Tea garden recognition method and device based on multi-source remote sensing big data and ensemble learning

Abstract

The application discloses a tea garden recognition method and device based on multi-source remote sensing big data and integrated learning, belongs to the technical field of tea garden image recognition, and solves the low tea garden recognition rate problem in the prior art. The application introduces a first marker, when the imaging time difference between optical and SAR images is too large, the gain step is blocked, and the situation of using heterogeneous big data that may have undergone state variation for filling is eliminated. The tea garden has frequent pruning, picking and other high dynamic change characteristics, and blind interpolation is avoided to introduce serious false features. Under the premise that the first marker indicates consistent time, by setting a determination threshold and a gain coefficient, when one party big data (such as optical) is severely polluted, the other party homogeneous high-quality observation (such as radar) is given an amplification weight. The application fully utilizes the complementary information of multi-modal big data under bad weather without deviating from the real physical observation, and improves the tea garden recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the field of tea garden image recognition technology, and specifically relates to a tea garden recognition method and device based on multi-source remote sensing big data and ensemble learning. Background Technology

[0002] Tea gardens are an important type of economic crop in my country, widely distributed in complex terrain areas such as hills and mountains. Accurate monitoring of the spatial distribution of tea gardens is of great significance for agricultural resource surveys and industrial planning. With the increasing maturity of collaborative observation systems using high spatial, temporal, and spectral resolution satellite constellations, tea garden monitoring has shifted from traditional "single-phase, small-area" interpretation to a dynamic mining paradigm based on remote sensing big data, encompassing the entire lifecycle and wide area. Remote sensing big data is characterized by its massive volume, diverse types, dynamic changes, varying quality (uncertainty), and high intrinsic value; it is a typical example of big data composed of massive, multi-source, and heterogeneous satellite observation datasets.

[0003] Currently, using remote sensing big data for surface vegetation identification has become an important research area. However, existing technologies still have the following shortcomings when processing tea garden identification: First, the mining of multi-source heterogeneous big data is insufficient. Existing studies are mostly based on single or a few optical satellite remote sensing images, using spectral bands, vegetation indices, and a small number of texture indicators for traditional monitoring. However, since tea gardens are mostly distributed in mountainous and hilly areas, they have a strong similarity to forests and orchards in spectral response, resulting in redundant ambiguity of "different objects with the same spectrum" in big data. In addition, mountain shadows and cloudy and foggy weather lead to a large amount of noise and data gaps in optical remote sensing big data, and relying solely on single-source optical data can easily lead to serious misjudgments. Second, the big data fusion mechanism lacks physical and logical constraints. Although some studies have attempted to introduce synthetic aperture radar (SAR) data to leverage its all-weather imaging advantages, when processing massive, multi-source, and non-stationary remote sensing big data fusion, existing technologies often rely only on simple mathematical interpolation or one-way numerical replacement to compensate for contaminated pixels. This approach ignores the spatiotemporal phenological consistency between heterogeneous data, which can easily lead to inconsistencies and accumulated biases in cross-source features. As a result, when processing remote sensing big data with complex dynamic features, the efficiency of feature-level collaboration is low, making it difficult to accurately depict the structural features of tea gardens. Ultimately, the accuracy and timeliness of identification cannot meet the needs of large-scale refined management.

[0004] Therefore, how to construct a remote sensing big data ensemble learning model that can efficiently process massive amounts of multi-source heterogeneous data and has strong robustness has become a key technical problem that urgently needs to be solved in the field of high-precision identification of tea gardens. Summary of the Invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a tea garden identification method and apparatus based on multi-source remote sensing big data and ensemble learning, so as to solve the problem of low identification accuracy in the prior art.

[0006] The objective of this invention is achieved as follows: On the one hand, a tea garden identification method based on multi-source remote sensing big data and ensemble learning is provided, including: Acquire multi-source remote sensing big data of the target study area; among which, multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; Based on the imaging date and preset time window, optical remote sensing data and SAR remote sensing data are aggregated to obtain multiple time window sets; The optical remote sensing data and SAR remote sensing data with the smallest date difference in any time window set are used as the main paired image of the time window set; If the date difference between the primary paired images is greater than a preset threshold, the primary paired image is marked with the first mark. Pixel screening is performed on optical remote sensing data and SAR remote sensing data. Based on the screening results, a second label is generated for the optical remote sensing data and a third label is generated for the SAR remote sensing data. Based on the first, second, and third labels, the main paired images are fused and compensated to obtain a fused feature vector and a quality label layer. Based on the fused feature vector, multi-temporal features at the pixel level are extracted; among them, multi-temporal features include spectral features, SAR backscattering features and texture features; Based on multi-temporal features and quality label layers, the classification model of the ensemble learning architecture is trained and predicted to obtain the classification results of the target study area; Based on the classification results, a spatial distribution map of tea gardens is output.

