Walnut Acreage Extraction and Growth Monitoring Based on the NDVI Time Series and Google Earth Engine

Abstract

Walnut (Juglans regia) planting is the main economic pillar industry in southern Xinjiang. Based on the Google Earth Engine (GEE) cloud platform, the NDVI maximum synthesis method was used to estimate changes in the walnut cultivation area in Ganquan Town, South Xinjiang, from 2017 to 2021. The simultaneous difference between NDVI and meteorological conditions was also used to monitor the growth and correlation analysis of walnuts from April to September 2021. To improve the classification accuracy of the extracted walnut plantation area, Sentinel-2 image data were selected, and features were trained using the random forest algorithm, and by combining topographic features, texture features, NDVI, and EVI. The results show that, compared with Statistical Yearbook data, the average error of the extracted walnut planted area is less than 10%, the overall classification accuracy is 92.828%, the average kappa coefficient is 90.344%, and the average walnut classification accuracy is 94.4%. The accuracy of the data was significantly improved by adding vegetation indices EVI and NDVI compared with the single vegetation index. An analysis of the results from monitoring comparative growth shows that the growth of walnuts in Ganquan was better during the hardcore and oil transformation stages compared with 2020, and in the fruit development stage, the growth was the same as in 2020, and overall, the growth of walnuts in 2021 was better than in previous years.

1. Introduction

Walnuts are rich in nutrients such as proteins and fats and have a high economic value, making them a major part of the economy in southern Xinjiang, where they are a characteristic forest fruit of the region [1]. Located at the north-west edge of the Taklamakan Desert, Ganquan Town has rich light and heat resources, and is mainly planted with walnuts. Its planting area accounts for 97% of the walnut planting area in the Alar City of the First Agricultural Division; moreover, the town has become a pillar of the industry for poverty alleviation in the area. It is important to promote the distribution and growth of walnuts in Ganquan Town in order to ensure their production.

Field surveys, as a traditional research method, are usually used to extract the crop’s planting area, and they have poor timeliness, lack accurate spatial information, and can be easily interfered with. With characteristics such as high timeliness, low costs, and wide coverage, remote sensing technology makes it possible to extract and monitor large areas, and it also provides crop information over a long-term period. It has become the main method for monitoring the distribution and growth of crops [2,3,4]. The most commonly used methods for extracting crop planting areas nationally and internationally today are detailed below.

(1) The threshold method sets thresholds for different crops via sequential variations of multiple vegetation indexes, such as EVI, VDVI, NDVI, etc., in combination with supervised classification, to extract the spatial distribution information of different crops [5]. Based on a MODIS-EVI image and phenological characteristics of maize, Zhang et al. [6] extracted the planting area of maize in north-east China, adopting the threshold method. The overall accuracy was 79%, which is higher than that of statistical data, with an R² of 0.82 and RMSE of 283.98. (2) The twi-difference algorithm is used to classify and extract different crops by identifying crop pixels with obvious crests [7]. Zhang X et al. [8] proposed identifying and extracting the winter wheat planting area in the Heilongjiang region from 2014 to 2017. Their work was based on MODIS remote sensing imagery, a technique based on the normalized difference vegetation index (NDVI) and time series coefficient of variation (NDVI-CV) combined with the NDVI curve features of different objects on the ground and second-order differencing (3) The supervised classification method: in this method, after classifying different ground objects in remote sensing images, the spatial distribution information of different crops is extracted with the use of decision trees [9], support vector machines [10,11], random forests [12,13,14], etc. Among the above-mentioned supervised classification algorithms, the random forest algorithm (RF) has a high degree of automation and low computational overhead, making it easy to implement [15], and it is widely used in remote sensing image classification and other research. Dong et al. [16] constructed the MODIS-NDVI data set, smoothed it with a harmonic analysis algorithm, and extracted and analyzed the planting area of winter wheat and summer corn within the study area from 2004 to 2016 using the random forest algorithm. It was verified that user accuracy was 94.5%, which was good and consistent with the Statistical Yearbook database.

