Invasive plant detection method based on fusion of satellite remote sensing image and ground observation sample

Abstract

The invention provides an invasive plant detection method based on fusion of a satellite remote sensing image and a ground observation sample, and relates to the technical field of ecological monitoring. According to the method, a technical route of'satellite remote sensing image-multi-screening feature-model training and detection 'is constructed under an extremely simple data framework of free multispectral satellite image + open biodiversity platform, and invasive plant mapping is completed with extremely low cost. According to the method, the data recorded in the open biodiversity platform is fully utilized, the labor cost required for acquiring the ground data is greatly reduced, and the data is provided with the longitude and latitude and the species label, so that the expensive data purchasing and manual drawing process is effectively avoided, and the low-cost, high-efficiency and replicable detection process is realized. According to the method, only the preset number of core features are selected for model training, the input dimension is controlled, and a model classification task is more stable when facing a small sample task.

Description

TECHNICAL FIELD

[0001] The embodiment of the application relates to the technical field of ecological monitoring, and particularly relates to an invasive plant detection method based on satellite remote sensing image and ground observation sample fusion. BACKGROUND

[0002] Satellite remote sensing image is a core tool for human beings to monitor the changes on the earth's surface. From resource investigation to ecological assessment, from disaster early warning to precision agriculture, remote sensing technology has become the infrastructure for global resource management and environmental governance with the advantages of macro coverage, dynamic update and objective quantification. However, due to the influence of clouds, spectral confusion and insufficient samples, the existing methods still have obvious performance bottlenecks in the fine monitoring of invasive plants.

[0003] Specifically in the field of ecological monitoring, remote sensing images mainly undertake the following tasks: Vegetation classification and mapping: distinguish different vegetation types, monitor forest cutting, grassland degradation, etc.

[0004] Invasive species monitoring: identify the spread range of alien invasive plants and assess the threat to the local ecology.

[0005] Growth state evaluation: inversion of plant health, biomass and phenology.

[0006] Among them, the detection of invasive plants based on satellite remote sensing mainly uses the following five schemes: 1. Multi-temporal multi-spectral + machine learning: This scheme uses medium resolution images and machine learning classification based on vegetation index (NDVI, EVI, red edge index).

[0007] This scheme has the following shortcomings: (1) The spatial resolution and the target size do not match, resulting in a common mixed pixel phenomenon: 10m-30m pixels often cover invasive patches, bare land, crops or water bodies at the same time, resulting in “dilution” of the spectral signal. Machine learning models can only learn “mixed” features and cannot accurately distinguish them.

[0008] (2) Vegetation index cannot distinguish local species with similar spectral curves in the phenology overlap period: NDVI / EVI is sensitive to total chlorophyll content and is not sensitive to “when green, when yellow”. When the invasive species and the local species have a difference of only a few days in the turning green or aging time, the vegetation index curve is almost overlapped, and the model loses the basis for discrimination.

[0009] (3) Simple classifiers are extremely sensitive to background noise and have poor stability: farmland plowing, fertilization, irrigation, shadow, etc. will cause the vegetation index to have a similar change amplitude as the invasive species; shallow models cannot capture multi-dimensional nonlinear boundaries, resulting in repeated false positives and false negatives, and poor stability.

[0010] 2. Commercial high-spectral / multi-spectral imagery + machine learning: This approach uses high spatial / spectral resolution data to extract invasive plants with SVM, etc.

[0011] This approach has the following disadvantages: (1) High cost of image acquisition and processing, difficult to use in large areas: Commercial satellites charge by km², and the price of commercial images is high, and the coverage is not continuous, making it difficult to bear the cost of large-area or long-time series research. For management departments, the cost of daily monitoring is too high.

[0012] (2) High-spectral data sources are scarce: satellite high-spectral missions have limited coverage and long revisit periods, and continuous acquisition is difficult due to high cloud cover at mid-latitudes. Airborne high-spectral data can provide centimeter-level details, but the cost of flight is even higher.

[0013] (3) Classification models rely on a large number of finely annotated samples, and the cost of manual annotation increases linearly with area: 0.3-2m resolution means that 1km² of image contains 250,000-100 million pixels; pixel-by-pixel or object-level annotation requires a large amount of manual work and requires high botanical knowledge.

[0014] 3. Deep learning target detection This approach is based on the above-mentioned 2nd approach, using CNN / UNet to segment multi-spectral images or time-series stacks.

[0015] This approach has the following disadvantages: (1) Strong dependence on large-scale, high-quality annotated data, performance drops sharply when migrating across regions: deep networks require a large amount of balanced and high-quality samples covering all spectral, angle, growth stage, and background combinations. When the image source, sensor, or region changes, data distribution drift causes the model performance to decay rapidly, requiring re-labeling or fine-tuning.

[0016] (2) Large number of model parameters and large amount of computation: deep segmentation models (UNet / Transformer) have a large number of parameters and require GPU support. For large-area (> thousand square kilometers) monitoring, the inference time is long and real-time performance is poor.

[0017] (3) False positives increase significantly when similar to background color or texture: crop flowering, leaf fall, or artificial yellow cloth, plastic greenhouses are highly similar in color and texture to invasive species; deep models are prone to misjudgment as invasive patches, especially when data is insufficient or classes are imbalanced.

[0018] 4. Time-series phenology method (NDVI / EVI time-series analysis) This scheme combines multi-seasonal images (spring green-up, autumn senescence / flowering) or constructs NDVI / red edge time series, uses the "window period of phenological difference" between communities to amplify spectral discrimination, and then trains RF / SVM models.

[0019] This scheme has the following disadvantages: (1) Cloud cover or data missing causes time series to break, making it difficult to obtain complete curves: Optical satellites are affected by clouds and rain, and in some years, the clear observations available are less than half of the theoretical number. Broken curves cannot reliably extract key phenological parameters such as inflection points and peaks, which in turn affects classification.

