A corn planting area corn identification method based on space-time feature selection

Abstract

The application provides a corn recognition method for a corn planting area based on space-time feature selection, and belongs to the technical field of remote sensing. The method comprises the following steps: acquiring a remote sensing image of a corn planting area, determining NDVI feature curves of each pixel in the corn planting area, screening out corn pixels meeting planting characteristics in a key growth period as corn sample points; determining the importance of different vegetation features for corn recognition based on the NDVI feature curves of the corn sample points and non-corn sample points; eliminating redundant features comprehensively according to the importance and feature correlation to obtain classification features; calculating the time-spectrum joint similarity of a to-be-classified pixel with the corn sample points and the non-corn sample points; and classifying the to-be-classified pixel into a category corresponding to the minimum time-spectrum joint similarity. The application can solve the problems of insufficient separability and feature redundancy in corn recognition of traditional classification methods and the problem of threshold dependence on agricultural statistics data of the TWDTW method.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of remote sensing, and particularly relates to a corn recognition method for a corn planting area based on spatiotemporal feature selection. BACKGROUND

[0002] Corn is one of the most important food and economic crops in the world, and plays an important role in agricultural production from human food, livestock feed to biofuel and industrial starch raw materials. Therefore, it is necessary to accurately grasp the spatial distribution and production dynamics of corn planting.

[0003] In recent years, breakthroughs in satellite remote sensing technology have provided a new paradigm for crop recognition. Compared with traditional methods that rely on manual investigation, it greatly saves labor costs and improves monitoring efficiency. Based on remote sensing images, spatiotemporal continuous observation data of the vegetation surface has become a core technical means for crop planting area recognition.

[0004] Existing crop recognition methods mainly include supervised learning-based classification methods and phenology curve similarity methods. Supervised learning methods train models by using labeled samples, extract remote sensing image features to generate crop distribution maps, and machine learning methods such as random forest (RF) and support vector machine (SVM) and supervised learning methods such as convolutional neural network (CNN) and long short-term memory network (LSTM) are widely used in time series remote sensing crop classification tasks, and have shown significant effects in the identification of typical crops such as wheat and rice. However, these supervised learning methods have strong dependence on the number of labeled samples, which leads to high cost constraints in large-area crop mapping. The phenology analysis method based on curve similarity provides a new idea, the Time-Weighted Dynamic Time Warping (TWDTW) method quantifies the morphological differences of spectral trajectories in the growth cycle of crops, and embeds the phenology time weight matrix, effectively realizing the fine recognition of crops in complex agricultural environments under different climate conditions. Existing research shows that TWDTW performs better than RF in the case of insufficient training data. However, the statistical representation of curve morphology in the small sample data scenario is prone to confidence interval expansion, which in turn affects the classification results.

[0005] Traditional TWDTW studies often rely solely on single spectral features for crop identification. This approach has limitations in regions with complex crop planting structures and high spectral similarity, especially during peak growing seasons when maize exhibits highly similar NDVI time-series curves to other crops like soybeans, sorghum, and peanuts. Relying solely on single vegetation indices like NDVI or EVI can easily lead to misclassification. In recent years, an increasing number of studies have combined multi-source remote sensing features for multi-crop identification. Studies using TWDTW to identify maize have found that combining two bands effectively improves identification accuracy. While the use of multi-source features can improve crop classification accuracy to some extent, simple feature overlay and indiscriminate addition can lead to information redundancy, increased computational burden, and reduced computational efficiency.

[0006] Existing TWDTW identification methods, when distinguishing between target and non-target crops, mostly rely on pre-set DTW distance thresholds based on regional agricultural statistical data, combined with the area of ​​the corn planting area, to determine corn pixels. This method first uses the calculated DTW distance to find the similarity between pixel points, and then uses statistical data as a benchmark to spatialize the overall quantity. This method has achieved high accuracy in mapping crops such as corn, rice, wheat, and sugarcane. However, the statistical data is usually regional macro-data, which is difficult to accurately reflect the spatial distribution complexity and heterogeneity of crops. Using it as a hard constraint or threshold derivation basis for pixel-level classification may introduce uncertainty in the spatial distribution of crops. At the same time, the acquisition of agricultural statistical data often has a certain lag, which limits the timeliness of classification results.

