Remote sensing monitoring system and method for non-agriculturalization of cultivated land based on multi-source feature fusion and XGBoost

Abstract

The application discloses a cultivated land non-agricultural remote sensing monitoring system and method based on multi-source feature fusion and XGBoost, and the system comprises a feature extraction module, a sample balancing module, a model training and prediction module and a post-processing module. Through the whole-process collaborative optimization of multi-source feature fusion, sample balancing, model improvement and spatial post-processing, under different terrains, on the one hand, the problem that the sample is extremely unbalanced, the missing detection rate is high and the small illegal graph spot is difficult to identify is solved; and on the other hand, the complex feature of ground objects, the interference of seasons, phenology and linear ground objects and the false alarm rate are reduced; based on the single feature extraction method, the ground object feature is constructed, the hyperparameter is adjusted by the artificial trial and error method, the collaborative optimization monitoring framework is formed, and the post-processing is not combined with the actual business logic of cultivated land protection, so that the monitoring result still cannot meet the actual application requirement.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing monitoring technology, specifically relating to a remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, and also relating to a remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and XGBoost. Background Technology

[0002] The rapid development of remote sensing technology has promoted the large-scale application of high-resolution remote sensing imagery in fields such as land resource management, urban construction, and emergency disaster reduction. As a core task of remote sensing data processing, land cover classification in remote sensing imagery aims to accurately assign pixels or regions in the image to the corresponding land cover categories, providing accurate spatial data support for subsequent decision analysis. Its classification accuracy and processing efficiency directly determine the application value of remote sensing data.

[0003] However, in the precise monitoring of farmland conversion to non-agricultural uses (farmland being converted into factories, storage yards, paved roads, illegal buildings, etc.), with the development of remote sensing technology, machine learning-based remote sensing image interpretation methods have become the mainstream approach for monitoring farmland conversion to non-agricultural uses. Specifically, the XGBoost model is used as an efficient ensemble learning algorithm for monitoring farmland conversion to non-agricultural uses. However, in actual large-scale farmland monitoring, the following technical challenges exist:

[0004] 1) The sample is extremely imbalanced. That is, in large-area remote sensing images, the proportion of areas where non-agriculturalization has occurred (positive samples) is usually less than 1%, while the ratio of positive to negative samples can reach 100:1 or even higher. The standard XGBoost model tends to favor the extremely high proportion of negative samples (unchanged cultivated land areas), resulting in a false negative rate of over 30% for non-agriculturalized areas. In particular, small and sporadic illegal non-agriculturalized patches are almost impossible to identify.

[0005] 2) The features of the land cover are complex and subject to a lot of interference. That is, the spectral characteristics of cultivated land change significantly with the seasons and phenology, and are easily confused with fallow or crop-harvested areas, resulting in the same spectrum of different features. The false alarm rate of traditional single-temporal feature monitoring models is over 25%. At the same time, linear features such as greenhouses and rural roads around cultivated land are easily confused with non-agricultural areas, further reducing the accuracy of monitoring.

[0006] 3) Hyperparameters are difficult to tune and the model has poor general applicability. That is, the monitoring effect of the XGBoost model depends on the selection of hyperparameters such as max_depth, learning_rate, and min_child_weight. With massive remote sensing data, manual parameter tuning is difficult to achieve the global optimum, and the monitoring accuracy of parameters tuned for a certain area fluctuates by 15%-20% in other terrain (plain / mountain / hill) areas.

[0007] 4) The monitoring efficiency is not matched with the hardware cost. That is, although existing deep learning models (such as U-Net and Transformer) can improve the monitoring accuracy, the training and monitoring are extremely time-consuming. It takes more than 24 hours to monitor 1,000 km² of cultivated land in a single batch, and it requires a high-performance GPU cluster, which results in high hardware costs and cannot meet the routine and rapid survey needs of grassroots land departments.

[0008] Based on the above difficulties, existing monitoring methods for the non-agriculturalization of arable land use include using a single sampling method to address sample imbalance, a single feature extraction method to construct land cover features, and a manual trial-and-error method to adjust hyperparameters. Although these methods can improve the monitoring effect to some extent, they are independent of each other and do not form a collaborative optimization monitoring framework. Furthermore, they do not incorporate the actual business logic of arable land protection for post-processing, resulting in monitoring results that still cannot meet the needs of practical applications. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide an improved remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost.

[0010] This invention also relates to a remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and XGBoost.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0012] A remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost includes a feature extraction module, a sample balancing module, a model training and prediction module, and a post-processing module. The feature extraction module acquires farmland remote sensing images from two different time phases, automatically extracts spectral features, exponential features, texture features, and temporal difference features, fuses them to construct and output a multi-dimensional feature vector. The sample balancing module incorporates the SMOTE-Tomek hybrid sampling algorithm, receives the sample data output by the feature extraction module, automatically performs 5-fold oversampling of minority class samples and 2-5% overlap sample removal, and outputs balanced training samples. The model training and prediction module incorporates an improved XGBoost model, equipped with a FocalLoss cost-sensitive loss function and an adaptive Bayesian hyperparameter optimization algorithm to complete automatic model training. It also includes a model validation unit, which uses five-fold cross-validation for accuracy verification. When the F1-Score is below 0.85, it automatically triggers hyperparameter re-optimization. After training, the module monitors the test samples and outputs a non-agricultural monitoring probability map. The post-processing module performs mode filtering and connected component analysis on the monitoring probability map, automatically removing small patches with an area of ​​less than 200m², and outputs the final farmland non-agricultural monitoring result map.

