An object-oriented orchard classification method based on Sentinel-2 time series and ReliefF

Abstract

The present invention relates to an object-oriented orchard classification method based on Sentinel-2 time series and ReliefF, comprising the following steps: obtaining Sentinel-2 data of a study area and performing preprocessing to obtain Sentinel-2 time series data; obtaining classification results of easily distinguishable crops and obtaining regions of interest of easily distinguishable crops; generating a classification feature set of easily distinguishable crops; screening to obtain an optimal feature set and an optimal classifier; performing SNIC image segmentation, and classifying easily distinguishable orchards in combination with the optimal feature set and the optimal classifier to obtain classification results of easily distinguishable crops; and evaluating the classification results of easily distinguishable crops. The present invention provides reference information for remote sensing multi-classification of other easily distinguishable crops by taking orchards with complex planting conditions as research objects; utilizing a multi-layer classification method, the present invention can significantly reduce the interference of some easily distinguishable crop types on the classification results of difficult-to-distinguish crops, thereby facilitating classification and obtaining more accurate and reliable classification results.

Description

Technical Field

[0001] The present invention relates to the technical field of orchard classification, in particular to an object-oriented orchard classification method based on Sentinel-2 time series and ReliefF. Background Art

[0002] Fruit is rich in nutrients and an essential component of the human diet. Accurately obtaining information on orchard planting area is crucial for policymaking, ecological function assessment, and agricultural economic management. Information on orchard planting area and spatial distribution serves as a foundational support for developing sustainable agricultural strategies. Studies have found that fruit trees contribute to carbon sequestration, and their water consumption patterns affect deep soil water content. Therefore, orchards play a crucial role in the carbon and water cycles of terrestrial ecosystems, and orchard mapping can aid in ecological function assessment. Furthermore, there is a positive correlation between orchard production scale and production efficiency; only moderate-scale production can promote the healthy and green development of the agricultural economy. Knowledge of orchard planting information is also a crucial prerequisite for intensive management in the face of a large population and limited land.

[0003] Remote sensing technology can cost-effectively provide large-scale crop distribution information, meeting the needs of various industries for large-scale crop statistics and providing a viable approach for orchard classification. While significant progress has been made in crop distribution extraction research both domestically and internationally, with remote sensing mapping techniques for major food crops such as rice, corn, wheat, and soybeans becoming increasingly mature, multi-classification of orchard crops remains a pressing need. Furthermore, multispectral imagery, with its large coverage area and short revisit periods, allows for dense time series imagery, compared to single images, to improve orchard classification accuracy. To address the "salt and pepper effect," object-oriented methods have been widely used in remote sensing crop classification. Therefore, for orchard classification, Sentinel-2 time series data was constructed and combined with object-oriented methods.

[0004] Currently, most research focuses on crop classification in large, simple scenarios, while research on orchard crop classification is limited. Furthermore, due to the inter-class similarity and intra-class variability of different fruit tree varieties, obtaining effective classification features is challenging. Furthermore, multi-class classification in orchards with complex planting structures suffers from low accuracy. Addressing these issues and achieving high-precision multi-class classification in orchards is a topic worthy of further exploration. Summary of the Invention

[0005] In order to solve the problem that orchards are easily confused and difficult to distinguish in the field of remote sensing identification, the purpose of the present invention is to provide an object-oriented orchard classification method based on Sentinel-2 time series and ReliefF, which is convenient for classification and can obtain more accurate and reliable classification results.

[0006] To achieve the above object, the present invention adopts the following technical solution: an object-oriented orchard classification method based on Sentinel-2 time series and ReliefF, the method comprising the following steps in sequence:

[0007] (1) Obtain Sentinel-2 data of the study area and perform preprocessing operations such as screening, cloud removal, cultivated land masking, and time series reconstruction on the Sentinel-2 data to obtain Sentinel-2 time series data;

[0008] (2) Use the multi-layer classification method to first classify easy-to-classify crops and obtain the easy-to-classify crop classification results. Use the easy-to-classify crop classification results to mask out the easy-to-classify crop pixels of the Setinel-2 time series data and obtain the easy-to-mix crop interest area;

[0009] (3) Based on the easily confused crops region of interest, generate a classification feature set of easily confused crops;

[0010] (4) The classification feature set of easily confused crops is sorted using the ReliefF algorithm, and the top 25%, 50%, 75%, and 100% features are input into three classifiers respectively to obtain the best feature set and the best classifier;

[0011] (5) Perform SNIC image segmentation and combine the best feature set and the best classifier to classify easily confused orchards and obtain the easily confused crop classification results;

[0012] (6) Evaluate the classification results of easily confused crops.

