A method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning
By combining multi-source remote sensing and environmental variable data, and using machine learning methods to monitor the flowering phenology of rapeseed, noise samples were removed, and a random forest regression model was constructed for rapeseed spatial mapping. This solved the problems of strong dependence on phenology and high sample collection costs in existing technologies, and realized the automation and large-scale monitoring of rapeseed remote sensing mapping.
Patent Information
- Application Number
- CN202410581801.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-05-11
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-05-11
AI Technical Summary
Existing technologies for rapeseed remote sensing mapping suffer from problems such as strong dependence on key phenological periods and high cost of training sample collection, resulting in a scarcity of large-scale, long-term rapeseed maps.
By combining multi-source remote sensing and environmental variable data, phenological and machine learning methods were used to monitor the flowering phenological period of rapeseed. Noise samples were removed by spectral angle mapping, and a random forest regression model was constructed to create spatial maps of rapeseed. Post-processing was performed using Matlab software.
It has automated the large-scale remote sensing spatial mapping of rapeseed, reduced the cost of manpower and materials, is suitable for monitoring rapeseed flowering phenology on a large spatial scale, and provides efficient information on the spatial distribution of rapeseed.
Smart Images

Figure CN118429468B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of remote sensing space information technology, and in particular to a rapeseed spatial mapping method based on multi-source remote sensing data and a machine learning algorithm and oriented to large-scale crop remote sensing identification. Background Art
[0002] rape( Brassica napus L. ) is an important cash crop, accounting for 13% to 15% of the world's vegetable oil production. It has been used for centuries as a source of edible vegetable oil, animal feed and protein meal. In recent years, rapeseed production has seen a significant increase due to the expansion of cultivated land and its potential for biofuels. However, in many underdeveloped countries and regions, labor-intensive smallholder rapeseed cultivation is closely related to the rural population structure. Due to rapid marketization and urbanization, rural economy and agricultural cultivation are undergoing major transformations. Understanding the spatial distribution and temporal dynamics of rapeseed can help agricultural producers, the scientific community and policymakers understand the potential of the vegetable oil market, implement precision agricultural management, and ensure national grain and oil security.
[0003] In order to meet the needs of precision agricultural production, many countries have developed and released various types of agricultural land cover products based on remote sensing data, which include specific rapeseed crop layers. For example, the 30-meter annual arable land data layer (Cropland Data Layer) produced by the U.S. Department of Agriculture and the U.S. National Agricultural Statistics Service using medium-resolution Landsat images; the 10-meter European crop type layer developed by the European Commission using Sentinel-1 SAR and LUCAS Copernicus station data; the 30-meter annual crop list product (Annual Crop Inventory) produced by the Canadian Department of Agriculture and Agri-Food using optical and radar images; the 10-meter UK annual crop map from 2016 to 2020 produced by the UK Rural Payments Agency using Sentinel-1 / 2 images; and the 2017-2019 10-meter rapeseed map dataset (RapeseedMap10) developed by Beijing Normal University in China for 33 countries in Europe and America using Sentinel-1 / 2 images. These rapeseed map products provide accurate field-scale rapeseed planting distribution information for government decision-making, international oilseed trade, farmers' production, agricultural insurance and global change research, greatly promoting the informatization and intelligence of global agricultural production. Despite this, in most countries and regions, large-scale and long-term rapeseed maps are still scarce.