The Google Earth Engine (GEE) cloud platform, developed by Google, integrates a large amount of satellite remote sensing data, and the parallel computing mechanism improves the efficiency of data processing [17]. It is also widely used in vegetation cover changes, the extraction of crop planting areas and classification, and growth monitoring. Shelestov et al. [18] used Landsat 8 as a data source, based on the GEE cloud platform and used neural networks, support vector machines, decision trees, and random forest classification methods for the large-scale classification and mapping of crops in Ukraine, comparing support vector machines and random forests for higher classification accuracy. Jin et al. [19] used the random forest algorithm to classify maize and non-maize areas in Tanzania and Kenya, based on the GEE cloud platform combined with Sentinel-1 and Sentinel-2 image data, with good accuracies of 79% and 63%, respectively.

In terms of crop area extraction and growth monitoring, previous studies have focused on food crops such as wheat and rice [20,21,22,23], and relatively few studies have been conducted on orchards such as walnut trees. Moreover, most studies chose MODIS image data for monitoring and analysis. The low spatial resolution of the MODIS image data can lead to a large number of mixed pixels. This study selected 10 m resolution Sentinel-2 image data as its data source, constructed a training feature set on the GEE cloud platform, and extracted the walnut planting area of Ganquan Town from 2017 to 2021 using the random forest algorithm. This study also combined these methods with MODIS image data to monitor and analyze the growth of walnuts in Ganquan Town from mid-April to September 2021, providing references and objective data support for realizing dynamic crop monitoring, which was further supported by high-frequency and medium- and high-resolution images, building a crop information extraction system based on the cloud platform, machine learning, and agricultural production.

2. Materials and Methods

2.1. Research Area and Data Source

2.1.1. Overview of the Research Area

Ganquan Town (40°22′30″ N, 80°03′45″ E) is located on the north-west edge of Taklimakan Desert under the jurisdiction of the First Agricultural Division of Alar City, and it experiences droughts and a warm, temperate, continental, and desert climate, with long sunshine durations and rich light and heat resources (Figure 1). In addition, the temperature varies substantially from daytime to nighttime, with an average annual temperature of 11.53 °C, a highest temperature of 43.9 °C, and a lowest temperature of −28.8 °C. The average frost-free period is 195 days; the annual accumulated temperature is 4620.8 °C; the annual total solar radiation is 142 kcal/cm²; the annual average wind speed is 21 m/s; the annual average sunshine duration is 2793.4 h; the annual average precipitation is 73.5 mm; the annual average evaporation is 1748.75 mm. In summary, the area is suitable for planting cotton, walnut, and other crops.

2.1.2. Image Data Set and Preprocessing

Taking Sentinel-2 and MODIS satellite image data stored on the GEE cloud platform as a data source, this study extracts the walnut planting area in Ganquan Town and monitors its growth. Equipped with an MSI multispectral imager, Sentinel-2 covers 13 wavebands, including 2 satellites A and B, with a reentry cycle of 5 days and a spatial resolution of 10 m [24]. The higher spatial resolution reduces the impact of mixed pixels on the extraction of the walnut planting area. MODIS MO9GQ is a product of surface reflectance, with a reentry cycle of 1 day and a spatial resolution of 250 m; it has high temporal resolution and can continuously track the growth of walnuts [25].

To improve the availability of the images, the image’s time range was selected with respect to the walnut phenology in Ganquan Town, as shown in

Table 1. The walnut growth cycle in Ganquan Town.

Month	Growth Period of Walnuts
Between April and June	Fruit development period
July	Hardcore period
Aug	Oil conversion period
Mid-September	Mature period

. The QA60 band cloud-bearing information of Sentinel-2 images was marked for de-clouding and then filtered to obtain images with less than 30% cloud content. Image pre-processing operations, such as image de-clouding, cropping, and fusion, were performed in GEE. After the de-clouding process was completed, the images were supplemented with images from adjacent years, a time window was set, the maximum value method was used to synthesize the images, and the images were mosaicked and cropped according to the extent of the study area; specific image information is shown in

Table 2. Details of selected Sentinel-2 images.