[0020] The area is too small in the early stages of invasion, and it cannot form distinguishable phenological signals in the curve: At the initial stage of invasion, the coverage is low, and the NDVI contribution is "overwhelmed" by the surrounding high-NDVI vegetation or bare land, with no significant performance on the curve, making it difficult to detect.

[0021] (2) When the peak difference is small and the local species phenology overlaps, the risk of misjudgment is high: When the invasive species and the local plants of the same genus differ in green-up, flowering, and senescence time by only a few days, the curve shape is almost identical; machine learning can only learn "the same curve", and the risk of misjudgment increases.

[0022] 5. Multi-source data fusion method (satellite + UAV + ground) This scheme combines satellite images, unmanned aerial vehicle multi-spectral (centimeter-level) and ground points to improve accuracy through feature-level or decision-level fusion.

[0023] This scheme has the following disadvantages: (1) The cost of synchronous acquisition and preprocessing of multi-source data is high: UAVs require airspace approval and on-site operation; ground surveys require manual on-site reconnaissance; multi-source data radiation, geometry, and time reference are different, and all need to be independently corrected, with a total cost much higher than that of a single data source.

[0024] (2) The coverage of UAVs is small, and the airspace approval is complex, making it impossible to quickly expand to regional scale: Limited by battery, regulations, and airspace, a single flight can only cover a few square kilometers; provincial or basin-level monitoring requires hundreds or even thousands of flights, which is not feasible in terms of time and manpower.

[0025] (3) Ground observation data is extremely limited and costly to obtain: Invasive patches are often located in wetlands, beaches, and steep slopes, making it difficult to sample on-site; even if it is possible, professionals need to carry GPS, spectrometer, and UAV equipment, making the cost of a single point much higher than that of traditional agricultural surveys.

[0026] (4) Fusion model is complex, and parameter space expands exponentially: after combination of spectral, spatial and temporal dimensions of different sensors, the feature space expands exponentially; a complex alignment, weighting, attention or fusion network needs to be designed, a large number of hyperparameters are trained, overfitting is easy to occur, and it is difficult to explain.

[0027] It can be seen that the related schemes all have insufficient spectral resolution or face the dilemma of high cost and small sample quantity and poor representativeness, and a new method for detecting invasive plants is urgently needed. SUMMARY

[0028] Embodiments of the present application provide a method for detecting invasive plants based on fusion of satellite remote sensing images and ground observation samples, to at least partially solve the above problems.

[0029] The first aspect of the embodiments of the present application provides a method for detecting invasive plants based on fusion of satellite remote sensing images and ground observation samples, the method comprising: acquiring multi-temporal image data of a target plant in a target area by a satellite, and acquiring a ground observation point of the target plant in the target area; preprocessing the multi-temporal image data to obtain satellite remote sensing images; determining a pixel point where the target plant is located from the satellite remote sensing images based on position information of the ground observation point, and generating a satellite remote sensing image sample set; acquiring image features of each satellite remote sensing image sample based on the satellite remote sensing image sample set; selecting a preset number of core features from the image features by cross comparison of a plurality of feature selection methods; training a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain an invasive plant detection network; detecting invasive plants based on the invasive plant detection network.

[0030] Optionally, acquiring the ground observation point of the target plant in the target area comprises: acquiring target plant observation information uploaded by a platform user through a personal terminal through an open biodiversity platform, the target plant observation information carrying position information aligned with a satellite pixel corresponding to the multi-temporal image data.

[0031] Optionally, preprocessing the multi-temporal image data comprises at least one of the following: identifying and removing cloud-covered pixels in the multi-temporal image data using a QA60 band provided by a satellite image; resampling all bands of the multi-temporal image data to 10m resolution by a bicubic interpolation method. The maximum synthesis method is used for time sequence synthesis of the multi-time phase image data.

[0032] Optionally, the image features of the satellite remote sensing image sample at least include: spectral features, vegetation index features and texture features, the spectral features at least include: original multi-spectral bands of the satellite remote sensing image, the vegetation index features at least include: NDVI, EVI and edge index, and the texture features at least include: entropy, contrast, variance and correlation.

[0033] Optionally, the pixel points where the target plants are located in the satellite remote sensing image are determined to generate a satellite remote sensing image sample set, including: The pixel points where the target plants are located in the satellite remote sensing image are marked as positive samples; The pixel points where other common plants are located in the satellite remote sensing image are marked as negative samples; The positive samples and the negative samples are subjected to outlier elimination based on a quartile range algorithm to obtain the satellite remote sensing image sample set.

[0034] Optionally, the preset network includes: a base model and a meta model, the base model includes: a random forest model and LightGBM, the meta model is a logistic regression model, the preset network is a Stacking integrated framework, the logistic regression is taken as a meta classifier to perform secondary learning on the prediction results of the base model to generate a final classification result; The satellite remote sensing image sample set and the core features of each satellite remote sensing image sample are used to train the preset network to obtain an invasive plant detection network, including: The random forest model and LightGBM are trained based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain the trained base model; The model parameters of the trained base model are frozen, the core features of each satellite remote sensing image sample are taken as input, the base model prediction results are obtained through the trained base model respectively, the base model prediction results are taken as meta features to input the meta model to obtain the final prediction results of the logistic regression model, and the parameters of the logistic regression model are trained based on the difference between the final prediction results and sample labels of each satellite remote sensing image sample; The trained base model and the trained meta model are taken as the invasive plant detection network.