[0007] In summary, existing crop identification methods have the following shortcomings:

[0008] 1. Although the method based on the similarity of phenological curves shows a classification advantage under small sample conditions, insufficient samples can easily lead to bias in the extraction of statistical features of curve morphology, which in turn expands the confidence interval of the results and ultimately affects the classification accuracy.

[0009] 2. Traditional maize identification methods have limited ability to distinguish maize from summer crops with similar phenology. When using multi-source features for classification, there is often information redundancy, which together restricts the accuracy and efficiency of identification.

[0010] 3. Traditional TWDTW identification methods mostly rely on the DTW distance threshold preset by regional agricultural statistics when distinguishing between target and non-target crops, which makes it difficult to accurately reflect the spatial distribution complexity of crops at the pixel scale and the heterogeneity within plots.

[0011] In summary, these inherent limitations restrict the performance of corn identification technology in practical applications, especially in areas with few samples and complex crop planting structures, often resulting in insufficient identification accuracy and low overall computational efficiency. SUMMARY

[0012] Therefore, the application provides a corn recognition method for corn planting areas based on spatio-temporal feature selection, which can solve the problems of insufficient separability and feature redundancy in traditional classification methods and the dependence of agricultural statistical data threshold in TWDTW method.

[0013] To solve the above technical problems, the application is implemented as follows.

[0014] A corn recognition method for corn planting areas based on spatio-temporal feature selection, comprising:

[0015] Step 1: Obtain remote sensing images of the corn planting area.

[0016] Step 2: Determine the corn sample points: according to the remote sensing images of the corn planting area, determine the NDVI feature curve of each pixel in the corn planting area; according to the NDVI feature curve, screen out the corn pixels meeting the planting characteristics in the key growth period as the corn sample points.

[0017] Step 3: Select classification features: calculate the vegetation feature time series curve of N kinds of feature categories for the corn sample points and non-corn sample points; based on the vegetation feature time series curve of the corn sample points and non-corn sample points, determine the importance of different vegetation features for corn recognition by using the Relief algorithm; based on the vegetation feature time series curve of the corn sample points, determine the correlation between the vegetation features; select a group of M vegetation features with high importance and low correlation as the classification features.

[0018] The vegetation feature time series curves of the M classification features of each corn sample point and non-corn sample point form a sample set.

[0019] Step 4: Classification: calculate the time-spectrum joint similarity of the sample points and non-corn sample points in the sample set and the classification pixels; the time-spectrum joint similarity of two pixels is expressed by the comprehensive TWDTW distance of the vegetation feature time series curves of the M classification features; classify the classification pixels into the category corresponding to the minimum time-spectrum joint similarity.

[0020] Preferably, in step 1, the remote sensing images of the corn planting area are obtained by:

[0021] In the case of knowing the location of the corn planting area in the research area, directly using the data of the corn planting area in the measured remote sensing images;

[0022] In the case of not knowing the location of the current corn planting area in the study area, the existing national corn dataset is used to determine the corn planting area in the last three consecutive years in the study area as the predicted current corn planting area; the data of the predicted current corn planting area in the measured remote sensing image is used.

[0023] Preferably, the method further comprises preprocessing the acquired remote sensing image of the corn planting area, including:

[0024] The remote sensing image is subjected to cloud removal processing to retain images with a cloud coverage rate lower than a set proportion;

[0025] The remote sensing image is subjected to band resampling;

[0026] The remote sensing image is subjected to linear interpolation based on a selected time window;

[0027] The remote sensing image is subjected to image smoothing using Savitzky-Golay filtering.

[0028] Preferably, the method further comprises using the cultivated land distribution provided by the land cover dataset CLCD as a mask file to mask the remote sensing image and exclude non-cultivated land pixels in the remote sensing image.

[0029] Preferably, in step 2, the corn pixels that meet the planting characteristics in the key growth period are selected as the corn sample points according to the NDVI characteristic curve, including:

[0030] Determining the key growth period: according to the remote sensing image of the corn planting area, the NDVI characteristic curve of each corn pixel is calculated; the data of the same time point of each NDVI characteristic curve is averaged to obtain the NDVI characteristic mean curve; the amplitude variation f between the maximum value and the minimum value of the NDVI characteristic mean curve is obtained, and the time interval between the time point when the NDVI amplitude starts to be greater than 20%f and the time point when the NDVI amplitude starts to be less than 20%f is taken as the key growth period;

[0031] Determining the key phenological fluctuation range: the data of the same time point of each NDVI characteristic curve is averaged to obtain the mean μ and the standard deviation σ, and [μ-σ, μ+σ] is taken as the allowed range of the NDVI characteristic curve fluctuation, i.e. the key phenological fluctuation range;

[0032] For the NDVI characteristic curve of each pixel in the corn planting area, the NDVI characteristic curve segment in the key growth period is intercepted using the key growth period; for each pixel, it is judged whether the fluctuation amplitude of the intercepted NDVI characteristic curve segment is within the key phenological fluctuation range, and if so, the current pixel is identified as a corn sample point.

