on information entropy; the smaller the information entropy, the greater the variability of the feature importance score and the more information it contains, thus requiring a higher weight. Specifically, the information entropy of the preliminary score vectors for spectral feature importance and spatial feature importance is calculated separately. Then, the weight coefficients of each vector are calculated based on the information entropy, and finally, the preliminary scores are weighted and fused to obtain the final feature importance score. Optionally, spectral reflectance features may include reflectance values ​​in blue, green, red, and near-infrared bands, or derived features such as vegetation indices calculated from these bands. Spatial structural features are extracted from remote sensing imagery using image segmentation or filtering algorithms. Mutual information can be understood as capturing nonlinear relationships between variables, making it suitable for evaluating continuous spectral reflectance features. The chi-square test is suitable for evaluating the correlation between discrete or discretized spatial structural features and categorical variables. The application of entropy weighting avoids the bias of subjective weighting, making feature importance assessment more objective. The final feature importance score will be used for subsequent feature weight adjustment and feature selection to optimize the performance of the land use identification model.

[0023] In some embodiments, feature weights are adjusted by combining topographic elevation data and soil physicochemical property data. The topographic elevation data is divided into multiple continuous elevation zones, and the distribution frequency of known land use types within each elevation zone is statistically analyzed. The distribution frequency reflects the probability of land use occurring under specific elevation conditions. Soil physicochemical property data, including pH value, organic matter content, and soil texture composition, are normalized to a range of zero to one, and the Pearson correlation coefficient between the normalized soil parameters and land use types is calculated. Topographic adjustment factors and soil adjustment factors are generated based on the distribution frequency and correlation coefficient. The topographic adjustment factor is proportional to the distribution frequency, and the soil adjustment factor is proportional to the absolute value of the correlation coefficient. The topographic adjustment factor and soil adjustment factor are multiplied by the final feature importance score to obtain the adjusted feature weights.

[0024] Optionally, the mutual information value is calculated based on the joint probability distribution and marginal probability distribution of spectral reflectance features and land use categories. A larger mutual information value indicates a greater contribution of spectral reflectance features to distinguishing land use categories. The chi-square statistic is calculated based on a contingency table composed of spatial structure features and land use categories. A larger chi-square statistic indicates a stronger correlation between spatial structure features and land use categories. The entropy weight method processes the preliminary score vectors of spectral feature importance and spatial feature importance. The entropy weight method determines the weight coefficient of each vector by calculating information entropy. The smaller the information entropy, the greater the information contained in the vector, and the higher its weight coefficient. The specific value of the topography adjustment factor is related to the concentration of dominant land use types within the elevation zone, and the specific value of the soil adjustment factor is related to the explanatory power of soil parameters for changes in land use types.

[0025] In practice, the final feature importance score, topography adjustment factor, and soil adjustment factor are all used in the calculation of feature weights. The adjusted feature weights are then used in subsequent feature fusion steps, giving higher weights to features that contribute more to land use identification. The feature weight adjustment process integrates multi-source information such as remote sensing image features, topographic constraints, and soil properties.

[0026] Example 2: See Figure 3In practical implementation, cluster analysis is performed on the multimodal feature set to identify potential land use patterns. The density clustering algorithm identifies cluster structure by calculating the distribution density of sample points in the feature space. Its core parameters include neighborhood radius and minimum number of points. The neighborhood radius defines the spherical region centered on each sample point for evaluating local density, and the minimum number of points specifies the minimum number of sample points required for a dense region. The algorithm first traverses each sample point in the multimodal feature set, calculating the number of neighboring points within its neighborhood radius. If the number of neighboring points of a sample point reaches or exceeds the minimum number of points, the sample point is marked as a core point. Feature clusters are formed based on core points. Starting from a core point, all sample points (including boundary points) within its neighborhood radius are assigned to the same feature cluster, and the neighborhood of newly added core points is recursively examined to expand the feature cluster until no new points can be added. Points that do not belong to the neighborhood of any core point are marked as noise points, and noise points do not belong to any feature cluster. Through this process, the multimodal feature set is divided into multiple feature clusters and a possible set of noise points. Each feature cluster corresponds to a densely distributed region of a potential land use pattern in the feature space. In some embodiments, the multimodal feature set needs to be standardized before being input into the density clustering algorithm to eliminate the influence of differences in the dimensions and numerical ranges of different features on distance calculation. The neighborhood radius parameter is determined by analyzing the distribution of k-nearest neighbor distances between sample points, for example, by selecting the distance value corresponding to the inflection point of the k-nearest neighbor distance curve. The minimum number of points parameter is usually set based on the total sample size of the multimodal feature set and the desired cluster size.