[0007] A preferred embodiment of the present invention includes a step of performing fusion compensation on master paired images based on a first marker, a second marker, and a third marker, comprising: According to the preset mapping rules, the second marker is mapped to an optical confidence level with a value between 0 and 1, and the third marker is mapped to a SAR confidence level with a value between 0 and 1. Extract raw optical features and raw SAR features from the master paired image; The original optical features are multiplied by the corresponding optical confidence level to obtain the modulated optical features, and the original SAR features are multiplied by the corresponding SAR confidence level to obtain the modulated SAR features. Set the gain coefficient; When the primary paired image does not have a first marker, if the corresponding optical confidence is lower than the decision threshold and the corresponding SAR confidence is higher than the decision threshold, the gain SAR feature is obtained by multiplying the gain coefficient with the modulated SAR feature; if the corresponding SAR confidence is lower than the decision threshold and the corresponding optical confidence is higher than the decision threshold, the gain optical feature is obtained by multiplying the gain coefficient with the modulated optical feature. Based on one or more of the modulated optical features, gain SAR features, gain optical features and modulated SAR features, feature stitching is performed to obtain a fused feature vector.

[0008] A preferred embodiment of the present invention further includes: When optical remote sensing data for a time window set is unavailable, the time window set is marked with a fourth label. When SAR remote sensing data in a time window set is unavailable, the time window set is marked with the fifth tag. Generate a quality marker layer based on the first, second, third, fourth, and fifth markers.

[0009] A preferred embodiment of the present invention includes the following steps for training and predicting a classification model based on an ensemble learning architecture, using multi-temporal features and a quality label layer: Construct and train multiple base learners; Multi-temporal features are input into each base learner for training to obtain multiple tea garden classification prediction probabilities; A meta-learner is constructed by inputting terrain factors, tea garden classification prediction probabilities, and quality label layers into the meta-learner to obtain classification results.

[0010] A preferred embodiment of the present invention further includes: Based on the tea garden classification prediction probability output by each base learner, the uncertainty index of the pixel is calculated; wherein, the uncertainty index can be characterized by one or more of the following: the variance of the base learner probability, the information entropy of the ensemble output, or the voting inconsistency rate. When the uncertainty index exceeds the set threshold, the corresponding cell is marked as a high-risk area, and the confidence backoff strategy is triggered.

[0011] In a preferred embodiment of the present invention, the confidence backoff strategy includes at least one of the following: When the label is characterized as low quality from any single data source, the weight of the base learner corresponding to that single data source is reduced. Robust model backoff: The classification results of high-risk regions are backed up to a more robust base learner; The consensus voting results of multiple base learners are used to replace the output of the meta learner.

[0012] A preferred embodiment of the present invention includes a step of pixel sorting of optical remote sensing data and SAR remote sensing data, comprising: Perform cloud cover and terrain shadow analysis on optical remote sensing data; Noise anomalies were investigated in SAR remote sensing data.

[0013] In a preferred embodiment of the present invention, the step of outputting a spatial distribution map of tea gardens based on classification results includes: Connectivity analysis was performed on the classification results to generate suspicious tea garden patches; Extract the slope or aspect values ​​of pixels in each suspicious tea garden patch, and calculate the topographic consistency index of the suspicious tea garden patch based on the slope or aspect values. If the terrain consistency index does not reach the preset verification threshold, the suspicious tea garden patches will be removed from the classification results to obtain the current classification result. The current classification results are vectorized to output a spatial distribution map of the tea gardens.

[0014] On the other hand, a tea garden identification device based on multi-source remote sensing big data and ensemble learning is also provided, including: The acquisition module is used to acquire multi-source remote sensing big data of the target study area; among which, multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; The aggregation module is used to aggregate optical remote sensing data and SAR remote sensing data according to the imaging date and preset time window to obtain multiple time window sets; The pairing module is used to select the optical remote sensing data and SAR remote sensing data with the smallest date difference in any time window set as the main paired image of the time window set. The first marking module is used to mark the main paired image with a first mark when the date difference between the main paired images is greater than a preset threshold. The second labeling module is used to perform pixel screening on optical remote sensing data and SAR remote sensing data, and generate a second label for optical remote sensing data and a third label for SAR remote sensing data based on the screening results. The fusion compensation module is used to perform fusion compensation on the main paired images based on the first, second, and third labels to obtain the fused feature vector and quality label layer; The extraction module is used to extract multi-temporal features at the pixel level based on the fused feature vector; among which, multi-temporal features include spectral features, SAR backscattering features and texture features; The classification module is used to train and predict the classification model of the ensemble learning architecture based on multi-temporal features and quality label layers to obtain the classification results of the target study area. The image generation module is used to output a spatial distribution map of tea gardens based on the classification results.