[0004] At present, a variety of methods have been developed in the research of rapeseed remote sensing mapping, which can be generally classified into two categories: unsupervised and supervised. The first is the unsupervised classification method based on spectroscopy and phenology. In the early spring period from February to March each year, the bright yellow color of rapeseed flowers makes it very easy to be visually identified and remotely monitored. Most of the current rapeseed remote sensing identification methods are based on the unique spectral or phenological characteristics of rapeseed when it blooms. For example, methods based on the position change of the red edge band, methods based on HSV color transformation, and methods based on the time-series phenological changes of the Normalized Difference Yellowness Index (NDYI). These methods are simple, efficient, and easy to implement. However, their dependence on key phenological periods limits their widespread application in areas lacking high-quality flowering period data. Then there are supervised classification methods based on machine learning, such as decision tree (DT) models, support vector machine (SVM) algorithms, random forest (RF) algorithms, and deep learning networks. This type of method can construct a mapping relationship between non-flowering period input data and training samples, realize automatic intelligent remote sensing interpretation of rapeseed, and become an effective way to get rid of the dependence on flowering period data. However, the collection process of training samples usually consumes a lot of manpower, material resources and time costs. Summary of the invention
[0005] In view of the difficulty of mapping the spatial distribution of rapeseed based on remote sensing data in the existing technology, this paper uses multi-source remote sensing and environmental variable data to propose a method for large-scale spatial mapping of rapeseed that can be quickly promoted. The main contents include: (1) analyzing the spectral and phenological characteristics of rapeseed before and after flowering, and based on multi-source environmental variables, using phenology and machine learning methods to achieve overall monitoring of rapeseed flowering phenology, and providing rapeseed flowering phenology priors for supervised classification algorithms; (2) based on the generated training samples, the noise samples are automatically removed by the method of spectral angle mapping, and Landsat images are used to achieve large-scale remote sensing spatial mapping of rapeseed in the entire test area.
[0006] The technical solution of the present invention provides a method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning, comprising the following steps:
[0007] Step 1: Collect multi-temporal Landsat optical satellite remote sensing images, perform preprocessing such as atmospheric correction and cloud detection on the collected optical satellite images, and obtain cloud-free time series optical surface reflectance data;
[0008] Step 2: Collect Global Land Data Assimilation System (GLDAS) data, select 10 characteristic variables, and synthesize the daily average of these characteristics;
[0009] Step 3: Based on field surveys and visual interpretation of high-resolution optical remote sensing images, representative rapeseed and non-rapeseed samples are obtained to provide basic data for rapeseed phenology monitoring and accuracy evaluation;
[0010] Step 4, construct a random forest regression (RFR) model to train the nonlinear fitting relationship between rapeseed flowering date and environmental variables;
[0011] Step 5, using the Canola Index (CI) method to automatically generate unsupervised rapeseed training samples for the study area;
[0012] Step 6, using the Spectral Angle Mapper (SAM) method to clean the training samples automatically generated in step 5 and remove noise samples;
[0013] Step 7, using the multi-RF (random forest) fusion mapping framework to perform spatial mapping of rapeseed;
[0014] Step 8: Post-process the mapping results through Matlab software and fill the holes in the mapping results with morphological methods.
[0015] In step 1, the time period for acquiring Landsat satellite images changes accordingly depending on the different winter and spring rapeseed planting areas. For example, the image acquisition time for winter rapeseed planting areas in southern China is from October to May of the following year, and the image acquisition time for spring rapeseed planting areas in northern China and the Qinghai-Tibet Plateau is from April to August of each year. When this method is extended to other regions, it is only necessary to modify the image acquisition time to the rapeseed growing season in the region. The image acquisition time period covers the phenological periods of rapeseed, such as sowing period, emergence period, opening period, bud period, flowering period, pod period, and harvest period. Landsat data can be obtained from the Geospatial Data Cloud website (https: / / www.gscloud.cn / home).
[0016] In step 2, GLDAS data includes three versions, namely GLDAS-2.0, GLDAS-2.1 and GLDAS-2.2, and the time periods of the three are January 1, 1948 to December 31, 2014, January 1, 2000 to February 9, 2023, and January 1, 2003 to October 31, 2022. The spatial resolution of GLDAS data is 0.25° and the temporal resolution is 3 hours. The 10 characteristic variables selected are: plant canopy surface moisture, canopy moisture evaporation, potential evaporation rate, soil moisture (0~10 cm), soil moisture (10~40 cm), soil temperature (0~10 cm), soil temperature (10~40 cm), cumulative air temperature, downward shortwave radiation, and total precipitation rate. GLDAS data can be downloaded from https: / / ldas.gsfc.nasa.gov / data.