[0027] Optionally, density clustering algorithms can discover feature clusters of arbitrary shapes without assuming convexity, making them suitable for the complex distribution patterns of land use features in high-dimensional space. The algorithm is insensitive to noise points and can effectively identify and eliminate the interference of outlier feature points on the grouping results. It can be understood that the partitioning results of feature clusters directly reveal the inherent clustering patterns in the multimodal feature set; sample points within each feature cluster have high feature similarity, corresponding to a relatively homogeneous land cover or use type. See Table 1 for examples of key parameters of the density clustering algorithm and their typical value ranges.

[0028] Table 1: Density Clustering Parameters and Cluster Selection Criteria Neighborhood radius Sample point neighborhood range 0.3–0.7 (determined by the inflection point of the k-distance curve) Influences the density of clusters Minimum points Number of neighboring points required for the core point 5–10 (adjusted based on sample size) Controlling noise point tolerance Contour coefficient threshold Intra-cluster aggregation index >0.5 (set based on historical data) Screening for high cohesive feature clusters In some embodiments, time dimension analysis is introduced to calculate the long-term rate of change and short-term fluctuation amplitude of eigenvalues. Time series modeling is performed on the multimodal feature set corresponding to multi-temporal image data, with the time series modeling targeting the multi-temporal eigenvalue sequence for each spatial unit. A trend curve for the eigenvalues ​​is fitted using a polynomial function or moving average method to capture the overall trend of eigenvalue changes over time. The average slope of the trend curve within a preset time window is calculated as the long-term rate of change, which characterizes the linear trend of eigenvalue changes over a longer time scale. The residual standard deviation between the eigenvalue and its trend curve is calculated as the short-term fluctuation amplitude, which measures the degree of dispersion of the eigenvalue around the trend curve.

[0029] Optionally, the neighborhood radius and minimum number of points parameters in the density clustering algorithm are determined through multiple experiments to obtain stable clustering results. The silhouette coefficient is calculated for each sample point in a feature cluster; the overall silhouette coefficient of a feature cluster is the average of the silhouette coefficients of all sample points within that cluster. The preset threshold is set based on the average silhouette coefficient of historical clustering results. It can be understood that the preset time window length used in time series modeling covers at least one complete crop growth cycle or annual variation cycle. The long-term rate of change is the arithmetic mean of the instantaneous slopes of the trend curve at each time point within the preset time window, and the short-term fluctuation amplitude is the standard deviation of the residuals between the feature value sequence and the trend curve sequence. The residual sequence reflects the degree to which the feature values ​​deviate from the long-term trend due to short-term factors.

[0030] In practical implementation, the identification of potential land use patterns relies on the geometric and statistical properties of feature clusters. The centroid of a feature cluster represents the typical feature vector of that land use pattern, and the feature distribution range is defined by the maximum and minimum values ​​of sample points in each feature dimension. The calculation of long-term change rate and short-term fluctuation amplitude generates independent time-series change indicators for each feature dimension of each spatial unit. The formula for calculating the long-term change rate is: in, Represents the long-term rate of change. This represents the total number of time points within the preset time window. The trend curve represents the time point. The instantaneous slope. The formula for calculating short-term fluctuation amplitude is: in, Represents the short-term fluctuation range. Represents the length of the time series. Representing a point in time The actual eigenvalues, Representing a point in time The trend curve fitting value.