[0015] A computer-readable storage medium storing a computer program, wherein the computer program is configured to perform the steps of the above-described method at runtime.

[0016] Compared with existing technologies, the tea garden identification method and device based on multi-source remote sensing big data and ensemble learning provided by this invention introduces a first marker. When the imaging time difference between optical and SAR images is too large, the gain step is blocked, preventing the use of heterogeneous big data that may have undergone state changes for filling in the gaps. Tea gardens have high dynamic change characteristics such as frequent pruning and harvesting, avoiding the serious false features introduced by blind interpolation. Under the premise that the first marker indicates consistent time, by setting a judgment threshold and gain coefficient, when one big data (such as optical) is severely contaminated, the other high-quality observation (such as radar) from the same source is given a magnification weight. Without deviating from real physical observation, it makes full use of the complementary information of multimodal big data under adverse weather conditions, ensuring that the model can still extract effective spatial structure features in cloudy, foggy, and highly heterogeneous mountainous areas, thus improving the identification accuracy.

[0017] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this specification or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings. Figure 1 This is a schematic flowchart illustrating the tea garden identification method based on multi-source remote sensing big data and ensemble learning provided by the present invention. Figure 2 A schematic flowchart illustrating the steps of fusing and compensating the main paired image based on the first marker, the second marker, and the third marker provided by the present invention; Figure 3 This is a schematic flowchart illustrating the steps for training and predicting a classification model based on multi-temporal features and quality label layers, as provided by the present invention. Figure 4This is a schematic flowchart illustrating the steps of outputting a spatial distribution map of tea gardens based on classification results, as provided by the present invention. Detailed Implementation

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

[0020] To facilitate understanding of the embodiments of this application, further explanation and description will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of this application. In the drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process sequence may be performed in a different order than that described. For example, two consecutively described processes may be performed substantially simultaneously or in the reverse order of their description. Furthermore, the same reference numerals denote the same components.

[0021] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms “a” and “the” are intended to include the plural forms as well. Furthermore, when the terms “comprising” and / or “including” and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms “substantially,” “about,” and other similar terms are used as approximate terms rather than as terms of degree, thus explaining the inherent biases in measurements, calculated values, and / or provided values ​​that would be recognized by one of ordinary skill in the art.

[0022] In one embodiment, such as Figure 1 As shown, a tea garden identification method based on multi-source remote sensing big data and ensemble learning is provided, including: S110, acquire multi-source remote sensing big data of the target study area; among which, multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; The target study area can be an administrative region, a natural geographic unit, or a user-defined area, with vector boundaries or raster ranges serving as unified spatial constraints. Optical remote sensing data is used to characterize the spectral reflectance characteristics of ground features, vegetation growth status, and chlorophyll / water related information. Optical remote sensing data includes, but is not limited to, Sentinel-2 multispectral data and Landsat series data. Sentinel-2 multispectral data preferably uses atmospherically corrected surface reflectance products (e.g., Level-2A or equivalent products). Landsat series data includes, but is not limited to, Landsat-5 / 7 / 8 / 9 surface reflectance products, used to supplement long-term series or provide alternative observations when Sentinel-2 data is missing. SAR (Synthetic Aperture Radar) remote sensing data is used to characterize ground feature structure scattering, dielectric properties, and canopy geometric differences. SAR data can be Sentinel-1 SAR data, and ground distance-corrected products or equivalent products can be selected.

[0023] Specifically, this can be achieved by accessing a satellite database.

[0024] S120: Based on the imaging date and preset time window, optical remote sensing data and SAR remote sensing data are aggregated to obtain multiple time window sets; The preset time window is a time span artificially defined based on the phenological period of tea trees (such as the spring tea sprouting period, summer tea growing period, autumn tea picking period, and winter dormancy period) or a fixed calendar cycle (such as every 15 days or every natural month).

[0025] Specifically, based on the phenological characteristics of tea trees, the entire year is divided into multiple preset time windows (e.g., 15 or 30 days per window). All optical remote sensing data and SAR remote sensing data whose imaging dates fall within the same window are aggregated into a single set.

