Farmland tree remote sensing image optimization method

Abstract

The invention discloses a farmland tree remote sensing image optimization method, relates to the technical field of remote sensing image processing, and solves the problems that in an existing optical image processing technology, a data processing flow is too complex in an optical image processing process, and wrong division and missing division are likely to occur, so that farmland tree images in optical images are not clear, and the image quality is poor. Therefore, the application of optical images in agricultural operation activities is restricted. A farmland tree remote sensing image optimization method comprises a data acquisition stage, a vegetation and non-vegetation distinguishing stage, a vegetation type distinguishing stage, a farmland tree and non-farmland tree distinguishing stage and a farmland tree remote sensing image optimization stage. The method is suitable for the fields of farmland tree remote sensing monitoring, ecology, remote sensing technology, geographic information systems and the like.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and also to the field of farmland tree monitoring technology. Background Technology

[0002] Currently, multi-source remote sensing imagery enables systematic monitoring of large-scale continuous forests. However, the dynamic changes and ecological value of trees widely distributed outside forests, especially scattered farmland trees in agricultural landscapes, including field ridge trees, farmland shelterbelts, hedgerows, ditch / roadside trees, isolated trees, and small clumps, are largely overlooked. In the context of the global search for sustainable development paths, agroforestry systems comprised of farmland trees offer significant potential as an important nature-based solution in carbon sequestration, biodiversity conservation, wind and water erosion reduction, microclimate improvement, and enhanced resilience of agro-ecosystems.

[0003] However, due to the inherent limitations of traditional remote sensing technology in identifying fragmented, linear, and discrete vegetation units, coupled with the continuous dynamic disturbance effect of agricultural operations such as cultivation, pruning, clearing, and crop rotation on tree cover, the directly obtained remote sensing images cannot clearly display the image information of farmland trees. Consequently, the application of remote sensing images in the field of farmland tree identification and monitoring is flawed.

[0004] Existing research on farmland trees focuses primarily on carbon sequestration, biodiversity conservation, and rural livelihood support, but the precise characterization of their spatial distribution patterns lags behind. Current techniques for accurately characterizing the spatial distribution patterns of farmland trees generally employ extremely high spatial resolution imagery or medium-to-low resolution optical imagery.

[0005] Among them, monitoring farmland trees using extremely high spatial resolution imagery has enabled the extraction of individual trees or tree crowns in small areas. However, this type of data acquisition is costly, has limited coverage of a single scene, and involves complex data processing procedures, making it difficult to achieve routine and repeatable monitoring on a large scale, thus limiting its universality as a regional census tool.

[0006] Monitoring farmland trees based on medium- and low-resolution optical images generally involves obtaining a macroscopic assessment of tree coverage at a global or continental scale through a hybrid pixel decomposition method. However, due to insufficient spatial resolution, misclassification and omission are prone to occur in the fragmented landscape of farmland tree distribution, making it difficult to accurately reflect the boundaries between farmland and forest land and meet the actual needs of precision agriculture and regional ecological management.

[0007] In optical image processing, pixel-based classification algorithms (such as maximum likelihood classification and decision tree classification) were widely used in early applications. However, their classification performance is highly dependent on the spectral separability of different land features. In areas with relatively homogeneous vegetation features, these methods can achieve basic classification, but in the context of fragmented spatial patterns and spectral heterogeneity in farmland trees, the recognition accuracy drops significantly. Because they ignore spatial context information, the algorithms struggle to maintain the geometric integrity of linear hedgerows or small tree patches, resulting in the processed optical images failing to clearly display the morphology of hedgerows or small trees.

[0008] In optical image processing techniques, object-oriented image analysis methods obtain homogeneous image objects through multi-scale segmentation and combine them with multi-dimensional features such as spectral, texture, and geometric characteristics. This effectively improves the classification accuracy of complex landscapes and has greater potential for preserving the structure of farmland tree patches in optical images. However, even if "tree" objects can be accurately identified, distinguishing farmland trees from adjacent conventional forests with similar spectral characteristics and assigning them accurate semantic labels remains a pressing technical challenge.

[0009] In summary, existing optical image processing technologies suffer from several drawbacks. The fragmented, linear, and discrete nature of farmland tree images, the continuous dynamic changes in tree cover caused by agricultural activities, the high cost of acquiring high spatial resolution images, and the limited coverage of individual scenes all contribute to overly complex data processing workflows. These drawbacks easily lead to misclassification and omissions, resulting in unclear farmland tree images and thus limiting the application of optical imagery in agricultural operations. Summary of the Invention

[0010] This invention alleviates the shortcomings of existing optical image processing technologies. These limitations stem from the fragmented, linear, and discrete nature of farmland tree images, the continuous dynamic changes in tree cover caused by agricultural activities, and the high cost and limited coverage of high-spatial-resolution images. These limitations lead to overly complex data processing workflows, prone to misclassification and omissions, and consequently, unclear farmland tree images, thus restricting the application of optical images in agricultural operations. This invention provides the following solution: A method for optimizing remote sensing images of farmland trees includes the following stages: Data acquisition phase: Collect remote sensing images and corresponding remote sensing observation data of the monitoring area throughout the year; Distinguishing between vegetated and non-vegetated stages: Based on the remote sensing images and the remote sensing observation data, vegetated areas are obtained using the vegetation frequency method; Vegetation type differentiation stage: Based on the remote sensing observation data of the vegetation area, the vegetation area is classified to obtain several initial candidate forest land objects; Distinguishing between farmland trees and non-farmland trees: Based on the aforementioned initial candidate forest land objects and corresponding remote sensing observation data, the morphological characteristics of each initial candidate forest land object are obtained; Based on the morphological characteristics of each initial candidate forest land object, the corresponding probability of farmland trees is obtained through the farmland tree probability prediction model; if the probability of farmland trees of the initial candidate forest land object is greater than the farmland tree determination threshold, it is determined to be farmland trees, otherwise it is determined to be non-farmland trees. Farmland Tree Remote Sensing Image Optimization Stage: Based on all the initial candidate forest land objects identified as farmland trees, the remote sensing image is optimized.

[0011] Furthermore, in one embodiment of the present invention, the remote sensing observation data includes vertical structure, scattering feature Sentinel-1, spectral feature Sentinel-2, vegetation index, phenological features, and auxiliary factors; The vertical structure includes a canopy height RH; The spectral features Sentinel-2 include the blue band B2_B, the green band B3_G, the red band B4_R, the red edge band B5_RE1, the red edge band B6_RE2, the red edge band B7_RE3, the near-red band B8_NIR1, the near-red band B8A_NIR2, the short-red band B11_SWIR1, and the short-red band B12_SWIR2. The auxiliary factors include elevation DEM and slope SLOPE.