[0031] See Figure 4 This chart revolves around the core logic of introducing a time dimension to analyze the dynamic characteristics of land use. Using months as the horizontal axis of the time series, it simultaneously presents two key indicators: long-term change rate and short-term fluctuation amplitude. The long-term change rate reflects the trend evolution of land use characteristics, embodying the continuous development patterns of types such as arable land and forest land over time. Short-term fluctuation amplitude, on the other hand, characterizes seasonal or sudden fluctuations in characteristics, corresponding to the drastic impact of short-term factors such as farming activities and phenological cycles on land use. By comparing the two axes, the difference between trend changes and fluctuations can be intuitively distinguished: a smooth increase in the long-term change rate corresponds to a gradual process of land use change, while sharp fluctuations in the short-term fluctuation amplitude correspond to periodic or temporary changes. This analysis provides quantitative support for subsequent classification of stable and variable land use regions. Periods with low long-term change rates and small short-term fluctuation amplitudes can be classified as stable regions, and vice versa, ultimately helping to accurately identify dynamic patterns of land use and providing highly discriminative time-series feature inputs for ensemble learning models.

[0032] Example 3: In specific implementation, land use is divided into stable and variable regions based on long-term change rate and short-term fluctuation amplitude. A long-term change rate threshold and a short-term fluctuation amplitude threshold are set. The long-term change rate threshold is used to determine the significance of long-term trend changes in characteristic values, while the short-term fluctuation amplitude threshold is used to measure the severity of short-term fluctuations in characteristic values. Regions with long-term change rates below the long-term change rate threshold and short-term fluctuation amplitudes below the short-term fluctuation amplitude threshold are marked as stable regions. In these regions, land cover and use remain relatively constant over time, and the time-series curves of characteristic values ​​exhibit smooth characteristics. Regions with long-term change rates above the long-term change rate threshold or short-term fluctuation amplitudes above the short-term fluctuation amplitude threshold are marked as variable regions. In these regions, land cover and use change significantly over time, and their characteristic values ​​may exhibit continuous trend changes or severe periodic or non-periodic fluctuations.

[0033] In some embodiments, the long-term rate of change threshold and the short-term fluctuation amplitude threshold are set using the percentile method. For example, the 75th percentile of the long-term rate of change value is used as the long-term rate of change threshold, and the 75th percentile of the short-term fluctuation amplitude value is used as the short-term fluctuation amplitude threshold. This method can adapt to the data characteristics of different regions. After the thresholds are set, the relationship between the long-term rate of change and the short-term fluctuation amplitude and the thresholds is compared for each spatial unit to complete the region division.

[0034] Optionally, representative features are extracted for both stable and variable regions. For stable regions, the mean and variance of the multimodal feature set are calculated as representative features, with the mean representing the average level of the features and the variance representing the range of fluctuation. For variable regions, the time-series extreme values ​​and gradients of the multimodal feature set are extracted as representative features. The time-series extreme values ​​include the minimum and maximum values ​​of the feature values ​​in the time series, and the gradient reflects the rate of change of the feature values ​​over time.

[0035] It is understandable that the mean and variance calculations for the stable region are based on the multimodal eigenvalue sequences at all time points within that region, with the calculation process performed independently for each eigenvalue dimension. Extremum extraction for the variability region involves identifying minimum and maximum values ​​by scanning time-series data, and gradient calculation involves analyzing the differences in eigenvalues ​​at consecutive time points.

[0036] The formula for calculating the gradient is: in: Represents the gradient of change. Representing a point in time eigenvalues, Representing a point in time eigenvalues, This represents the time interval between adjacent time points. The extraction results of representative features generate a set of feature vectors for each region, which are then used as input to the subsequent model.