[0026] S130, take the optical remote sensing data and SAR remote sensing data with the smallest date difference in any time window set as the main paired image of the time window set; Specifically, all optical and SAR data combinations within a given time window are iterated, and the absolute difference between their imaging dates is calculated. The combination with the smallest difference is selected as the primary paired image for that window to ensure that the physiological state of tea trees and the physical structure of the land surface observed by both sensors are consistent. When multiple candidate images exist, the one with higher quality is selected first (e.g., optical images with a higher effective pixel ratio behind cloud cover, or SAR images with lower geometric risk).

[0027] S140, if the date difference between the primary paired images is greater than a preset threshold, mark the primary paired image with the first mark; Specifically, a preset threshold is set (such as 3, 5, or 7 days). If the difference between the imaging dates of optical and SAR images in the primary pair is greater than this threshold, it indicates that the tea garden is at risk of being manually harvested, pruned, or experiencing sudden changes in physical condition due to rainfall during this period. In this case, a first label is assigned to the primary pair image.

[0028] S150 performs pixel screening on optical remote sensing data and SAR remote sensing data, and generates a second label for optical remote sensing data and a third label for SAR remote sensing data based on the screening results. Specifically, the screening of optical remote sensing data can involve combining the QA cloud mask band of optical imagery to identify clouds and cloud shadows, while simultaneously incorporating solar altitude angle, azimuth angle, and DEM (Digital Elevation Model) data to calculate mountain terrain shadows. Areas with clouds, cloud shadows, and mountain terrain shadows are marked as contaminated pixels, generating corresponding secondary labels.

[0029] The investigation of SAR remote sensing data can combine satellite orbital incident angle and DEM data to identify distorted pixels with abnormal echoes, edge noise, overlay and geometric shadows, and generate corresponding third labels.

[0030] S160, based on the first, second, and third markers, perform fusion compensation on the main paired images to obtain fused feature vectors and quality marker layers; Specifically, the first, second, and third labels can be mapped as weighting factors to perform quality weighting on optical and SAR features before concatenation, resulting in a fused feature vector. This automatically reduces the contribution of low-quality data sources. The first, second, and third labels are packaged to output a quality label layer. The fusion compensation includes at least confidence mapping and feature modulation, and optionally includes gain processing; when the first label is present, the gain processing is not performed.

[0031] Furthermore, such as Figure 2 As shown, the steps for fusion compensation of the master paired image based on the first marker, the second marker, and the third marker include: S210, according to the preset mapping rules, the second marker is mapped to an optical confidence level with a value between 0 and 1, and the third marker is mapped to a SAR confidence level with a value between 0 and 1; Specifically, a nonlinear mapping function is constructed to convert the second label into an optical confidence level of [0, 1] (e.g., 1 for no clouds, 0 for thick clouds, and 0.5 for thin clouds). This can be based on the cloud / shadow probability values ​​output by the cloud detection algorithm. An upper and lower threshold are set. When the cloud / shadow probability value is greater than the upper threshold, the optical confidence level is 0; when it is less than the lower threshold, the optical confidence level is 1. If the cloud / shadow probability value is between the upper and lower thresholds, the confidence level can be calculated using the exponential decay formula, or other methods can be employed, as long as the confidence level rapidly and smoothly drops towards 0 as the cloud probability increases.

[0032] The method for converting the third marker to a SAR confidence level of [0, 1] can be similar to the optical confidence level settings described above. Specifically, the SAR confidence level is obtained based on the distortion mask value and the local incident angle. When the distortion mask value is 1, it indicates that the pixel is in a radar blind zone or a severely folded area, and the SAR confidence level is 0. When the distortion mask value is 0, the confidence level is calculated based on the degree of deviation of the local incident angle. SAR imaging has an optimal observation angle, typically between 35° and 40°. The greater the deviation of the local incident angle from the optimal observation angle, the closer the confidence level is to 0.

[0033] S220, extract the raw optical features and raw SAR features from the master paired image; S230, the original optical features are multiplied by the corresponding optical confidence level to obtain the modulated optical features, and the original SAR features are multiplied by the corresponding SAR confidence level to obtain the modulated SAR features; Specifically, soft attenuation of contamination is achieved by multiplying the original optical features and original SAR features by the corresponding confidence levels.