[0012] Furthermore, in one embodiment of the present invention, the scattering feature Sentinel-1 includes VV polarization, VH polarization, polarization ratio Rt, polarization difference Dt, amplitude Amp, coefficient of variation CV, texture entropy T_E, and texture contrast T_C.

[0013] Furthermore, in one embodiment of the present invention, the vegetation index includes the normalized vegetation index (NDI). Enhance vegetation index and red-edge index .

[0014] Furthermore, in one embodiment of the present invention, the phenological characteristics include the maximum value of the vegetation index. Average vegetation index and vegetation index standard deviation .

[0015] Furthermore, in one embodiment of the present invention, the vegetation frequency method for distinguishing between vegetated and non-vegetated stages is as follows: Based on the remote sensing observation data, the normalized vegetation index of each pixel in the remote sensing image is obtained. ; The normalized vegetation index of each pixel throughout the year is calculated. The proportion exceeding the NDVI threshold T is used as the vegetation frequency for the corresponding pixel. ; Each pixel of the remote sensing image is processed as follows to obtain vegetation pixels, and the vegetation pixels are combined into vegetation regions: If the vegetation frequency of the pixel Less than the vegetation frequency threshold If the condition is met, the pixel is determined to be a non-vegetation pixel; otherwise, the pixel is determined to be a vegetation pixel.

[0016] Furthermore, in one embodiment of the present invention, in the stage of distinguishing between farmland trees and non-farmland trees, the morphological characteristics include shape index SI, elongation EL, and adjacent farmland ratio APR.

[0017] Furthermore, in one embodiment of the invention, during the vegetation type differentiation stage, Based on the scattering feature Sentinel-1 and the spectral feature Sentinel-2, the remote sensing image of the vegetation area is segmented using the Simple Non-Iterative Clustering (SNIC) method to obtain several superpixel objects. Based on the remote sensing observation data corresponding to the aforementioned superpixel objects, the superpixel objects are classified using a vegetation type classification model to obtain several superpixel objects whose classification results are initial candidate forest land objects.

[0018] Furthermore, in one embodiment of the present invention, the vegetation type classification model is obtained through the following method: Construct a dataset, which includes the remote sensing observation data of the superpixel object and the corresponding vegetation type; the vegetation type includes initial candidate forest objects and non-initial candidate forest objects. Using the remote sensing observation data as input and the vegetation type as output, a machine learning regression model is trained using a gridded search optimization method, and the obtained machine learning regression model serves as a vegetation type classification model.

[0019] Furthermore, in one embodiment of the present invention, in the stage of distinguishing between farmland trees and non-farmland trees, the farmland tree probability prediction model is obtained by the following method: Construct a dataset that includes morphological features of forest objects and corresponding binary labels for farmland trees; Using the morphological features as input and the binary labels of farmland trees as output, a multiple linear regression model is trained, and the obtained multiple linear regression model serves as a probability prediction model for farmland trees.

[0020] The method for optimizing remote sensing images of farmland trees described in this invention is based on multimodal remote sensing data collaborative spatial semantic features and a machine learning architecture. It effectively alleviates the shortcomings of existing optical image processing techniques. These limitations stem from the fragmented, linear, and discretized nature of farmland tree images, the continuous dynamic changes in tree cover caused by agricultural activities, and the high cost and limited coverage of high-spatial-resolution images. These limitations lead to overly complex data processing procedures, prone to misclassification and omissions, and consequently, unclear farmland tree images, thus restricting the application of optical images in agricultural operations. Specific beneficial effects include: 1. The farmland tree remote sensing image optimization method described in this invention constructs a real-time monitoring method for farmland trees that integrates time-series phenological analysis, object-oriented classification, morphological features and spatial semantic features. It adopts a coarse-to-fine hierarchical processing strategy to achieve large-scale and highly heterogeneous farmland tree image optimization.

[0021] In general, the method first generates a vegetation mask using time-series vegetation indices to quickly separate vegetation from non-vegetated land features. Second, object-oriented classification is implemented within the vegetated areas to distinguish major vegetation types (farmland, woodland, wetland, and grassland). Subsequently, identification is performed based on the significant differences in morphological and spatial semantic features between farmland trees and non-farmland trees. This approach, while fully considering the distribution characteristics of trees under complex farmland conditions, further mines the spatial semantic features, spectral characteristics, phenological features, and structural physicochemical differences of farmland trees, achieving accurate and robust optimization of remote sensing images of farmland trees in large-scale and highly heterogeneous farmland landscapes.

[0022] Furthermore, the present invention can also utilize reserved test sets to evaluate user accuracy, producer accuracy, overall accuracy, and F1 score at each level.

[0023] 2. The farmland tree remote sensing image optimization method described in this invention uses GEDI LiDAR, Sentinel-1 synthetic aperture radar (SAR), and Sentinel-2 multi-remote sensing images as data sources for collaborative monitoring of the spatial distribution of farmland trees. These images complement each other in terms of vertical structure, physical scattering characteristics, and spectral phenological information. This data can capture vegetation vertical structure information, surface backscattering characteristics, spectral differences, and seasonal phenological rhythms, achieving a comprehensive characterization of land cover in a four-dimensional space of "height-structure-spectrum-time." This significantly improves the initial identification accuracy of the model in distinguishing between tall crops, shrubs, and trees, and reduces the impact of continuous dynamic disturbances caused by agricultural activities on tree cover.

[0024] 3. The farmland tree remote sensing image optimization method of the present invention introduces spatial semantic constraints and landscape function discrimination mechanism, and combines machine learning algorithm to perform semantic classification on the initially identified tree patches.

[0025] By quantifying patch morphometric indicators, setting area thresholds (to effectively filter large continuous forest patches), and analyzing spatial context information such as their spatial topological relationships with adjacent farmland plots (e.g., adjacency, inclusion, and separation), this method distinguishes farmland trees from conventional forests based on the essential nature of landscape patterns and ecological functions. It maintains the geometric integrity of linear hedgerows or small tree patches and distinguishes farmland trees from adjacent conventional forests with similar spectral characteristics. This solves the semantic confusion problem caused by traditional methods that rely solely on pixel attributes, thereby enabling large-scale, high-precision thematic mapping and dynamic monitoring of farmland trees.