Year	Data	Filter Cloud Amount	Number of Images
2017–2018	1 April to 1 July	<20%	18
	1 July to 15 September	<20%	13
2018–2019	1 April to 1 July	<20%	18
	1 July to 15 September	<20%	13
2019–2020	1 April to 1 July	<10%	18
	1 July to 15 September	<10%	13
2020–2021	1 April to 1 July	<10%	18
	1 July to 15 September	<10%	13

.

2.1.3. Basis of Sample Data and Authentication Data

Classification accuracy is affected by the sample number and the accuracy; thus, the number of training samples should be sufficiently large and representative. The ground objects in the research area are classified into five categories: impervious surface, water body, walnut, other orchards, and other crops. This research selects 225 sample points via a visual interpretation of high-resolution images in Google Earth to obtain sample data. In addition, sample data are evenly distributed, from which 110 walnut sample points are selected. The other sample points are water bodies, impervious surfaces, and other vegetation. The stratified random sampling method is selected by comprehensive consideration: 70% of samples are set as the training set of the classification model, and 30% of samples are set as the test set to verify the classification accuracy. The basis for sample selection is shown in

Table 3. Basis for sample selection.

Specimen Type	Interpretation Signs	Description
Impervious surface		The impervious surfaces in the study area consist of town buildings, roads, and bare ground, which are identifiable on Google Earth by their texture information and occur in patches.
Water		The water body area in the study area is mainly composed of reservoirs and rivers, etc. The edges are obvious with respect to texture, and with respect to color, the water body area is cyan and dark cyan.
Walnuts		The walnuts in the study area are mainly distributed around the area. On Google Earth, the texture is clearer, occurs in patches, exhibits a more regular shape (rectangular), and is dark green in color.
Other orchards		Other orchards in the study area are mainly located in the rural periphery of the plain zone, exhibiting clearer textures, more regular shapes (rectangular), and light green color in Google Earth.
Other crops		The study area is also planted with cotton, pepper, wheat, corn, and many other crops in addition to orchards. On Google Earth, it exhibits clearer textures, more regular shapes (rectangles), and bright green color.

, and authentication data were obtained from the walnut planting area data released in the Statistical Yearbook of Alar City.

2.2. Research Methods

2.2.1. Technical Process

Taking the walnut planting area in Ganquan Town as the research object and the Sentinel-2 image data provided by the GEE cloud platform as the data source, this thesis classifies land types in the research area by combining spectral features, terrain features, textural features, and the vegetation index. It then extracts the walnut planting area with the use of a random forest classification algorithm. This was combined with MODIS image data, and the walnut growth in the research area from April to mid-September 2021 was analyzed with the use of the NDVI difference model. The specific research process is as follows: (1) first, the obtained Sentinel-2 image data set is preprocessed by using methods such as image cloud removal, clipping, and fusion; (2) the feature set is constructed via spectral features, terrain features, textural features, and vegetation index features; (3) ground objects are classified, and the planting area of walnut is extracted; (4) the extracted area is compared with the Statistical Yearbook data to verify its accuracy; (5) NDVI with MODIS image data are calculated during the monitoring period, the NDVI maximum value for every ten-day image is synthesized, and then the difference value is calculated; (6) walnut growth in the same month of 2020 is monitored with the method of contemporary comparisons. The flow chart is shown in Figure 2.

2.2.2. Information Acquisition of the Walnut Planting Area

Taking Sentinel-2 image as the data source, in this study, the NDVI sequential variation of different ground objects from April to September was calculated in GEE (as shown in Figure 3). Figure 3 shows that the NDVI value of walnuts in the research area is significantly higher than that of the other four types from April to the first and middle ten days of July. The months from April to July comprise the fruit development period and core-hardening period of walnuts, and the leaves of walnut trees are constantly turning green. In order to highlight the characteristics of walnuts, this study selects image data from April to the first and middle ten days of September, calculates the NDVI value of each image, sorts each pixel of the image set according to the NDVI value from small to large in GEE, and extracts the synthesized image according to the maximum NDVI value (including original spectral information). The NDVI calculation formula is as follows [26].

$$NDVI=\frac{NIR-\mathit{Red}}{NIR+\mathit{Red}}$$

(1)

where Red is the red waveband, and NIR is the near-infrared waveband.