[0013] Preferably, spectral features extract the spectral reflectance of multiple bands (red, green, blue, and near-infrared) from remote sensing images to provide basic information for land cover identification; and / or, index features calculate improved vegetation index, building index, and bare soil index to characterize the vegetation cover of cultivated land, artificial buildings, and hardened bare soil, respectively, to achieve preliminary screening of non-agricultural areas; and / or, texture features extract the mean, variance, information entropy, and contrast based on the gray-level co-occurrence matrix, and use the difference between the natural irregular texture of cultivated land and the regular texture of artificial non-agricultural land cover to distinguish easily confused land cover; and / or, temporal difference features calculate the difference in normalized vegetation index between two images, and △NDVI=NDVI_t2-NDVI_t1, as a core indicator to characterize drastic changes in cultivated land vegetation cover to eliminate the interference of seasonal phenological changes.

[0014] According to a specific embodiment and preferred aspect of the present invention, the SMOTE-Tomek hybrid sampling algorithm includes the SMOTE algorithm for synthesizing minority class samples through linear interpolation in the feature space, and the TomekLinks algorithm for removing overlapping sample pairs located on inter-class boundaries. The specific synthesis formula of the SMOTE algorithm is as follows: Let a certain non-agricultural minority class sample be x. i =[x i1 x i2 , ..., x id ] T Its k-nearest neighbor sample in the feature space is x j =[x j1 x j2 , ..., x jd ] T Then the synthesized new sample x new Generate using the following formula: x new =x i +λ(x j -x i ), where λ∈(0,1) are random interpolation coefficients, randomly generated from a uniform distribution; d is the feature dimension; x j -x i Let x be the vector difference between the samples in the feature space; TomekLinks algorithm formula and definition: Let x be any two different class samples in the sample space. i Cultivated land type, x j For non-agricultural categories, the Euclidean distance between the two is... Where d is the feature dimension, if x i With x j If the following TomekLinks conditions are met, x i It is x j nearest neighbor; x j It is x i nearest neighbor; x i With x jIf they belong to different categories, then they are called (x i x j) A set of TomekLinks sample pairs, which are located at the classification boundary and overlap, are considered noise samples.

[0015] In some specific implementations, the SMOTE algorithm is first used to oversample the non-agricultural minority class samples by 5 times in the feature space to supplement the number of minority class samples and solve the problem of the model being biased towards negative samples. The oversampling factor of 5 is the optimal value verified by multiple experiments. If the factor is lower than 3, the sample supplementation will be insufficient, and if the factor is higher than 7, overfitting is likely to occur. Secondly, the TomekLinks algorithm is used to remove overlapping sample pairs located on the inter-class boundaries. The overlapping samples removed account for about 3 to 4% of the total number of samples, cleaning up the inter-class overlapping noise and making the classification decision boundary clearer.

[0016] According to a specific embodiment and preferred aspect of the present invention, the logarithmic loss function of standard XGBoost is replaced with the FocalLoss cost-sensitive loss function, and the objective function is: The adjustment factor γ = 2.0, and the positive / negative sample weight ratio α = 0.75. To predict probabilities for the model, This is the regularization term; the loss function is obtained through... It automatically reduces the weight of easily classified samples, allowing the model to focus on edge samples that are difficult to distinguish.

[0017] Preferably, a cost function with F1-Score as the objective is established. F1-Score is more suitable for precision evaluation of imbalanced samples. Its calculation formula is: F1-Score = 2 × precision × recall ÷ (precision + recall), where precision = number of true positive samples ÷ (number of true positive samples + number of false positive samples) and recall = number of true positive samples ÷ (number of true positive samples + number of false negative samples).

[0018] Furthermore, a surrogate model for hyperparameters and model performance is constructed using Gaussian processes. With maximizing the F1-Score as the optimization objective, the max_depth, min_child_weight, and eta hyperparameters of the XGBoost model are automatically optimized. The specific process is as follows: Hyperparameter search space definition: Set the range of values ​​for the hyperparameters to be optimized, tree depth max_depth ∈ [3, 12], minimum subtree weight min_child_weight ∈ [1, 10], and learning rate eta ∈ [0.01, 0.3]. Gaussian process surrogate model construction: Using the sampled hyperparameter combinations and their corresponding F1-Scores as training data, the mapping relationship between hyperparameters and F1-Score is fitted using Gaussian processes to quickly predict the model performance of unsampled points.

[0019] In some specific implementations, the expected boost function is used as the sampling criterion for the selection of the acquisition function, as follows: EI(x)=E[max(F1(x)−F1_best,0)], where F1(x) is the predicted F1-Score corresponding to the hyperparameter combination x; F1_best is the currently found optimal F1-Score; and / or, an iterative optimization process is used, randomly sampling 10 sets of hyperparameter combinations, training the model and calculating the F1-Score; in each round, the surrogate model is updated using a Gaussian process, and the next set of hyperparameters is selected through the EI function; the optimization stops after 50 iterations. The max_depth, min_child_weight, and eta hyperparameters of the XGBoost model are automatically optimized using a Gaussian process, with 50 iterations, and the optimal hyperparameter combination is finally determined to be: learning_rate=0.05, max_depth=8, min_child_weight=1; this avoids the subjectivity and limitations of manual parameter tuning and improves the model's versatility. This method replaces manual parameter tuning and automatically searches for the optimal combination through a probabilistic proxy model, reducing subjectivity and regional limitations, and significantly improving the model's generalization ability and versatility under different terrains and different types of cultivated land.