[0013] The step (1) specifically includes the following steps:

[0014] (1a) Obtain Sentinel-2Level-2A data of the study area on the GEE platform;

[0015] (1b) The Sentinel-2 Level-2A data were screened for image data with cloud cover less than 10%, and then the cloud fraction method was used to remove the cloud for the image data with cloud cover less than 10%, and the cultivated land mask was performed on the cloud-removed data;

[0016] (1c) The Sentinel-2 data after the cultivated land mask is reconstructed into a time series. The time series reconstruction includes 10-day median synthesis, linear interpolation and SG filtering, and finally the Sentinel-2 time series data is obtained.

[0017] The step (2) specifically includes the following steps:

[0018] The multi-layer classification method (2a) is to classify crops into easy-to-classify crops and easy-to-confuse crops based on the time series differences of the Normalized Difference Vegetation Index (NDVI) of crops. The first layer is to classify easy-to-classify crops. The easy-to-classify crops classification uses the first-order difference of NDVI time series data and its third harmonic fitting parameters, as well as the red edge, near-infrared and short-wave infrared bands as the classification features of easy-to-classify crops. The calculation formulas for the first-order difference of NDVI time series data and its third harmonic fitting parameters are shown in formulas (1) and (2):

[0019] ΔNDVI k =NDVI K+1 -NDVI k (1)

[0020]

[0021] Among them, ΔNDVI k is the first-order difference of the kth image in the NDVI time series data; NDVI k+1 is the k+1th image in the NDVI time series data, k=0,1,...,28; NDVI t It is a third harmonic fitting of the NDVI time series data of a pixel at time t, a0 is the harmonic remainder, a i and b i is the coefficient of each harmonic, θ i is the initial phase of the i-th harmonic; t is the observation time, t∈[0,1], 0 represents January 1 of the year, and 1 represents December 31 of the year;

[0022] (2b) The classifier for easy-to-classify crops uses decision tree CART and random forest RF, gradually searches for the optimal parameters of decision tree CART and random forest RF, and inputs the easy-to-classify crop classification features into the classifier to obtain the easy-to-classify crop classification results;

[0023] (2c) The easy-to-classify crop classification result mask is used to remove the easy-to-classify crop pixels in the Setinel-2 time series data to obtain the easy-to-classify crop area of interest.

[0024] In step (3), the classification feature set of the easily confused crops includes band features, vegetation index and texture features in Setinel-2 time series data.

[0025] The step (4) specifically includes the following steps:

[0026] (4a) Select the pixel feature values covered by the easily confused crop samples from the classification feature set of easily confused crops, calculate the feature importance and sort them using the ReliefF algorithm, and use the feature importance as the feature weight to cumulatively analyze the key features of easily confused crops;

[0027] (4b) The top 25%, 50%, 75% and 100% of the feature importance are input into the decision tree CART, random forest RF and support vector machine SVM classifiers respectively, and the best classifier and the best feature set are screened.

[0028] The step (5) specifically includes the following steps:

[0029] (5a) Use the SNIC algorithm to segment the image at different scales based on the optimal feature set to obtain the optimal scale segmentation result;

[0030] (5b) performing mean calculation on the best feature set for each object in the segmentation result to obtain the object-oriented best feature set;

[0031] (5c) The optimal feature set of the facing object is input into the optimal classifier to obtain the classification result of easily confused crops.

[0032] In step (6), the methods for evaluating the classification results of easily confused crops include three methods: the first method is to evaluate the classification results of easily confused crops through mapping accuracy PA, user accuracy UA and overall accuracy OA; the second method is to superimpose and compare the sample areas collected on the spot with the classification results of easily confused crops. The closer the boundaries of the classified plots are, the better the classification effect; the third method is to compare the classified areas of the main crops in the classification results of easily confused crops with the actual areas shown in the survey report of the local government to verify the accuracy of the classification.