[0033] Preferably, the vegetation feature time series curve of the corn sample point and the vegetation feature time series curve of the pixel to be classified are intercepted during the classification feature selection and classification.

[0034] Preferably, in step 3, the vegetation feature time series curve of the corn sample point is used to determine the correlation between the vegetation features as follows:

[0035] For each vegetation feature, the mean curve of the vegetation feature time series of all corn sample points is calculated, and the Pearson correlation coefficient is calculated for the mean curve of the vegetation feature time series of the N vegetation features, serving as the correlation between the vegetation features.

[0036] Preferably, in step 3, the M vegetation features with high importance and low correlation are selected as classification features as follows:

[0037] The N vegetation features are sorted according to the importance from high to low, and each vegetation feature is selected one by one. It is determined whether the Pearson correlation coefficient between the selected vegetation feature and the determined classification feature is greater than the set correlation threshold. If it is greater, it is determined that the correlation is too large, and the selected vegetation feature is not used as a classification feature. If it is less than or equal to the set correlation threshold, the selected vegetation feature is determined as a classification feature. This process is repeated until M classification features are selected.

[0038] Preferably, the time series-spectrum combined similarity is obtained as follows:

[0039] For the pixel to be classified and one corn sample point or non-corn sample point in the sample set, the TWDTW distance of each classification feature is calculated, and the TWDTW distances of the M classification features are combined to obtain a comprehensive TWDTW distance expression.

[0040] Preferably, the N feature categories include vegetation index features and red edge vegetation index features:

[0041] The vegetation index features include normalized difference vegetation index NDVI, enhanced vegetation index EVI, difference vegetation index DVI, optimized soil-adjusted vegetation index OSAVI, spectral ratio index SR, green normalized difference vegetation index GNDVI, normalized difference greenness index NDGI, green canopy cover index GCVI, ratio vegetation index RVI, normalized difference pollution index NDPI, land surface water index LSWI, and photochemical vegetation reflectance index PSRI.

[0042] The red edge vegetation index features include: a first normalized red edge difference index NREDI1, a second normalized red edge difference index NREDI2, a third normalized red edge difference index NREDI3, a near-infrared / red edge ratio index NIR / RE, a red band / red edge ratio index RED / RE, and a red edge position index REP.

[0043] Advantages:

[0044] (1) The sample representativeness is improved, and a standard growth curve is constructed: Based on morphological characteristics, sample points meeting planting characteristics in key growth periods are screened, a more representative corn standard growth curve is constructed, a more reliable basis is provided for subsequent TWDTW matching, and TWDTW matching based on minimum distance is made possible.

[0045] (2) The corn separability features are screened, and the dependence on a single or preset feature is overcome: By constructing a Relief filtering type feature importance evaluation method, the importance of different features for corn identification is calculated, redundant features are removed comprehensively according to feature correlation, and a corn identification feature set considering feature importance and independence is constructed. The missing phenomenon commonly seen in single NDVI method is significantly reduced, the separability between corn and other crops is effectively improved, and the adaptability and robustness under complex topography and planting structure are enhanced.

[0046] (3) The dependence of traditional TWDTW agricultural statistical data threshold setting and timeliness limitation are overcome: The present application quantifies the time-spectrum joint similarity of the classified pixel and the standard curve by calculating the DTW distance, and constructs a corn identification model based on minimum distance classification by comprehensively considering feature distance. The present application does not need to be constrained by prior statistical area data, avoiding the uncertainty caused by subjective threshold setting. Pepper and salt noise is effectively suppressed, the homogeneity and spatial continuity of the graph spot are improved, and the boundary of the identified corn field is clearer and more accurate. Since statistical area data is not needed, the lag of statistical area acquisition is avoided.