Ganquan Town is surrounded by the Gobi Desert, and the field is mainly comprised of sandy soil; thus, identifying the walnut planting area is difficult. In addition to the spectral features, textural features and terrain features are added to build an image feature set, in order to improve the classification accuracy of ground objects in Ganquan Town and better extract the walnut planting area. This research selects DEM data drawn by SRTM (shuttle radar topography mission) as data for building the topographic features of the research area, with a spatial resolution of 30 m. Different ground objects have different textural information, and textural features reflect the distribution characteristics of image pixels. In summary, classification and recognition can be conducted according to the textural information of the research object [27]. The methods commonly used to extract texture features are as follows: gray-level co-occurrence matrix, local binary patterns, Gabor transform feature extraction, etc. This study builds textural features with a gray-level co-occurrence matrix (GLCM), and the calculation function of the gray-level co-occurrence matrix [28] is provided by the GEE cloud platform. To avoid redundancy, this study selects four types of texture information with low correlations for calculations [29,30]: angular second moment (ASM), inverse difference moment (IDM), contrast ratio (CON), and correlation (CORR). The calculation formula is as follows:

$$ASM={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}M{(i,j)}^2}}$$

(2)

$$IDM={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}\frac{M(i,j)}{1+{(i-j)}^2}}}$$

(3)

$$CON={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}M(i,j){(i-j)}^2}}$$

(4)

$$CORR={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}\frac{(i-\mu )(j-\mu )M{(i,j)}^2}{\epsilon }}}$$

(5)

$$\mu ={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}M(i,j)}}$$

(6)

$$\epsilon ={\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}M(i,j)\times {(i-\mu )}^2}}$$

(7)

where k is the gray level of the original image, M (i, j) is the gray-level co-occurrence matrix generated based on the original image, $\mathsf{\mu }$ is the mean value, and $\mathsf{\epsilon }$ is the variance.

With respect to the selection of classification methods, based on the GEE cloud platform, this study classifies remote sensing images in the research area with the random forest algorithm, and extracts the planting area of walnuts in the study area from 2017 to 2021. There are multiple decision trees in the random forest algorithm (RF), and each decision tree is a classifier. When a sample is inputted, any tree in the random forest will generate a corresponding classification result; the random forest algorithm collects the classification results of all subtrees and takes the category with the most covered trees as the final outputted result [14].

2.2.3. Walnut Growth Information Extraction

After completing the classification of ground objects and the extraction of the planting area of walnuts in the research area, a mask of the extracted walnut planting area would be generated, and the mask area was calculated based on the MODIS image, to monitor the growth of walnuts in the research area annually from 2017 to 2021. In order to ensure data integrity and reduce the impact of clouds, in this study, Sentinel-2 and MODIS images were selected for combination. Firstly, images of Ganquan Town from April to mid-September were obtained by the projection method of Sentinel-2 images from 2017 to 2021; secondly, the maximum value of NDVI was synthesized with image data from April to mid-September every year; finally, walnut growth from April to mid-September in 2021 was compared to the NDVI value within the same period in 2020 using the NDVI difference model [29]. The NDVI difference model is shown in Formula (8):

$$NDV{I}_n=NDV{I}_N-NDV{I}_{N-1}$$

(8)

where NDVI_N and NDVI_N−1 are the maximum NDVI values of the corresponding year; N and N − 1 represent the current year and the previous year, respectively.

