Machine Learning Methods and Systems for Extracting Crop Straw from Multispectral Remote Sensing Images

Abstract

A machine learning-based method and system for extracting crop straw from multispectral remote sensing images addresses the problems of human error and measurement error in existing methods of monitoring straw removal area and progress, and the inability to accurately control the location of areas not yet removed from the field. The method includes the following steps: Step S1: Acquire multispectral image data during the crop growth period, delineate and establish a polygon sample set based on crop spectral information, and build a crop classification model from this set. The established crop classification model predicts the crop distribution results. Step S2: Acquire multispectral image data after the crop harvest period, delineate polygon samples for the corresponding categories, and perform sampling processing to obtain a polygon sample set of removed plots. Step S3: Divide the polygon sample set of removed plots using a five-fold partitioning method. The resulting five training and validation sets are used to train five base models. These base models are used to predict the image data, and the prediction results are used to determine the removed plots through probability mean fusion.

Description

Technical Field

[0001] This invention relates to the field of multispectral satellite remote sensing image data mining technology, specifically to a method and system for extracting off-field plots from multispectral remote sensing images using machine learning. Background Technology

[0002] With the rapid development of agriculture in my country, crop straw management has gradually become an important part of agriculture. Currently, the main method of straw management is straw removal from the field. There are two main methods: one is to chop or crush the straw, spread it evenly on the surface, and then use a high-powered tractor to plow or rotary till it deep into the soil; the other is to directly bale and transport the straw. Removing straw from the field not only reduces air pollution caused by direct burning of straw but also achieves the recycling of straw, which is of great significance for agricultural environmental protection and promoting increased production and income for farmers.

[0003] Currently, agricultural and environmental protection departments have implemented various methods to supervise the removal of farmland straw from the fields. By urging and supervising the removal process, the utilization rate and removal rate of straw have been effectively improved, reducing illegal burning. However, the main method for obtaining information on straw removal is currently through farmers or cooperatives compiling and reporting data themselves. The main drawback of this method is:

[0004] 1) It is impossible to accurately measure the area and progress of straw removal from the field, and there are certain human and measurement errors.

[0005] 2) The inability to locate areas that have not left the fields makes it impossible to achieve precise control.

[0006] In summary, although the existing methods for regulating straw removal from the field can effectively improve the utilization rate and removal rate of straw and reduce the phenomenon of illegal burning of straw, there are certain human and measurement errors in monitoring the area and progress of straw removal from the field, and it is impossible to accurately control the location of areas that have not been removed from the field. Summary of the Invention

[0007] This invention solves the problems of existing methods for monitoring straw removal from the field, which suffer from human error and measurement error in terms of the area and progress of straw removal, and are unable to achieve precise control over the location of areas that have not yet been removed from the field.

[0008] The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning as described in this invention includes the following steps:

[0009] Step S1: Obtain multispectral image data of crops during their growth period, delineate and establish a polygonal sample set based on the crop spectral information, establish a crop classification model based on it, and use the established crop classification model to predict the crop distribution results.

[0010] Step S2: Obtain multispectral image data after crop harvest, delineate polygon samples of the corresponding categories, and perform sampling processing to obtain a polygon sample set of land parcels away from the field.

[0011] Step S3: Divide the polygonal sample set of the land parcels away from the field into five groups of training sets and validation sets. After training, five base models are obtained. Use the base models to predict the image data. The prediction results are used to determine the land parcels away from the field by probability mean fusion.

[0012] Furthermore, in one embodiment of the present invention, in step S1, acquiring multispectral image data of the crop growth period, the multispectral image data refers to data with the following spectral bands:

[0013] The reflectance data for 10 bands, namely B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, are provided, and the multispectral image data is in the form of unsigned 16-bit integers.

[0014] Furthermore, in one embodiment of the present invention, step S1, which involves delineating and establishing a polygonal sample set based on crop spectral information,

[0015] The extracted crops include four categories: corn, rice, other crops, and other land cover crops.

[0016] Delineating and establishing a polygonal sample set refers to delineating and establishing 50 to 100 polygonal samples for each of the above-mentioned crop categories to form a sample set.

[0017] Furthermore, in one embodiment of the present invention, the establishment of the crop classification model in step S1 is based on the LightGBM framework.

[0018] Furthermore, in one embodiment of the present invention, in step S2, the polygon sample corresponding to the category is drawn, and the category includes:

[0019] The land is categorized into three types: land away from the field, land not away from the field, and other land features.