[0034] S240, set the gain factor; S250, when the primary paired image does not have a first marker, when the corresponding optical confidence is lower than the decision threshold and the corresponding SAR confidence is higher than the decision threshold, the gain SAR feature is obtained by multiplying the gain coefficient with the modulated SAR feature; when the corresponding SAR confidence is lower than the decision threshold and the corresponding optical confidence is higher than the decision threshold, the gain optical feature is obtained by multiplying the gain coefficient with the modulated optical feature. Specifically, the gain coefficient can be 1.2, 1.5, etc. Thresholds include low and high thresholds, with the low threshold being 0.2-0.4 and the high threshold being 0.6-0.8. For example, if the first marker is not triggered, and the optical confidence is <0.3 and the SAR confidence is >0.7, it indicates optical contamination but radar reliability, and SAR gain is applied; conversely, if the SAR confidence is <0.3 and the optical confidence is >0.7, optical gain is applied.

[0035] S260 uses one or more of modulation optical features, gain SAR features, and modulation SAR features to stitch together features to obtain a fused feature vector.

[0036] Specifically, if the optical features have undergone the gain process (i.e., step S250), then the gain optical features are used for feature stitching; if the gain process has not been performed, then the modulation optical features are used for feature stitching. The same applies to SAR features.

[0037] S170 extracts multi-temporal features at the pixel level based on fused feature vectors; among which, multi-temporal features include spectral features, SAR backscattering features and texture features; Among them, spectral features can be obtained by extracting the reflectance of each band of optical image and calculating the normalized vegetation index, enhanced vegetation index and surface water index.

[0038] Specifically, spectral features, SAR backscattering features, and texture features are extracted from the repaired fused feature vector. Topographic factors can be calculated based on a digital elevation model.

[0039] S180, based on multi-temporal features and quality label layers, trains and predicts the classification model of the ensemble learning architecture to obtain the classification results of the target study area; The classification model can be XGBoost, LightGBM, or Random Forest. The locations of tea garden and non-tea garden samples with known labels are used as indices, and the above multi-temporal features and quality label layers are input into the classification model.

[0040] S190, based on the classification results, output a spatial distribution map of tea gardens.

[0041] Specifically, by post-processing the classification results, a spatial distribution map of the tea gardens can be obtained.

[0042] The aforementioned tea garden identification method based on multi-source remote sensing big data and ensemble learning introduces a first marker. When the imaging time difference between optical and SAR images is too large, the gain step is blocked, preventing the use of heterogeneous data that may have undergone state changes for filling in the gaps. Tea gardens have highly dynamic characteristics such as frequent pruning and harvesting, avoiding the introduction of serious false features by blind interpolation. Under the premise that the first marker indicates consistent time, by setting a judgment threshold and gain coefficient, when one type of data (such as optical) is severely contaminated, the other type of high-quality observation (such as radar) is given a magnification weight. Without deviating from real physical observations, it makes full use of the complementary information of multimodal data under adverse weather conditions, ensuring that the model can still extract effective spatial structure features in cloudy, foggy, and highly heterogeneous mountainous areas, thus improving the identification accuracy.

[0043] In one embodiment, it further includes: When optical remote sensing data for a time window set is unavailable, the time window set is marked with a fourth label. When SAR remote sensing data in a time window set is unavailable, the time window set is marked with the fifth tag. Generate a quality marker layer based on the first, second, third, fourth, and fifth markers.

[0044] Specifically, when only optical or only SAR (other source missing) exists within the time window set, the window is marked with the fourth and fifth marks respectively.

[0045] In one embodiment, such as Figure 3 As shown, the steps for training and predicting a classification model based on multi-temporal features and a quality label layer include: S310, build and train multiple base learners; Specifically, models with complementary learning mechanisms are selected as base learners, including but not limited to: random forest, support vector machine, gradient boosting tree, XGBoost, LightGBM, etc.

[0046] Parameter spaces are set for different base learners and training / validation selections are made. Parameters may include, but are not limited to: tree depth, number of leaf nodes, learning rate, subsampling ratio, feature subsampling ratio, regularization coefficient, and kernel function parameters. The parameters may be determined by validation set performance or cross-validation.

[0047] S320 inputs multi-temporal features into each base learner for training, and obtains multiple tea garden classification prediction probabilities; Specifically, after each base learner completes training, it outputs the corresponding tea garden classification prediction probability for each pixel / object within the study area. Furthermore, to support subsequent stacked ensemble learning and error diagnosis, at least one of the following intermediate results can be saved: a) the prediction probabilities generated by each base learner during training / validation (used to construct OOF probabilities and train meta-learners); b) the feature importance or contribution information of each base learner (used for subsequent feature selection and error analysis).

[0048] S330, construct a meta-learner by inputting terrain factors, tea garden classification prediction probabilities, and quality label layers into the meta-learner to obtain classification results.

[0049] Specifically, the classification probabilities output by each base learner are used as input features of the meta-learner to construct a stacked ensemble model; the meta-learner can be logistic regression, linear SVM, Naive Bayes, or other lightweight models.