[0037] Example 4: In specific implementation, a representative feature-based ensemble learning model is trained, constructing a hybrid model of random forest and gradient boosting decision tree. Random forest is an ensemble learning algorithm used to handle features in stable regions. By constructing multiple decision trees and aggregating their predictions, the model's accuracy and robustness are improved. Each decision tree is trained using randomly selected feature and sample subsets, and the final prediction is output through voting or averaging. Gradient boosting decision tree is an ensemble learning algorithm that sequentially constructs multiple decision trees, each new decision tree focusing on correcting the prediction errors of the preceding trees, and finally outputting the prediction through weighted summation. The hybrid model uses the random forest component to handle features in stable regions and the gradient boosting decision tree component to handle features in variable regions. Features from stable regions are input to the random forest model, and features from variable regions are input to the gradient boosting decision tree model. The training process uses representative features extracted from stable and variable regions as input data. Representative features include the mean and variance of stable regions, and the extreme values ​​and gradients of variable regions.

[0038] Gradient boosting decision trees are used to handle features in variable regions. Features from stable regions are input into a random forest model, while features from variable regions are input into a gradient boosting decision tree model. Ten-fold cross-validation is used to evaluate the performance of the ensemble model and select the optimal model parameters. Ten-fold cross-validation randomly divides the training dataset into ten mutually exclusive subsets. Nine subsets are used sequentially to train the ensemble model, and the remaining subset is used to validate the model performance. This process is repeated ten times, and the average performance index is calculated. A grid search method is used to optimize the model hyperparameters. The number of trees and the maximum depth search range for the random forest, and the learning rate and iteration count search range for the gradient boosting decision tree are defined. The search range is set based on prior knowledge or pre-experiments. Parameter combinations are traversed within the search range, and the cross-validation accuracy for each parameter group is calculated. The traversal process enumerates all possible parameter combinations. The parameter combination with the highest cross-validation accuracy is selected as the optimal hyperparameters, which are used to configure the final ensemble learning model.

[0039] In some embodiments, the search range for the number of trees in a random forest is set to an integer sequence, the search range for the maximum depth is set to an integer sequence, the search range for the learning rate in a gradient boosting decision tree is set to a decimal sequence, and the search range for the number of iterations is set to an integer sequence. When evaluating the performance of the hybrid model using ten-fold cross-validation, classification accuracy is used as the performance metric; classification accuracy is the ratio of correctly classified samples to the total number of samples. During the optimization of model hyperparameters using grid search, the cross-validation accuracy is calculated based on the average accuracy of ten-fold cross-validation. Optionally, the outputs of the random forest model and the gradient boosting decision tree model are integrated; the integration method can be weighted voting or average probability. The classification results of stable and variable regions are ultimately merged into a unified land use classification map. The subset partitioning for ten-fold cross-validation employs a stratified sampling strategy to ensure that the distribution of land use categories in each subset is consistent with the overall dataset.

[0040] It is understandable that the parameter combination traversal of the grid search method involves a large amount of computation, so parallel computing techniques can be used to accelerate the search process. The formula for calculating the cross-validation accuracy is: in: Represents the accuracy of cross-validation. This represents the total number of parameter combinations in the grid search. Representing the The average classification accuracy of the group parameters on 10-fold cross-validation. See Table 2 for the specific values ​​of the hyperparameter search range.

[0041] Table 2: Hyperparameter Search Range Table Random Forest Number of trees [50,100,150] Random Forest Maximum depth [5,10,15] Gradient boosting decision tree Learning rate [0.05,0.1,0.2] Gradient boosting decision tree Number of iterations [100,200,300] See Figure 5This graph, using the maximum depth of the random forest as the x-axis and classification accuracy as the y-axis, focuses on the impact of hyperparameter combinations on model performance, visually presenting the model's classification performance under different combinations of tree numbers and maximum depths. The trend shows that the classification accuracy reaches its optimal level when a specific number of trees is combined with a medium maximum depth; the accuracy corresponding to a lower number of trees shows a slight increase with increasing maximum depth; another higher number of trees achieves suboptimal performance at a specific maximum depth, slightly lower than the accuracy of the optimal combination. As a key visualization of the results of optimizing model hyperparameters using the grid search method, this graph effectively verifies the rationality of the preset hyperparameter search range, provides an intuitive reference for subsequently selecting the parameter combinations with the highest cross-validation accuracy, and ultimately helps the ensemble learning model process feature data in stable regions more efficiently, ensuring the accuracy of land use identification.