[0017] In step 3, the time series ultra-high resolution optical remote sensing images are accessed with the help of the water micro-map software. According to prior knowledge, rapeseed is bright yellow in flowering period, which is obviously different from other crops of the same period. It is judged whether the image contains rapeseed plots by visual interpretation. The present invention adopts a hexagonal uniform sampling strategy to realize the visual interpretation and manual labeling of samples. Taking into account the surface coverage and cultivated land morphology of the study area, the size of a single grid is determined, such as the size of a single grid is determined to be 10 km 2 The sample points were randomly labeled in each grid to prevent spatial autocorrelation of the samples. For grids that did not contain rapeseed plots, the number of non-rapeseed samples was 5; for grids that contained rapeseed plots, 1 to 10 rapeseed samples (5 on average) were randomly labeled according to the proportion of rapeseed plots.
[0018] In step 4, the accumulated temperature is calculated based on the difference between the average daily temperature and the benchmark temperature within the flowering period of winter rapeseed and spring rapeseed in the monitoring area, which is used as one of the inputs of the RFR model; for example, the accumulated temperature is obtained by accumulating the difference between the average daily temperature and the benchmark temperature within the time range of November 1st to February 1st of the following year in Hubei Province (the benchmark temperature for rapeseed growth is 3°C); the total number of input features of the RFR model is 93 ((10 GLDAS feature variables) × (9 deca-synthesized dates) + 3 (longitude, latitude and elevation)), the model's mtry is set to the square root of the total number of input variables, and the number of RF classifier trees is set to 100.
[0019] In step 5, the expression of CI index is as follows:
[0020] CI =( ρGreen + ρRed )× ρNIR (1)
[0021] in, ρGreen , ρRed and ρNIR Represent the surface reflectance values of green band, red band and near infrared band pixels respectively;
[0022] The CI index is used to automatically segment rapeseed and non-rapeseed pixels, and the segmentation threshold is set to 0.07. When the CI of a pixel is greater than 0.07, the pixel is judged to be rapeseed, otherwise it is non-rapeseed.
[0023] In step 6, the SAM method is used to eliminate the noise components in the training samples. This method regards the spectrum of each pixel in the image as a high-dimensional vector, and measures the similarity between their spectra by calculating the angle between two vectors. The smaller the angle, the higher the similarity, and the greater the possibility that the two belong to the same category of objects. Therefore, the potential category of the generated training samples can be determined based on the size of the spectral angle between the generated rapeseed sample and the ground reference rapeseed sample. The spectral angle between the two vectors can be calculated by the cosine formula, as shown in formula (2):
[0024] (2)
[0025] in, i represents pixels, and the spectrum of the ground reference rapeseed sample is r , the spectrum of the pseudo rapeseed sample generated is u , the spectral angle between them is α , nb is the number of bands. If the spectral angle between the generated rapeseed sample and the reference rapeseed sample is α Below the set threshold T , it is classified as a pure rapeseed pixel, otherwise it is a noise pixel, as shown in formula (3):
[0026] (3)
[0027] in, Pseudo_rapeseed is the pseudo rapeseed sample generated. Pure_rapeseed For pure rapeseed samples, std(r) is the standard deviation of the ground reference rapeseed sample, taking into account T As the geographical environment of rapeseed cultivation varies, each sub-region T The spectrum of the ground reference rapeseed sample in this area r The standard deviation of is used to clean and enhance the quality of training samples in a region-adaptive manner.