[0026] Remote sensing images processed by the optimization method described in this invention can be applied to agricultural management activities, providing more accurate remote sensing images for activities such as precisely quantifying the carbon sequestration potential of agricultural landscapes, assessing their ecosystem service value, and planning farmland data.

[0027] The optimization method described in this invention can quickly and accurately map the spatial distribution and changes of farmland trees. Applied to agricultural management, it can help formulate management activities to address climate change and support regional agricultural management and ecological governance. It is applicable to fields such as farmland tree remote sensing monitoring, ecology, remote sensing technology, and geographic information systems. Attached Figure Description

[0028] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of the farmland tree remote sensing image optimization method described in Implementation Method 1; Figure 2This is a flowchart of the farmland tree remote sensing image optimization method described in Implementation Method Eleven; Figure 3 This is a diagram illustrating the effect of vegetation threshold determination based on ROC curves as described in Implementation Method Eleven. Figure 4 This is a numerical range representation diagram of the morphological characteristics described in Implementation Method Eleven in farmland trees; Figure 5 This is a numerical range representation diagram of the morphological characteristics described in Implementation Method Eleven in other forest lands; Figure 6 It is the morphological feature importance assessment diagram described in Implementation Method Eleven; Figure 7 This is a diagram illustrating the effect of the optimal threshold determination based on F1-score as described in Implementation Method Eleven. Figure 8 This is a classification result diagram of typical vegetation types as described in Implementation Method Eleven; Figure 9 This is a classification effect diagram of typical vegetation types as described in Implementation Method Eleven; Figure 10 This is a classification and proportion diagram of typical vegetation types as described in Implementation Method Eleven; Figure 11 This is a spatial distribution pattern diagram of the typical vegetation types described in Implementation Method Eleven; Figure 12 This is a spatial distribution diagram of the study area described in Implementation Method Eleven, where a represents region a, b represents region b, c represents region c, and d represents region d. Figure 13 This is a map showing the classification results of farmland trees in region a as described in Implementation Method Eleven; Figure 14 This is a map showing the classification results of farmland trees in region b as described in Implementation Method Eleven; Figure 15 This is a map showing the classification results of farmland trees in region c as described in Implementation Method Eleven; Figure 16 This is a map showing the classification results of farmland trees in region d as described in Implementation Method Eleven. Detailed Implementation

[0029] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0030] Implementation Method 1: This implementation method describes a method for optimizing remote sensing images of farmland trees, such as... Figure 1 As shown, it includes the following stages: Data acquisition phase: Collect remote sensing images and corresponding remote sensing observation data of the monitoring area throughout the year; Distinguishing between vegetated and non-vegetated stages: Based on the remote sensing images and the remote sensing observation data, vegetated areas are obtained using the vegetation frequency method; Vegetation type differentiation stage: Based on the remote sensing observation data of the vegetation area, the vegetation area is classified to obtain several initial candidate forest land objects; Distinguishing between farmland trees and non-farmland trees: Based on the aforementioned initial candidate forest land objects and corresponding remote sensing observation data, the morphological characteristics of each initial candidate forest land object are obtained; Based on the morphological characteristics of each initial candidate forest land object, the corresponding probability of farmland trees is obtained through the farmland tree probability prediction model; if the probability of farmland trees of the initial candidate forest land object is greater than the farmland tree determination threshold, it is determined to be farmland trees, otherwise it is determined to be non-farmland trees. Farmland Tree Remote Sensing Image Optimization Stage: Based on all the initial candidate forest land objects identified as farmland trees, the remote sensing image is optimized.

[0031] This embodiment describes a method for optimizing remote sensing images of farmland trees. Through a data acquisition stage, a vegetation / non-vegetation differentiation stage, a vegetation type differentiation stage, a farmland tree / non-farmland tree differentiation stage, and a farmland tree remote sensing image optimization stage, it achieves accurate and robust optimization of farmland tree remote sensing images in large-scale and highly heterogeneous farmland landscapes. In the vegetation / non-vegetation differentiation stage, a vegetation frequency (VF) method based on annual time series is used for vegetation identification, effectively improving classification efficiency and avoiding interference from non-research objects (such as buildings, roads, water bodies, and bare land).

[0032] In the vegetation type differentiation stage, based on remote sensing observation data of the vegetation area, the vegetation area is classified to obtain several initial candidate forest land objects.

[0033] In the stage of distinguishing between farmland trees and non-farmland trees, given that the differences in morphological structure and landscape configuration between farmland and non-farmland trees mainly lie in their morphological features and spatial semantic relationships, this stage explores these differences in morphological and spatial semantic features. Based on the "forest land" category, geometric morphology and neighborhood semantic constraints are introduced to achieve refined extraction of farmland trees. A probabilistic prediction model for farmland trees is used to output the probability that each forest land object belongs to farmland trees. A threshold for classifying farmland trees is used to distinguish between farmland and non-farmland trees, achieving secondary identification of farmland trees from the overall forest land. Based on the principle of reducing average impurity, this stage evaluates the contribution of each feature and quantifies the role of different semantic features in distinguishing farmland trees.

[0034] Implementation Method 2: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method 1. In this implementation method, the remote sensing observation data includes vertical structure GEDI, scattering feature Sentinel-1, spectral feature Sentinel-2, vegetation index, phenological features, and auxiliary factors. The vertical structure includes a canopy height RH; The spectral features Sentinel-2 include the blue band B2_B, the green band B3_G, the red band B4_R, the red edge band B5_RE1, the red edge band B6_RE2, the red edge band B7_RE3, the near-red band B8_NIR1, the near-red band B8A_NIR2, the short-red band B11_SWIR1, and the short-red band B12_SWIR2. The auxiliary factors include elevation DEM and slope SLOPE.

[0035] In this embodiment, the remote sensing observation data includes a total of six modules: vertical structure, SAR scattering, spectrum, vegetation index, phenology, and topographic aids. Specific characteristics are shown in Table 1. Table 1

[0036] These features can be used, after resampling and registration, to input into a subsequent random forest classifier for vegetation type classification, in order to achieve preliminary tree area delineation.