[0047] (4) The present application is an organic combination of various steps: The prior art needs to determine the TWDTW threshold in combination with the corn planting area. Due to small sample size and unreasonable classification feature selection, it may lead to difficulty in determining the TWDTW threshold and inaccuracy. In order to solve this problem, the present application combines the sample point screening of key growth period morphological characteristics and the importance+correlation classification feature selection strategy, so that the similarity matching scheme based on minimum distance becomes possible. By optimizing the quality of the input sample and the spectral-time sequence feature set, the precision and robustness of corn identification are significantly improved, and the dependence of the traditional method on statistical prior knowledge and subjective threshold is overcome. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1A flow chart of the corn recognition method based on the spatio-temporal feature selection of the corn planting area of the application.

[0049] Fig. 2(a) is the corn phenology key growth period extraction of region 1.

[0050] Fig. 2(b) is the corn phenology key growth period extraction of region 2.

[0051] Fig. 2(c) is the corn phenology key growth period extraction of region 3.

[0052] Fig. 3(a) is the different feature separability diagram of region 1.

[0053] Fig. 3(b) is the different feature separability diagram of region 2.

[0054] Fig. 3(c) is the different feature separability diagram of region 3.

[0055] Fig. 4(a) is the correlation diagram between the time series of region 1.

[0056] Fig. 4(b) is the correlation diagram between the time series of region 2.

[0057] Fig. 4(c) is the correlation diagram between the time series of region 3.

[0058] Fig. 5(a) is the corn recognition result diagram of region 1.

[0059] Fig. 5(b) is the corn recognition result diagram of region 2.

[0060] Fig. 5(c) is the corn recognition result diagram of region 3. DETAILED DESCRIPTION

[0061] The application will be described in detail below with reference to the accompanying drawings and examples.

[0062] The application provides a corn recognition method based on spatio-temporal feature selection of a corn planting area, as shown in the figure, which comprises the following steps: Figure 1

[0063] Step 1: Obtain the remote sensing image of the corn planting area.

[0064] In this step, two situations may occur. One is to obtain the remote sensing image of the research area, and the position of the corn planting area in the year is known, that is, it is known which pixels in the remote sensing image are corn pixels and which are non-corn pixels, so the data of the corn planting area in the measured remote sensing image can be directly used.

[0065] ​The second case is that the position of the corn planting area in the research area in the current year is unknown, and the embodiment adopts a prediction method based on historical data. The existing national corn data set is used to determine the corn planting area in the research area in the three consecutive years, that is, the union of the corn planting area in the past three years, as the predicted corn planting area in the current year. Then, the data of the predicted position of the corn planting area in the current year is extracted from the measured remote sensing image to participate in the processing of the subsequent steps.

[0066] In a preferred scheme, in order to exclude the influence of non-cultivated land pixels on the crop classification investigation and research, the cultivated land distribution provided by the historical annual China land cover data set (CLCD) can also be used as a mask file to mask the remote sensing image, so as to obtain the overall distribution data of farmland vegetation, thereby excluding non-cultivated land pixels, and the subsequent steps are carried out on this basis.

[0067] In practice, the acquired remote sensing image may also have noise, resolution and other problems. Therefore, preprocessing can be performed, including: first, performing cloud removal processing on the remote sensing image, and retaining the image with a cloud coverage rate of less than 20%; then, resampling the remote sensing image band to 10 m with a spatial resolution of 20 m, so as to unify the spatial resolution. Secondly, linear interpolation is performed on the image with a time window of 10 days. Finally, image smoothing is performed using the Savitzky-Golay filter.

[0068] In the specific embodiment, the remote sensing data can adopt the Sentinel-2 image obtained in the Google Earth Engine.

[0069] Step 2: Corn sample point determination.

[0070] The present application classifies based on the method of phenology curve similarity (time series curve of vegetation characteristics), but insufficient samples can cause deviation in the extraction of curve shape statistical characteristics, thereby expanding the confidence interval of the result and ultimately affecting the classification accuracy. Therefore, by screening sample points based on morphological characteristics, a more representative corn standard growth curve can be constructed, which provides a more reliable basis for subsequent TWDTW matching.

[0071] Therefore, in this step, the NDVI feature curve of each pixel in the corn planting area is determined according to the remote sensing image of the corn planting area; and the corn pixels meeting the planting characteristics in the key growth period are screened out as the corn sample points according to the NDVI feature curve.

