A cultivated land planting artificial turf identification method and system based on time-series SAR images

Abstract

The application discloses a cultivated land planting artificial turf identification method and system based on time sequence SAR images, and the identification prediction model is jointly constructed by using multi-temporal SAR remote sensing images and field sample polygons obtained by field investigation, the model requires less sample amount for target input, and the time-consuming and labor-consuming traditional artificial method and the deep learning method requiring a large number of sample accumulations are avoided; meanwhile, since the method uses multi-temporal SAR remote sensing images which can be all-day, all-weather and full-coverage as data support, not only the dependence on high-resolution optical images which are easy to be lost in traditional optical images in south China and other cloudy and rainy areas can be avoided, but also the application range can be easily extended to other areas.

Description

Technical Field

[0001] This invention relates to the field of remote sensing identification technology, and in particular to a method and system for identifying artificial turf planted in cultivated land based on time-series SAR imagery. Background Technology

[0002] The planting of artificial turf on cultivated land is a key focus of satellite imagery monitoring of cultivated land. Timely and accurate understanding of the distribution and area of ​​artificial turf is crucial for monitoring the "non-grain" conversion of cultivated land. Current automated remote sensing monitoring of cultivated land conversion focuses primarily on monitoring and extracting typical broad-based changes, such as building houses on cultivated land or planting trees and digging ponds on basic farmland. There is little attention paid to the more specific change of planting artificial turf on cultivated land, and no research or application has been found using SAR imagery data for artificial turf extraction and monitoring. This lack of research is not due to the scarcity of artificial turf planting on cultivated land; on the contrary, this phenomenon is widespread in many rural areas, but it is often overlooked because researchers rarely conduct field surveys. Artificial turf on cultivated land is often visually classified by remote sensing image interpreters as rice paddies, grasslands, or bare land, rather than being studied and differentiated separately.

[0003] Currently, most monitoring methods for the conversion of arable land to non-grain crops are still limited to optical imagery. However, simply transferring this method to the monitoring of artificial turf fields has significant limitations. The visible light characteristics of artificial turf are very similar to traditional land types such as rice paddies, grasslands, and bare land. Planted turf often appears as green, uniform rectangular fields in optical images, while harvested turf often appears as brown rectangular fields similar to bare land. When such fields appear in large quantities, experienced professionals can visually interpret them to distinguish them from rice paddies and grasslands. However, manually viewing optical remote sensing images to identify arable land converted to artificial turf can only obtain initial change patches, requiring further manual field verification for confirmation. This method has the following problems: First, although manual field verification has a high accuracy rate, it is extremely labor-intensive, resource-intensive, and time-consuming, failing to maximize economic and social benefits. Secondly, traditional methods require manual visual extraction of optical remote sensing images for initial change patch acquisition. In South China, where visibility is easily affected by low-visibility weather conditions such as clouds, rain, and fog, data acquisition is unstable and cannot obtain optical images in a timely manner, hindering timely monitoring of illegal occupation of farmland for artificial turf planting. Finally, manual visual interpretation of optical images in the office faces problems such as the easy misidentification of artificial turf image features by humans, and inconsistent optical image quality leading to misidentification. Additionally, while better deep learning methods can achieve accuracy similar to human judgment, the limited accumulation of relevant turf samples makes training related deep learning models difficult. Summary of the Invention

[0004] To address the aforementioned problems, this invention proposes a method and system for identifying artificial turf planted in cultivated land based on temporal SAR imagery, primarily resolving the issues raised in the background technology.

[0005] To address the aforementioned technical problems, the first aspect of this invention proposes a method for identifying artificial turf planted on cultivated land based on temporal SAR imagery, comprising the following steps:

[0006] S1, acquire multi-temporal SAR remote sensing images of the target area, and obtain SAR images after preprocessing;

[0007] S2, acquire high-resolution visible light images of the target area during a clear, cloudless period, and use a field segmentation program to segment all fields within the target area;

[0008] S3, Field survey to obtain field sample maps of the target area;

[0009] S4, calculate the centroid of the field plots within the target area, and label the field plots whose centroids fall within the range of the field sample plots as the corresponding plot categories to form field plot samples;

[0010] S5, extract the statistical feature attributes of each period of the SAR image within the field sample;

[0011] S6, the dataset composed of the statistical feature attributes is divided into a training set and a validation set. The validation set is used to evaluate the performance of XGBoost models with different hyperparameters. The evaluation index of the validation set is used to select the optimal model to form a recognition and prediction model.