[0037] This implementation further refines the method for optimizing remote sensing images of farmland trees, providing an example of remote sensing observation data. By integrating GEDI LiDAR-derived vertical canopy height, Sentinel-1 polarization, and Sentinel-2 multispectral data, this implementation constructs a multimodal feature set including GEDI LiDAR, Sentinel-1 synthetic aperture radar (SAR), Sentinel-2 multi-remote sensing image data, vegetation indices, phenological characteristics, and auxiliary factors. It comprehensively characterizes vegetation attributes from three dimensions: vertical structural height, microwave scattering characteristics, and spectral phenological dynamics, thereby improving the accuracy of farmland tree identification in complex farmland landscapes. It quantifies the differences in annual growth cycle, canopy structure, and spatial heterogeneity of vegetation from three dimensions: phenological characteristics, canopy structure, and spatial texture, improving classification robustness and reducing interference from cloud cover and surface complexity, achieving robust identification of farmland trees in large-scale heterogeneous landscapes.

[0038] Among them, GEDI LiDAR data can extract vertical structure indicators such as vegetation canopy height, leaf height diversity, and canopy cover. These indicators directly reflect the three-dimensional morphology, vertical distribution of biomass, and spatial heterogeneity of vegetation, and are the core basis for distinguishing between tall trees and low-growing crops. This implementation method uses 1-meter resolution tree canopy height predictions as the main vertical structure indicator. These predictions are based on high-resolution (Maxar) optical imagery, trained using self-supervised learning and a vision transformer model, combined with airborne LiDAR data, and calibrated using spaceborne GEDI LiDAR data, thereby generating a high-precision tree canopy height map, effectively compensating for the lack of spatial sampling density in GEDI data.

[0039] Sentinel-1 SAR dual polarization (VV / VH) time series are sensitive to changes in vegetation branch structure, canopy roughness and water content, and are unaffected by clouds and rain. They are suitable for capturing scattering intensity fluctuations caused by farmland crops on an annual scale, and can be contrasted with the relatively stable scattering behavior of perennial woody vegetation.

[0040] Furthermore, this implementation integrates SRTM digital elevation model (DEM) data to extract auxiliary factors such as altitude and slope. These factors help the model understand the preferred distribution of trees on slopes or at high altitudes, further improving the overall accuracy and generalization ability of the identification, and taking into account the potential impact of terrain on vegetation distribution.

[0041] Implementation Method 3: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method 2. In this implementation method, the scattering feature Sentinel-1 includes VV polarization, VH polarization, polarization ratio Rt, polarization difference Dt, amplitude Amp, coefficient of variation CV, texture entropy T_E, and texture contrast T_C.

[0042] In this embodiment, the polarization ratio Rt is obtained through...

[0043] Obtain, among which, For image timing, For VV polarization in The backscattering coefficient at time t. For VH polarization in The backscattering coefficient at time t.

[0044] In this embodiment, the polarization difference Dt is obtained through...

[0045] get.

[0046] In this embodiment, the amplitude Amp is obtained through...

[0047] Obtain, among which, This represents the backscattering coefficient sequence of the target pixel at each time phase, specifically the VV and VH polarization backscattering coefficient sequences.

[0048] In this embodiment, the coefficient of variation CV is obtained through...

[0049] Obtain, among which, For the target pixel in time phase VV polarization and VH polarization backscattering coefficients Standard deviation, This is the mean.

[0050] This implementation further refines the remote sensing observation data, illustrating the Sentinel-1 scattering feature scheme. Within the growing season and year-round time windows, this implementation constructs features such as backscattering intensity, polarization channel ratio, and differences, while simultaneously calculating dynamic statistics such as amplitude, mean, and standard deviation to characterize temporal stability. Furthermore, this implementation introduces gray-level co-occurrence matrix (GLCM) texture features, texture entropy, and contrast to characterize local spatial heterogeneity and structural complexity, thereby reducing confusion caused by highly scattering bare land or building edges, and enhancing sensitivity to typical linear / zonal patterns of farmland trees (protective forest belts, field ridge trees, roadside forest networks) and their boundary abrupt changes.

[0051] Implementation Method Four: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method Two. In this implementation method, the vegetation index includes the Normalized Difference Vegetation Index (NDVI). Enhance vegetation index and red-edge index .

[0052] In this embodiment, the normalized vegetation index pass

[0053] Obtain, among which, Near-infrared band 1, It is the red band.

[0054] In this embodiment, the enhanced vegetation index pass

[0055] Obtain, among which, It is the blue band.

[0056] In this embodiment, the red-edge index pass

[0057] Obtain, among which, For red-edge band 1, For red-edge band 2, It is the red-edge band 3.

[0058] This implementation further defines the remote sensing observation data and provides an example of the vegetation index scheme. Based on the spectral feature Sentinel-2, this implementation extracts the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), and Red Edge Index (IRECI). These indicators can reflect the chlorophyll content, biomass, and health status of vegetation.

[0059] Implementation Method 5: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method 2. In this implementation method, the phenological characteristics include the maximum value of the vegetation index. Average vegetation index and vegetation index standard deviation .

[0060] In this embodiment, the maximum value of the vegetation index pass

[0061] Obtain, among which, for Time series vegetation index characteristics at any given time.

[0062] In this embodiment, the average vegetation index pass

[0063] Obtain, among which, This refers to the temporal characteristics of the image.

[0064] In this embodiment, the vegetation index standard deviation pass

[0065] Obtain, among which, This represents the average value of the vegetation index characteristics of the target pixel over the entire time series.

[0066] This embodiment further defines the remote sensing observation data and provides examples of phenological characteristics. This embodiment performs time-series analysis, statistically analyzing the maximum, average, and standard deviation of the vegetation index and other time-series statistical characteristics throughout the annual growth cycle to obtain the phenological characteristics, including the maximum value of the vegetation index. Average vegetation index and vegetation index standard deviation This time-series analysis is used to filter out seasonal noise and highlight the persistent green cover characteristics of trees to capture the longer growing season, higher peak biomass, and more stable phenological dynamics of farmland trees relative to crops.

[0067] Implementation Method Six: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method One. In this implementation method, the vegetation frequency method for distinguishing between vegetated and non-vegetated stages is as follows: Based on the remote sensing observation data, the normalized vegetation index of each pixel in the remote sensing image is obtained. ; The normalized vegetation index of each pixel throughout the year is calculated. The proportion exceeding the NDVI threshold T is used as the vegetation frequency for the corresponding pixel. ; Each pixel of the remote sensing image is processed as follows to obtain vegetation pixels, and the vegetation pixels are combined into vegetation regions: If the vegetation frequency of the pixel Less than the vegetation frequency threshold If the condition is met, the pixel is determined to be a non-vegetation pixel; otherwise, the pixel is determined to be a vegetation pixel.