[0020] According to another specific embodiment and preferred aspect of the present invention, a 3×3 mode filter is used to smooth the monitoring probability map and eliminate isolated noise points. The calculation method is as follows: taking any pixel (i,j) in the probability map as the center, the monitoring probability values ​​of 9 pixels in its 3×3 neighborhood are taken to form a set: S={P(i−1,j−1),P(i−1,j),P(i−1,j+1),P(i,j−1),P(i,j),P(i,j+1),P(i+1,j−1),P(i+1,j),P(i+1,j+1)}. The neighborhood probability values ​​are statistically analyzed, and the probability value that appears most frequently (the mode) is taken as the output value of the center pixel after filtering: P′(i,j)=Mode(S) where: P(i,j) is the monitoring probability of the center pixel before filtering; P′(i,j) is the monitoring probability of the center pixel after filtering. The smoothing probability of elements; Mode(·) is the operation of taking the mode of the set; 3×3 neighborhood can effectively suppress isolated noise points, making the monitoring results more continuous and smooth; and / or, connected component analysis, to calculate the area of ​​the monitored non-agricultural patches, specifically by using the 4-connected or 8-connected component labeling rule to group spatially adjacent pixels with probability values ​​higher than the threshold into the same independent patch; using the actual ground area corresponding to a single pixel as the basic unit, the area of ​​the non-agricultural patch is calculated according to the following formula: Patch area S=N×S0, where S is the actual ground area of ​​the current connected component; N is the total number of pixels contained in the current connected component; S0 is the actual ground area of ​​a single pixel. After calculating the area of ​​all connected components according to the above method, small patches with an area of ​​less than 200m² are removed, and valid non-agricultural patches that meet the farmland supervision standards are retained to obtain the final monitoring results.

[0021] Another technical solution of the present invention is: a remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, which adopts the above-mentioned remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, and includes the following steps:

[0022] S1. Construction of Multi-Source Heterogeneous Features

[0023] Two remote sensing images of farmland at different time phases were acquired, and spectral features, exponential features, texture features, and temporal difference features were extracted to construct a 3D feature vector. The exponential features include the improved vegetation index, building index, and bare soil index. The texture features are the mean, variance, and information entropy extracted based on the gray-level co-occurrence matrix. The temporal difference feature is the NDVI difference ΔNDVI between the two images.

[0024] S2, Sample Balancing Process

[0025] The SMOTE-Tomek hybrid sampling algorithm was used to process the cultivated land samples and non-agricultural samples. First, the non-agricultural minority samples were oversampled by 5 times using the SMOTE algorithm, and then the TomekLinks algorithm was used to remove the inter-class overlapping sample pairs that accounted for 2-5% of the total sample volume, so as to obtain the balanced training samples.

[0026] S3, Improved XGBoost Model Training and Prediction

[0027] The balanced training samples obtained in step S2 are input into the improved XGBoost model. The model replaces the original log loss function with the FocalLoss cost-sensitive loss function, and automatically optimizes the max_depth, min_child_weight, and eta hyperparameters of the XGBoost model with F1-Score as the objective function through an adaptive Bayesian optimization algorithm. After the model training is completed, the test samples are monitored and a non-agricultural monitoring probability map is output.

[0028] S4. Spatial Consistency Post-processing

[0029] The monitoring probability map output in step S3 is subjected to 3×3 mode filtering and connected component analysis in sequence. Monitoring patches with an area of ​​less than 200m² are removed to obtain the final monitoring results of farmland non-agriculturalization.

[0030] Due to the implementation of the above technical solutions, the present invention has the following advantages compared with the prior art:

[0031] This invention addresses the core pain points of existing farmland non-agriculturalization monitoring technologies through a collaborative optimization process encompassing multi-source feature fusion, sample balancing, model improvement, and spatial post-processing, achieving the following significant technical effects:

[0032] 1. High monitoring accuracy, significantly reducing the rate of missed detections and false alarms.

[0033] The present invention achieves a monitoring accuracy of 92.4%, a recall rate of 90.8%, and an F1-Score of 0.916. Compared with the traditional XGBoost model, the recall rate is improved by about 16%, and it can accurately identify small and scattered illegal non-agricultural patches (such as simple sheds on farmland and narrow hardened roads). At the same time, by fusing multi-temporal difference features and texture features, it effectively eliminates interference from fallow land, crop harvesting, and greenhouse facilities, reducing the false alarm rate by more than 12%.

[0034] 2. High monitoring efficiency and low hardware requirements.

[0035] Compared to the deep learning model U-Net, this invention improves training speed by nearly 17 times while maintaining higher monitoring accuracy, requiring only 210 seconds to monitor a single batch of 1200 km² of cultivated land. It is also compatible with general-purpose servers (Intel i9 CPU + 64GB RAM + NVIDIA RTX 3090 GPU), and even ordinary office computers can complete small- to medium-sized area monitoring, reducing hardware costs by 90%, making it more suitable for large-scale and rapid surveys by grassroots land and resources departments.

[0036] 3. The model has strong versatility and can be adapted to different application scenarios.

[0037] Through adaptive Bayesian hyperparameter optimization, the model can be adapted to farmland monitoring in different terrains such as plains, mountains, and hills without manual debugging, with accuracy fluctuations of less than 5%. At the same time, the parameters of each module of the system are configurable, and the sampling multiple, patch removal threshold, etc. can be adjusted according to the farmland management needs of different regions, making it highly adaptable.

[0038] 4. Monitoring results align with actual business needs.

[0039] Spatial consistency post-processing incorporates industry standards for land and space planning, eliminating noise from small, meaningless patches. The monitoring results can be directly used for the supervision and enforcement of non-agricultural use of arable land, eliminating the need for secondary manual screening and improving work efficiency. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the core technology flow chart for monitoring the non-agricultural use of cultivated land in this embodiment;

[0041] Figure 2 The images show a comparison of the monitoring effects of farmland non-agriculturalization in this embodiment, including the original remote sensing image (a), the monitoring results of the traditional XGBoost model (b), and the monitoring results of this invention (c). Specifically, image a shows key features such as farmland, rural paved roads, greenhouse facilities, and illegal construction sites, clearly demonstrating the actual distribution of features in the experimental area; image b shows areas with missed detections (narrow paved roads, small simple sheds) and false alarms (crop harvesting areas, greenhouse facilities), demonstrating the monitoring deficiencies of the traditional method; image c shows areas without missed detections or false alarms, where farmland non-agriculturalization patches (paved roads, illegal construction sites) are completely and accurately identified, demonstrating the monitoring effect of this invention. Detailed Implementation

[0042] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0043] like Figure 1 and Figure 2 As shown, the remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost involved in this embodiment includes a feature extraction module, a sample balancing module, a model training and prediction module, and a post-processing module.