[0072] The present application screens corn sample points from two aspects: one is the time selection of the key growth period; and the other is the restriction of the key phenological fluctuation range of the phenological spectrum characteristics; through the restriction of the time and the phenological spectrum characteristics, the corn pixels meeting the planting characteristics of the key growth period are selected as the standard sample points.

[0073] The method for determining the key growth period comprises the following steps: calculating the NDVI characteristic curve of each corn pixel according to the remote sensing image of the corn planting area; obtaining the NDVI characteristic mean curve by averaging the data of the same time point of each NDVI characteristic curve; obtaining the amplitude variation range f between the maximum value and the minimum value of the NDVI characteristic mean curve; determining the time point t1 at which the amplitude of the NDVI characteristic curve starts to be greater than 20%f, and the time point t2 at which the amplitude starts to be less than 20%f, and the interval [t1, t2] is the key growth period.

[0074] In the following steps of classification feature selection and classification, the preferred scheme is to use the key growth period to intercept the vegetation characteristic time curve of the corn sample points and the pixels to be classified, so as to ensure that the characteristic curve for feature selection and similarity comparison can accurately represent the planting characteristics of the key growth period.

[0075] The method for determining the key phenological fluctuation range comprises the following steps: obtaining the mean value μ and the standard deviation σ of the data of the same time point of each NDVI characteristic curve, both the mean value μ and the standard deviation σ are curves, taking [μ-σ, μ+σ] as the allowed range of the NDVI characteristic curve fluctuation, then superimposing and calculating the mean value curve and the standard deviation curve according to μ-σ and μ+σ to obtain two curves, and the curve between the two curves is the allowed range of the NDVI characteristic curve fluctuation, that is, the key phenological fluctuation range.

[0076] In the screening of the corn sample points, the NDVI characteristic curve segment of the key growth period is intercepted from the NDVI characteristic curve of each pixel in the corn planting area; for each pixel, it is judged whether the NDVI characteristic curve segment has a part exceeding the key phenological fluctuation range, if yes, the pixel is removed and is not identified as a corn sample point, through the operation, only the key growth period NDVI characteristic curve with the fluctuation amplitude in the key phenological fluctuation range is reserved, and the corresponding pixel is the corn sample point.

[0077] Step 3: Classification feature selection.

[0078] In this step, M (M

[0079] In this step, the vegetation feature time series curves of the N types of features are calculated for the corn sample points and the non-corn sample points. Based on the vegetation feature time series curves of the corn sample points and the non-corn sample points, the importance of different vegetation features for corn recognition is determined by using the Relief algorithm. Based on the vegetation feature time series curves of the corn sample points, the correlation between the vegetation features is determined. A group of M vegetation features with high importance and low correlation are selected as the classification features.

[0080] Based on the selected classification features, the vegetation feature time series curves of the M classification features of each corn sample point and each non-corn sample point are obtained to form a sample set.

[0081] In an example, M can be selected as M = 5.

[0082] In a preferred scheme, the specific implementation scheme of step 3 includes the following sub-steps:

[0083] Step 301: Calculate the spectral features and then calculate the 34 types of vegetation features for the corn sample points and the non-corn sample points. The present application selects 34 types of vegetation features including vegetation index features and red edge vegetation index features. Table 1 shows the calculation methods of the 34 types of vegetation features selected in the embodiment of the present application.

[0084] Table 1: Spectral indices used

[0085]

[0086] Step 302: Calculate the feature importance between the sample points by using the Relief algorithm for the corn sample points and the non-corn sample points to obtain the classification importance of the selected 34 features for corn and other crops in the key growth period.

[0087] The calculation formula of the Relief algorithm is:

[0088] In the Relief model, the distance between any two sample points (the corn sample point) and (the non-corn sample point) in the vegetation feature is represented as represents the first Individual vegetation characteristics:

[0089]

[0090] In the formula, yes exist The value on, yes exist The value on, Indicates in The maximum value on, Indicates in The minimum value of the Relief feature importance. The iterative calculation formula for the Relief feature importance is expressed as:

[0091]

[0092] In the formula, NH and NM represent the nearest neighbor samples of the same class and the nearest neighbor samples of different classes of x, respectively. Indicates x and NH in The distance above, Indicates x and NM in f i The distance above, express The feature weight iteration results.