[0035] The second aspect of the embodiment of the application provides an invasive plant detection device based on fusion of satellite remote sensing images and ground observation samples, and the device includes: A data acquisition module is configured to acquire multi-time phase image data of target plants in a target region by a satellite and to acquire ground observation points of target plants in the target region. a preprocessing module, configured to preprocess the multi-temporal image data to obtain satellite remote sensing images; a sample generation module, configured to determine pixel points where target plants are located from the satellite remote sensing images based on position information of the ground observation points, and generate a satellite remote sensing image sample set; a feature acquisition module, configured to acquire image features of each satellite remote sensing image sample based on the satellite remote sensing image sample set; a feature selection module, configured to select a preset number of core features from the image features by cross comparison of a plurality of feature selection methods; a training module, configured to train a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain an invasive plant detection network; a detection module, configured to perform invasive plant detection on a satellite remote sensing image to be detected based on the invasive plant detection network.

[0036] In a third aspect, an electronic device is provided, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the invasive plant detection method based on fusion of satellite remote sensing images and ground observation samples according to the first aspect.

[0037] In a fourth aspect, a computer readable storage medium is provided, which stores a computer program executable by a processor to implement the invasive plant detection method based on fusion of satellite remote sensing images and ground observation samples according to the first aspect.

[0038] In a fifth aspect, a computer program product is provided, which includes computer programs / instructions executable by a processor to implement the steps of the invasive plant detection method based on fusion of satellite remote sensing images and ground observation samples according to the first aspect.

[0039] The present application constructs a technical route of “satellite remote sensing images → multiple screening features → model training and detection” under the extremely simple data framework of free multispectral satellite images + open biodiversity platform, and completes invasive plant mapping at extremely low cost. The present application makes full use of the data recorded in the open biodiversity platform, greatly reduces the labor cost of obtaining ground data, and effectively avoids expensive data procurement and manual delineation process by using the latitude and longitude and species label from the data, thereby realizing a low-cost, efficient and reproducible detection process. The present application only selects a preset number of core features for model training, controls the input dimension, and makes the model classification task more robust when facing small sample tasks. BRIEF DESCRIPTION OF DRAWINGS

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of the steps of the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples provided in the embodiments of the present invention; Figure 2 This is a schematic diagram of the multi-temporal image data preprocessing process in the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples provided in the embodiments of the present invention.

[0042] Figure 3 This is a schematic diagram of the feature construction and sample management process in the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the overall process of model training and integration in the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples provided in the embodiments of the present invention; Figure 5 This is a hardware structure diagram of any device with data processing capabilities, which is the invasive plant detection device based on the fusion of satellite remote sensing images and ground observation samples provided in the embodiments of the present invention. Detailed Implementation

[0043] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] Currently, the main difficulties in the practical application of satellite remote sensing imagery for detecting invasive plants include: 1. Cloud interference: rendering 67% of imagery unusable. Specifically, on average, about 67% of optical remote sensing imagery globally is contaminated by clouds, resulting in degraded image quality; clouds cause abnormal surface reflectivity, leading to spectral distortion. Cloud shadows impair the discriminative power of texture features, causing texture destruction.

[0045] 2. Spectral Confusion: The "Camouflage" Problem of Invasive Plants. Specifically, plants of the same genus exhibit similar behaviors under various characteristics. Soil background, atmospheric humidity, and topographic shading further exacerbate spectral variations, leading to a high false detection rate with traditional threshold methods.

[0046] 3. Sample bottleneck: Ground truth data is scarce and unbalanced. Specifically, manual field surveys require 2-3 people per day per square kilometer, and it is difficult to reach areas such as wetlands and steep slopes.

[0047] 4. Spatio-temporal mismatch: Remote sensing and field data do not match, leading to spatial resolution difference and time window deviation.

[0048] Based on this, the embodiment of the present application proposes a technical solution of "low cost, anti-interference and small sample applicability".

[0049] Specifically, the embodiment of the present application proposes a minimalist data framework of "free multi-spectral satellite images + open biodiversity platform", and constructs a technical route of "satellite remote sensing image -> multiple screening features -> model training and detection", which completes the detection of invasive plants at a very low cost.

[0050] In the embodiment of the present application, the observation records collected by the open biodiversity platform are uploaded by scientific researchers, nature lovers and citizen scientists, covering photos, spatial and temporal coordinates and ecological metadata of animals and plants worldwide, forming a massive species database across regions and ecosystems; for invasive plant research, the multi-source observation information provided by the platform can be directly called to replace traditional ground surveys, significantly reducing the cost of manual data collection.

[0051] In the embodiment of the present application, the data recorded in the open biodiversity platform is fully utilized, greatly reducing the manual cost required for obtaining ground data.

[0052] In the face of the problems of high cost, scarce archives and expensive labeling of "commercial high-resolution / high-spectral + machine learning", the embodiment of the present application relies entirely on public data sources: satellite images are free, and data from the open biodiversity platform are free and come with latitude and longitude and species labels, effectively avoiding expensive data procurement and manual tracing process, thereby realizing a low-cost, efficient and replicable detection process.

[0053] Compared with "deep learning target detection" which relies on a large number of high-quality labeled samples and has the defects of poor migration in small sample scenarios and large computational cost, the embodiment of the present application abandons deep learning and only uses a preset number (for example: 16) of robust features selected by Boruta-LightGBM-mutual information multiple screening and lightweight stacking, which not only reduces the dependence on large-scale labeling, but also significantly reduces the model calculation complexity. At the same time, the introduction of feature selection controls the input dimension, making the classification more robust when facing small sample tasks.

[0054] In view of the disadvantages of "time series phenology method" such as sequence breakage due to data missing, weak early patch signal and misjudgment of phenology overlap, the embodiment of the present application uses time series synthesis to make up for the data gap, ensures the effective use of the key window period, and assists with texture and multi-vegetation index combination at the feature level to replace complex time series curves, so as to maintain a relatively high discrimination in the early diffusion stage.