[0044] The feature extraction module acquires remote sensing images of cultivated land from two different time periods, automatically extracting spectral features, exponential features, texture features, and temporal difference features. These features are then fused to construct and output a multi-dimensional feature vector. Spectral features extract the spectral reflectance of multiple bands (red, green, blue, and near-infrared) from the remote sensing images, providing basic information for land cover identification. Exponential features calculate improved vegetation index, building index, and bare soil index to characterize cultivated land vegetation cover, artificial buildings, and hardened bare soil, respectively, enabling preliminary screening of non-agricultural areas. Texture features extract mean, variance, information entropy, and contrast based on the gray-level co-occurrence matrix, utilizing the difference between the natural irregular texture of cultivated land and the regular texture of artificial non-agricultural features to distinguish easily confused features. Temporal difference features calculate the difference in normalized vegetation index between the two images, where ΔNDVI = NDVI_t2 - NDVI_t1, serving as a core indicator to characterize drastic changes in cultivated land vegetation cover and eliminate interference from seasonal phenological variations.

[0045] The sample balancing module incorporates the SMOTE-Tomek hybrid sampling algorithm. It receives sample data output from the feature extraction module and automatically performs 5-fold oversampling of minority class samples and 2-5% removal of overlapping samples, outputting balanced training samples. First, the SMOTE algorithm is used to perform 5-fold oversampling of non-agricultural minority class samples in the feature space to supplement the number of minority class samples and solve the problem of model bias towards negative samples. Setting the oversampling factor to 5 is the optimal value verified by multiple experiments. A factor lower than 3 results in insufficient sample supplementation, while a factor higher than 7 is prone to overfitting. Second, the TomekLinks algorithm is used to remove overlapping sample pairs located on the inter-class boundaries. The removed overlapping samples account for about 3-4% of the total samples, cleaning up inter-class overlap noise and making the classification decision boundary clearer.

[0046] The SMOTE-Tomek hybrid sampling algorithm includes the SMOTE algorithm, which synthesizes minority class samples through linear interpolation in the feature space, and the TomekLinks algorithm, which removes overlapping sample pairs located on inter-class boundaries. The specific synthesis formula of the SMOTE algorithm is as follows: Let a certain non-agricultural minority class sample be x. i =[x i1 x i2 , ..., x id ] T Its k-nearest neighbor sample in the feature space is x j =[x j1 x j2 , ..., x jd ] T Then the synthesized new sample x new Generate using the following formula: x new =x i +λ(x j -x i ), where λ∈(0,1) are random interpolation coefficients, randomly generated from a uniform distribution; d is the feature dimension; x j -x i Let x be the vector difference between the samples in the feature space; TomekLinks algorithm formula and definition: Let x be any two different class samples in the sample space. i Cultivated land type, x j For non-agricultural categories, the Euclidean distance between the two is... Where d is the feature dimension, if x i With x j If the following TomekLinks conditions are met, x i It is x j nearest neighbor; x j It is x i nearest neighbor; x i With x j If they belong to different categories, then they are called (x i x j) A set of TomekLinks sample pairs, which are located at the classification boundary and overlap, are considered noise samples.

[0047] The model training and prediction module incorporates an improved XGBoost model, equipped with a FocalLoss cost-sensitive loss function and an adaptive Bayesian hyperparameter optimization algorithm to complete automatic model training. It also includes a model validation unit, which uses five-fold cross-validation for accuracy verification. When the F1-Score is below 0.85, it automatically triggers hyperparameter re-optimization. After training, the module monitors the test samples and outputs a non-agricultural monitoring probability map. The post-processing module performs mode filtering and connected component analysis on the monitoring probability map, automatically removing small patches with an area of ​​less than 200m², and outputs the final farmland non-agricultural monitoring result map.

[0048] Replace the standard XGBoost log loss function with the FocalLoss cost-sensitive loss function, and the objective function is: The adjustment factor γ = 2.0, and the positive / negative sample weight ratio α = 0.75. To predict probabilities for the model, This is the regularization term; the loss function is obtained through... The weights of easily classified samples are automatically reduced, allowing the model to focus on marginal samples that are difficult to distinguish. A cost function is established with F1-Score as the objective. F1-Score is more suitable for evaluating the accuracy of imbalanced samples, and its calculation formula is: F1-Score = 2 × Precision × Recall ÷ (Precision + Recall), where Precision = Number of true positive samples ÷ (Number of true positive samples + Number of false positive samples) and Recall = Number of true positive samples ÷ (Number of true positive samples + Number of false negative samples). A surrogate model for hyperparameters and model performance is constructed using Gaussian processes. Maximizing the F1-Score is the optimization objective. The max_depth, min_child_weight, and eta hyperparameters of the XGBoost model are automatically optimized. The specific process is as follows: Hyperparameter search space definition: Set the range of values ​​for the hyperparameters to be optimized. Tree depth max_depth ∈ [3, 12], minimum subtree weight min_child_weight ∈ [1, 10], and learning rate eta ∈ [0.01, 0.3]. Gaussian process surrogate model construction: Using sampled hyperparameter combinations and their corresponding F1-Scores as training data, a Gaussian process is used to fit the mapping relationship between hyperparameters and F1-Score, quickly predicting the model performance for unsampled points. The expected boost function is used as the sampling criterion for selection of the sampling function, as follows: EI(x)=E[max(F1(x)−F1_best,0)], where F1(x) is the predicted F1-Score corresponding to the hyperparameter combination x; F1_best is the currently found optimal F1-Score; and / or, an iterative optimization process is used, randomly sampling 10 sets of hyperparameter combinations, training the model and calculating the F1-Score; in each round, the surrogate model is updated using a Gaussian process, and the next set of hyperparameters is selected through the EI function; the optimization stops after 50 iterations. The max_depth, min_child_weight, and eta hyperparameters of the XGBoost model are automatically optimized using a Gaussian process, with 50 iterations, and the optimal hyperparameter combination is finally determined to be: learning_rate=0.05, max_depth=8, min_child_weight=1; this avoids the subjectivity and limitations of manual parameter tuning and improves the model's versatility. This method replaces manual parameter tuning and automatically searches for the optimal combination through a probabilistic proxy model, reducing subjectivity and regional limitations, and significantly improving the model's generalization ability and versatility under different terrains and different types of cultivated land.