[0093] After the Relief algorithm, each vegetation feature has a feature weight, which represents the importance of the feature, and the features are ranked according to the feature weight.

[0094] Step 303: For maize sample points, determine the correlation between vegetation features based on the vegetation feature time series curve.

[0095] This invention incorporates a feature correlation screening perspective, increasing the independence of the selected classification features. It solves the information redundancy problem that often exists when using multi-source features for classification, and helps to construct a corn identification feature set that takes into account both feature importance and independence. This effectively improves the separability of corn from other crops and enhances the accuracy and efficiency of identification.

[0096] Specifically, this step calculates the time-series mean curve of vegetation features for all maize sample points for each vegetation feature; for the time-series mean curves of N vegetation features, the Pearson correlation coefficient is calculated pairwise to represent the correlation between vegetation features. In this example, the Pearson correlation coefficients of 34 features can yield a 34×34 Pearson correlation coefficient matrix.

[0097] Step 304: Select a set of M vegetation features that are prioritized in importance and whose correlation with each other does not exceed a set threshold, as classification features.

[0098] The step can be specifically: 34 vegetation characteristics are sorted according to importance, and are selected one by one in order from high to low importance; for each selected vegetation characteristic, it is determined whether the Pearson correlation coefficient between the selected vegetation characteristic and the determined classification characteristic is greater than a set correlation threshold; if yes, it is determined that the correlation is too large, and the selected vegetation characteristic is not used as a classification characteristic; if no, the selected vegetation characteristic is determined as a classification characteristic; until 5 classification characteristics are selected.

[0099] In a preferred embodiment, the correlation threshold can be 0.9.

[0100] Step 305: Obtain the vegetation characteristic time sequence curves of the 5 classification characteristics of each corn sample point and the 5 classification characteristics of the non-corn sample points, form a multi-dimensional feature time sequence of the sample, and jointly constitute a standard sample, and store it in a sample set to form a reference library. The reference library provides a standard feature time variation mode set for corn and non-corn categories.

[0101] Step 4: Classify the to-be-classified pixel.

[0102] The two distinguishing points of this step are: one is to use the comprehensive TWDTW distance to express the similarity, and the other is to abandon the scheme of using the set TWDTW threshold and the corn planting area to jointly determine the pixel classification in the traditional scheme, avoiding the use of the corn planting area obtained from the yearbook data, and solving the problem of inaccurate determination of the TWDTW threshold combined with the corn planting area. The prior art needs to determine the TWDTW threshold combined with the corn planting area, because the small sample size and unreasonable feature selection lead to the problem of difficult and inaccurate TWDTW threshold. The present application combines the sample point screening of key growth period morphology and the classification characteristic selection strategy of importance + correlation, so that the similarity matching scheme based on the minimum distance becomes possible.

[0103] In this step, the time sequence-spectrum joint similarity of the to-be-classified pixel and the corn sample points and the non-corn sample points in the sample set is calculated; the time sequence-spectrum joint similarity of the two pixels is expressed by the comprehensive TWDTW distance of the vegetation characteristic time sequence curves of the M classification characteristics; and the to-be-classified pixel is classified into the category corresponding to the minimum time sequence-spectrum joint similarity.

[0104] This step uses a time-weighted dynamic time warping algorithm (TWDTW) based on a Logistic function weighting, which has obvious advantages in crop classification accuracy compared with a linear weighting method. The logical weights of the steepness α and the midpoint β are determined in the TWDTW method.

[0105]

[0106] wherein is the elapsed time in days between and , and α and β represent the logistic weights for steepness and midpoint, respectively, responsible for shape matching and time alignment, respectively. The logistic weighting function used in this step builds a non-linear time penalty mechanism by setting α = -0.1 and β = 50, giving lower penalty to time warping less than 50 days, and higher penalty to larger time warping.

[0107] The crop classification procedure based on TWDTW includes the following steps:

[0108] (1) Calculate the multi-dimensional weighted dynamic time warping distance.

[0109] For each classification feature, the TWDTW distance between the to-be-classified pixel and each category sample point is calculated, i.e., the similarity of the feature curve under the classification feature is calculated; the TWDTW distances of the five classification features are added or weighted to obtain the comprehensive TWDTW distance expression, i.e., the multi-dimensional weighted dynamic time warping distance.