[0033] It can be seen from the above technical solution that the beneficial effects of the present invention are: First, by taking orchards with complex planting conditions as research objects, the present invention provides reference information for remote sensing multi-classification of other easily confused crops; Second, the present invention uses a multi-layer classification method to significantly reduce the interference caused by some simply distinguished crop types on the classification results of difficult-to-distinguish crops, facilitates classification, and can obtain more accurate and reliable classification results; Third, the present invention evaluates the importance of features based on the ReliefF algorithm, avoids the "Hughes effect" problem in time series, and screens out important classification features and key classification time windows; Fourth, the orchard classification method of the present invention based on Sentinel-2 time series data and GEE platform has simple data source acquisition, and with the help of the computing advantages of the GEE platform, it can draw orchard planting distribution maps of large areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is a flow chart of the method of the present invention;

[0035] Figure 2 This is the NDVI time series difference diagram of various crops in the present invention;

[0036] Figure 3This is the feature importance map of easily confused crops based on the ReliefF algorithm in the present invention;

[0037] Figure 4 This is a comparison diagram of easily confused crop areas based on pixel-based and object-based classification results in the present invention. DETAILED DESCRIPTION

[0038] like Figure 1 As shown, an object-oriented orchard classification method based on Sentinel-2 time series and ReliefF includes the following steps in sequence:

[0039] (1) The Sentinel-2 data of the study area were obtained through the Google Earth Engine (GEE) platform. The Sentinel-2 data were preprocessed by screening, cloud removal, farmland masking, and time series reconstruction to obtain the Sentinel-2 time series data.

[0040] (2) Use the multi-layer classification method to first classify easy-to-classify crops and obtain the easy-to-classify crop classification results. Use the easy-to-classify crop classification results to mask out the easy-to-classify crop pixels of the Setinel-2 time series data and obtain the easy-to-mix crop interest area;

[0041] (3) Based on the easily confused crops region of interest, generate a classification feature set of easily confused crops;

[0042] (4) The classification feature set of easily confused crops is sorted using the ReliefF algorithm, and the top 25%, 50%, 75%, and 100% features are input into three classifiers respectively to obtain the best feature set and the best classifier;

[0043] (5) Perform SNIC image segmentation and combine the best feature set and the best classifier to classify easily confused orchards and obtain the easily confused crop classification results;

[0044] (6) Evaluate the classification results of easily confused crops.

[0045] The step (1) specifically includes the following steps:

[0046] (1a) Obtain Sentinel-2Level-2A data of the study area on the GEE platform;

[0047] (1b) The Sentinel-2 Level-2A data were screened for image data with cloud cover less than 10%. The cloud fraction method was then used to remove the cloud from the image data with cloud cover less than 10%. The cloud-removed data were masked with cultivated land products provided by the European Space Agency (ESA).

[0048] (1c) Time series reconstruction of the Sentinel-2 data after the cultivated land mask was performed. This reconstruction includes a 10-day median composite, linear interpolation, and SG filtering to obtain the Setinel-2 time series data. The SG filter observation moving window is set to 7, and the filter order is set to 3. The final Setinel-2 time series data from March 1 to December 6 is 29 images.

[0049] The step (2) specifically includes the following steps:

[0050] (2a) The multi-layer classification method is to classify crops that are easy to classify and crops that are easy to confuse based on the time series differences of the Normalized Difference Vegetation Index (NDVI) of crops. The easy-to-classify crops include jujube, grape, corn, apricot and cherry, and the easily confuse crops include apple, peach, pear, persimmon and pomegranate. The first layer is to classify the easy-to-classify crops. The easy-to-classify crops are classified using the first-order difference of the NDVI time series data and its third harmonic fitting parameters, as well as the red edge, near infrared and short-wave infrared bands as the classification features of the easy-to-classify crops. The calculation formulas for the first-order difference of the NDVI time series data and its third harmonic fitting parameters are shown in formulas (1) and (2):

[0051] ΔNDVI k =NDVI K+1 -NDVI k (1)

[0052]

[0053] Among them, ΔNDVI k is the first-order difference of the kth image in the NDVI time series data; NDVI k+1 is the k+1th image in the NDVI time series data, k=0,1,...,28; NDVI t It is a third harmonic fitting of the NDVI time series data of a pixel at time t, a0 is the harmonic remainder, a i and b i is the coefficient of each harmonic, θ i is the initial phase of the i-th harmonic; t is the observation time, t∈[0,1], 0 represents January 1 of the year, and 1 represents December 31 of the year;