[0055] In view of the difficulty of synchronization of unmanned aerial vehicles and ground observations, mismatch of scales, and complexity of fusion models in multi-source data fusion, the embodiment of the present application uniformly uses 10 m resolution, and the citizen science point naturally aligns with the satellite pixel, without the need for unmanned aerial space approval; the preset network selected in the embodiment of the present application is a two-layer Stacking containing only 3 models, and the complexity is linearly controllable, thereby simplifying the multi-source fusion project originally with heavy equipment and heavy process into a script job that can run on ordinary electronic devices (for example: personal notebook computers). In this way, not only the technical difficulties of multi-platform fusion are avoided, but also the overall implementation cost is reduced, making it more suitable for being widely used in actual ecological management and invasion monitoring.

[0056] In summary, the plant invasion detection system based on fusion of satellite remote sensing images and ground observation samples provided by the embodiment of the present application can solve the problems of mixed pixel phenomenon, phenology overlap, insufficient samples, high cost, and insufficient model generalization ability in traditional invasive plant detection.

[0057] Specifically, as shown in Figure 1 the step flowchart of the invasive plant detection method based on fusion of satellite remote sensing images and ground observation samples provided by the embodiment of the present application is shown, and the method comprises the following steps: S101, acquiring multi-temporal image data of a target plant in a target area by a satellite, and acquiring a ground observation point of the target plant in the target area.

[0058] Multi-temporal remote sensing imagery refers to a set of multiple remote sensing images of the same geographical area acquired at different time nodes. It introduces the "time dimension" to make up for the limitation of single temporal image that can only reflect the "state of ground objects at a certain time", and becomes the core data basis for capturing dynamic changes of ground objects and mining time sequence rules, and is widely used in fields such as ecological monitoring, agricultural yield estimation, and disaster assessment.

[0059] In the embodiment of the present application, satellite images are used as the core data source to provide multi-temporal and multi-spectral information; and the open biodiversity platform provides ground sample points as the label source for classification of invasive plants and local plants.

[0060] Specifically, in the embodiment of the present application, the target plant observation information uploaded by the platform user through the personal terminal is acquired through the open biodiversity platform, and the target plant observation information carries position information aligned with the satellite pixel corresponding to the multi-temporal image data.

[0061] For the convenience of understanding, the following description will take Solidago canadensis as an example. Specifically, in the embodiments of the present application, satellite multi-temporal image data of the flowering period (August-September peak flowering period) of Solidago canadensis in the target region can be automatically obtained through the GEE platform, and the species observation points in the region can be obtained through the OpenBiodiversity platform.

[0062] In the actual application process, for other invasive plants, the target region can be selected according to the actual situation.

[0063] In the actual application process, for other invasive plants, the target region can be selected according to the actual situation.

[0064] In the embodiments of the present application, for the time range of the obtained multi-temporal image data, the plants in the peak flowering period have more prominent spectral response and special texture characteristics according to the characteristics of the peak flowering period of Solidago canadensis. The multi-temporal image data in the peak flowering period of August-September is selected to facilitate the distinction from other vegetation species. In the actual application process, for other invasive plants, the target time range can be selected according to the actual situation.

[0065] S102, pre-processing the multi-temporal image data to obtain satellite remote sensing image.

[0066] In the embodiments of the present application, the multi-temporal image data can be processed according to the mature pre-processing operation in the related art.

[0067] Specifically, in the embodiments of the present application, the pre-processing operation includes at least one of the following: Cloud coverage pixels in the multi-temporal image data are identified and removed using the QA60 band provided by the satellite image; All bands of the multi-temporal image data are resampled to 10m resolution by bicubic interpolation method; The multi-temporal image data is time-series synthesized by maximum value composition method.

[0068] In the embodiments of the present application, the cloud coverage pixels are removed, which can avoid abnormal reflection values in subsequent analysis. All bands are resampled to 10m resolution, which can ensure the consistency of feature extraction. The multi-temporal image data is time-series synthesized, which can highlight the peak value of vegetation growth, thereby enhancing the spectral characteristics of invasive plants. After time-series synthesis, the image fuses information of many time phases, reduces the interference caused by short-term weather changes and different observation conditions, and makes the data more reliable.

[0069] The QA60 band is a dedicated band in optical satellite imagery such as Sentinel-2 for cloud and cloud shadow quality control. Through preset pixel quality coding, it automatically identifies and removes cloud-covered pixels. Its core objective is to eliminate abnormally high reflectance values ​​caused by clouds (such as bright white cloud areas) or abnormally low reflectance values ​​caused by shadows, resulting in imagery that retains only "cloud-free / shadow-free ground features" for subsequent analysis (e.g., preventing inflated NDVI indices due to high cloud reflectance). This ensures the accuracy of subsequent analyses such as ground feature classification and vegetation index calculations.

[0070] Bicubic interpolation is an interpolation method based on cubic convolution kernels. It determines the value of a new pixel by considering the information of 16 pixels in a 4×4 area around the target pixel and using cubic convolution kernels to calculate weights.

[0071] Maximum Value Composite (MVC) is a core preprocessing technique for multi-temporal remote sensing images. By extracting the maximum value of each pixel within a specific time series (such as 1 month or 1 growing season), it achieves the dual effect of "enhancing target features and weakening interference factors". It is widely used in scenarios such as invasive plant monitoring and vegetation growth dynamic analysis.

[0072] The essence of MVC is to compare the reflectance (or vegetation index, such as NDVI) of pixels at the same spatial location but different time points, and finally retain the maximum value within that time series. Its core logic is based on the following two points: Target Feature Enhancement: During peak growth periods, vegetation (especially invasive plants) exhibits high chlorophyll content and extensive leaf coverage, resulting in significantly higher near-infrared reflectance (or NDVI value) compared to the early growth / decline phases. Conversely, interference factors such as clouds, shadows, and aerosols can lead to lower pixel values. By maximizing the value, observational data from the peak vegetation growth phase can be automatically selected, indirectly "amplifying" the spectral characteristics of invasive plants (e.g., spectral differences compared to native plants at their peak).