[0110] wherein the category sample points include corn and non-corn sample points.

[0111] For each to-be-classified remote sensing image pixel in the study area, the multi-dimensional feature time series of the pixel in all selected classification features are extracted by using the same method as that for constructing the reference library. For the multi-dimensional feature time series of each to-be-classified pixel, the weighted dynamic warping distance between the pixel and the corn and non-corn categories in the reference library is calculated to obtain the distance image set of corn and non-corn.

[0112] (2) Perform category determination based on the minimum distance principle.

[0113] For each pixel in the study area, the TWDTW distance values of the pixel to the corn sample points and the non-corn sample points are compared, and the pixel is assigned to the category with the closest distance, i.e., the category corresponding to the minimum TWDTW distance value.

[0114] The effect of the present application is verified as follows.

[0115] To evaluate the effectiveness of the research method, the overall accuracy (OA), Kappa coefficient, producer's accuracy (PA), and user's accuracy (UA) are used for precision verification.

[0116] Three typical corn planting areas with significant climate differences and terrain characteristics are selected as research objects. The method proposed in the application is used for screening samples in the experimental area, and the characteristics of corn in the key growth period of the experimental area are screened. The first five characteristics screened are added one by one to the classification model.

[0117] Fig. 2 (a), Fig. 2 (b) and Fig. 2 (c) are schematic diagrams of corn growth periods in three research areas respectively, the green curve is the smoothed NDVI time curve, the green point indicates the beginning of the phenological period, and the red point indicates the end of the phenological period; Fig. 3 (a), Fig. 3 (b) and Fig. 3 (c) are respectively the feature importance presentation results of three research areas, the darker the color, the higher the importance; Fig. 4 (a), Fig. 4 (b) and Fig. 4 (c) are respectively the feature correlation analysis, the darker the color, the stronger the correlation; Fig. 5 (a), Fig. 5 (b) and Fig. 5 (c) are respectively the corn recognition results of three research areas, A1-A3 represent three local areas in region 1, B1-B3 represent three local areas in region 2; C1-C3 represent three local areas in region 3. In the second row of figures, orange is the corn planting area, and white is the non-corn area.

[0118] The results show that the classification accuracy of multi-feature combination is usually better than that of single feature. Due to the differences in terrain conditions and crop planting structure of each research area, the number of features and their combinations that achieve the optimal classification accuracy are different. The recognition results of different feature combinations are shown in Table 2.

[0119] Table 2 Corn mapping accuracy evaluation

[0120]

[0121] The method of the application systematically optimizes the input sample quality and spectral-time feature set, not only significantly improves the accuracy and robustness of corn recognition, but also overcomes the dependence of traditional methods on statistical prior knowledge and subjective threshold.

[0122] The above specific embodiments only describe the design principles of the application, and the shapes and names of the components in the description can be different and are not limited. Therefore, the person skilled in the art of the application can modify or replace the technical solutions described in the foregoing embodiments; and these modifications and replacements do not deviate from the purpose and technical solutions of the application, and should all belong to the protection scope of the application.