[0049] The post-processing module uses a 3×3 mode filter to smooth the monitoring probability map and eliminate isolated noise points. The calculation method is as follows: taking any pixel (i,j) in the probability map as the center, the monitoring probability values ​​of 9 pixels in its 3×3 neighborhood are taken to form a set: S={P(i−1,j−1),P(i−1,j),P(i−1,j+1),P(i,j−1),P(i,j),P(i,j+1),P(i+1,j−1),P(i+1,j),P(i+1,j+1)}. The neighborhood probability values ​​are statistically analyzed, and the probability value that appears most frequently (the mode) is used as the output value of the center pixel after filtering: P′(i,j)=Mode(S) where: P(i,j) is the monitoring probability of the center pixel before filtering; P′(i,j) is the smoothed probability of the center pixel after filtering. Mode(·) is the operation for taking the mode of the set; a 3×3 neighborhood can effectively suppress isolated noise points, making the monitoring results more continuous and smooth; and / or, connected component analysis is used to calculate the area of ​​the monitored non-agricultural patches. Specifically, the calculation method is to use the 4-connected or 8-connected component labeling rule to group spatially adjacent pixels with a probability value higher than the threshold into the same independent patch; using the actual ground area corresponding to a single pixel as the basic unit, the area of ​​the non-agricultural patch is calculated according to the following formula: Patch area S = N × S0, where S is the actual ground area of ​​the current connected component; N is the total number of pixels contained in the current connected component; and S0 is the actual ground area of ​​a single pixel. After calculating the area of ​​all connected components using the above method, small patches with an area less than 200m² are removed, and valid non-agricultural patches that meet the farmland supervision standards are retained to obtain the final monitoring results.

[0050] In summary, the remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and XGBoost in this embodiment adopts the aforementioned remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, and includes the following steps:

[0051] S1. Construction of Multi-Source Heterogeneous Features

[0052] Two remote sensing images of farmland at different time phases were acquired, and spectral features, exponential features, texture features, and temporal difference features were extracted to construct a 3D feature vector. The exponential features include the improved vegetation index, building index, and bare soil index. The texture features are the mean, variance, and information entropy extracted based on the gray-level co-occurrence matrix. The temporal difference feature is the NDVI difference ΔNDVI between the two images.

[0053] S2. Sample balancing process:

[0054] The SMOTE-Tomek hybrid sampling algorithm was used to process the cultivated land samples and non-agricultural samples. First, the non-agricultural minority samples were oversampled by 5 times using the SMOTE algorithm, and then the TomekLinks algorithm was used to remove the inter-class overlapping sample pairs that accounted for 2-5% of the total sample volume, so as to obtain the balanced training samples.

[0055] S3. Improved XGBoost model training and prediction:

[0056] The balanced training samples obtained in step S2 are input into the improved XGBoost model. The model replaces the original log loss function with the FocalLoss cost-sensitive loss function, and automatically optimizes the max_depth, min_child_weight, and eta hyperparameters of the XGBoost model with F1-Score as the objective function through an adaptive Bayesian optimization algorithm. After the model training is completed, the test samples are monitored and a non-agricultural monitoring probability map is output.

[0057] S4. Spatial Consistency Post-processing

[0058] The monitoring probability map output in step S3 is subjected to 3×3 mode filtering and connected component analysis in sequence. Monitoring patches with an area of ​​less than 200m² are removed to obtain the final monitoring results of farmland non-agriculturalization.

[0059] To make the technical solution, experimental process and technical effects of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments. These embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention.

[0060] Example 1: Remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and improved XGBoost

[0061] 1. Experimental Environment and Data Preparation

[0062] 1.1 Experimental Area: A county-level administrative region in a province of my country was selected as the experimental area, with a total area of ​​approximately 1200 km². The area includes plains and hills, with dry land and paddy fields as the main types of cultivated land. There are also a small number of non-agricultural types such as rural paved roads and simple factory buildings, which have typical characteristics of cultivated land non-agricultural monitoring scenarios.

[0063] 1.2 Data source: Two sets of domestic Gaofen-2 (GF-2) remote sensing images from 2022 and 2023 were used, with a spatial resolution of 0.8 meters. The images were unobstructed by clouds and covered the entire cultivated land area of ​​the experimental area.

[0064] 1.3 Sample Construction: Three professionals with over five years of remote sensing interpretation experience constructed a sample set through a combination of manual interpretation and field surveys. A total of 50,000 negative samples of cultivated land and 1,200 positive samples of non-agricultural land were obtained, with a positive-to-negative sample ratio of approximately 41.7:1. The consistency of sample labeling reached 98.5%, and the labeling results were corrected through field verification, with a verification accuracy rate of 99.2%. The sample set was divided into training samples and test samples at a ratio of 7:3.

[0065] 1.4 Hardware environment: Intel i9-13900K CPU, 64GB DDR5 RAM, NVIDIA RTX3090 24GB GPU;

[0066] 1.5 Software environment: Python 3.8, XGBoost 1.7.5, Scikit-learn 1.2.2, OpenCV 4.7.0, GDAL 3.6.2, Matplotlib 3.7.1.

[0067] 2. Algorithm Implementation Steps

[0068] 2.1. Construction of Multi-Source Heterogeneous Features

[0069] Based on GF-2 remote sensing images, 6-dimensional spectral features, 3-dimensional exponential features (EVI, NDBI, BSI), 8-dimensional GLCM texture features, and 7-dimensional temporal difference features (with △NDVI as the core) are extracted and fused to construct a 24-dimensional feature vector, which is used as the model input.