[0020] Furthermore, in one embodiment of the present invention, the sampling process described in step S2 includes the following steps:

[0021] Step S201: The band value of each pixel in a polygonal sample is calculated as the basic band feature value, and the corresponding category spectral index feature value is calculated based on the band feature value.

[0022] Step S202: Normalize the spectral index eigenvalues.

[0023] Furthermore, in one embodiment of the present invention, the method for dividing the polygonal sample set of land parcels using five-fold partitioning in step S3 is as follows:

[0024] The polygonal sample set of land parcels was divided into a training set and a validation set in a 4:1 ratio.

[0025] Furthermore, in one embodiment of the present invention, the prediction result in step S3 is fused using probability mean as follows:

[0026] The block probability matrices predicted by the five base models are re-merged, and the probability matrices are added together and the mean is calculated to obtain the fused total probability matrix.

[0027] The machine learning-based multispectral remote sensing image crop straw off-field extraction system of the present invention includes the following modules:

[0028] Module S1 is used to acquire multispectral image data of crops during their growth period, delineate and establish polygonal sample sets based on crop spectral information, build a crop classification model from these samples, and then use the established crop classification model to predict the crop distribution results.

[0029] Module S2 is used to acquire multispectral image data after crop harvest, delineate polygon samples of the corresponding categories, perform sampling processing, and obtain polygon sample sets of land parcels away from the field.

[0030] Module S3 is used to divide the polygonal sample set of off-field plots using five-fold partitioning. The resulting five training sets and validation sets are used to train five base models. It is also used to use the base models to predict image data. The prediction results are used to determine off-field plots through probability mean fusion.

[0031] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the method described in any of the above methods.

[0032] The present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, when the processor runs the computer program stored in the memory, it executes the steps of any one of the methods described above.

[0033] This invention solves the problems of existing methods for monitoring straw removal from fields, which suffer from human and measurement errors in terms of area and progress, and cannot achieve precise control over the location of areas that have not yet been removed from the fields. Specific beneficial effects include:

[0034] 1. This invention is the first to propose using remote sensing satellite imagery to monitor both off-field and non-off-field plots, opening up a new avenue for monitoring straw removal from fields.

[0035] 2. This invention extracts data from remote sensing satellite images of land parcels that have been removed from the field and those that have not, providing crucial data support for monitoring the area and progress of land removal. This solves the problem of measurement errors in the area and progress of land removal using existing methods for monitoring straw removal.

[0036] 3. By extracting the areas of land removed from and not removed from remote sensing satellite images, this invention not only achieves precise monitoring of the areas of land removed from and not removed from the fields, reducing the cost of existing manual monitoring methods, but also effectively monitors the progress of straw removal from the fields and prevents illegal burning of straw. At the same time, it is of great significance for the efficient utilization of agricultural resources and environmental protection.

[0037] 4. This invention extracts spectral and exponential features from multispectral satellite remote sensing images and uses machine learning to establish a straw removal extraction model, solving the problem of being unable to distinguish between off-field and non-off-field plots.

[0038] This invention is applicable to the field of remote sensing large-area straw removal monitoring technology. By establishing a straw removal detection model, it achieves automated extraction, solving the technical problem of low straw removal interpretation efficiency over large areas, greatly reducing the amount of manual intervention, and enabling rapid extraction of straw removal data after data acquisition to achieve effective control of the removal progress. Attached Figure Description

[0039] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein:

[0040] Figure 1 This is a flowchart illustrating the construction of a method for extracting crop straw off-field plots based on machine learning from multispectral remote sensing images, as described in this invention.

[0041] Figure 2 It is the band composite color image described in Implementation Method 5.

[0042] Figure 3 The images shown are prediction results of the straw removal plot classification model described in Implementation Method 10: (a) Original Image 1 (color), (b) Original Image 1 (removal result), (c) Original Image 2 (color), and (d) Original Image 2 (removal result). Detailed Implementation

[0043] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0044] Implementation Method 1: The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning described in this implementation method is based on... Figure 1 To better understand this implementation method, the following steps are included:

[0045] Step S1: Obtain multispectral image data of crops during their growth period, delineate and establish a polygonal sample set based on the crop spectral information, establish a crop classification model based on it, and use the established crop classification model to predict the crop distribution results.

[0046] Step S2: Obtain multispectral image data after crop harvest, delineate polygon samples of the corresponding categories, and perform sampling processing to obtain a polygon sample set of land parcels away from the field.

[0047] Step S3: Divide the polygonal sample set of the land parcels away from the field into five groups of training sets and validation sets. After training, five base models are obtained. Use the base models to predict the image data. The prediction results are used to determine the land parcels away from the field by probability mean fusion.