[0050] In one embodiment, it also includes: Based on the tea garden classification prediction probability output by each base learner, the uncertainty index of the pixel is calculated; wherein, the uncertainty index can be characterized by one or more of the following: the variance of the base learner probability, the information entropy of the ensemble output, or the voting inconsistency rate. When the uncertainty index exceeds the set threshold, the corresponding cell is marked as a high-risk area, and the confidence backoff strategy is triggered.

[0051] Confidence backoff strategies include at least one of the following: When the label is characterized as low quality from any single data source, the weight of the base learner corresponding to that single data source is reduced. Robust model backoff: The classification results of high-risk regions are backed up to a more robust base learner; The consensus voting results of multiple base learners are used to replace the output of the meta learner.

[0052] Among these, uncertainty is a feature that measures how confident the model is in its predictions. Variance refers to the dispersion of the predicted probabilities among multiple base learners. High variance indicates a lack of consensus among the models. Information entropy is the degree of disorder calculated based on the probability distribution. Higher entropy indicates higher uncertainty. Voting inconsistency rate is the proportion of inconsistent voting results among the base learners.

[0053] For example, the first strategy: If the optical data indicates clouds, and model A (pure optical) predicts the pixel as a tea garden, but model B (pure SAR) predicts it as not a tea garden, in the ensemble voting, the weight of model A is multiplied by 0.1, and the weight of model B is increased. The second strategy: Pre-evaluate the stability of each base learner on the validation set. Select the model with the strongest stability as the more robust base learner. The third strategy: Ignore the complex output of the meta-learner and directly use the voting results of the K base learners as the final classification result. That is: if more than 50% of the base learners identify it as a tea garden, it is classified as a tea garden.

[0054] In one embodiment, the step of pixel sorting of optical remote sensing data and SAR remote sensing data includes: Perform cloud cover and terrain shadow analysis on optical remote sensing data; Specifically, cloud cover analysis involves generating cloud probability maps using the Sentinel-2 QA60 band or the S2cloudless algorithm. Shading maps are generated using the Hillshade function in ArcGIS or GDAL.

[0055] Noise anomalies were investigated in SAR remote sensing data.

[0056] Specifically, the mean and standard deviation of the backscattering coefficients of a local window (e.g., 5x5) are calculated, and the mean and standard deviation are used to determine whether there is noise anomaly. Judgment can also be made based on anomalous echo masks, edge noise masks, geometric shadow / overlap risk masks, and noise intensity.

[0057] In one embodiment, such as Figure 4 As shown, the steps for outputting a spatial distribution map of tea gardens based on the classification results include: S410, perform connected component analysis on the classification results to generate suspicious tea garden patches; Specifically, the classification result is a preliminary binary image of tea garden classification (0 for background, 1 for tea garden). All independent tea garden patches identified are considered suspicious tea garden patches. The identification method can be implemented using any algorithm in this field.

[0058] S420: Extract the slope or aspect values ​​of pixels in each suspicious tea garden patch, and calculate the topographic consistency index of the suspicious tea garden patch based on the slope or aspect values. Specifically, slope or aspect values ​​can be calculated using digital elevation model data. Topographic consistency indices include slope consistency and aspect consistency.

[0059] Slope consistency can be measured by the standard deviation of slope within a patch. Aspect consistency can be measured by converting the aspect to sin / cos components and then calculating the dispersion, or by classifying the aspect and then calculating the proportion of the principal direction.

[0060] S430, if the terrain consistency index does not reach the preset verification threshold, the suspicious tea garden patches will be removed from the classification results to obtain the current classification result; The conditions under which the terrain consistency index fails to reach the preset verification threshold are: the standard deviation of slope within a patch exceeds a certain threshold, or the proportion of pixels whose slope falls within the main range is lower than the first preset verification threshold; or the proportion of the main direction calculated after dispersion or slope aspect classification is less than the second preset verification threshold.

[0061] If a suspected tea plantation patch is misidentified as a tea plantation by the model, it is removed.

[0062] S440: Vectorize the current classification results and output a spatial distribution map of the tea gardens.

[0063] Specifically, GIS software was used to convert the removed binary map into a spatial distribution map of the tea gardens.