Claims

1. A method for corn identification in corn-growing areas based on spatiotemporal feature selection, characterized in that, include: Step 1: Acquire remote sensing images of the corn-growing area; Step 2: Determining maize sample points: Based on the remote sensing images of the maize planting area, determine the Normalized Difference Vegetation Index (NDVI) characteristic curve for each pixel in the maize planting area. Based on the NDVI characteristic curve, maize pixels that meet the planting characteristics of the key growth period are selected as maize sample points. The step of selecting maize pixels that meet the planting characteristics of the key growth period based on the NDVI feature curve as maize sample points includes: Determine the critical growth period: Based on remote sensing images of the maize planting area, calculate the NDVI characteristic curve of each maize pixel; calculate the mean value of the data at the same time point of each NDVI characteristic curve to obtain the NDVI characteristic mean curve; obtain the amplitude change range f between the maximum and minimum values ​​in the NDVI characteristic mean curve, and take the time interval between the time point when the NDVI amplitude starts to be greater than 20%f and the time point when it starts to be less than 20%f as the critical growth period; Determine the range of key phenological fluctuations: Calculate the mean μ and standard deviation σ of the data at the same time point for each NDVI characteristic curve, and take [μ-σ,μ+σ] as the allowable range of fluctuations in the NDVI characteristic curve, i.e., the range of key phenological fluctuations; For the NDVI characteristic curves of each pixel in the maize planting area, the NDVI characteristic curve segments of the key growth period are extracted using the key growth period; for each pixel, it is determined whether the fluctuation range of the extracted NDVI characteristic curve segment is entirely within the key phenological fluctuation range. If so, the current pixel is identified as a maize sample point. Step 3: Categorical Feature Selection: For maize sample points and non-maize sample points, calculate the time-series curves of vegetation features for N feature categories; based on the time-series curves of vegetation features for maize and non-maize sample points, use the Relief algorithm to determine the importance of different vegetation features for maize identification; for each vegetation feature, calculate the time-series mean curve of vegetation features for all maize sample points; calculate the Pearson correlation coefficient for each pair of the time-series mean curves of the N vegetation features as the correlation between vegetation features; sort the N vegetation features according to their importance, selecting them one by one in descending order of importance; for each selected vegetation feature, determine whether its Pearson correlation coefficient with the determined categorical feature is greater than a set correlation threshold; if it is greater, the correlation is too high, and the currently selected vegetation feature is not used as a categorical feature; if it is less than or equal to the set correlation threshold, the currently selected vegetation feature is determined as a categorical feature; until M categorical features are selected. The time-series curves of vegetation characteristics of M classification features of each maize sample point and non-maize sample points form a sample set; The N feature categories include vegetation index features and red-edge vegetation index features: The vegetation index features include: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Differential Vegetation Index (DVI), Soil-Optimized Adjusted Vegetation Index (OSAVI), Spectral Ratio Index (SR), Normalized Difference Greenness Index (GNDVI), Normalized Difference Greenness Index (NDGI), Green Leaf Cover Index (GCVI), Ratio Vegetation Index (RVI), Normalized Difference Pollution Index (NDPI), Land Surface Moisture Index (LSWI), and Photochemical Vegetation Reflectance Index (PSRI). The red-edge vegetation index features include: the first normalized red-edge difference index NREDI1, the second normalized red-edge difference index NREDI2, the third normalized red-edge difference index NREDI3, the near-infrared / red-edge ratio index NIR / RE, the red band / red-edge ratio index RED / RE, and the red-edge position index REP. Step 4: Classification: Calculate the temporal-spectral joint similarity between the pixel to be classified and the maize sample points and non-maize sample points in the sample set; the temporal-spectral joint similarity between two pixels is expressed by the comprehensive TWDTW distance of the vegetation feature temporal curves of M classification features; classify the pixel to be classified into the category corresponding to the minimum temporal-spectral joint similarity; The method for obtaining the time-spectral joint similarity is as follows: For a pixel to be classified and a sample point in the sample set, the TWDTW distance of each classification feature is calculated separately. The TWDTW distances of the M classification features are then combined and used as the comprehensive TWDTW distance expression.

2. The corn identification method for corn planting areas based on spatiotemporal feature selection as described in claim 1, characterized in that, In step 1, acquiring remote sensing images of the corn planting area involves: Given the known location of the corn-growing areas in the study area that year, we directly used the data of the corn-growing areas from the measured remote sensing images. Without knowing the location of the corn planting area in the study area for that year, the existing national corn dataset is used to determine the corn planting area for the most recent three consecutive years in the study area, which is then used as the predicted corn planting area for that year; the data of the predicted corn planting area for that year are used from measured remote sensing images.

3. The corn identification method for corn planting areas based on spatiotemporal feature selection as described in claim 1, characterized in that, The method further includes preprocessing the acquired remote sensing images of the maize planting area, including: Perform cloud removal processing on remote sensing images, retaining images with cloud coverage below a set ratio; Band resampling of remote sensing images; Linear interpolation of remote sensing images is performed based on a selected time window; Savitzky-Golay filtering was used for image smoothing of remote sensing images.

4. The corn identification method for corn planting areas based on spatiotemporal feature selection as described in claim 1, 2, or 3, characterized in that, The method further includes: using the cultivated land distribution provided by the land cover dataset CLCD as a mask file to mask the remote sensing image and exclude non-cultivated land pixels in the remote sensing image.

5. The corn identification method for corn-growing areas based on spatiotemporal feature selection as described in claim 1, characterized in that, In both feature selection and classification, the vegetation feature time-series curves of maize sample points and pixels to be classified were truncated during the key growth period.

Images