[0073] Reduced interference: Short-term weather changes (such as cloudy days or haze) and differences in observation angles can cause fluctuations in pixel values ​​of the same ground feature at different times (e.g., vegetation near-infrared reflectance decreases on cloudy days). MVC effectively reduces such random interference by removing low-value "abnormal times" and retaining high-value "high-quality times," making the data more stable.

[0074] Specifically, such as Figure 2 As shown, it illustrates the multi-temporal image data preprocessing flow provided by an embodiment of the present invention. In this embodiment, the raw satellite image data obtained without any processing contains the initial information of the study area, but there are problems that need to be processed, such as cloud coverage, inconsistent band resolution, and multiple temporal phases.

[0075] For the satellite image raw data, first, the QA60 band specially used for cloud and cloud shadow quality control in the satellite image can be used to automatically identify and remove the cloud coverage and cloud shadow pixels by analyzing the pixel quality code, to complete the cloud removal processing, avoid errors caused by abnormal reflection value of the cloud in subsequent analysis, and ensure the accuracy of the ground object information in the image.

[0076] Since different bands of the satellite image can have different spatial resolutions, in order to ensure the consistency of subsequent feature extraction and other operations, the embodiment of the application performs band resampling by using a bicubic interpolation method or the like, and adjusts the spatial resolution of all bands to 10m, so that the bands remain consistent in the spatial scale.

[0077] In the embodiment of the application, the maximum synthesis method is used to process the multi-temporal (photographed at different times) images, to highlight the peak flowering period. Because the vegetation has a high reflectivity in the relevant bands (such as the near-infrared band) during the peak flowering period, by taking the maximum value of each pixel in the multi-temporal images, the peak state of the vegetation growth can be highlighted, and the interference caused by short-term weather changes and different observation conditions can be reduced, so that the data is more reliable.

[0078] In the preprocessing process in the embodiment of the application, the vegetation index (NDVI) and other vegetation indexes can also be calculated, and the spectral features (reflectivity of each band) and texture features (GLCM) of the remote sensing image can also be extracted.

[0079] Specifically, the NDVI is calculated by the reflectivity of the near-infrared band and the red band, to reflect the coverage and growth status of the vegetation, and is an important index for vegetation monitoring. Different ground objects have unique spectral reflection characteristics in different bands, and the reflectivity data of each band of the image is extracted, which is an important basis for distinguishing the types of ground objects (such as vegetation, water body, building, etc.). The gray level co-occurrence matrix (GLCM) is used to extract the texture features of the image. The texture features reflect the spatial distribution law of the pixel gray values in the image, and different ground objects such as vegetation and urban land have obvious differences in texture, and the features are helpful for more accurate identification of the ground objects.

[0080] In the embodiment of the application, the vegetation index, the spectral features and the texture features can all be used as image features in the subsequent process.

[0081] Finally, in the embodiment of the application, the image can also be cropped in the geometric range according to the research requirements, only the specific area to be researched is retained, the irrelevant area is removed, the data amount is reduced, and the subsequent analysis of the research area is facilitated.

[0082] After the above series of processing steps, the satellite image raw data is processed into standardized preprocessed data suitable for subsequent in-depth analysis.

[0083] S103, determining a pixel point where the target plant is located in the satellite remote sensing image based on the position information of the ground observation point, and generating a satellite remote sensing image sample set.

[0084] In the embodiment of the application, on the basis of the pre-processing being completed, a multi-dimensional feature sample set can be further generated based on the satellite remote sensing image, specifically including: Data alignment: the longitude and latitude points obtained by statistics in the open biodiversity platform data are accurately matched with the spatial coordinates of the satellite image, and a sample set is automatically generated. The pixel center in the satellite image is located according to the longitude and latitude of the corresponding point, and the corresponding pixel point is found out from the satellite remote sensing image to obtain the following features: spectral feature, vegetation index feature and texture feature.

[0085] Specifically, in the embodiment of the application, the step S103 includes: S1031, marking the pixel point where the target plant is located in the satellite remote sensing image as a positive sample; S1032, marking the pixel point where other common plants in the target area are located as a negative sample; S1033, performing outlier elimination on the positive sample and the negative sample based on the interquartile range algorithm, and obtaining a satellite remote sensing image sample set.

[0086] In the embodiment of the application, the pixel point where the target plant is located can be marked as a positive sample, and other common local plants in the target area are marked as negative samples.

[0087] In actual application, due to measurement error or interference of some environmental factors, under most features, there are a large number of outliers in the sample. Therefore, the embodiment of the application proposes to use the IQR algorithm to eliminate outliers. After processing, the data distribution becomes more concentrated, and the overall dispersion degree is significantly reduced, laying a solid foundation for subsequent model training and feature analysis.

[0088] The IQR (Interquartile Range) algorithm is a non-parametric outlier detection method based on the quartile of data distribution. The core idea is to define the “normal data boundary” through the “middle 50%” distribution range of data, and the values exceeding the boundary are determined as outliers (abnormal values). This method does not depend on the data obeying a specific distribution (such as normal distribution), and is robust to extreme values, and is widely used in data preprocessing (such as sensor data, experimental data, business indicators, etc.).

[0089] S104, obtaining the image features of each satellite remote sensing image sample based on the satellite remote sensing image sample set.

[0090] In the embodiment of the present application, the image features of the satellite remote sensing image sample at least include: spectral features, vegetation index features and texture features, the spectral features at least include: original multi-spectral bands of the satellite remote sensing image, the vegetation index features at least include: NDVI, EVI and edge index, and the texture features at least include: entropy, contrast, variance and correlation.