[0068] In this embodiment, the vegetation frequency pass

[0069] Obtain, among which, This is an indicator function that takes the value 1 when the condition is met and 0 otherwise. This represents the total number of valid observations throughout the year. For the first i The normalized vegetation index of the first effective observation; T is the threshold of the normalized vegetation index.

[0070] In this embodiment, the vegetation frequency method preferably performs vegetation masking processing on the images of the study area.

[0071] In this embodiment, the normalized vegetation index The preferred method is to synthesize Sentinel-2 cloud-free images every half month, for a total of 24 Sentinel-2 images throughout the year. The Normalized Difference Vegetation Index (NDVI) is calculated for each image, and the 24 NDVI values ​​are used as the vegetation frequency input parameters. A 10-day NDVI time series (36 images throughout the year) is constructed based on the Sentinel-2 data.

[0072] In this embodiment, the vegetation frequency ∈ [0,1].

[0073] In this embodiment, the total number of valid observations N throughout the year is 36.

[0074] In this embodiment, the optimal threshold The preferred value is 0.194.

[0075] In this embodiment, the non-vegetation pixels include bare soil, water bodies, impermeable surfaces, and permanent ice and snow.

[0076] In this embodiment, the non-vegetation pixels are preferably removed from the remote sensing image of the area to be detected by using a mask.

[0077] In this embodiment, the vegetation frequency threshold The ROC curve (Receiver Operating Characteristic) is obtained through evaluation. ROC is a graphical tool used to evaluate the discriminative performance of binary classification models. It systematically iterates through all possible classification thresholds, plotting a curve with the False Positive Rate (FPR) on the horizontal axis and the True Positive Rate (TPR) on the vertical axis, thus comprehensively depicting the trade-off between the classifier's sensitivity and specificity at different decision thresholds. Specifically: The NDVI threshold T is traversed within the interval [0.1, 0.5] with a step size of 0.05, and the vegetation frequency of each pixel is calculated at each fixed T value. Subsequently As a continuous discriminant variable, the system change classification threshold Plot the corresponding ROC curve and calculate the area under the curve. Select the curve that maximizes the AUC value. and The combination is used as the optimal parameter, and the Youden's Index is used as the parameter. Maximize the corresponding The value is used as the optimal decision point for the final vegetation / non-vegetation binary classification.

[0078] This embodiment further defines the distinction between vegetated and non-vegetated stages and provides an example of the vegetation frequency method. This embodiment uses the vegetation frequency (VF) method based on annual time series for vegetation identification. This method utilizes high-frequency remote sensing observations throughout the year to count the proportion of effective observations where the NDVI value of each pixel exceeds the threshold T (VF = number of observations where NDVI>T / total number of effective observations). This transforms vegetation identification from judging the instantaneous state at a single point in time to statistical analysis of complete phenological rhythms, effectively overcoming the limitations of single-phase images affected by seasons, cloud shadows, and atmospheric interference, improving classification efficiency, and avoiding interference from non-research objects (such as buildings, roads, water bodies, and bare land).

[0079] Implementation Method Seven: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method One. In this implementation method, during the stage of distinguishing farmland trees from non-farmland trees, the morphological features include shape index. elongation and the proportion of adjacent farmland .

[0080] In this embodiment, the shape index pass

[0081] Obtain, among which, The perimeter of the initial candidate forest land object. The area of ​​the initial candidate forest land object.

[0082] In this embodiment, the elongation pass

[0083] Obtain, among which, Let be the length of the longer side of the minimum bounding rectangle of the initial candidate forest land object. The length of the short side of the minimum bounding rectangle of the initial candidate forest land object.

[0084] In this embodiment, the proportion of adjacent farmland pass

[0085] Obtain, among which, It is the sum of the common boundary lengths of the initial candidate forest land object and all adjacent farmland category objects.

[0086] This implementation further refines the stage of distinguishing between farmland trees and non-farmland trees, providing illustrative examples of morphological characteristics. It extracts initial candidate objects from the "woodland" category and obtains key morphological characteristics for each object, including shape indices. elongation and the proportion of adjacent farmland .

[0087] When the initial candidate woodland objects are close to circular or regular block shapes The value is relatively large; when the object is long and narrow, strip-shaped, or has a complex and irregular outline, The value decreases. When the object is linear or elongated, The value is significantly greater than 1, while the value of blocky woodland is significantly greater than 1. The value is relatively low. A higher value indicates that the forest land object has more contact with farmland along the boundary, and is more likely to be embedded in the farmland landscape in the form of shelterbelts, field ridges, or field roadside trees. Through these characteristics, it is possible to morphologically distinguish strip-shaped or shelterbelt-like farmland trees from large areas of contiguous forest land, and spatially differentiate farmland trees adjacent to fields from isolated forest patches or mountain forest land far from cultivated land. By exploring the differences in morphological and spatial semantic features between farmland trees and non-farmland trees, precise extraction of farmland trees can be achieved.

[0088] Implementation Method Eight: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method Two. In this implementation method, during the vegetation type differentiation stage... Based on the scattering feature Sentinel-1 and the spectral feature Sentinel-2, the remote sensing image of the vegetation area is segmented using the Simple Non-Iterative Clustering (SNIC) method to obtain several superpixel objects. Based on the remote sensing observation data corresponding to the aforementioned superpixel objects, the superpixel objects are classified using the random forest method to obtain several superpixel objects whose classification results are initial candidate forest objects.

[0089] In this embodiment, the classification results also include grassland, wetland and farmland.

[0090] In this embodiment, the Simple Non-Iterative Clustering (SNIC) method generates uniform superpixel objects by initializing seed points and calculating the spatial-spectral distance between pixels. The calculation formula is as follows:

[0091] in, These are the differences in the horizontal and vertical coordinates of the space, respectively. For the first Characteristic values ​​of each band, This is the spatial distance weighting coefficient, used to adjust the degree of influence of spatial proximity in distance metrics; This is the spectral distance weighting coefficient, used to adjust the degree of influence of spectral feature differences in distance measurement; The pixels to be classified; The seed point is either the cluster center pixel or a neighboring pixel whose distance to the pixel to be classified is calculated.