[0028] In step 7, each RF classifier in the RF fusion mapping framework ntree The parameter is defined as 100, mtryThe parameter is set to the square of the total number of input features. The input feature is the Landsat monthly median composite image. The core of the multi-RF fusion mapping framework is to integrate the rapeseed classification probability of all phases to calculate the average probability of each pixel being classified as rapeseed, as shown in formula (4). i The probability of being classified as rapeseed is greater than the probability of being classified as non-rapeseed. i For rapeseed category:
[0029] (4)
[0030] in, P i Represents pixels i The probability of being classified as rapeseed, m is the number of cloud-free phases in the time series image, p i,j It's pixels i In phase j The probability of being classified as rapeseed.
[0031] In step 8, the specific method of post-processing the mapping results using Matlab software is as follows: 1) removing non-cultivated land pixels in the rapeseed map by masking the cultivated land layer; 2) masking pixels with a slope greater than 25° in the rapeseed map by using digital elevation model (DEM) data; 3) filling holes in the rapeseed map based on morphological methods.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] (1) This paper couples remote sensing and environmental variable data and proposes a method for estimating rapeseed flowering phenology information based on multi-source data and machine learning algorithms. It effectively breaks through the temporal and spatial limitations of traditional field survey methods and site recording methods in the field of crop phenology monitoring and can provide technical support and methodological guidance for other vegetation phenology monitoring applications.
[0034] (2) The present invention combines the advantages of unsupervised classification and supervised classification in remote sensing identification tasks. Through an integrated framework such as sample adaptive generation, quality enhancement, and model training, end-to-end automated crop spatial mapping is achieved, thereby reducing the manpower and material costs in the traditional sample collection process. The proposed method is suitable for optical remote sensing images and has strong applicability.
[0035] (3) The method proposed in the present invention can be directly used for the large-scale spatial distribution mapping task of rapeseed flowering phenology monitoring, which greatly expands the spatiotemporal scale of research and application in related fields. It can realize the annual dynamic monitoring of rapeseed crops in recent decades based on Landsat images, and provide basic information services for government agricultural departments, farmers and researchers. BRIEF DESCRIPTION OF THE DRAWINGS
[0036] Figure 1 is a flow chart of the method of the present invention;
[0037] Figure 2 (a)-Figure 2 (b) is an overview of the study area, specifically Figure 2 (a) is a climate feature distribution map; Figure 2 (b) is a river distribution map; climate data comes from the Köppen-Geiger climate classification map, where Am: tropical monsoon; Aw: tropical grassland; BWk: arid, desert, cold; BSK: arid, grassland, cold; Cwa: temperate, dry in winter, hot in summer; Cwb: temperate, dry in winter, warm in summer; Cfa: temperate, no dry season, hot in summer; Cfb: temperate, no dry season, warm in summer; Dwa: cold and dry in winter, hot in summer; Dwb: cold and dry in winter, warm in summer; Dwc: cold and dry in winter, cold in summer; Dfc: cold, no dry season, cold in summer; ET: polar, tundra; EF: polar, frost;
[0038] Figure 3 This is the spatial distribution map of rapeseed samples based on visual interpretation;
[0039] Figure 4 This is the rapeseed flowering phenology monitoring map in Hubei Province (2022, DOY: Day of Year);
[0040] Figure 5 Schematic diagram of training sample quality enhancement;
[0041] Figure 6 The results of spatial mapping of rapeseed in Hubei Province (2022);
[0042] Figure 7 Schematic diagram of post-processing of remote sensing identification results. DETAILED DESCRIPTION
[0043] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the following examples further illustrate the method of the present invention in combination with the regional distribution of rapeseed planting in Hubei Province, and describe the specific steps of the rapeseed mapping method in detail in combination with the accompanying drawings. Figure 1 As shown, the rapeseed spatial mapping method based on multi-source remote sensing data and machine learning of the present invention comprises the following steps:
[0044] Step 1: Collect multi-temporal Landsat-5, 7, 8, and 9 optical satellite remote sensing images of the entire Hubei Province on the geospatial data cloud website, perform atmospheric correction and cloud detection on the collected optical satellite images on the ENVI software, and obtain cloud-free time series optical surface reflectance data. The study area is shown in Figure 2. According to the growing season and vernalization time, the rapeseed in Hubei is mainly winter rapeseed, that is, it is sown around November in winter and matured and harvested in May of the following year.