[0091] In the embodiment of the present application, the spectral features refer to: directly using original multi-spectral bands (B2-B12) of the satellite image, these bands cover the main spectral range from visible light to near infrared, which can effectively represent the spectral characteristics of vegetation.

[0092] The vegetation index features refer to: calculating common vegetation indexes such as NDVI, EVI and red edge index, which fully reflect the growth state of vegetation from different angles, such as chlorophyll, health status and soil background, and can improve the discrimination ability of the classification model for different plant types.

[0093] The texture features refer to: calculating common texture features such as entropy, contrast, variance and correlation based on the gray level co-occurrence matrix (GLCM), which can capture the texture differences of the vegetation community.

[0094] In the embodiment of the present application, as many image features of the satellite remote sensing image sample as possible are collected, so that each sample has more than 100 features, which causes the data set to have the characteristics of high dimension and small sample, and the high-dimensional feature space brings many problems to model training. Firstly, it increases the complexity of calculation and the amount of resources required for training; secondly, it is easy to produce overfitting; thirdly, high-dimensional data will increase the redundancy between features, thereby interfering with the model to obtain and learn key features, further affecting the classification effect.

[0095] Therefore, in the embodiment of the present application, a preset number of core features are selected from a large number of image features for subsequent network training.

[0096] S105, a preset number of core features are selected from the image features by cross comparison of a plurality of feature selection methods.

[0097] In the embodiment of the present application, in view of the problem that the multi-temporal medium resolution + machine learning is easily affected by mixed pixels, the image of the blooming period of the invasive plant is used in the embodiment of the present application, the mixed pixel problem is converted into a learnable feature weight problem by point-pixel accurate registration, more features are introduced, the difficulty in distinguishing when the spectral features are similar is avoided, and then the integrated model automatically learns the classification.

[0098] Specifically, in the embodiment of the present application, the redundant features are reduced by feature selection, and the dimension is reduced. In order to overcome the problem of instability of single feature selection method, the embodiment of the present application adopts multiple feature selection methods for cross comparison. The feature selection methods adopted in the embodiment of the present application include: Boruta feature selection (recursive feature importance evaluation based on random forest), LightGBM (capture feature interaction) and mutual information method (feature scoring based on statistical correlation).

[0099] In the embodiment of the present application, by cross comparison of the results of different methods, a preset number of core features which are stably present under multiple methods are selected, and the final training sample matrix is constructed based on the core features.

[0100] In the embodiment of the present application, for the specific implementation case of Canada goldenrod, after exploration and research, it is finally determined that the preset number is 16, which can not only ensure the accuracy of the subsequent network prediction result, but also maintain the low calculation amount of the network. In the actual application process, for different invasive plants, the setting of the preset number can be determined according to the actual situation.

[0101] In the embodiment of the present application, after completing the feature selection, the data set is divided into a training set and a test set in a ratio of 7:3. Subsequently, data normalization is performed to unify the scale of the features, so as to eliminate the influence of the difference in dimension and order of magnitude on the model training and classification result. In order to solve the problem of unbalanced data samples, the embodiment of the present application also proposes that the SMOTE algorithm is used to expand the positive samples and perform data enhancement, and the data set processing is completed.

[0102] SMOTE (Synthetic Minority Oversampling Technique) is a classic algorithm for solving the problem of unbalanced data categories. The core idea is to synthesize new minority class (such as positive sample) data instead of simply copying, to achieve sample balance while avoiding overfitting. This algorithm is widely used in the data preprocessing stage of classification tasks (such as disease diagnosis, fraud detection, fault identification, etc.), and is particularly suitable for scenarios where the number of positive samples is much less than that of negative samples.

[0103] Specifically, Figure 3 The feature construction and sample management process schematic diagram in the embodiment of the present application is shown. As Figure 3 As shown in the embodiment of the present application, first, the cleaned data set (i.e. the preprocessed satellite remote sensing image sample) is obtained to provide high-quality basic data for subsequent processing.

[0104] Subsequently, feature selection is performed: from the influence features of the preprocessed satellite remote sensing image sample, the features valuable for subsequent network training are selected, irrelevant or redundant information is filtered out, the data dimension is reduced, and the processing efficiency and accuracy are improved.

[0105] In the feature selection process, a plurality of feature selection methods are adopted for cross comparison (including: Baccala algorithm, LightGBM feature importance and mutual information method) to obtain a feature candidate set, and then 16 key features with the highest scores are selected from the feature candidate set. After the foregoing feature selection operation, the satellite remote sensing image sample including 16 key features is obtained.

[0106] Further, the data is divided into different subsets, including a training set and a test set, for subsequent training, testing and other different links of the preset network.

[0107] In the embodiment of the application, the training set and the test set can also be normalized. The data is mapped to a specific range (such as [0, 1]), and the dimensional difference of different features is eliminated, so that the data is more suitable for model training, the stability and convergence speed of the model are improved, and the consistency of the data in the training and testing stages is ensured.

[0108] In the embodiment of the application, the training data can also be subjected to a data enhancement operation to generate more diverse training data, enrich the data samples, and help improve the generalization ability of the model and reduce the risk of overfitting. Finally, based on the normalized test set and the enhanced training set, the final data set is obtained for training and testing the preset network.

[0109] S106, training the preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain an invasive plant detection network.

[0110] In the embodiment of the application, a multi-model fusion network based on ensemble learning is proposed. A plurality of classification algorithms suitable for small sample problems are integrated to improve the classification accuracy of feature complexity and target class imbalance. There are two common algorithms for ensemble learning models, one is Bagging algorithm, and the other is Boosting algorithm. The typical machine learning of Bagging algorithm is random forest, and the typical machine learning of Boosting algorithm includes GBDT, XGBoost and LightGBM, etc.