[0048] In this embodiment, extraction is performed on the main straw-producing crops in cultivated land, with corn and rice being the two main straw-producing crops.

[0049] The growing season for corn and rice straw is from July to August, and the harvest season for corn and rice is from October to November. By October and November, the corn and rice harvest has been completed. Collecting corresponding multispectral image data based on these growth cycles is more conducive to the extraction of data from areas that have not yet left the fields.

[0050] Implementation Method Two: This implementation method is based on the machine learning-based multispectral remote sensing image crop straw removal plot extraction method described in Implementation Method One. In this implementation method, in step S1, acquiring multispectral image data of the crop growth period, the multispectral image data refers to data with the following bands:

[0051] The reflectance data for 10 bands, namely B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, are provided, and the multispectral image data is in the form of unsigned 16-bit integers.

[0052] In this embodiment, reflectance products from 10 bands of multispectral imagery are selected because the spectral characteristics of straw removed from the field and not removed from the field differ within these bands. Furthermore, the machine learning algorithm used in this method performs well with multiple features and can automatically reduce the weight of features that do not improve model accuracy. By extracting features from these 10 bands and combining them, the overall accuracy of straw removal prediction can be improved.

[0053] Implementation Method 3: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method 1. In this implementation method, step S1 involves delineating and establishing a polygon sample set based on crop spectral information.

[0054] The extracted crops include four categories: corn, rice, other crops, and other land cover crops.

[0055] Delineating and establishing a polygonal sample set refers to delineating and establishing 50 to 100 polygonal samples for each of the above-mentioned crop categories to form a sample set.

[0056] In this embodiment, the band feature value of each pixel is extracted as a training set based on the drawn sample polygon.

[0057] Implementation Method 4: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method 1. In this implementation method, the crop classification model established in step S1 is based on the LightGBM framework.

[0058] In this embodiment, a crop classification model is established based on the LightGBM framework, and this model is used to predict the distribution of corn and rice crops in the target area using satellite imagery, resulting in the original straw distribution. This crop distribution is then used to mask the straw extraction results in subsequent steps, ensuring the accuracy of the extracted straw.

[0059] Implementation Method 5: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method 1. In this implementation method, the delineation of polygonal samples corresponding to the category in step S2 is as follows:

[0060] The land is categorized into three types: land away from the field, land not away from the field, and other land features.

[0061] In this implementation, "other land features" need to be delineated to include primary categories such as water bodies, impermeable surfaces, and woodlands, in order to exclude the influence of irrelevant land features in the classification model.

[0062] Specifically, to enhance the difference between straw removed from the field and straw still in the field, and to make the delineation results more accurate, a true-color image was reconstructed using band combination. The specific method was as follows: band B12 was subtracted from band B11 to obtain the red band, B9 was used as the green band, and B10 as the blue band. Then, a stretching process was performed using a standard deviation of n=2 to obtain the image. Figure 2 The image shown is a composite band diagram. In the diagram, the bright yellow areas represent plots of land that have not yet been removed from the field, while the dark yellow areas represent plots of land that have been removed to varying degrees. Using this composite band diagram, it is possible to clearly distinguish between plots that have been removed from the field and those that have not.

[0063] Implementation Method Six: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method One. In this implementation method, the sampling processing in step S2 includes the following steps:

[0064] Step S201: The band value of each pixel in a polygonal sample is calculated as the basic band feature value, and the corresponding category spectral index feature value is calculated based on the band feature value.

[0065] Step S202: Normalize the spectral index eigenvalues.

[0066] In this embodiment, it is first necessary to read each polygon sample in the vector data, obtain the image coordinates of all pixels within each sample polygon, calculate the band value of each pixel within a sample, and use it as the basic band feature. Then, various spectral index features are calculated based on the band feature.

[0067] 1) Normalized Diversified Tillage Index (NDTI)

[0068]

[0069] 2) Normalized Difference Ratio (NDRI)

[0070]

[0071] 3) Simple Tillage Index (STI)

[0072]

[0073] 4) Normalized Difference Index (NDI7)

[0074]

[0075] In the formula, B11 and B12 represent the shortwave infrared bands that are more sensitive to off-field plots. B4 and B8 represent the red band and near-infrared band, respectively. Adding the following spectral index features will effectively improve the accuracy of the model. Combining the band features with the spectral index model yields a multi-feature sample dataset.