[0042] Example 5: In specific implementation, the feature data of the area to be identified is input into the optimized ensemble learning model to obtain the multimodal feature set of the area in real time. The multimodal feature set includes spectral reflectance features and spatial structure features extracted from remote sensing images, as well as auxiliary topographic elevation data and soil physicochemical property data. The same preprocessing and feature extraction as the training data are performed. Preprocessing operations include radiometric correction, geometric fine correction, and eigenvalue standardization. The feature extraction process involves spectral reflectance feature calculation, spatial structure feature calculation, and multi-source data fusion. The preprocessed feature data is input into the optimized ensemble learning model, which is a hybrid model of random forest and gradient boosting decision tree whose optimal hyperparameters have been determined through grid search. The model outputs a land use probability distribution, which is a vector where each element represents the probability that the area to be identified belongs to a predefined land use category. The final land use category is determined based on the maximum value of the probability distribution; the final land use category is the category label corresponding to the highest probability value in the probability distribution vector.

[0043] In some embodiments, real-time acquisition of the multimodal feature set of the area to be identified relies on continuous observation data from remote sensing satellites and UAV platforms. Remote sensing satellite platforms provide large-scale, periodic Earth observation data; their orbital parameters and sensor configurations ensure revisit capability for specific areas, enabling the acquisition of multispectral or hyperspectral imagery at preset time intervals. UAV platforms, as a flexible supplementary observation method, conduct low-altitude flights in specific areas according to mission requirements, acquiring high spatial resolution remote sensing data, particularly suitable for responding to sudden monitoring needs or compensating for data loss caused by cloud cover in satellite imagery. Continuous observation data is transmitted in real-time or near real-time via ground receiving stations or data links to the data processing center. The data processing center performs a series of preprocessing operations on the received raw remote sensing observation data. These preprocessing operations include radiometric correction to eliminate sensor errors and atmospheric scattering and absorption effects, geometric fine correction to correct geometric distortions caused by sensor attitude and terrain undulations, and precise registration of the imagery to a unified geodetic coordinate system. The preprocessed image data then enters the feature extraction stage. Broad-spectral reflectance features and macroscopic spatial structure features are extracted from satellite imagery, while more detailed texture features and finer spatial structure features are extracted from UAV imagery. Simultaneously, topographic elevation data and soil physicochemical property data acquired synchronously or quasi-synchronously with remote sensing observations are introduced as auxiliary data sources. Topographic elevation data typically originates from digital elevation models, while soil physicochemical property data can be obtained through field sampling or soil survey databases.

[0044] In some embodiments, the remote sensing satellite platform follows a preset imaging plan, sending downlink observation data to the data processing center. Upon data reception, an automated preprocessing process is triggered. The observation tasks of the UAV platform can be dynamically triggered by the data processing center based on the quality of the satellite data or specific identification needs. Data acquired by the UAV is transmitted to the data processing center via a wireless network. Optionally, the generation of the multimodal feature set relies on a data fusion process, integrating features from different platforms and time phases within a unified geographic grid framework. Spectral reflectance features include reflectance values ​​for specific bands and vegetation indices calculated from them; spatial structure features include texture metrics and shape metrics. Topographic elevation data is converted into derived features such as slope and aspect; soil physicochemical properties include attributes such as pH value and organic matter content. Real-time acquisition means that the time delay from data collection to the generation of a multimodal feature set usable for model inference is controlled within the task requirements, forming a continuously updated data stream. The content and structure of the multimodal feature set are completely consistent with the feature set used in the model training phase, ensuring that the ensemble learning model can perform correct inference operations. Finally, this real-time generated multimodal feature set is immediately fed into the optimized ensemble learning model to output the land use classification result for the current moment. See Table 3 for examples of the specific composition of the multimodal feature set and its data sources.