[0070] 2.2. SMOTE-Tomek Sample Balancing

[0071] The training samples were processed by oversampling 1200 non-agricultural positive samples by 5 times using the SMOTE algorithm to obtain 6000 positive samples; then, the TomekLinks algorithm was used to remove overlapping sample pairs that accounted for 3% of the total samples, and finally, approximately 63000 balanced training samples were obtained.

[0072] 2.3. Improved XGBoost model training

[0073] The FocalLoss adjustment factor γ was set to 2.0, and the positive and negative sample weight ratio α was set to 0.75. An adaptive Bayesian optimization algorithm was used with F1-Score as the objective function to optimize the hyperparameters max_depth, min_child_weight, and eta for 50 iterations. The optimal hyperparameters were determined to be: learning_rate=0.05, max_depth=8, and min_child_weight=1. The model was trained using balanced training samples, and five-fold cross-validation was used to verify the accuracy. The F1-Score on the validation set reached 0.908.

[0074] 2.4. Model Prediction and Post-processing

[0075] The test samples are input into the trained model, which outputs a non-agricultural monitoring probability map. The probability map is then subjected to 3×3 mode filtering and smoothing, followed by connected component analysis to remove small patches with an area of ​​less than 200m², thus obtaining the final monitoring results.

[0076] 3. Experimental Results and Comparative Analysis

[0077] To verify the superiority of this invention, a comparative experiment was conducted with existing mainstream algorithms for monitoring farmland non-agriculturalization (Random Forest RF, traditional XGBoost, and deep learning U-Net). The hyperparameters of all comparison algorithms were adjusted to their respective optimal values. The experimental results are shown in Table 1.

[0078] Table 1. Performance comparison of different algorithms in the task of monitoring farmland non-agriculturalization.

[0079] Algorithm Model Accuracy Recall F1-Score Training time (s) Random Forest (RF) 82.3% 71.5% 0.765 180 Traditional XGBoost 85.1% 74.8% 0.796 120 U-Net (Deep Learning) 89.5% 85.4% 0.874 3600 This embodiment 92.4% 90.8% 0.916 210

[0080] As shown in Table 1, the accuracy, recall, and F1-Score of the method of this invention are significantly higher than those of other comparative algorithms. The recall is 16% higher than that of the traditional XGBoost, proving that the present invention can effectively capture small and hidden non-agricultural patches and solve the problem of missed detection in the prior art. The training time of the present invention is only 210 seconds, which is much lower than that of deep learning U-Net (3600 seconds). Although it is slightly higher than that of the traditional XGBoost, the accuracy is significantly improved, achieving a balance between efficiency and accuracy. Compared with random forest, the present invention improves the F1-Score by 0.151 with similar training time, showing a clear advantage in accuracy.

[0081] Meanwhile, a qualitative comparative analysis was conducted on three typical monitoring scenarios of farmland conversion to non-agricultural uses, and the results are as follows:

[0082] Scenario A (linear features): For newly added narrow, paved rural roads, traditional XGBoost often fails to detect road breaks due to unclear road features, resulting in a false negative rate of 45%. This invention achieves complete road identification through ΔNDVI differential features and connected component analysis, with an identification rate of 98%.

[0083] In Scenario B (complex terrain interference), in an area where greenhouse facilities and construction sites are mixed, the false alarm rate of traditional methods reaches 28%; this invention utilizes GLCM texture features to distinguish the transparent texture of the greenhouse from the hardened cement texture of the construction site, reducing the false alarm rate to 16%;

[0084] Scenario C (Seasonal Impact): For areas where vegetation cover has decreased due to crop harvesting, traditional methods are prone to misclassifying them as non-agricultural areas, with a false alarm rate of 32%. This invention eliminates seasonal interference by using multi-temporal difference features and accurately locates permanent farmland destruction caused by illegal dumping sites, reducing the false alarm rate in this scenario to 5%.

[0085] Example 2: Remote Sensing Monitoring System for Farmland Non-Agriculturalization Based on Multi-Source Feature Fusion and Improved XGBoost

[0086] This embodiment provides a monitoring system that implements the method described in Embodiment 1. The system adopts a modular design, including a feature extraction module, a sample balancing module, a model training and prediction module, and a post-processing module. Each module is developed based on Python 3.8, uses a B / S architecture, and supports web-based operation. The specific implementation is as follows:

[0087] The feature extraction module integrates GDAL and OpenCV libraries, supports batch import of domestic remote sensing images such as GF-2 and GF-6, automatically extracts multi-source heterogeneous features and constructs 24-dimensional feature vectors, and outputs the feature extraction results in CSV format for visualization.

[0088] The sample balancing module integrates the SMOTE and TomekLinks algorithms from the Scikit-learn library. It supports importing feature extraction results, automatically completes sample oversampling and noise removal, and the balanced samples can be directly imported into the model training module. It also supports custom adjustment of sampling parameters.

[0089] The model training and prediction module integrates the XGBoost library, with a built-in improved XGBoost model and adaptive Bayesian optimization algorithm. It supports automatic model training and five-fold cross-validation, and displays the accuracy curve in real time during training. The trained model can be saved as a model file, supports batch monitoring of test samples, and outputs a monitoring probability map.

[0090] The post-processing module integrates the mode filtering and connected component analysis algorithms from the OpenCV library. It supports importing monitoring probability maps, automatically performs smoothing and small patch removal, and outputs the final monitoring result map of farmland non-agriculturalization. The result map supports overlaying remote sensing image base maps for easy visualization and manual verification.

[0091] The system is deployed on a general-purpose server with an Intel i9 CPU and 64GB RAM. It can support 10 people to operate online at the same time. A single batch can monitor 1,000-2,000 km² of farmland remote sensing images. The monitoring results can be directly exported to the SHP vector format commonly used by land and resources departments without secondary format conversion, which is in line with actual business applications.