[0111] In the embodiment of the application, the preset network includes a base model and a meta model, the base model includes a random forest model and a LightGBM, and the meta model is a logistic regression model. The preset network is a Stacking ensemble framework, and the logistic regression is used as a meta classifier to perform secondary learning on the prediction results of the base model to generate a final classification result.

[0112] In the embodiment of the application, the meta-model is a logistic regression, which is used as a reference model and can integrate the prediction results of the sub-models through linear combination, thereby avoiding overfitting problems caused by complex decision-making processes. The base models include random forest (RF) and LightGBM, which are good at processing high-dimensional nonlinear features and small sample training, respectively. The fusion model adopts a stacking integration framework, uses logistic regression as a meta-classifier, performs secondary learning on the prediction results of the base models, and generates the final classification result.

[0113] The step S106 specifically includes: S1061, training the random forest model and LightGBM based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample, to obtain the trained base model; S1062, freezing the model parameters of the trained base model, taking the core features of each satellite remote sensing image sample as input, and obtaining base model prediction results through the trained base model, taking the base model prediction results as meta-features to input into the meta-model, obtaining the final prediction result of the logistic regression model, and training the parameters of the logistic regression model based on the difference between the final prediction result and the sample label of each satellite remote sensing image sample; S1064, taking the trained base model and the trained meta-model as the invasive plant detection network.

[0114] To ensure the robustness of the model, the KFold cross-validation and GridSearchCV hyperparameter optimization are adopted in the embodiment of the application, and a sample balancing method (such as SMOTE) is introduced to alleviate the class imbalance problem. The overall process of model training and integration is as shown in Figure 4 .

[0115] In the embodiment of the application, the input features (core features selected in the above step) are first processed, and a stratified K-fold cross-validation strategy is adopted to avoid the bias caused by data division and ensure the stability and generalization ability of subsequent model training.

[0116] Subsequently, the integrated model (random forest RF + LightGBM) is used for network search to obtain the corresponding base model prediction results, which are used as meta-features to provide prior knowledge for subsequent logistic regression LR, and the final prediction is obtained by logistic regression LR.

[0117] Wherein, the random forest RF and the LightGBM are trained respectively based on a satellite remote sensing image sample set and core features of each satellite remote sensing image sample in advance. Then, the network parameters of the random forest RF and the LightGBM are fixed, and the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample are input into the trained base model again to obtain the prediction results of the base model as meta-features, which are input into a meta-model. In the training process, the model parameters of the meta-model are adjusted, and finally the trained invasive plant detection network is obtained.

[0118] In S107, the satellite remote sensing image to be detected is detected for invasive plants based on the invasive plant detection network.

[0119] In the embodiment of the application, after the data is preprocessed and the features are constructed, combined with multi-method feature selection, the data is finally input into a preset network to train an invasive plant detection network. The network can generate high-precision invasive detection results.

[0120] In the embodiment of the application, the invasive plant detection result can be output in the form of a spatial distribution map. Specifically, based on the invasive plant detection result, the positions of the detected invasive plants can be marked on the map corresponding to the target area to obtain a spatial distribution map of the invasive plants.

[0121] Based on the same inventive concept, the embodiment of the application also provides an invasive plant detection device based on fusion of satellite remote sensing images and ground observation samples. The device comprises: A data acquisition module is configured to acquire multi-temporal image data of a target plant in a target area by a satellite and acquire a ground observation point of the target plant in the target area. A preprocessing module is configured to preprocess the multi-temporal image data to obtain satellite remote sensing images. A sample generation module is configured to determine pixel points where the target plant is located from the satellite remote sensing images based on position information of the ground observation point to generate a satellite remote sensing image sample set. A feature acquisition module is configured to acquire image features of each satellite remote sensing image sample based on the satellite remote sensing image sample set. A feature selection module is configured to select a preset number of core features from the image features by cross comparison of a plurality of feature selection methods. A training module is configured to train a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain an invasive plant detection network. A detection module is configured to detect an invasive plant in a satellite remote sensing image to be detected based on the invasive plant detection network.

[0122] The embodiment of the invasive plant detection device based on satellite remote sensing image and ground observation sample fusion provided by the application can be applied to any device with data processing capability, which can be a device or apparatus such as a computer. The device embodiment can be realized by software, or realized by hardware or a combination of software and hardware. Taking software realization as an example, as a device in a logical sense, it is formed by reading the corresponding computer program instructions in the non-volatile memory into the memory and running by the processor of the device with data processing capability. From the hardware level, as shown in Figure 5 Fig. 3 is a hardware structure diagram of the device with data processing capability in which the invasive plant detection device based on satellite remote sensing image and ground observation sample fusion provided by the application is located, in addition to the processor, the memory, the network interface, and the non-volatile memory shown in Figure 5 The device with data processing capability in which the device is located in the embodiment can also include other hardware according to the actual functions of the device with data processing capability, and details are not repeated.

[0123] Based on the same inventive concept, the embodiment of the application further provides an electronic device, which includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps in the invasive plant detection method based on satellite remote sensing image and ground observation sample fusion when executed.

[0124] Based on the same inventive concept, the embodiment of the application further provides a computer readable storage medium, which stores a computer program, and the program implements the steps in the invasive plant detection method based on satellite remote sensing image and ground observation sample fusion when executed by a processor.

[0125] Based on the same inventive concept, the embodiment of the application provides a computer program product, which includes a computer program / instruction, and the computer program / instruction implements the steps in the invasive plant detection method based on satellite remote sensing image and ground observation sample fusion when executed by a processor.