According to the calculation results of ${\mathrm{N}\mathrm{D}\mathrm{V}\mathrm{I}}_{\mathrm{n}}$, walnut growth in Ganquan Town is classified into three categories based on normal statistical rules: when f(NDVI_n) = −1, it indicates that growth is worse than in previous years; when f(NDVI_n) = 0, it indicates that growth is the same as that of previous years; when f(NDVI_n) = 1, it indicates that growth is better than in previous years.

2.2.4. Accuracy Verification

After extracting the walnut planting area in the research area, the extracted walnut planting area from 2017 to 2020 is compared, using remote sensing, to data from the local Statistical Yearbook, and the absolute error and relative error are calculated. The relative error can objectively describe extraction accuracy and better reflect the reliability of the extracted walnut planting area; the smaller the relative error, the higher the extraction accuracy. The calculation formulas for the absolute error and relative error are as follows:

$$e=\left|I_{rs}-I_{re}\right|$$

(9)

$$p=\frac{e}{I_{re}}\times 100\%$$

(10)

where e is the absolute error, I_rs is the remote sensing extraction area, I_re is the statistical planting area, and p is the relative error.

After image classification, this study conducts tests on the recognition results of authentication data using a confusion matrix, and carries out calculations on the overall classification accuracy (OA), kappa coefficient, and walnut classification accuracy (W) [30]. The calculation formula is as follows:

$$OA=\frac{{\displaystyle {\sum }_{i=0}^{k-1}C(i,j)}}{{\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}C(i,j)}}}$$

(11)

$$Kappa=\frac{OA-{\displaystyle {\sum }_{x=0}^{k-1}[{\displaystyle {\sum }_{i=0}^{k-1}C(i,x)\times {\displaystyle {\sum }_{j=0}^{k-1}C(i,j)}}]}/[{\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}C(i,j)}}]}{1-{\displaystyle {\sum }_{x=0}^{k-1}[{\displaystyle {\sum }_{i=0}^{k-1}C(i,x)\times {\displaystyle {\sum }_{j=0}^{k-1}C(i,j)}}]/[{\displaystyle {\sum }_{i=0}^{k-1}{\displaystyle {\sum }_{j=0}^{k-1}C(i,j)}}]}}$$

(12)

$$W=\frac{C(a,a)}{{\displaystyle {\sum }_{j=0}^{k-1}C(a,j)}}$$

(13)

where k is the number of sample types, C (i, j) is the confusion matrix, and the main diagonal position is the number of correct classifications relative to each category in the sample’s verification data.

3. Results

3.1. Area Monitoring Results

Based on the GEE cloud platform, this thesis performs feature construction with Sentinel-2 images as the data source, extracts and calculates the walnut planting area in Ganquan Town from 2017 to 2021 with the random forest classification algorithm, and calculates the relative error via a comparison with statistical data (as shown in

Table 4. Comparison of the walnut planting area extracted by remote sensing and the statistical walnut planting area.

Classification Feature	Year	Extraction Area (hm2)	Statistical Area (hm2)	Absolute Error (hm2)	The Relative Error (%)
Spectrum + Terrain + Texture + NDVI	2017	7700	8800	1100	12.5
	2018	7000	8800	1800	20.4
	2019	4900	5333.3	433.3	8
	2020	5200	5333.3	133.3	2
	2021	4600	5333.3	733.3	13
Spectrum + Topography + Texture + NDVI + EVI	2017	7700	8800	1100	12.5
	2018	7000	8800	1800	20.4
	2019	5000	5333.3	333.3	6
	2020	5400	5333.3	66.7	1
	2021	4700	5333.3	633.3	11

). Additionally, when performing feature construction, this study adds vegetation index EVI; compared with the non-adding vegetation index EVI, the relative error of the walnut planting area extraction is reduced, and the average relative error is within 10%. The relative error of the walnut planting area was reduced from 8% to 6% in 2018–2019, from 2% to 1% in 2019–2020, and from 13% to 11% in 2020–2021, with good extraction effects. In addition to the above verification, this thesis calculates the overall classification accuracy, kappa coefficient, and walnut classification accuracy (

Table 5. Comparison of the extraction accuracy of the walnut planting area based on the confusion matrix in the research area.