[0012] S7. Input the field sample into the identification and prediction model to obtain the patch category corresponding to the field sample.

[0013] In some implementations, in S1, the SAR image is VH backscattering coefficient data.

[0014] In some implementations, in S4, a field with an area of ​​400 square meters or more is selected from the field sample as the final field sample.

[0015] In some implementations, in S5, the statistical feature attributes include the maximum pixel value, minimum pixel value, and average pixel value of the SAR image.

[0016] A second aspect of this invention proposes a system for identifying artificial turf planted on cultivated land based on temporal SAR imagery, comprising:

[0017] The multi-temporal SAR remote sensing image acquisition module is used to acquire multi-temporal SAR remote sensing images of the target area, and obtain SAR images after preprocessing.

[0018] The field segmentation module is used to acquire high-resolution visible light images of the target area during clear, cloudless periods, and to segment all the fields within the target area using the field segmentation program;

[0019] The field sample map collection module is used to collect field sample maps of the target area obtained from field surveys;

[0020] The field sample labeling module is used to calculate the centroid of the field within the target area, and label the field with the centroid falling within the range of the field sample patch as the corresponding patch category to form a field sample;

[0021] The feature extraction module is used to extract the statistical feature attributes of each period of the SAR image within the field sample;

[0022] The model building module is used to divide the statistical feature attributes into a training set and a validation set, classify the field samples using the training set based on the XGBoost model, and train the XGBoost model using the statistical feature attributes to form a recognition and prediction model.

[0023] An automatic identification module is used to input the field sample into the identification and prediction model to obtain the patch category corresponding to the field sample.

[0024] In some implementations, in the multi-temporal SAR remote sensing image acquisition module, the SAR image is VH backscattering coefficient data.

[0025] In some implementations, in the field sample labeling module, fields with an area of ​​more than 400 square meters are selected from the field samples as the final field samples.

[0026] In some implementations, the statistical feature attributes in the feature extraction module include the maximum pixel value, minimum pixel value, and average pixel value of the SAR image.

[0027] The beneficial effects of this invention are as follows: by using multi-temporal SAR remote sensing images combined with field sample patches obtained from field surveys to jointly construct an identification and prediction model, the target input of this model requires a small number of samples, avoiding the time-consuming and labor-intensive traditional manual methods and the deep learning methods that require a large number of samples to accumulate; at the same time, since this method uses multi-temporal SAR remote sensing images that can cover all weather conditions all day long as data support, it can not only avoid the dependence of traditional optics on high-resolution optical images that are easily missing in areas such as South China with cloudy and rainy weather, but also make it easy to extend the application scope to other regions. Attached Figure Description

[0028] Figure 1This is a flowchart illustrating the method for identifying artificial turf planted in cultivated land based on temporal SAR imagery disclosed in this invention.

[0029] Figure 2 The flowchart for the XGBoost algorithm;

[0030] Figure 3 This is a flowchart illustrating a verification example in Embodiment 1 of the present invention;

[0031] Figure 4 To verify the category proportion of the field samples in Jiexi County, Guangdong Province in the example;

[0032] Figure 5 To verify the category proportion of the field samples in Xinhui District, Guangdong Province in the example;

[0033] Figure 6 To verify the classification average time series plot of the SAR image feature mean of field samples within the cultivated land area in the example;

[0034] Figure 7 The confusion matrix diagram of the validation set for the model based on SAR image features in Jiexi County, Guangdong Province, in the validation example;

[0035] Figure 8 The confusion matrix diagram of the validation set of the model based on SAR image features in Xinhui District, Guangdong Province, is used to verify the example.