[0076] To ensure the model's generalization ability, the feature values ​​need to be normalized. Specifically, the band features are normalized to a range of 0 to 1 using a maximum-minimum method. The normalization formula is as follows:

[0077]

[0078] In the formula, X max X minThese represent the maximum and minimum spectral values ​​of the region, respectively, and X represents each spectral value. new This represents the normalized spectral value. Since the spectral index feature is in ratio form, it does not require normalization. Based on the above steps, a sample set containing 14 features across multiple categories can be obtained.

[0079] Implementation Method Seven: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method One. In this implementation method, the method for dividing the polygonal sample set of off-field plots using a five-fold division in step S3 is as follows:

[0080] The polygonal sample set of land parcels was divided into a training set and a validation set in a 4:1 ratio.

[0081] In this embodiment, in order to make full use of all data, a five-fold partitioning method is used to create the dataset. This method can ensure that all datasets participate in the model training process.

[0082] Implementation Method Eight: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method Seven. In this implementation method, the method for obtaining five sets of training and validation sets in step S3 is as follows:

[0083] We take five training sets and validation sets. Each time, we take 4 sets as the training set and 1 set as the validation set, resulting in five sets of training and validation sets.

[0084] Implementation Method Nine: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method Seven. In this implementation method, the method for obtaining the five base models in step S3 is as follows:

[0085] The five training-validation sets were used to build classification models for each set using the LightGBM framework. The models were trained for 150 epochs with a learning rate of 0.1, a feature sampling probability of 50%, and a sample sampling probability of 50%.

[0086] In this embodiment, five base models are obtained after training. After model training, the five base models are used to predict the image data separately. Since the full-domain image is large, it cannot be directly read into memory for processing. Therefore, a block-based prediction method is needed to reduce memory consumption. First, after reading the original remote sensing image, according to the preset block size and block overlap rate (generally, a block size of 4096 and an overlap rate of 10% are selected), the remote sensing image data within the block is generated into a smaller image, and the dataset is flattened into a single pixel, single sample. Then, it is input into the model to predict the probability matrix of each block.

[0087] Implementation Method 10: This implementation method is based on the machine learning method for extracting crop straw off-field plots from multispectral remote sensing images described in Implementation Method 1. In this implementation method, the prediction results in step S3 are fused using probability mean as follows:

[0088] The block probability matrices predicted by the five base models are re-merged, and the probability matrices are added together and the mean is calculated to obtain the fused total probability matrix.

[0089] In this embodiment, the block probability matrices predicted by the five models are re-merged, and the sum of the probability matrices is calculated to obtain the fused total probability matrix. The category value of each pixel is determined according to the index of the maximum value of the probability matrix, and the off-field plots in the target area are determined according to the category. A masking process is applied to the extracted corn and rice crop results to remove non-corn and rice crop patches, resulting in more accurate off-field plot results, such as... Figure 3 The extraction results of the two sample images shown are as follows: (a) original image 1 (color), (b) original image 1 (extraction result after leaving the field), (c) original image 2 (color), and (d) original image 2 (extraction result after leaving the field).

[0090] In Figures (b) and (d), the black area represents the off-field plots extracted by the model, and the white area represents the off-field plots and other non-cultivated land areas.

[0091] Select a suitable projection coordinate system based on the location of the target area, calculate the area of ​​the off-farm plots, and obtain the off-farm rate for each village and town according to the off-farm rate calculation formula, where the off-farm rate is calculated as follows:

[0092]

[0093] Implementation Method 11: The machine learning-based multispectral remote sensing image crop straw removal plot extraction system described in this implementation method includes the following modules:

[0094] Module S1 is used to acquire multispectral image data of crops during their growth period, delineate and establish polygonal sample sets based on crop spectral information, build a crop classification model from these samples, and then use the established crop classification model to predict the crop distribution results.

[0095] Module S2 is used to acquire multispectral image data after crop harvest, delineate polygon samples of the corresponding categories, perform sampling processing, and obtain polygon sample sets of land parcels away from the field.

[0096] Module S3 is used to divide the polygonal sample set of off-field plots using five-fold partitioning. The resulting five training sets and validation sets are used to train five base models. It is also used to use the base models to predict image data. The prediction results are used to determine off-field plots through probability mean fusion.

[0097] Implementation Method Twelve: A computer-readable storage medium according to this implementation method stores a computer program thereon. In this implementation method, when the computer program is executed by a processor, it implements the steps of the method described in any one of the implementation methods one to ten.