[0054] (2b) The classifiers for easy-to-classify crops use decision trees (CART) and random forests (RF). The optimal parameters of decision trees (CART) and random forests (RF) are gradually searched for. The classification features of easy-to-classify crops are input into the classifiers to obtain the classification results of easy-to-classify crops. The optimal parameters for the minimum number of samples required for a leaf node in the decision tree (minLeafPopulation) and the optimal parameters for the number of decision trees in the random forest (numberOfTrees) are gradually searched for. Table 1 shows the classification results of the two classifiers under different parameters. The highest accuracy is obtained when the numberOfTrees is 800. To save computational effort, it is decided to use RF with a parameter of 200 as the classifier for easy-to-classify crops.

[0055] Table 1 Classification accuracy of easy-to-classify crops by two classifiers under different parameters

[0056]

[0057] (2c) The easy-to-classify crop classification result mask is used to remove the easy-to-classify crop pixels in the Setinel-2 time series data to obtain the easy-to-classify crop area of interest.

[0058] In step (3), the classification feature set of the easily confused crops includes band features, vegetation index and texture features in Setinel-2 time series data.

[0059] The band features described are the nine commonly used raw bands of the Sentinel-2 time series. The red edge index was primarily selected as the vegetation index, and five characteristic indices were ultimately introduced: NDVI, NDMI, RESI, A, and REP. The formulas are shown in Table 2. Among them, the RESI index highlights the sensitivity of the red edge band to changes in forest moisture content and canopy density. The A index is a red edge index used to distinguish citrus perennials and has also achieved good results in orchard classification. The NDMI index is highly sensitive to leaf moisture and soil moisture, and there are significant differences in moisture content between fruit trees. The REP index is used to estimate nitrogen content. Nitrogen is used in photosynthesis in fruit tree leaves and is an indispensable element in leaves.

[0060] Table 2 Calculation formula for vegetation index of easily confused crops

[0061]

[0062] Texture features were calculated using the gray-level co-occurrence matrix (GLCM), and five of these parameters were selected for classification: angular second moment, contrast, inverse moment, correlation, and energy. NDVI images were used to calculate GLCM texture features, with a kernel size of 3x3. The NDVI time series was roughly divided into regions based on the growth period. Images from the flowering period (March 1st to April 1st), the vigorous growth period (June 1st to September 1st), and the leaf-fall period (November 1st to December 6th) were selected for median synthesis and then texture features were calculated. The feature symbols are denoted as NDVIpre, NDVIpeak, and NDVIpost, respectively. As shown in Table 3, a total of 421 features were obtained.

[0063] Table 3 Classification feature set of easily confused crops

[0064]

[0065] The step (4) specifically includes the following steps:

[0066] (4a) Select the pixel feature values covered by the easily confused crop samples from the classification feature set of easily confused crops, calculate the feature importance and sort them using the ReliefF algorithm, and use the feature importance as the feature weight to cumulatively analyze the key features of easily confused crops;

[0067] (4b) The top 25%, 50%, 75% and 100% of the feature importance are input into the decision tree CART, random forest RF and support vector machine SVM classifiers respectively, and the best classifier and the best feature set are screened.

[0068] Table 4 shows the classification accuracy of easily confused crops of the three classifiers with different input importance features. The highest accuracy is 83.82%, which is obtained when all features are input into the RF classifier. Therefore, the best classifier of the present invention is RF, and the best feature set is all features.

[0069] Table 4 Classification accuracy of easily confused crops by three classifiers with different input importance features

[0070]

[0071] The step (5) specifically includes the following steps:

[0072] (5a) Using the simple non-iterative clustering (SNIC) algorithm to segment the image at different scales based on the optimal feature set, the optimal scale segmentation result is obtained;

[0073] Since the local plot width requires an average of 5 pixels to cover, the position spacing of the superpixel seeds is set to 3, 5, 10, and 15 pixels, respectively. The appropriate segmentation scale is determined based on the segmentation result and the relative size of the plot. In addition, the compactness parameter of the SNIC algorithm is set to 0, the connectivity parameter is set to 8, and the neighborhood size is set to 256. The result shows that the optimal segmentation result of 5 can better describe the heterogeneous objects in the fragmented farmland.