[0036] Figure 9 The confusion matrix diagram of the validation set for the model based on optical imagery + SAR imagery features in Jiexi County, Guangdong Province, in the validation example;

[0037] Figure 10 The confusion matrix diagram of the validation set for the model based on optical imagery + SAR imagery features in Xinhui District, Guangdong Province, in the validation example;

[0038] Figure 11 To verify the spatial distribution map and local magnified schematic diagram of the artificial turf extraction results in Jiexi County, Guangdong Province in the example;

[0039] Figure 12 To verify the spatial distribution map and local magnified schematic diagram of the artificial turf extraction results in Xinhui District, Guangdong Province in the example. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the content of this invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to this invention are shown in the accompanying drawings, not all of them.

[0041] Example 1

[0042] like Figure 1 As shown in the figure, this embodiment proposes a method for identifying artificial turf planted on cultivated land based on temporal SAR imagery, including the following steps:

[0043] S1: Acquire multi-temporal SAR remote sensing images of the target area, and obtain SAR images after preprocessing.

[0044] In S1, the SAR imagery consists of VH backscattering coefficient data. Specifically, GRD data from C-band Level 1 interferometric wide swath (IW) observations conducted by Sentinel-1A satellite from January to October at a resolution of 10 meters were acquired. This data has a 12-day revisit period and comprises 25 scenes. After orbit file correction, thermal noise removal, radiometric calibration, speckle noise filtering (Lee-Sigma), and terrain correction, the required VH backscattering coefficient (dB) data was obtained.

[0045] S2. Acquire high-resolution visible light images of the target area during a clear, cloudless period (not necessarily limited to the SAR observation period), and use a field segmentation program to segment all fields within the target area. Since the timeliness requirement for optical images is not strong in this scheme, observations from last year or previous years can be used.

[0046] S3, field survey to obtain field sample patches of the target area.

[0047] S4. Calculate the centroid of the fields within the target area, and label the fields whose centroids fall within the range of the field sample patches as the corresponding patch categories to form field samples.

[0048] Data processing in S4 converts the field sample patch data collected in S3 into field samples with category labels for subsequent feature extraction. To ensure the integrity and representativeness of the samples, fields larger than 400 square meters (i.e., at least 4 Sentinel-1A SAR image pixels) are selected as the final field samples.

[0049] S5 extracts the statistical feature attributes of each period of SAR image within the field sample.

[0050] In S5, statistical feature attributes include the maximum, minimum, and average pixel values ​​of SAR images.

[0051] S6 divides the dataset composed of statistical feature attributes into a training set and a validation set. Based on the XGBoost model, the training set is used to classify field samples, the validation set is used to evaluate the performance of XGBoost models with different hyperparameters, and the evaluation index of the validation set is used to select the optimal model to form a recognition and prediction model.

[0052] In S6, the statistical feature attributes are divided into training and validation sets in a 7:3 ratio. The only hyperparameter adjustment for the XGBoost model is the number of trees: 10, 20, 50, 100, 300, 500, and 1000. All other hyperparameters use the default values ​​from the Python third-party library XGBoost. The XGBoost algorithm is a prediction algorithm that uses multiple decision trees. Its flowchart is shown below. Figure 2 As shown, the basic prediction steps are as follows: the first decision tree (Tree1) takes X as input and predicts θ1, obtaining the first predicted value f1 and the residual f2 = θ1 - f1. The second decision tree (Tree2) then predicts θ2, obtaining the predicted value f2 and the new residual f3 = θ2 - f2. This process is repeated for the residuals of the previous decision tree until all decision trees have been predicted. At this point, the predicted value of the XGBoost algorithm is the sum of the predicted values ​​of all individual decision trees, i.e., ∑f k (X,θ k ).

[0053] S7. Input the field sample into the recognition and prediction model to obtain the patch category corresponding to the field sample.

[0054] The effectiveness of this solution is illustrated below through two sets of verification examples. (See attached document.) Figure 3 The framework diagram shown is as follows:

[0055] S1. GRD data from C-band Level 1 interferometric wide swath (IW) observations of two target areas, Jiexi County and Xinhui District, Guangdong Province, from January to October 2022, with a multi-temporal resolution of 10 meters, were acquired. The data had a repeat period of 12 days and consisted of 25 scenes. After orbit file correction, thermal noise removal, radiometric calibration, speckle noise filtering (Lee-Sigma), and terrain correction, the VH backscattering coefficient (dB) data required for the study were obtained.