[0045] Step 2: Download the GLDAS land assimilation data, select 10 environmental variables closely related to rapeseed growth, and synthesize the daily average of these characteristics. These variables are: plant canopy surface moisture, canopy moisture evaporation, potential evaporation rate, soil moisture (0-10 cm), soil moisture (10-40 cm), soil temperature (0-10 cm), soil temperature (10-40 cm), cumulative air temperature, downward shortwave radiation, and total precipitation rate.
[0046] Step 3: Based on field surveys and visual interpretation of high-resolution remote sensing images, representative rapeseed and non-rapeseed samples are obtained to provide basic data for rapeseed phenology monitoring and accuracy evaluation. Figure 3 As shown in the figure, the rapeseed samples visually interpreted are distributed throughout Hubei Province. The collected rapeseed samples include the rapeseed flowering date recorded by the image. For example, if an image is taken on March 25, and the rapeseed in the image is in full bloom (bright yellow in the corresponding color image), the flowering time of the rapeseed sample recorded based on the image is set to March 25.
[0047] Step 4: construct an RFR model to train the nonlinear fitting relationship between rapeseed flowering date and environmental variables. The input variables of the model and their spatiotemporal resolution are shown in Table 1. The training samples of the model are the rapeseed flowering dates obtained in Step 3. The rapeseed flowering phenology monitoring map of Hubei Province predicted by the model is shown in Figure 4 As shown (taking 2022 as an example). In general, rapeseed in Hubei Province mainly reaches its flowering peak (75-90 DOY) in mid-March.
[0048]
[0049] Step 5: Use the CI method to automatically generate unsupervised rapeseed training samples for the study area. Table 2 shows the statistical analysis results of the four vegetation indices under different Landsat combinations. The statistical results show that for winter rapeseed, CI and Winter Rapeseed Index (WRI) perform well, with high t and p < 0.01, indicating that they can significantly characterize the phenological differences of rapeseed in the pre-flowering and flowering stages. In contrast, the Normalized Difference Vegetation Index (NDVI) and NDYI have small t values and p > 0.1 in some cases, indicating that their robustness is poor. In addition, WRI performs poorly in the spring rapeseed scene, indicating that WRI is not suitable for spring rapeseed planting information identification. Overall, CI performs well in all Landsat satellite combinations, with the highest or second highest t values. Therefore, the CI-based rapeseed identification method is used to automatically generate training samples for the supervised classifier. The CI index is used to automatically segment rapeseed and non-rapeseed pixels, and the segmentation threshold is set to 0.07. When the CI of a pixel is greater than 0.07, the pixel is judged to be rapeseed, otherwise it is non-rapeseed.
[0050]
[0051] Step 6: Use the SAM method to clean the training samples automatically generated in step 5 and remove the noise samples. Figure 5 As shown in the figure, this method can effectively clean the training samples automatically generated based on the unsupervised strategy and remove the noise samples.
[0052] Step 7, construct multiple RF classifiers to achieve spatial mapping of rapeseed. The present invention divides the study area into several 2°×2° grids, and trains RF classifiers in each grid separately. During the classification process, in order to maintain the integrity of the local block land cover type, the training samples of each local grid are provided by 3×3 grids (i.e., 6°×6° grids) centered on the grid. Figure 6 As shown, the present invention can realize the spatial distribution mapping of rapeseed with a resolution of 30 m in the whole Hubei Province based on Landsat images. The results show that rapeseed is mainly distributed in the Jianghan Plain areas such as Jingzhou, Jingmen, and Tianmen.