[0092] The key parameters of the Simple Non-Iterative Clustering (SNIC) method include the seed point spacing in pixels, the compactness factor, and the connectivity parameter. Specifically, the seed point spacing is set to 10 pixels, the compactness factor is set to a low value (0.1), and the connectivity parameter is set to 8.

[0093] This parameter setting results in an average area of ​​approximately 100-200 pixels for the segmented superpixel objects. This range effectively balances the precision of the segmentation with computational efficiency, ensuring both the precision of the segmentation results and the smoothness of the boundaries.

[0094] In this embodiment, the scattering features used include VV / VH polarization features in Sentinel-1 and B2 (blue light), B3 (green light), B4 (red light), and B8 (near-infrared) bands in Sentinel-2. These features are designed to make full use of information from different bands to enhance the accuracy of segmentation and the ability to capture details.

[0095] In this embodiment, the remote sensing observation data corresponding to the plurality of superpixel objects are statistically analyzed using metrics such as mean, standard deviation, and median to obtain statistical features. These features are used to quantify internal homogeneity, preserve spatial details, and suppress SAR speckle noise and optical spectral variability. Furthermore, morphological indicators such as aspect ratio and area-to-perimeter ratio are combined to capture the spatial structural characteristics of linear forest belts or isolated trees, thereby providing object primitives for subsequent hierarchical classification.

[0096] In this embodiment, the random forest method performs hyperparameter tuning on the input features (such as the number of decision trees, maximum depth, and minimum number of split samples). By combining grid search with ten-fold cross-validation, the optimal combination of parameters is applied to vegetation mask to generate a spatial distribution map of vegetation types, thereby avoiding overfitting.

[0097] In this embodiment, the key hyperparameters of the random forest method are tuned based on the remote sensing observation data, with a focus on parameters that significantly affect the model's fitting ability and generalization performance. These parameters mainly include the number of decision trees, the maximum depth of a single tree, and the minimum number of samples required for a node to continue splitting.

[0098] The tuning was performed using a gridded search approach, which is a systematic hyperparameter optimization strategy. Its core idea is to predefine a set of discrete candidate values ​​for each hyperparameter to be tuned, and then perform Cartesian product combinations on all candidate values ​​to exhaustively generate a complete parameter configuration space. For each parameter combination in the space, its classification performance is evaluated on the training set through cross-validation. Finally, the parameter combination that optimizes the validation evaluation index is selected as the final hyperparameter configuration of the model.

[0099] The grid search range is set to n_estimators: [50, 100, 200, 500], max_depth: [10,20, None], and min_samples_split: [2, 5, 10]. Internal optimization is performed on the training set using 10-fold hierarchical cross-validation. Each fold of validation uses a 70% and 30% training and validation split. The average overall accuracy of cross-validation is used as the primary evaluation metric, and the Kappa coefficient is used as a consistency auxiliary metric to ensure robust generalization of the model in heterogeneous landscapes across regions.

[0100] In this embodiment, classifying the vegetation area can also accurately distinguish typical vegetation types such as grassland, woodland, wetland and farmland.

[0101] This implementation further refines the stage for distinguishing vegetation types, providing an example. The method, based on Sentinel-1 polarimetric and Sentinel-2 multispectral data, employs the Simple Non-Iterative Clustering (SNIC) method for segmentation, achieving object-level classification. A random forest algorithm is then used for object-based vegetation type classification. Differences in annual growth cycle, canopy structure, and spatial heterogeneity of vegetation are quantified from three dimensions: phenological characteristics, canopy structure, and spatial texture, improving classification robustness and reducing interference from cloud cover and surface complexity.

[0102] Random forest, as a typical ensemble learning method, constructs multiple decision trees on random subsamples and random feature subsets and outputs the final class by majority voting. It can effectively handle high-dimensional, multi-source, and noisy remote sensing feature data without making strict assumptions about the data distribution, and has a strong ability to suppress overfitting. Therefore, it has been widely used in multi-source remote sensing land cover classification tasks.

[0103] Implementation Method Nine: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method Eight. In this implementation method, the vegetation type classification model is obtained through the following method: Construct a dataset, which includes the remote sensing observation data of the superpixel object and the corresponding vegetation type; the vegetation type includes initial candidate forest objects and non-initial candidate forest objects. Using the remote sensing observation data as input and the vegetation type as output, a machine learning regression model is trained using a gridded search optimization method, and the obtained machine learning regression model serves as a vegetation type classification model.

[0104] In this embodiment, the machine learning regression model is preferably a random forest model.

[0105] In this embodiment, the vegetation type also includes several typical vegetation types, including grassland, woodland, wetland and farmland.

[0106] In this embodiment, samples are randomly selected and manually labeled (70% for training set and 30% for test set), and a machine learning regression model is used for training.

[0107] Implementation Method 10: This implementation method further defines the farmland tree remote sensing image optimization method described in Implementation Method 1. In this implementation method, during the stage of distinguishing farmland trees from non-farmland trees, the farmland tree probability prediction model is obtained through the following method: Construct a dataset that includes morphological features of forest objects and corresponding binary labels for farmland trees; Using the morphological features as input and the binary labels of farmland trees as output, a multiple linear regression model is trained, and the obtained multiple linear regression model serves as a probability prediction model for farmland trees.

[0108] In this embodiment, representative samples are randomly selected from all candidate forest land objects at a certain proportion, taking into account different spatial locations, different morphological feature ranges, and different landscape backgrounds (such as forest patches near villages, rivers, roads, or far from farmland). With the assistance of high-resolution imagery or field surveys, each sample object is manually labeled as a training sample. That is, objects that are adjacent to farmland, distributed in strips or rows, and clearly belong to types such as farmland shelterbelts, field ridge trees, roadside trees, and canal forests are labeled as "farmland trees" (referred to as 1), while other objects (such as independent forest land, mountain forest, residential shelterbelts, park green space forests, etc.) are labeled as "non-farmland trees" (referred to as 0).

[0109] In this embodiment, the dataset is randomly divided into a training set and a test set, with the training set accounting for 70% and the test set accounting for 30%. The proportion of the two types of samples in training and testing is kept as consistent as possible to avoid the impact of class imbalance on model training and evaluation.

[0110] In this embodiment, a multiple linear regression model capable of outputting continuous probability values ​​is used to learn the mapping relationship between the above-mentioned object-level features and farmland tree attributes. Using features such as SI, EL, and APR as input, the binary label (encoded as 0 and 1) indicating whether a tree belongs to farmland is treated as an approximately continuous target variable for modeling.