[0092] In summary, this invention addresses the core pain points of existing farmland non-agriculturalization monitoring technologies through a comprehensive optimization process encompassing multi-source feature fusion, sample balancing, model improvement, and spatial post-processing, achieving the following significant technical effects:

[0093] 1. High monitoring accuracy, significantly reducing the rate of missed detections and false alarms.

[0094] The present invention achieves a monitoring accuracy of 92.4%, a recall rate of 90.8%, and an F1-Score of 0.916. Compared with the traditional XGBoost model, the recall rate is improved by about 16%, and it can accurately identify small and scattered illegal non-agricultural patches (such as simple sheds on farmland and narrow hardened roads). At the same time, by fusing multi-temporal difference features and texture features, it effectively eliminates interference from fallow land, crop harvesting, and greenhouse facilities, reducing the false alarm rate by more than 12%.

[0095] 2. High monitoring efficiency and low hardware requirements.

[0096] Compared to the deep learning model U-Net, this invention improves training speed by nearly 17 times while maintaining higher monitoring accuracy, requiring only 210 seconds to monitor a single batch of 1200 km² of cultivated land. It is also compatible with general-purpose servers (Intel i9 CPU + 64GB RAM + NVIDIA RTX 3090 GPU), and even ordinary office computers can complete small- to medium-sized area monitoring, reducing hardware costs by 90%, making it more suitable for large-scale and rapid surveys by grassroots land and resources departments.

[0097] 3. The model has strong versatility and can be adapted to different application scenarios.

[0098] Through adaptive Bayesian hyperparameter optimization, the model can be adapted to farmland monitoring in different terrains such as plains, mountains, and hills without manual debugging, with accuracy fluctuations of less than 5%. At the same time, the parameters of each module of the system are configurable, and the sampling multiple, patch removal threshold, etc. can be adjusted according to the farmland management needs of different regions, making it highly adaptable.

[0099] 4. Monitoring results align with actual business needs.

[0100] Spatial consistency post-processing incorporates industry standards for land and space planning, eliminating noise from small, meaningless patches. The monitoring results can be directly used for the supervision and enforcement of non-agricultural use of arable land, eliminating the need for secondary manual screening and improving work efficiency.

[0101] The present invention has been described in detail above, with the aim of enabling those skilled in the art to understand and implement the invention. However, this description should not be construed as limiting the scope of protection of the invention. All equivalent changes or modifications made in accordance with the spirit and essence of the invention should be included within the scope of protection of the invention.

Claims

1. A remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, comprising a feature extraction module, a sample balancing module, a model training and prediction module, and a post-processing module, characterized in that, The feature extraction module is used to acquire remote sensing images of farmland from two different time periods, automatically extract spectral features, exponential features, texture features, and temporal differential features, fuse them to construct and output a multidimensional feature vector; The sample balancing module has a built-in SMOTE-Tomek hybrid sampling algorithm. It receives the sample data output by the feature extraction module, automatically performs 5-fold oversampling of minority class samples and 2-5% elimination of overlapping samples, and outputs balanced training samples. The model training and prediction module incorporates an improved XGBoost model, equipped with a FocalLoss cost-sensitive loss function and an adaptive Bayesian hyperparameter optimization algorithm to complete automatic model training; it also includes a model validation unit, which uses five-fold cross-validation for accuracy verification, and automatically triggers hyperparameter re-optimization when the F1-Score is below 0.

85. After training is completed, the test samples are monitored, and a non-agricultural monitoring probability map is output. The post-processing module sequentially performs mode filtering and connected component analysis on the monitoring probability map, automatically removes small patches with an area of ​​less than 200m², and outputs the final monitoring result map of farmland non-agriculturalization.

2. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost as described in claim 1, characterized in that, Spectral features extract the spectral reflectance of multiple bands (red, green, blue, and near-infrared) from remote sensing images, providing basic information for land cover identification; and / or, index features calculate improved vegetation index, building index, and bare soil index to characterize the vegetation cover of cultivated land, artificial buildings, and hardened bare soil, respectively, enabling preliminary screening of non-agricultural areas; and / or, texture features extract mean, variance, information entropy, and contrast based on the gray-level co-occurrence matrix, utilizing the difference between the natural irregular texture of cultivated land and the regular texture of artificial non-agricultural features to distinguish easily confused land features; and / or, temporal difference features calculate the difference in normalized vegetation index between two images, where ΔNDVI = NDVI_t2 - NDVI_t1, serving as a core indicator for characterizing drastic changes in cultivated land vegetation cover to eliminate interference from seasonal phenological changes.

3. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost as described in claim 1, characterized in that, The SMOTE-Tomek hybrid sampling algorithm includes the SMOTE algorithm, which synthesizes minority class samples through linear interpolation in the feature space, and the TomekLinks algorithm, which removes overlapping sample pairs located on inter-class boundaries. The SMOTE algorithm... The specific synthesis formula is as follows: Let x be a non-agricultural minority class sample. i =[x i1 x i2 , ..., x id ] T Its k-nearest neighbor sample in the feature space is x j =[x j1 x j2 , ..., x jd ] T Then the synthesized new sample x new Generate using the following formula: x new =x i +λ(x j -x i ), where λ∈(0,1) are random interpolation coefficients, randomly generated from a uniform distribution; d is the feature dimension; x j -x i The vector difference between the samples in the feature space; TomekLinks Algorithm Formula and Definition: Let x be any two distinct class samples in the sample space. i Cultivated land type, x j For non-agricultural categories, the Euclidean distance between the two is... Where d is the feature dimension, if x i With x j If the following TomekLinks conditions are met, x i It is x j nearest neighbor; x j It is x i nearest neighbor; x i With x j If they belong to different categories, then they are called (x i x j) A set of TomekLinks sample pairs, which are located at the classification boundary and overlap, are considered noise samples.

4. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 3, characterized in that, First, the SMOTE algorithm is used to oversample the non-agricultural minority class samples by 5 times in the feature space to supplement the number of minority class samples and solve the problem of the model being biased towards negative samples. The oversampling factor of 5 is the optimal value verified by multiple experiments. If the factor is lower than 3, the sample supplementation will be insufficient, and if the factor is higher than 7, overfitting is likely to occur. Second, the TomekLinks algorithm is used to remove overlapping sample pairs located on the inter-class boundaries. The overlapping samples removed account for about 3-4% of the total sample, cleaning up the inter-class overlapping noise and making the classification decision boundary clearer.

5. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 1, characterized in that, Replace the standard XGBoost log loss function with the FocalLoss cost-sensitive loss function, and the objective function is: The adjustment factor γ = 2.0, and the positive / negative sample weight ratio α = 0.

75. To predict probabilities for the model, This is the regularization term; the loss function is obtained through... It automatically reduces the weight of easily classified samples, allowing the model to focus on edge samples that are difficult to distinguish.

6. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 1, characterized in that, Establish a cost function with F1-Score as the objective. F1-Score is more suitable for precision evaluation of imbalanced samples. Its calculation formula is: F1-Score = 2 × precision × recall ÷ (precision + recall), where precision = number of true positive samples ÷ (number of true positive samples + number of false positive samples) and recall = number of true positive samples ÷ (number of true positive samples + number of false negative samples).

7. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 1, characterized in that, A surrogate model for hyperparameters and model performance is constructed using Gaussian processes. Maximizing the F1-Score is the optimization objective. The max_depth, min_child_weight, and eta hyperparameters of the XGBoost model are automatically optimized. The specific process is as follows: Hyperparameter search space definition: Set the range of values ​​for the hyperparameters to be optimized. Tree depth max_depth ∈ [3, 12], minimum subtree weight min_child_weight ∈ [1, 10], and learning rate eta ∈ [0.01, 0.3]. Gaussian process surrogate model construction: Using sampled hyperparameter combinations and their corresponding F1-Scores as training data, a Gaussian process is used to fit the mapping relationship between hyperparameters and F1-Score, quickly predicting the model performance for unsampled points.

8. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 7, characterized in that, The sampling function is selected using the expected boost function as the sampling criterion, as follows: EI(x)=E[max(F1(x)−F1_best,0)], where F1(x) is the predicted F1-Score corresponding to the hyperparameter combination x; F1_best is the currently found optimal F1-Score; and / or, the iterative optimization process randomly samples 10 sets of hyperparameter combinations, trains the model and calculates the F1-Score; in each round, the surrogate model is updated using a Gaussian process, and the next set of hyperparameters is selected through the EI function; the optimization stops after 50 iterations.

9. The remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost according to claim 1, characterized in that, A 3×3 mode filter is used to smooth the monitoring probability map and eliminate isolated noise points. The calculation method is as follows: taking any pixel (i,j) in the probability map as the center, the monitoring probability values ​​of 9 pixels in its 3×3 neighborhood are taken to form a set: S={P(i−1,j−1),P(i−1,j),P(i−1,j+1),P(i,j−1),P(i,j),P(i,j+1),P(i+1,j−1),P(i+1,j),P(i+1,j+1)}. The probability values ​​of the neighborhood are statistically analyzed, and the probability value that appears most frequently (the mode) is used as the output value after filtering the center pixel: P′(i,j)=Mode(S) where: P(i,j) P'(i,j) represents the monitoring probability of the center pixel before filtering; P'(i,j) represents the smoothing probability of the center pixel after filtering; Mode(·) is the operation of taking the mode of the set; a 3×3 neighborhood can effectively suppress isolated noise points, making the monitoring results more continuous and smooth; and / or, connected component analysis, to calculate the area of ​​the monitored non-agricultural patches, specifically by using the 4-connected or 8-connected component labeling rules to group spatially adjacent pixels with probability values ​​higher than the threshold into the same independent patch; using the actual ground area corresponding to a single pixel as the basic unit, the area of ​​the non-agricultural patch is calculated according to the following formula: Patch area S=N×S0, where S is the actual ground area of ​​the current connected component; N is the total number of pixels contained in the current connected component; S0 represents the actual ground area of ​​a single pixel.

10. A remote sensing monitoring method for farmland non-agriculturalization based on multi-source feature fusion and XGBoost, characterized in that, It employs the remote sensing monitoring system for farmland non-agriculturalization based on multi-source feature fusion and XGBoost as described in any one of claims 1 to 9, and includes the following steps: S1. Construction of Multi-Source Heterogeneous Features Two remote sensing images of farmland at different time phases were acquired, and spectral features, exponential features, texture features, and temporal difference features were extracted to construct a 3D feature vector. The exponential features include the improved vegetation index, building index, and bare soil index. The texture features are the mean, variance, and information entropy extracted based on the gray-level co-occurrence matrix. The temporal difference feature is the NDVI difference ΔNDVI between the two images. S2. Sample balancing process: The SMOTE-Tomek hybrid sampling algorithm was used to process the cultivated land samples and non-agricultural samples. First, the non-agricultural minority samples were oversampled by 5 times using the SMOTE algorithm, and then the TomekLinks algorithm was used to remove the inter-class overlapping sample pairs that accounted for 2-5% of the total sample volume, so as to obtain the balanced training samples. S3. Improved XGBoost model training and prediction: The balanced training samples obtained in step S2 are input into the improved XGBoost model. The model replaces the original log loss function with the FocalLoss cost-sensitive loss function, and automatically optimizes the max_depth, min_child_weight, and eta hyperparameters of the XGBoost model with F1-Score as the objective function through an adaptive Bayesian optimization algorithm. After the model training is completed, the test samples are monitored and a non-agricultural monitoring probability map is output. S4. Spatial Consistency Post-processing The monitoring probability map output in step S3 is subjected to 3×3 mode filtering and connected component analysis in sequence. Monitoring patches with an area of ​​less than 200m² are removed to obtain the final monitoring results of farmland non-agriculturalization.

Images