[0074] (5b) performing mean calculation on the best feature set for each object in the segmentation result to obtain the object-oriented best feature set;

[0075] (5c) The optimal feature set of the facing object is input into the optimal classifier to obtain the classification result of easily confused crops.

[0076] In step (6), the methods for evaluating the classification results of easily confused crops include three methods: the first method is to evaluate the classification results of easily confused crops through mapping accuracy PA, user accuracy UA, and overall accuracy OA. Table 5 compares the pixel-based and object-based methods. The UA of apples, peaches, and persimmons for the object-based classification method is higher than that of the pixel-based method, and the PA of peaches and persimmons is higher than that of the pixel-based method. From the OA point of view, the accuracy of the object-based method is slightly lower than that of the pixel-based method. This is because the inspection and evaluation method is not suitable for object-based result verification. More inspection methods for plot boundaries should be added to make the comprehensive evaluation more accurate.

[0077] Table 5 Classification accuracy of easily confused crops based on pixel-based and object-based methods

[0078]

[0079]

[0080] The second method is to superimpose and compare the sample areas collected on the spot with the classification results of easily confused crops. The closer the boundaries of the classified plots are, the better the classification effect. The third method is to compare the classified areas of the main crops in the classification results of easily confused crops with the actual areas shown in the local government's survey report to verify the accuracy of the classification.

[0081] like Figure 2 As shown in the figure, based on the field survey samples in the study area, a mean curve with plots as units was generated, and then the mean statistics of each type of plot were performed to obtain the NDVI time series difference map. There is a lot of aliasing in the spectra of the fruit tree category in the NDVI time series map, so the fruit tree classification task is a challenge.

[0082] like Figure 3 As shown, Figure 3(a) shows that the weight of each feature category is accumulated according to the date of the image. The top 6 features are SWIR1, SWIR2, RED1, RED2, RED3, RESI, and A. Figure 3 (b) in the figure shows that the weight of each image date is accumulated according to the current date feature category. Therefore, it is believed that the dates that are more important for classification are March 1 to April 10, September 17 to October 7, and November 6 to December 6.

[0083] like Figure 4 As shown in the figure, the orchard distribution results obtained by pixel-based and object-based methods were statistically analyzed for the classified areas of easily confused crops, and the consistency of crop distribution was verified by comparing the statistical classified areas with the reference areas. The reference area for persimmon planting was provided by the Linyi County Statistics Bureau in 2019, and the reference area for apple, peach, pear, and pomegranate planting was obtained based on the 2020 Linyi County Orchard Development Survey and Research Report. The area of apples is not much different; the classified areas of peaches, pears, and pomegranates are quite different from the reference areas.

[0084] In summary, the present invention takes orchards with complex planting conditions as research objects, and provides reference information for remote sensing multi-classification of other easily confused crops; the present invention utilizes a multi-layer classification method, which can significantly reduce the interference caused by some simply distinguished crop types on the classification results of difficult-to-distinguish crops, facilitates classification, and can obtain more accurate and reliable classification results; the present invention evaluates the importance of features based on the ReliefF algorithm, avoids the "Hughes effect" problem in time series, and screens out important classification features and key classification time windows; fourth, the orchard classification method of the present invention based on Sentinel-2 time series data and GEE platform has simple data source acquisition, and with the help of the computing advantages of the GEE platform, it can draw orchard planting distribution maps of large areas.