[0111] By training the model on the training set, the model learns the contribution patterns of different morphological features and adjacent farmland relationships to whether an object belongs to "farmland trees". Thus, in the prediction stage, it can output a probability value between 0 and 1 for each unlabeled forest object, representing the confidence level that the object belongs to farmland trees.

[0112] This implementation further defines the stage for distinguishing between farmland trees and non-farmland trees, and provides an example of a probabilistic prediction model for farmland trees. By combining manual sample labeling with supervised learning regression modeling, the question of "whether it is a farmland tree" is transformed into a probabilistic prediction problem, thereby improving the flexibility and adjustability of classification.

[0113] Implementation Method Eleven: The method for optimizing remote sensing images of farmland trees used in this implementation method is as follows: Figure 2 As shown, this paper presents an optimization method for farmland tree remote sensing images based on Embodiment 1, combined with optimizations to Embodiments 2 to 10.

[0114] In this embodiment, multiple indicators such as overall accuracy based on the confusion matrix, kappa coefficient, and F1-score are used to evaluate accuracy. Using reserved test sample points, the accuracy of the vegetation / non-vegetation spatial distribution mask and the basic classification maps of the four types of vegetation (grassland, woodland, wetland, and farmland) generated in the vegetation-distinguishing stage and the vegetation-type distinguishing stage are verified to quantify the classification performance of the basic model at the macro level.

[0115] Independent validation was performed on the vegetation type differentiation stage. Evaluation metrics included overall precision of the overall classification performance and the Kappa coefficient considering random consistency. In addition, the F1 score was calculated, which comprehensively reflects the classification performance of each class by integrating precision and recall.

[0116] These metrics work together to form a systematic quantitative evaluation of the classification algorithm's performance, providing a reliable indicator of model reliability. The specific calculation formula is as follows: Overall Accuracy (OA):

[0117] Where N is the total number of categories, represents the elements on the diagonal of the confusion matrix (i.e., the number of samples correctly classified into class i), and N is the total number of validation samples.

[0118] Kappa coefficient:

[0119] in, This is the sum of the i-th row in the confusion matrix (i.e., the number of true samples in the i-th class). It is the sum of the i-th column (i.e., the total number of samples predicted as class i).

[0120] F1 score:

[0121]

[0122]

[0123] Wherein, TP (True Positive): the number of correctly predicted positive samples; FP (False Positive): the number of incorrectly predicted positive samples (the sum of the off-diagonal elements in the column containing the class in the confusion matrix); FN (False Negative): the number of positive samples incorrectly predicted as other classes (the sum of the off-diagonal elements in the row containing the class in the confusion matrix).

[0124] In this embodiment, 22,238 ground verification points were set up in a typical sample area (17,431 of which were vegetation, including farmland / forest / grassland, and 4,807 of which were non-vegetation, including roads / bare soil / buildings), and the farmland trees in the typical sample area were optimized using the remote sensing images of this embodiment.

[0125] In the stage of distinguishing between vegetation and non-vegetation, ROC analysis was used to assess threshold sensitivity. For example... Figure 3 The figure shows the effect of vegetation threshold determination based on ROC curve. The optimal threshold point is 0.194, and the corresponding true positive rate is 0.951, while the false positive rate is 0.076.

[0126] Based on vegetation frequency Binary classification, vegetation frequency Pixels with a VF value below 0.194 are identified as non-vegetation features (such as bare soil, water bodies, impermeable surfaces, and permanent snow and ice) and are masked out, while pixels with a VF value greater than 0.194 are classified as vegetation areas. Based on this, a high-confidence vegetation spatial distribution mask is generated and proceeds to the subsequent fine classification process of vegetation types.

[0127] Differentiating vegetation types: such as Figure 8The image shows the classification results for typical vegetation types. Experimental tests conducted in farmland with extensive intensive cultivation in Northeast China demonstrate that this implementation method achieves an overall classification accuracy of 97.11% for vegetation and non-vegetation, with a Kappa coefficient of 0.954. Specifically, the user accuracy and producer accuracy for the vegetation category are as high as 98.75% and 98.12%, respectively; while for the non-vegetation category, the user accuracy and producer accuracy are 96.87% and 98.11%, respectively. Based on these results, the proposed vegetation classification model was further quantitatively evaluated. The model's overall classification accuracy reached 85.75%, with a Kappa coefficient of 0.82. The model exhibits excellent classification performance across all categories, with producer accuracy (PA) and user accuracy (UA) both exceeding 90%, and F1 scores ranging from 0.82 to 0.87. In particular, although there are slight differences in model performance between different categories, the overall performance is excellent.

[0128] like Figure 9 and Figure 10 As shown, farmland accounts for 60.1%, grassland 22.2%, forest 13.5%, and wetland 4.2%. Farmland performed best (F1=0.87), with both PA and UA approaching 94%, indicating that the model kept both misclassification and missed classification errors at a low level, resulting in highly reliable classification results. Similarly, forest and grassland also performed well (F1=0.84 for both), with similar PA and UA values, indicating that the model maintained balanced control over misclassification and missed classification errors for these two land cover types, demonstrating high classification stability. In contrast, wetland performed the worst among the four vegetation categories (F1=0.82).

[0129] Further analysis of the spatial distribution pattern, such as Figure 11 As shown, Figure 11 The upper part is an image, and the lower part is a classification image. The left side shows that farmland and protective forests in farmland can be distinguished well. The middle image can separate farmland and grassland well, with farmland having the highest classification accuracy. The right side image clearly shows the classification differences between woodland and grassland.

[0130] In the stage of distinguishing between farmland trees and non-farmland trees, the numerical range of morphological characteristics in farmland trees and other forest lands is represented as follows: Figure 4 and Figure 5 As shown, the three spatial semantic features exhibit significant differences in their numerical range representations across farmland trees and other woodlands, thus making them well-suited for the subsequent construction of a multiple linear regression model. In the feature importance assessment of the three semantic features, such as... Figure 6 As shown, further findings Its discrimination is better than and Morphological characteristics include shape index. elongation and the proportion of adjacent farmland .