Classification Feature	Year	Overall Accuracy (%)	Kappa (%)	Walnut Classification Accuracy (%)
Spectrum + Terrain + Texture + NDVI	2017	89.95	86.75	92.40
	2018	90.43	87.41	90.90
	2019	93.29	90.70	89.41
	2020	94.47	92.38	88.37
	2021	94.47	92.40	92.77
Spectrum + Topography + Texture + NDVI + EVI	2017	89.95	86.84	94.66
	2018	91.38	88.65	92.30
	2019	93.29	90.72	89.28
	2020	95.02	93.11	89.53
	2021	94.50	92.40	92.80

) after classification, based on the confusion matrix for the verification samples. It was observed that, after adding EVI, the overall accuracy, kappa coefficient, and walnut classification accuracy improved. It was observed from the analysis above that the method adopted in this research has higher accuracy and better effects when extracting information regarding walnut planting areas in Ganquan Town.

The extraction result of the walnut planting area shows that, based on the GEE cloud platform, the random forest classification algorithm can preprocess remote sensing image data quickly and quickly map the spatial distribution information of walnut planting areas in Ganquan Town. Spatial information on the walnut planting area in Ganquan Town from 2017 to 2021 is shown in Figure 4, and it can be observed that walnuts are mainly distributed around the center of Ganquan Town. On account of the special climate in Southern Xinjiang, Ganquan Town has been dry for a long time, and is rich in water resources near the center of the town, which is conducive to walnut growth. This shows that, in Figure 4, the walnut planting area measured by remote sensing in Ganquan Town from 2017 to 2021 is generally stable; the walnut planting area measured by remote sensing in 2020 increased compared with that of 2019, which is consistent with the area’s change measured by the Statistical Bureau of Alar City.

3.2. Growth Monitoring Results

As discussed in Section 3.1, the spatial distribution information of walnut planting in the research area extracted by the method is reliable and accurate. This research carries out a statistical analysis of the annual walnut planting area in Ganquan Town from 2018 to 2021 and calculates the number of pixels corresponding to the NDVI increment from 2018 to 2021 with the use of the NDVI difference model compared to the same period in previous years (as shown in Figure 5). It can be observed from Figure 5a that, although the NDVI increment in 2018 increased at −0.3 compared with the same period in 2017, the number of pixels within the NDVI increment range from 0.6 to 0.7 is significantly higher than the number of pixels within the NDVI increment range from 0 to −0.3, indicating that the NDVI value of most regions planted with walnut in this period is higher than that of previous years, and walnut growth is better than that of previous years. Compared with the same period in 2019, the number of NDVI increment pixels in 2020 is concentrated between 0.5 and 0.6, indicating that overall growth is good; the growth of more than 90% of walnut planting areas is the same as that of previous years. Compared with the same period in 2020, the number of NDVI increment pixels in 2021 is concentrated between 0.6 and 0.7, and the overall growth of the planting areas is better than that of 2020.

Walnut growth is greatly influenced by climate change. As a result, this thesis conducts statistical calculations on meteorological weekly report data from April to September during the period from 2017 to 2021, as released by the Alar Meteorological Bureau and analyzes walnut growth in Ganquan Town. Figure 5 and Figure 6 show that, compared to the same period in previous years, the growth of walnuts in 2020 is better than that of 2019. Figure 7 shows that the temperature in July 2020 is higher than that of 2019, and higher than the same period in previous years. In July, walnuts enter the fruit core-hardening stage, and a sufficient light duration is conducive to walnut growth, so walnut growth is better than that of 2019. The temperature in July and August 2019 is higher than that of 2018, and the temperature drops from the end of August to September as a whole, which is lower than the same period in 2018. Then, walnuts enter the mature stage, and light and temperature have little impact on walnut growth. Overall, walnut growth in 2019 is still better than that of the previous years and in 2018.