[0064] In one embodiment, a tea garden identification device based on multi-source remote sensing big data and ensemble learning is also provided, comprising: The acquisition module is used to acquire multi-source remote sensing big data of the target study area; among which, multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; The aggregation module is used to aggregate optical remote sensing data and SAR remote sensing data according to the imaging date and preset time window to obtain multiple time window sets; The pairing module is used to select the optical remote sensing data and SAR remote sensing data with the smallest date difference in any time window set as the main paired image of the time window set. The first marking module is used to mark the main paired image with a first mark when the date difference between the main paired images is greater than a preset threshold. The second labeling module is used to perform pixel screening on optical remote sensing data and SAR remote sensing data, and generate a second label for optical remote sensing data and a third label for SAR remote sensing data based on the screening results. The fusion compensation module is used to perform fusion compensation on the main paired images based on the first, second, and third labels to obtain the fused feature vector and quality label layer; The extraction module is used to extract multi-temporal features at the pixel level based on the fused feature vector; among which, multi-temporal features include spectral features, SAR backscattering features and texture features; The classification module is used to train and predict the classification model of the ensemble learning architecture based on multi-temporal features and quality label layers to obtain the classification results of the target study area. The image generation module is used to output a spatial distribution map of tea gardens based on the classification results.

[0065] Specific limitations regarding the tea garden identification device based on multi-source remote sensing big data and ensemble learning can be found in the limitations of the tea garden identification method based on multi-source remote sensing big data and ensemble learning mentioned above, and will not be repeated here. Each module in the aforementioned tea garden identification device based on multi-source remote sensing big data and ensemble learning can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module. It should be noted that the module division in this embodiment is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0066] In one embodiment, a computer-readable storage medium is also provided, on which a computer program is stored, which, when executed by a processor, performs the following steps: Acquire multi-source remote sensing big data of the target study area; among which, multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; Based on the imaging date and preset time window, optical remote sensing data and SAR remote sensing data are aggregated to obtain multiple time window sets; The optical remote sensing data and SAR remote sensing data with the smallest date difference in any time window set are used as the main paired image of the time window set; If the date difference between the primary paired images is greater than a preset threshold, the primary paired image is marked with the first mark. Pixel screening is performed on optical remote sensing data and SAR remote sensing data. Based on the screening results, a second label is generated for the optical remote sensing data and a third label is generated for the SAR remote sensing data. Based on the first, second, and third labels, the main paired images are fused and compensated to obtain a fused feature vector and a quality label layer. Based on the fused feature vector, multi-temporal features at the pixel level are extracted; among them, multi-temporal features include spectral features, SAR backscattering features and texture features; Based on multi-temporal features and quality label layers, the classification model of the ensemble learning architecture is trained and predicted to obtain the classification results of the target study area; Based on the classification results, a spatial distribution map of tea gardens is output.

[0067] It is understood that the embodiments described herein can be implemented in hardware, software, firmware, middleware, microcode, or a combination thereof. For hardware implementation, the processing unit can be implemented in one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), general-purpose processors, controllers, microcontrollers, microprocessors, other electronic units for performing the functions described herein, or combinations thereof.

[0068] For software implementation, the techniques described herein can be implemented by units that perform the functions described herein. The software code can be stored in memory and executed by a processor. The memory can be implemented in the processor or external to the processor.

[0069] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0070] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0071] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0072] If the aforementioned function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of this application, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks. It should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0073] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A tea garden identification method based on multi-source remote sensing big data and ensemble learning, characterized in that, include: Acquire multi-source remote sensing big data of the target study area; wherein, the multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; Based on the imaging date and a preset time window, the optical remote sensing data and the SAR remote sensing data are aggregated to obtain multiple time window sets; The optical remote sensing data and SAR remote sensing data with the smallest date difference in any of the time window sets shall be used as the main paired image of the time window set; If the date difference between the main paired images is greater than a preset threshold, the main paired images are marked with a first mark. Pixel screening is performed on the optical remote sensing data and the SAR remote sensing data, and a second label is generated for the optical remote sensing data and a third label is generated for the SAR remote sensing data based on the screening results; Based on the first marker, the second marker, and the third marker, the main paired image is fused and compensated to obtain a fused feature vector and a quality marker layer; Based on the fused feature vector, pixel-level multi-temporal features are extracted; wherein, the multi-temporal features include spectral features, SAR backscattering features, and texture features; Based on the multi-temporal features and the quality label layer, the classification model of the ensemble learning architecture is trained and predicted to obtain the classification result of the target study area; Based on the classification results, a spatial distribution map of tea gardens is output.

2. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 1, characterized in that, The step of performing fusion compensation on the primary paired image based on the first marker, the second marker, and the third marker includes: According to the preset mapping rules, the second marker is mapped to an optical confidence level with a value between 0 and 1, and the third marker is mapped to a SAR confidence level with a value between 0 and 1. Extract the raw optical features and raw SAR features from the main paired image; The original optical features are multiplied by the corresponding optical confidence scores to obtain the modulated optical features, and the original SAR features are multiplied by the corresponding SAR confidence scores to obtain the modulated SAR features. Set the gain coefficient; When the first marker is not present in the primary paired image, if the corresponding optical confidence is lower than the determination threshold and the corresponding SAR confidence is higher than the determination threshold, the gain SAR feature is obtained by multiplying the gain coefficient with the modulation SAR feature; if the corresponding SAR confidence is lower than the determination threshold and the corresponding optical confidence is higher than the determination threshold, the gain optical feature is obtained by multiplying the gain coefficient with the modulation optical feature. The fused feature vector is obtained by concatenating features based on one or more of the modulation optical features, the gain SAR features, the gain optical features, and the modulation SAR features.

3. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 1, characterized in that, Also includes: When the optical remote sensing data of the time window set is unavailable, the time window set is marked with a fourth tag. When the SAR remote sensing data in the time window set is unavailable, the time window set is marked with a fifth tag. The quality marker layer is generated based on the first marker, the second marker, the third marker, the fourth marker, and the fifth marker.

4. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 1, characterized in that, The steps for training and predicting the classification model of the ensemble learning architecture based on the multi-temporal features and the quality label layer include: Construct and train multiple base learners; The multi-temporal features are input into each of the base learners for training to obtain multiple tea garden classification prediction probabilities. A meta-learner is constructed by inputting the terrain factors, the tea garden classification prediction probability, and the quality label layer into the meta-learner to obtain the classification result.

5. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 4, characterized in that, Also includes: Based on the tea garden classification prediction probability output by each of the base learners, the uncertainty index of the pixel is calculated; wherein, the uncertainty index can be characterized by one or more of the following: the variance of the base learner probability, the information entropy of the ensemble output, or the voting inconsistency rate. When the uncertainty index exceeds the set threshold, the corresponding cell is marked as a high-risk area, and a confidence fallback strategy is triggered.

6. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 5, characterized in that, Confidence backoff strategies include at least one of the following: When a label indicates that a single data source is of low quality, the weight of the base learner corresponding to that single data source is reduced. Robust model backoff: The classification results of high-risk regions are backed up to a more robust base learner; The output of the meta-learner is replaced by the consensus voting result of multiple base learners.

7. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 1, characterized in that, The steps for pixel sorting of the optical remote sensing data and the SAR remote sensing data include: The optical remote sensing data was used for cloud cover and terrain shadow screening. The SAR remote sensing data was subjected to noise anomaly investigation.

8. The tea garden identification method based on multi-source remote sensing big data and ensemble learning according to claim 1, characterized in that, Based on the classification results, the steps for outputting a spatial distribution map of tea gardens include: Connectivity analysis was performed on the classification results to generate suspicious tea garden patches; Extract the slope or aspect values ​​of the pixels in each of the suspected tea garden patches, and calculate the topographic consistency index of the suspected tea garden patches based on the slope or aspect values; If the terrain consistency index does not reach the preset verification threshold, the suspicious tea garden patch will be removed from the classification result to obtain the current classification result. The current classification results are vectorized to output a spatial distribution map of the tea gardens.

9. A tea garden identification device based on multi-source remote sensing big data and ensemble learning, characterized in that, include: The acquisition module is used to acquire multi-source remote sensing big data of the target study area; wherein, the multi-source remote sensing big data includes optical remote sensing data and SAR remote sensing data; The aggregation module is used to aggregate the optical remote sensing data and the SAR remote sensing data according to the imaging date and a preset time window to obtain multiple time window sets; The pairing module is used to select the optical remote sensing data and SAR remote sensing data with the smallest date difference in any of the time window sets as the main paired image of the time window set; The first marking module is used to mark the main paired image with a first mark when the date difference between the main paired images is greater than a preset threshold. The second labeling module is used to perform pixel screening on the optical remote sensing data and the SAR remote sensing data, and generate a second label for the optical remote sensing data and a third label for the SAR remote sensing data based on the screening results. The fusion compensation module is used to perform fusion compensation on the main paired image based on the first marker, the second marker, and the third marker to obtain a fused feature vector and a quality marker layer; An extraction module is used to extract multi-temporal features at the pixel level based on the fused feature vector; wherein, the multi-temporal features include spectral features, SAR backscattering features, and texture features; The classification module is used to train and predict the classification model of the ensemble learning architecture based on the multi-temporal features and the quality label layer, so as to obtain the classification result of the target study area; The image generation module is used to output a spatial distribution map of tea gardens based on the classification results.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein the computer program is configured to perform the steps of the method described in any one of claims 1 to 8 when executed.

Images