[0056] S101 involves acquiring high-resolution remote sensing images of the target areas in Jiexi County and Xinhui District, taken from January to October after preprocessing optical remote sensing satellite data. For Jiexi County, high-resolution images of 0.5m resolution across four bands from the Jilin-1 satellite from May 2022 are used. Field features of the target areas are extracted based on the acquired optical images. In S101, the acquired optical remote sensing satellite images and field features are compared and verified with the VH backscattering coefficient data acquired in S1 above.

[0057] S2: Acquire high-resolution remote sensing images of the target areas in Jiexi County and Xinhui District from January to October after preprocessing optical remote sensing satellite data. Use a field segmentation program to segment all fields within the target area. Extract field features from the acquired optical images of the target area. In S101, the acquired optical remote sensing satellite images and field features are compared and verified with the VH backscattering coefficient data acquired in S1 above.

[0058] S3, obtain field sample patches based on the sample data collected in the field survey.

[0059] S4 reclassifies various land cover types from the field survey into three categories: artificial turf, rice, and others. Then, the field samples are converted into field sample data. The conversion method assigns the corresponding field sample patch category to field plots whose centroids fall within the field sample patch area. Based on this, 11,208 field samples were obtained in Jiexi County and 3,815 in Xinhui District, with each category accounting for a certain percentage. Figure 4 and Figure 5 As shown, to ensure the integrity and representativeness of the samples, patches of more than 400 square meters (i.e., at least 4 Sentinel-1 ASAR image pixels) were selected, resulting in 3,747 field samples in Jiexi County and 3,311 field samples in Xinhui District.

[0060] S5 extracts the maximum, minimum, and mean pixel values ​​from each of the 25 periods of SAR images within each type of sample field, totaling 75 items (3 items × 25 periods). Among these, the time series of the mean SAR images for the three major categories of samples is included. Figure 6 As shown.

[0061] S501: To verify the effectiveness of SAR image features, and since the optical image acquired in S2 happened to be within the SAR observation period, its field statistical features were extracted for control experiments.

[0062] Here, field features (hereinafter collectively referred to as optical image features) consisting of three parts, including pixel statistics, pixel histogram statistics, and gray-level co-occurrence matrix statistics, were selected for the control experiment.

[0063] These include:

[0064] Pixel statistics: maximum, minimum, mean, mode, median, number of pixels, deviation, 10th percentile, 90th percentile, and standard deviation of optical remote sensing images within a field.

[0065] Cell histogram statistics: histogram skewness and kurtosis;

[0066] Statistical values ​​of the gray-level co-occurrence matrix: homogeneity, energy, and correlation of the gray-level co-occurrence matrix.

[0067] The grayscale co-occurrence matrix selects the RGB composite band and the near-infrared band. In order to reduce feature redundancy, the RGB composite band is obtained by combining the three bands into a grayscale band using the classic formula 0.299×Red+0.587×Green+0.114×Blue.

[0068] Two directions and three distances were selected for gray-scale co-occurrence calculation, namely 90° and 180°; the three distances were 1, 3 and 5 pixel units in length.

[0069] The gray-level co-occurrence matrix features only select three indicators: homogeneity, energy, and correlation, in order to reduce feature redundancy.

[0070] S6 divides the dataset consisting of field sample data into a training set and a validation set in a 7:3 ratio.

[0071] The XGBoost algorithm was used to classify the statistical attributes of the samples. The only hyperparameter of the XGBoost model was adjusted, which was 10, 20, 50, 100, 300, 500, and 1000. All other hyperparameters used their default values.

[0072] S601, Experiment with controlled variables:

[0073] ① The confusion matrix results for the two regions are as follows, using only the statistical attribute features of time-series SAR image data to train the model. Figure 7 and Figure 8 As shown: ② The confusion matrix results for the two regions, constructed by combining SAR imagery and optical imagery features, are as follows. Figure 9 and Figure 10 As shown. ③ The model is trained using only optical image features.