[0131] The threshold for classifying farmland trees is determined, and the model performance is independently evaluated on the test set. Continuous probability outputs are converted into binary class labels using an F1-score discrimination threshold (0.31), and the overall accuracy is calculated to comprehensively test the model's discriminative ability and stability under different thresholds. Figure 7 As shown, the maximum F1 score can be obtained when the threshold is 0.31, and the corresponding F1 score is 0.784 when the threshold is 0.784. Building upon the stable identification of vegetation types, this implementation further conducted independent accuracy evaluations for farmland trees and other trees to verify the effectiveness of the refined extraction method for farmland trees. Based on the validation sample set and using 0.31 as the threshold for determining farmland trees, the overall accuracy of the binary classification of farmland trees and non-farmland trees reached 0.79, with a Kappa coefficient of 0.73, indicating that this implementation achieved a relatively reliable discrimination effect in the subdivision of tree categories. For the target category of farmland trees, the producer accuracy was 74.68%, and the user accuracy was 78.67%, demonstrating that this implementation can correctly identify most real farmland tree objects.

[0132] like Figure 12 As shown, Figure 12 To show the spatial distribution pattern of the study area, from Figures 13 to 16 It can be seen from the data that the most obvious distinction is made between the protective forests in the farmland ridges (yellow stripes), followed by the farmland trees intertwined with villages and farmland, which also have a good identification effect.

[0133] Based on the above indicators, the multi-stage classification and fine extraction system constructed in this embodiment achieves high-precision identification of vegetation / non-vegetation and major vegetation types. It can effectively separate farmland trees from large-scale heterogeneous landscapes, providing highly reliable basic data support for regional-scale farmland shelterbelt surveys and related ecological effect assessments.

Claims

1. A method for optimizing remote sensing images of farmland trees, characterized in that, Includes the following stages: Data acquisition phase: Collect remote sensing images and corresponding remote sensing observation data of the monitoring area throughout the year; Distinguishing between vegetated and non-vegetated stages: Based on the remote sensing images and the remote sensing observation data, vegetated areas are obtained using the vegetation frequency method; Vegetation type differentiation stage: Based on the remote sensing observation data of the vegetation area, the vegetation area is classified to obtain several initial candidate forest land objects; Distinguishing between farmland trees and non-farmland trees: Based on the aforementioned initial candidate forest land objects and corresponding remote sensing observation data, the morphological characteristics of each initial candidate forest land object are obtained; Based on the morphological characteristics of each initial candidate forest land object, the corresponding probability of farmland trees is obtained through the farmland tree probability prediction model; if the probability of farmland trees of the initial candidate forest land object is greater than the farmland tree determination threshold, it is determined to be farmland trees, otherwise it is determined to be non-farmland trees. Farmland Tree Remote Sensing Image Optimization Stage: Based on all the initial candidate forest land objects identified as farmland trees, the remote sensing image is optimized.

2. The method for optimizing remote sensing images of farmland trees according to claim 1, characterized in that, The remote sensing data includes vertical structure, scattering characteristics Sentinel-1, spectral characteristics Sentinel-2, vegetation index, phenological characteristics, and auxiliary factors; The vertical structure includes a canopy height RH; The spectral features Sentinel-2 include the blue band B2_B, the green band B3_G, the red band B4_R, the red edge band B5_RE1, the red edge band B6_RE2, the red edge band B7_RE3, the near-red band B8_NIR1, the near-red band B8A_NIR2, the short-red band B11_SWIR1, and the short-red band B12_SWIR2. The auxiliary factors include elevation DEM and slope SLOPE.

3. The method for optimizing remote sensing images of farmland trees according to claim 2, characterized in that, The scattering feature Sentinel-1 includes VV polarization, VH polarization, polarization ratio Rt, polarization difference Dt, amplitude Amp, coefficient of variation CV, texture entropy T_E, and texture contrast T_C.

4. The method for optimizing remote sensing images of farmland trees according to claim 2, characterized in that, The vegetation index includes the normalized vegetation index. Enhance vegetation index and red-edge index .

5. The method for optimizing remote sensing images of farmland trees according to claim 2, characterized in that, The phenological characteristics include the maximum value of the vegetation index. Average vegetation index and vegetation index standard deviation .

6. The method for optimizing remote sensing images of farmland trees according to claim 1, characterized in that, The vegetation frequency method for distinguishing between vegetated and non-vegetated stages is as follows: Based on the remote sensing observation data, the normalized vegetation index of each pixel in the remote sensing image is obtained. ; The normalized vegetation index of each pixel throughout the year is calculated. The proportion exceeding the NDVI threshold T is used as the vegetation frequency for the corresponding pixel. ; Each pixel of the remote sensing image is processed as follows to obtain vegetation pixels, and the vegetation pixels are combined into vegetation regions: If the vegetation frequency of the pixel Less than the vegetation frequency threshold If the condition is met, the pixel is determined to be a non-vegetation pixel; otherwise, the pixel is determined to be a vegetation pixel.

7. The method for optimizing remote sensing images of farmland trees according to claim 1, characterized in that, In the stage of distinguishing between farmland trees and non-farmland trees, the morphological characteristics include shape index SI, elongation EL, and adjacent farmland ratio APR.

8. The method for optimizing remote sensing images of farmland trees according to claim 2, characterized in that, In the stage of distinguishing vegetation types, Based on the scattering feature Sentinel-1 and the spectral feature Sentinel-2, the remote sensing image of the vegetation area is segmented using the Simple Non-Iterative Clustering (SNIC) method to obtain several superpixel objects. Based on the remote sensing observation data corresponding to the aforementioned superpixel objects, the superpixel objects are classified using a vegetation type classification model to obtain several superpixel objects whose classification results are initial candidate forest land objects.

9. The method for optimizing remote sensing images of farmland trees according to claim 8, characterized in that, The vegetation type classification model was obtained through the following method: Construct a dataset, which includes the remote sensing observation data of the superpixel object and the corresponding vegetation type; the vegetation type includes initial candidate forest objects and non-initial candidate forest objects. Using the remote sensing observation data as input and the vegetation type as output, a machine learning regression model is trained using a gridded search optimization method, and the obtained machine learning regression model serves as a vegetation type classification model.

10. The method for optimizing remote sensing images of farmland trees according to claim 1, characterized in that, In the stage of distinguishing between farmland trees and non-farmland trees, the probability prediction model for farmland trees is obtained through the following method: Construct a dataset that includes morphological features of forest objects and corresponding binary labels for farmland trees; Using the morphological features as input and the binary labels of farmland trees as output, a multiple linear regression model is trained, and the obtained multiple linear regression model serves as a probability prediction model for farmland trees.

Images