[0126] Each embodiment in the specification is described in a progressive manner, and each embodiment focuses on the difference from other embodiments. The same and similar parts between the embodiments can be referred to each other.

[0127] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, apparatus, or computer program product. Accordingly, embodiments of the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, embodiments of the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, and the like) embodying computer program code thereon for use by or in connection with an instruction execution system. For the purposes of this description, a computer-usable or computer readable storage medium can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device.

[0128] Embodiments of the present application are described herein with reference to the drawings, which are as follows: Figure 1 one or more functions specified by one or more blocks Figure 1 one or more functions specified by one or more blocks

[0129] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 one or more functions specified by one or more blocks Figure 1 one or more functions specified by one or more blocks

[0130] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the Figure 1 one or more functions specified by one or more blocks Figure 1 one or more functions specified by one or more blocks

[0131] While preferred embodiments of the present application have been described, additional variations and modifications can be made to these embodiments by those skilled in the art once they have the benefit of the foregoing description. Therefore, the appended claims are intended to encompass within their scope all such variations and modifications as are included within the scope of the present application.

[0132] Finally, it is to be understood that the phraseology or terminology such as "first" and "second" etc. used herein is merely intended to differentiate one entity or operation from another entity or operation, without necessarily requiring or implying any actual such relationship or order between such entities or operations. Moreover, the terms "comprising", "including", or any other variation thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element preceded by "comprises a" does not, without more constraints, foreclose the existence of additional identical elements in the process, method, article, or apparatus that comprises the recited element.

[0133] The above describes in detail the method for detecting invasive plants based on satellite remote sensing images and ground observation samples provided by the present application. The principles and implementation manners of the present application are described by using specific examples. The above description of the examples is only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the specification should not be understood as a limitation of the present application.

Claims

1. A method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples, characterized in that, The method includes: Acquire multi-temporal image data of target plants in the target area from satellite, and acquire ground observation points of target plants in the target area; The multi-temporal image data is preprocessed to obtain satellite remote sensing images; Based on the location information of the ground observation points, the pixel points where the target plants are located are determined from the satellite remote sensing images, and a satellite remote sensing image sample set is generated. Based on the satellite remote sensing image sample set, the image features of each satellite remote sensing image sample are obtained; By cross-comparing multiple feature selection methods, a preset number of core features are selected from the image features; An invasive plant detection network is obtained by training a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample. The invasive plant detection network is used to detect invasive plants in the satellite remote sensing images to be detected.

2. The method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples according to claim 1, characterized in that, Obtaining ground observation points of the target plants in the target area includes: By opening a biodiversity platform, target plant observation information uploaded by platform users through their personal terminals is obtained. The target plant observation information carries location information aligned with satellite pixels corresponding to the multi-temporal image data.

3. The method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples according to claim 1, characterized in that, Preprocessing of the multi-temporal image data includes at least one of the following: Using the QA60 band provided by satellite imagery, cloud-covered pixels in the multi-temporal image data are identified and removed; All bands of the multi-temporal image data were resampled to 10m resolution using bicubic interpolation. The multi-temporal image data is time-series synthesized using the maximum value synthesis method.

4. The method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples according to claim 1, characterized in that, The image features of satellite remote sensing image samples include at least: spectral features, vegetation index features and texture features. The spectral features include at least: the original multispectral bands of the satellite remote sensing image. The vegetation index features include at least: NDVI, EVI and edge index. The texture features include at least: entropy, contrast, variance and correlation.

5. The method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples according to claim 1, characterized in that, The target plant is located in the pixel of the satellite remote sensing image, and a satellite remote sensing image sample set is generated, including: Mark the pixels in the satellite remote sensing image where the target plant is located as positive samples; Mark the pixels containing other common plants in the satellite remote sensing image as negative samples; The positive and negative samples are purged using the interquartile range algorithm to obtain a satellite remote sensing image sample set.

6. The method for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples according to any one of claims 1-5, characterized in that, The preset network includes a base model and a meta-model. The base model includes a random forest model and LightGBM. The meta-model is a logistic regression model. The preset network is a Stacking ensemble framework that uses logistic regression as a meta-classifier to perform secondary learning on the prediction results of the base model to generate the final classification result. An invasive plant detection network is obtained by training a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample, including: The random forest model and LightGBM are trained based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain the trained base model. The model parameters of the trained base model are frozen. The core features of each satellite remote sensing image sample are used as input. The base model prediction results are obtained after training. The base model prediction results are used as meta-features to input into the meta-model to obtain the final prediction result of the logistic regression model. Based on the difference between the final prediction result and the sample label of each satellite remote sensing image sample, the parameters of the logistic regression model are trained. The trained base model and trained meta-model are used as the invasive plant detection network.

7. A device for detecting invasive plants based on the fusion of satellite remote sensing imagery and ground observation samples, characterized in that, The device includes: The data acquisition module is used to acquire multi-temporal image data of target plants in the target area from satellites, and to acquire ground observation points of target plants in the target area; The preprocessing module is used to preprocess the multi-temporal image data to obtain satellite remote sensing images; The sample generation module is used to determine the pixel where the target plant is located from the satellite remote sensing image based on the location information of the ground observation point, and generate a satellite remote sensing image sample set. The feature acquisition module is used to acquire the image features of each satellite remote sensing image sample based on the satellite remote sensing image sample set. The feature selection module is used to select a preset number of core features from the image features by cross-comparing multiple feature selection methods; The training module is used to train a preset network based on the satellite remote sensing image sample set and the core features of each satellite remote sensing image sample to obtain an invasive plant detection network. The detection module is used to detect invasive plants in the satellite remote sensing image to be detected based on the invasive plant detection network.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the invasive plant detection method based on the fusion of satellite remote sensing images and ground observation samples as described in any one of claims 1-6.

Images