[0074] The experimental results with controlled variables are shown in Tables 1 and 2. It can be seen that the overall accuracy of the model trained solely on optical image features for artificial turf identification is relatively low in both regions. However, the overall accuracy of artificial turf identification using time-series SAR images is similar to and relatively higher than that of artificial turf identification using a combination of optical and SAR images. Therefore, the use of SAR image features has the greatest impact on the results. Considering that it is difficult to guarantee the acquisition of cloudless optical images in the cloudy and rainy areas of the south, the final model selected is the one trained solely on the statistical attribute features of time-series SAR image data.

[0075] Table 1. Accuracy of Artificial Turf Identification in Xinhui District

[0076]

[0077]

[0078] Table 2. Accuracy of Artificial Turf Identification in Jiexi County

[0079]

[0080] S7 utilizes the model trained in S601 using only statistical attribute features of temporal SAR image data to identify artificial turf in Jiexi County and Xinhui District of Guangdong Province, obtaining the final identification results. The spatial distribution of artificial turf is shown below. Figure 11 and Figure 12 As shown.

[0081] Example 2

[0082] This embodiment proposes a system for identifying artificial turf planted on cultivated land based on temporal SAR imagery, including:

[0083] The multi-temporal SAR remote sensing image acquisition module is used to acquire multi-temporal SAR remote sensing images of the target area, and obtain SAR images after preprocessing.

[0084] The field segmentation module is used to acquire high-resolution visible light images of the target area during clear, cloudless periods (not necessarily limited to the SAR observation period), and to segment all the fields within the target area using the field segmentation program.

[0085] The field sample map collection module is used to collect field sample maps of the target area obtained from field surveys;

[0086] The field sample labeling module is used to calculate the centroid of the fields within the target area, and to label the fields whose centroids fall within the range of the field sample patches as the corresponding patch categories, thus forming field samples;

[0087] The feature extraction module is used to extract the statistical feature attributes of each SAR image corresponding to a field sample.

[0088] The model building module is used to divide the dataset composed of the statistical feature attributes into a training set and a validation set, classify the field samples using the training set based on the XGBoost model, evaluate the performance of XGBoost models with different hyperparameters using the validation set, and select the optimal model to form a recognition and prediction model using the evaluation index of the validation set.

[0089] The automatic identification module is used to input field samples into the identification and prediction model to obtain the corresponding patch category of the field sample.

[0090] The functions of the above modules are as described in S1-S7 of Embodiment 1.

[0091] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for identifying artificial turf planted on cultivated land based on temporal SAR imagery, characterized in that, Includes the following steps: S1: Acquire multi-temporal SAR remote sensing images of the target area, and obtain SAR images after preprocessing; In S1, the SAR images are C-band VH backscattering coefficient data acquired by Sentinel-1A satellite, and the preprocessing includes orbit file correction, thermal noise removal, radiometric calibration, speckle noise filtering and terrain correction. S2, acquire a high-resolution visible light image of the target area during a clear, cloudless period, and use a field segmentation program to segment all the fields within the target area. The resolution of the high-resolution visible light image is 0.5m. S3, Field survey to obtain field sample maps of the target area; S4, calculate the centroid of the field plots within the target area, and mark the field plots whose centroids fall within the range of the field sample plots as the corresponding plot categories to form field plot samples; in S4, select field plots with an area of ​​more than 400 square meters from the field plot samples as the final field plot samples, and the field plots with an area of ​​more than 400 square meters correspond to at least 4 pixels of SAR images. S5, extract the statistical feature attributes of each period of the SAR image within the field sample; In S5, the statistical feature attributes include the maximum pixel value, minimum pixel value, and average pixel value of the SAR image; S6, the dataset composed of the statistical feature attributes is divided into a training set and a validation set in a 7:3 ratio. The training set is used to classify the field samples based on the XGBoost model. The validation set is used to evaluate the performance of the XGBoost model with different hyperparameters. The evaluation index of the validation set is used to select the optimal model to form the recognition and prediction model. S7. Input the field sample into the identification and prediction model to obtain the patch category corresponding to the field sample.

Images