Claims

1. An object-oriented orchard classification method based on Sentinel-2 time series and ReliefF, characterized by: The method comprises the following steps in sequence: (1) Obtain Sentinel-2 data of the study area and perform preprocessing operations such as screening, cloud removal, cultivated land masking, and time series reconstruction on the Sentinel-2 data to obtain Sentinel-2 time series data; (2) Use the multi-layer classification method to first classify easy-to-classify crops and obtain the easy-to-classify crop classification results. Use the easy-to-classify crop classification results to mask out the easy-to-classify crop pixels of the Setinel-2 time series data and obtain the easy-to-mix crop interest area; (3) Based on the easily confused crops region of interest, generate a classification feature set of easily confused crops; (4) The classification feature set of easily confused crops is sorted using the ReliefF algorithm, and the top 25%, 50%, 75%, and 100% features are input into three classifiers respectively to obtain the best feature set and the best classifier; (5) Perform SNIC image segmentation and combine the best feature set and the best classifier to classify easily confused orchards and obtain the easily confused crop classification results; (6) Evaluate the classification results of easily confused crops; The step (1) specifically includes the following steps: (1a) Obtain Sentinel-2Level-2A data of the study area on the GEE platform; (1b) The Sentinel-2 Level-2A data were screened for image data with cloud cover less than 10%, and then the cloud fraction method was used to remove the cloud for the image data with cloud cover less than 10%, and the cultivated land mask was performed on the cloud-removed data; (1c) performing time series reconstruction on the Sentinel-2 data after the cultivated land mask, wherein the time series reconstruction includes 10-day median synthesis, linear interpolation and SG filtering, and finally obtaining the Sentinel-2 time series data; The step (2) specifically includes the following steps: The multi-layer classification method (2a) is to classify crops into easy-to-classify crops and easy-to-confuse crops based on the time series differences of the Normalized Difference Vegetation Index (NDVI) of crops. The first layer is to classify easy-to-classify crops. The easy-to-classify crops classification uses the first-order difference of NDVI time series data and its third harmonic fitting parameters, as well as the red edge, near-infrared and short-wave infrared bands as the classification features of easy-to-classify crops. The calculation formulas for the first-order difference of NDVI time series data and its third harmonic fitting parameters are shown in formulas (1) and (2): ΔNDVI k =NDVI K+1 -NDVI k (1) Among them, ΔNDVI k is the first-order difference of the kth image in the NDVI time series data; NDVI k+1 is the k+1th image in the NDVI time series data, k=0,1,...,28; NDVI t It is a third harmonic fitting of the NDVI time series data of a pixel at time t, a0 is the harmonic remainder, a i and b i is the coefficient of each harmonic, θ i is the initial phase of the i-th harmonic; t is the observation time, t∈[0,1], 0 represents January 1 of the year, and 1 represents December 31 of the year; (2b) The classifier for easy-to-classify crops uses decision tree CART and random forest RF, gradually searches for the optimal parameters of decision tree CART and random forest RF, and inputs the easy-to-classify crop classification features into the classifier to obtain the easy-to-classify crop classification results; (2c) The easy-to-classify crop classification result mask is used to remove the easy-to-classify crop pixels in the Setinel-2 time series data to obtain the easy-to-classify crop area of interest.

2. The object-oriented orchard classification method based on Sentinel-2 time series and ReliefF according to claim 1 is characterized in that: In step (3), the classification feature set of the easily confused crops includes band features, vegetation index and texture features in Setinel-2 time series data.

3. The object-oriented orchard classification method based on Sentinel-2 time series and ReliefF according to claim 1 is characterized in that: The step (4) specifically includes the following steps: (4a) Select the pixel feature values covered by the easily confused crop samples from the classification feature set of easily confused crops, calculate the feature importance and sort them using the ReliefF algorithm, and use the feature importance as the feature weight to cumulatively analyze the key features of easily confused crops; (4b) The top 25%, 50%, 75% and 100% of the feature importance are input into the decision tree CART, random forest RF and support vector machine SVM classifiers respectively, and the best classifier and the best feature set are screened.

4. The object-oriented orchard classification method based on Sentinel-2 time series and ReliefF according to claim 1, characterized in that: The step (5) specifically includes the following steps: (5a) Use the SNIC algorithm to segment the image at different scales based on the optimal feature set to obtain the optimal scale segmentation result; (5b) performing mean calculation on the best feature set for each object in the segmentation result to obtain the object-oriented best feature set; (5c) The optimal feature set of the facing object is input into the optimal classifier to obtain the classification result of easily confused crops.

5. The object-oriented orchard classification method based on Sentinel-2 time series and ReliefF according to claim 1, characterized in that: In step (6), the methods for evaluating the classification results of easily confused crops include three methods: the first method is to evaluate the classification results of easily confused crops through mapping accuracy PA, user accuracy UA and overall accuracy OA; the second method is to superimpose and compare the sample areas collected on the spot with the classification results of easily confused crops. The closer the boundaries of the classified plots are, the better the classification effect; the third method is to compare the classified areas of the main crops in the classification results of easily confused crops with the actual areas shown in the survey report of the local government to verify the accuracy of the classification.

Images