[0045] Table 3: Example of Multimodal Feature Set Construction Spectral characteristics Reflectivity of blue, green, red, and near-infrared bands Multispectral satellite imagery Radiation correction, band calculation Spatial features Texture entropy, contrast, shape index High-resolution drone imagery Image segmentation and filtering Topographic features Elevation, slope, aspect DEM data Resampling to a uniform grid Soil characteristics pH value, organic matter content, texture Soil Survey Database Normalize to the [0,1] interval Optionally, the land use probability distribution vector is generated by aggregating the prediction results of all base learners in the ensemble learning model. The aggregation method employs a soft voting mechanism, i.e., a weighted average of the class probabilities output by each base learner. The final land use category determination criterion is expressed in mathematical form: ,in The function returns the index of the category with the highest probability value. It's understandable that preprocessed feature data needs format conversion before being input into the model, such as flattening multidimensional features into one-dimensional feature vectors to meet the model's input interface requirements. The sum of all elements in the land use probability distribution vector is 1, and the magnitude of the probability value directly reflects the model's confidence in the classification result. Once the final land use category is determined, it can be visualized in conjunction with a geographic information system.

[0046] In practice, determining the maximum value of the probability distribution involves vector comparison operations, and the calculation formula is as follows: in: Represents the final land use category, Represents the total number of predefined land use categories. The region to be identified in the model output belongs to the first... The probability values ​​of each category, The function returns the probability value. Reaching the maximum category index The entire classification process achieves automated mapping from multi-source data to specific land use categories.

[0047] See Figure 6 This study presents the temporal characteristics of long-term change rates and short-term fluctuation amplitudes for different land use types (arable land, forest land, grassland, construction land, and water bodies) at 24 time points (monthly intervals). The figure employs a dual-axis design: the left axis represents the long-term change rate, and the right axis represents the short-term fluctuation amplitude. Long-term change rate thresholds (red horizontal lines) and short-term fluctuation amplitude thresholds (black horizontal lines) are also marked to delineate areas of stability / variability in land use. Specifically, the long-term change rate of construction land (yellow broken line) is significantly higher than other types and consistently remains above the long-term change rate threshold, reflecting a clear expansion trend during the observation period. The long-term change rates of arable land, forest land, and grassland are at lower levels, mostly below the threshold. In the short-term fluctuation amplitude dimension, arable land (red dashed line) shows the most significant fluctuations, repeatedly approaching or exceeding the short-term fluctuation amplitude threshold, while the fluctuation amplitudes of forest land, grassland, and water bodies are relatively gentle, generally remaining below the threshold. Combining the dual-threshold standard, it can be seen that: construction land belongs to a typical variable area, with both its long-term trend and short-term fluctuations showing strong dynamism; although the long-term change rate of cultivated land is stable, its short-term fluctuations are significant, requiring close attention to its periodic anomalies; forest land, grassland, and water bodies are generally in a stable area, with relatively stable spatiotemporal characteristics. By comparing the temporal characteristics of multiple types of elements, the differences in the dynamic evolution of different land use types are intuitively presented, providing a quantitative basis for the division of land use stability / variability areas.

[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent land use identification, characterized in that, The method is implemented through the following steps: It receives multi-temporal image data from remote sensing satellites and UAVs, and performs radiometric and geometric correction on the image data; Spectral reflectance features and spatial structure features are extracted from the corrected image data, while topographic elevation data and soil physicochemical property data are obtained. A feature selection algorithm was used to evaluate the importance scores of spectral reflectance features and spatial structure features, and the feature weights were adjusted by combining topographic elevation data and soil physicochemical property data. Feature fusion is performed based on the adjusted feature weights to generate a multimodal feature set, and cluster analysis is performed on the multimodal feature set to identify potential land use patterns. By introducing time dimension analysis, the multimodal feature sets corresponding to multi-temporal image data are arranged in chronological order, and the long-term rate of change and short-term fluctuation amplitude of the feature values ​​are calculated. Based on the long-term rate of change and the short-term fluctuation range, land use stability areas and change areas are divided, and representative features are extracted for each. An ensemble learning model is trained using representative features, and its hyperparameters are optimized using a grid search method. The feature data of the area to be identified is input into the optimized ensemble learning model, and the land use classification results are output.