[0098] Implementation Method Thirteen: A computer device includes a memory and a processor. The memory stores a computer program. In this implementation method, when the processor runs the computer program stored in the memory, it executes the steps of the method described in any one of Implementation Methods One to Ten.

[0099] The above embodiments are illustrative examples of the method for extracting crop straw off-field plots based on machine learning from multispectral remote sensing images according to the present invention. The scope of protection of the present invention also includes reasonable combinations of the features defined in the above embodiments.

Claims

1. A machine learning-based method for extracting crop straw off-field plots from multispectral remote sensing images, characterized in that, Includes the following steps: Step S1: Acquire multispectral image data of the crop growth period, delineate and establish a polygonal sample set based on the crop spectral information, and build a crop classification model from it. The established crop classification model predicts the crop distribution results. In step S1, the multispectral image data refers to data with the following spectral bands: The reflectance data for 10 bands, including B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, are provided, and the multispectral image data is in the form of unsigned 16-bit integers. Step S2 involves acquiring multispectral image data after crop harvest, outlining polygonal samples of the corresponding categories, and then performing sampling processing to obtain a polygonal sample set of off-field plots. Step S2 also includes reconstructing a true-color image using a band combination method to enhance the difference between off-field straw and non-off-field straw and make the outlining results more accurate. Specifically, band B12 is subtracted from band B11 to obtain the red band, B9 is used as the green band, and B10 is used as the blue band. The image is then stretched using a standard deviation of n=2 to obtain a composite band image. This composite band image is then used to distinguish between off-field plots and non-off-field plots. Step S3: Divide the polygonal sample set of the land parcels away from the field into five groups of training sets and validation sets. After training, five base models are obtained. Use the base models to predict the image data. The prediction results are used to determine the land parcels away from the field by probability mean fusion.

2. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, The step S1, which involves delineating and establishing a polygonal sample set based on crop spectral information, includes four categories of crops: corn, rice, other crops, and other land features. Delineating and establishing a polygonal sample set means delineating and establishing 50 to 100 polygonal samples for each of the above categories of crops to form a sample set.

3. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, The crop classification model established in step S1 is based on the LightGBM framework.

4. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, In step S2, the polygon samples corresponding to the categories are outlined, and the categories include: The land is categorized into three types: land away from the field, land not away from the field, and other land features.

5. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, The sampling process described in step S2 includes the following steps: Step S201: The band value of each pixel in a polygonal sample is calculated as the basic band feature value, and the corresponding category spectral index feature value is calculated based on the band feature value. Step S202: Normalize the spectral index eigenvalues.

6. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, The method for dividing the polygonal sample set of off-field plots using five-fold partitioning in step S3 is as follows: The polygonal sample set of land parcels was divided into a training set and a validation set in a 4:1 ratio.

7. The method for extracting crop straw off-field plots from multispectral remote sensing images using machine learning according to claim 1, characterized in that, The prediction results in step S3 are fused using probability mean as follows: The block probability matrices predicted by the five base models are re-merged, and the probability matrices are added together and the mean is calculated to obtain the fused total probability matrix.

8. A machine learning-based system for extracting crop straw from off-field plots in multispectral remote sensing images, characterized in that... The system includes the following modules: Module S1 is used to acquire multispectral image data of crops during their growth period, delineate the claims based on the crop spectral information, establish a polygonal sample set, build a crop classification model from it, and predict the crop distribution results using the established crop classification model; in the acquisition of multispectral image data of crops during their growth period by module S1, the multispectral image data refers to reflectance data of 10 bands: B2, B3, B4, B5, B6, B7, B8, B8A, B11, and B12, and the multispectral image data is in the form of unsigned 16-bit integers; Module S2 is used to acquire multispectral image data after crop harvest, delineate polygonal samples of the corresponding categories, and perform sampling processing to obtain a polygonal sample set of off-field plots. Module S2 also includes a method to reconstruct true-color images using band combination to enhance the difference between off-field straw and non-off-field straw and make the delineation results more accurate. Specifically, band B12 is subtracted from band B11 to obtain the red band, B9 is used as the green band, and B10 is used as the blue band. The images are then stretched using a standard deviation of n=2 to obtain a composite band image. This composite band image is then used to distinguish between off-field plots and non-off-field plots. Module S3 is used to divide the polygonal sample set of off-field plots using five-fold partitioning. The resulting five training sets and validation sets are used to train five base models. It is also used to use the base models to predict image data. The prediction results are used to determine off-field plots through probability mean fusion.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method described in any one of claims 1 to 7.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program stored in the memory, it performs the steps of the method described in any one of claims 1 to 7.

Images