[0053] Step 8: Post-process the mapping results using local Matlab software and fill the holes in the mapping results using morphological methods. This mainly includes three steps:
[0054] The cultivated land layer was obtained by merging GlobeLand30 and ESA WorldCover data, and the non-cultivated land pixels in the rapeseed map were removed by masking. The former was obtained from http: / / www.globallandcover.com / home.html?type=data, and the latter was obtained from https: / / esa-worldcover.org / en / data-access;
[0055] (2) Calculate the terrain slope using 30-meter DEM data and mask pixels with slopes greater than 25° in the rapeseed map to remove noise pixels in the complex mountainous geographical environment. DEM data can be obtained from the geospatial data cloud;
[0056] (3) Filling holes in rapeseed maps based on morphological methods. The bwareaopen function in MATLAB R2022a was used to fill holes smaller than 1 hectare in rapeseed plots, preserving the spatial integrity of the rapeseed plots (https: / / www.mathworks.com / help / images / ref / bwareaopen.html).
[0057] The post-processing results of rapeseed remote sensing identification in the case of this invention are as follows: Figure 7 shown.
[0058] The above is only a preferred embodiment of the present invention, but the present invention is not limited to the described embodiments. For those skilled in the art, changes, modifications, substitutions and variations of the embodiments without departing from the principles and spirit of the present invention should all fall within the scope of protection of the present invention.
Claims
1. A method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning, characterized in that: The following steps are involved: Step 1: Collect multi-temporal Landsat optical satellite remote sensing images on the GEE remote sensing cloud computing platform, perform atmospheric correction and cloud detection preprocessing on the collected optical satellite remote sensing images, and obtain cloud-free time series optical surface reflectance data; Step 2: GLDAS data were collected on the GEE remote sensing cloud computing platform, 10 environmental variables closely related to rapeseed growth were selected as characteristic variables, and daily mean values of these characteristic variables were synthesized; Step 3: Based on field surveys and visual interpretation of high-resolution optical remote sensing images, representative rapeseed and non-rapeseed samples are obtained to provide basic data for rapeseed phenology monitoring and accuracy evaluation. The rapeseed samples obtained include the flowering date of rapeseed; Step 4, constructing an RFR model to train the nonlinear fitting relationship between the rapeseed flowering date obtained in step 3 and the environmental variables, estimating the rapeseed flowering period, and obtaining the rapeseed flowering phenology prior information; Step 5: Based on the rapeseed flowering phenology prior information obtained in step 4, the rapeseed index-CI method is used to automatically generate unsupervised rapeseed training samples for the study area; Step 6: Use the spectral angle mapping-SAM method to clean the training samples automatically generated in step 5 and remove the noise samples; Step 7: Based on the cleaned training samples obtained in step 6, spatial mapping of rapeseed is performed using a multi-RF fusion mapping framework on the GEE remote sensing cloud platform; Step 8: Post-process the mapping results through local Matlab software and use morphological methods to fill the holes in the mapping results.
2. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 1, the time period for acquiring the Landsat optical satellite remote sensing image is the rapeseed growing season of the area to be measured, including the image acquisition time of the winter rapeseed planting area and the image acquisition time of the spring rapeseed planting area. The image acquisition time period covers the sowing period, emergence period, opening period, bud period, flowering period, pod period, and harvest period of rapeseed.
3. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 2, the 10 characteristic variables selected from the GLDAS data are: Plant canopy surface water, Canopy water evaporation, Potential evaporation rate, Soil moisture (0-10 cm), Soil moisture (10-40 cm), Soil temperature (0-10 cm), Soil temperature (10-40 cm), Accumulated air temperature, Downward shortwave radiation, and Total precipitation rate.
4. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In the step 3, the time series ultra-high resolution optical remote sensing images are accessed with the help of Google Earth software. According to prior knowledge, rapeseed appears bright yellow in flowering period, which is obviously different from other crops of the same period. Whether the image contains rapeseed plots is determined by visual interpretation. A hexagonal uniform sampling strategy is used to achieve visual interpretation and manual labeling of samples. The size of a single grid is determined according to the surface coverage and cultivated land morphology of the study area, and sample points are randomly labeled in each grid. For grids that do not contain rapeseed plots, the number of non-rapeseed samples is 5; and for grids that contain rapeseed plots, 1 to 10 rapeseed samples are randomly labeled according to the proportion of rapeseed plots.
5. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 4, the accumulated temperature is calculated based on the difference between the cumulative daily average temperature and the benchmark temperature within the flowering period of winter rapeseed and spring rapeseed in the study area, which is used as one of the inputs of the RFR model; the total number of input features of the RFR model is 93, the mtry of the model is set to the square root of the total number of input variables, and the number of RF classifier trees is set to 100.
6. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 5, the expression of CI index is as follows: CI=(ρGreen+ρRed)×ρNIR (1) Among them, ρGreen, ρRed and ρNIR represent the surface reflectance values of pixels in the green band, red band and near-infrared band, respectively. The CI index is used to automatically segment rapeseed and non-rapeseed pixels, and the segmentation threshold is set to 0.
07. When the CI of a pixel is greater than 0.07, the pixel is judged to be rapeseed, otherwise it is non-rapeseed.
7. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 6, the SAM method is used to eliminate the noise components in the training samples. The spectrum of each pixel in the image is regarded as a high-dimensional vector. The similarity between the spectra of two vectors is measured by calculating the angle between them. The smaller the angle, the higher the similarity, and the greater the possibility that the two belong to the same category of objects. The potential category of the generated training samples is determined according to the size of the spectral angle between the generated rapeseed sample and the ground reference rapeseed sample. The spectral angle between the two vectors can be calculated by the cosine formula, as shown in formula (2): Where i represents a pixel, the spectrum of the ground reference rapeseed sample is r, the spectrum of the generated pseudo rapeseed sample is u, the spectral angle between them is α, and nb is the number of bands. If the spectral angle α between the generated rapeseed sample and the reference rapeseed sample is lower than the set threshold T, it is classified as a pure rapeseed pixel, otherwise it is a noise pixel, as shown in formula (3): Pseudo_rapeseed=Pure_rapeseed, if α<T=std(r) (3) Considering that T varies with the geographical environment of rapeseed cultivation, T of each sub-region is set to the standard deviation of the spectrum r of the ground reference rapeseed sample in the region to achieve cleaning and quality enhancement of training samples in a regionally adaptive manner.
8. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 7, the ntree parameter of each RF classifier in the RF fusion mapping framework is defined as 100, and the mtry parameter is set to the square of the total number of input features. The input features are the Landsat monthly median synthetic image and the derived VIs. The core of the multi-RF fusion mapping framework is to integrate the rapeseed classification probabilities of all phases to calculate the average probability of each pixel being classified as rapeseed, as shown in formula (4). If the probability of pixel i being classified as rapeseed is greater than the non-rapeseed probability, the pixel i is considered to be rapeseed: Among them, P i represents the probability that pixel i is classified as rapeseed, m is the number of cloud-free phases in the time series image, and p i,j is the probability that pixel i is classified as rapeseed in phase j.
9. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: In step 8, the specific method of post-processing the mapping results using Matlab software is as follows: a) removing non-cultivated land pixels in the rapeseed map by masking the cultivated land layer; b) masking pixels in the rapeseed map with a slope greater than 25° by using digital elevation model data; c) filling holes in the rapeseed map based on a morphological method.
10. The method for spatial mapping of rapeseed based on multi-source remote sensing data and machine learning as claimed in claim 1, characterized in that: The Landsat optical satellite remote sensing images collected in step 1 are optical satellite remote sensing images from Landsat-5, 7, 8, and 9 respectively.
Citation Information
Patent Citations
Rape planting plot remote sensing automatic identification method based on optical satellite image
CN111209871A
Mining area long-time-sequence surface soil water content remote sensing data production method and system
CN113591759A