2. The method for intelligent land use identification according to claim 1, characterized in that, The importance scores for evaluating spectral reflectance features and spatial structure features using the feature selection algorithm include: Calculate the mutual information value between each spectral reflectance feature and the land use category to obtain a preliminary score of the importance of the spectral features; Calculate the chi-square statistic between each spatial structure feature and land use category to obtain a preliminary score of spatial feature importance; The initial scores for the importance of spectral features and spatial features are combined, and the final score for feature importance is determined using the entropy weighting method.

3. The method for intelligent land use identification according to claim 1, characterized in that, The adjustment of feature weights by combining topographic elevation data and soil physicochemical property data includes: The topographic elevation data was divided into multiple elevation zones, and the distribution frequency of land use types within each elevation zone was statistically analyzed. Normalize the soil physicochemical property data to the same dimension and calculate the correlation coefficient between soil parameters and land use type; Topographic adjustment factors and soil adjustment factors are generated based on distribution frequency and correlation coefficient; The adjusted feature weights are obtained by multiplying the topography adjustment factor and the soil adjustment factor by the feature importance score.

4. The intelligent land use identification method according to claim 1, characterized in that, The clustering analysis of the multimodal feature set to identify potential land use patterns includes: Density clustering algorithm is used to group the multimodal feature set to obtain multiple feature clusters; Calculate the centroid and contour coefficient of each feature cluster, and select feature clusters with contour coefficients greater than the threshold as potential land use patterns; Each potential land use pattern is assigned a temporary label, and the distribution range of its characteristics is recorded.

5. The intelligent land use identification method according to claim 1, characterized in that, The calculation of the long-term rate of change and short-term fluctuation range of the characteristic values ​​includes: Time series modeling is performed on the multimodal feature set corresponding to multi-temporal image data, and the trend curve of the feature values ​​is fitted. Calculate the average slope of the trend curve within a preset time window as the long-term rate of change; The residual standard deviation of the eigenvalues ​​and their trend curves is calculated as the short-term volatility.

6. The method for intelligent land use identification according to claim 1, characterized in that, The delineation of land use stability zones and change zones includes: Set long-term rate of change thresholds and short-term fluctuation range thresholds; Regions with long-term rate of change below the threshold and short-term fluctuation amplitude below the threshold are marked as stability regions. Regions with long-term rates of change exceeding a threshold or short-term fluctuations exceeding a threshold are marked as volatile regions.

7. The method for intelligent land use identification according to claim 1, characterized in that, The extracted representative features include: For the stable region, the mean and variance of the multimodal feature set are calculated as representative features; For regions with variability, the time series extrema and gradient of the multimodal feature set are extracted as representative features.

8. The intelligent land use identification method according to claim 1, characterized in that, The method of training the ensemble learning model using representative features includes: A hybrid model of random forest and gradient boosting decision tree is constructed, in which random forest is used to handle features of stable regions and gradient boosting decision tree is used to handle features of variable regions. The performance of the hybrid model was evaluated using ten-fold cross-validation, and the optimal model parameters were selected. The optimization of model hyperparameters using the grid search method includes: Define the number of trees and maximum depth search range for random forests, and the learning rate and number of iterations search range for gradient boosting decision trees; Within the search range, perform parameter combination traversal and calculate the cross-validation accuracy for each parameter group; Choose the parameter combination that yields the highest cross-validation accuracy as the optimal hyperparameter.

9. The method for intelligent land use identification according to claim 1, characterized in that, The step of inputting the feature data of the region to be identified into the optimized ensemble learning model includes: The multimodal feature set of the region to be identified is acquired in real time, and the same preprocessing and feature extraction are performed as on the training data. The preprocessed feature data is input into the optimized ensemble learning model, and the model outputs the land use probability distribution. The final land use category is determined based on the maximum value of the probability distribution.

10. A smart land use identification system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the intelligent land use identification method according to any one of claims 1 to 9.

Images