Via a comparative analysis between 2020 and the same period in previous years, it can be observed that walnut growth in Ganquan Town in 2020 is better than that of 2019, and walnut growth in 2019 is better than that of 2018. In terms of walnut yield, research data in this study show that from the germination stage to the fruit expansion stage, or from the oil conversion stage to the mature stage, the regions with an NDVI increment greater than 0 are larger than the regions with an NDVI increment lower than 0, and the regions with good walnut growth are greater in number than regions with poor walnut growth. In short, the overall walnut yield in Ganquan Town in 2021 is better than that of previous years.

4. Discussion

4.1. Accuracy Evaluation on the Extraction of Walnut Planting Areas in the Research Area

This study obtained the Sentinel-2 image data set of Ganquan Town from 2017 to 2021 by obtaining a large number of remote sensing image data stored on the GEE cloud platform, using Sentinel-2 image data with a spatial resolution of 10 m. The thesis of this study is to perform image pre-processing, such as cloud removal, inset, and clipping, in order to avoid disturbances such as image loss, cloud cover, etc., and to ensure the quality of remote sensing image data. Previous studies have shown that the random forest algorithm has strong anti-interference and anti-overfitting capabilities. Compared with other machine learning algorithms, it has high accuracy and good effects on crop classification [31,32,33]. This study extracts the walnut planting area in Ganquan Town from 2017 to 2021 using the random forest algorithm, and the characteristics of walnut planting are highlighted via the EVI time series curve in image synthesis. When building the feature data set, the thesis combines textural features, terrain features, and vegetation indices NDVI and EVI, reducing the number of mixed pixels of the remote sensing image.

After adding the enhanced vegetation index (EVI), classification accuracy is improved compared with the accuracy of spectrum + terrain + texture + NDVI. Subdividing orchard types with NDVI alone is difficult, and NDVI can be easily saturated in areas with high vegetation coverage, while the addition of blue-light wavebands to EVI reduces the impact of soil and the atmosphere [34], which is more suitable for monitoring dense vegetation areas. Compared with the NDVI construction feature data of a single vegetation index, the EVI accuracy added to the vegetation index significantly improved; the average overall accuracy increased from 92.522% to 92.828%, and the average Kappa coefficient increased from 89.928% to 90.344%. Compared with the statistical data released in the yearbook, the overall relative error is also small. Although the overall extraction effect of this study is good, there are still years exhibiting large errors in the extracted area. For example, in 2018, the difference between the remote sensing extraction of the planting area in the research area and statistical data is 1800 ha, with an absolute error of 20.4%.

4.2. Remote Sensing Monitoring of Walnut Planting in the Research Area

Based on the extracted walnut planting information in the research area from 2017 to 2021, this study compares the planting areas of adjacent years with the NDVI difference method. Taking Landsat TM and DEM images as data sources, Gandhi et al. [35] used the NDVI threshold method to extract the spatial information of rice, wheat, and maize in the Jiangsu Province by building different NDVI threshold vegetation models. To summarize, many mixed pixels in the walnut planting area are extracted by this method. In order to extract more accurate planting areas of walnuts and achieve increasingly accurate monitoring with respect to walnut growth, this study extracts walnut planting areas and monitors walnut growth using the random forest algorithm, combined with the NDVI difference method, and adopts S-G filtering to correct the time series curves of NDVI and EVI, so as to reduce monitoring errors and improve accuracy. It can be observed from Figure 5 that the number of pixels in the NDVI increment is inclined toward the positive value range, indicating that growth is better than the previous year.

Numerous previous studies have shown that GEE can be effectively applied to crop classification and area monitoring [36,37,38]. In this study, walnuts were extracted from 2017 to 2021 in Ganquan Town, and the GEE cloud platform was used to quickly map the walnut planting area. From Figure 5, it can be observed that, although NDVI increased in the −0.3 range in 2018 compared to 2017, the overall increase was concentrated within the 0.6 to 0.7 range. In 2019, compared to 2018, the NDVI increase was concentrated within the 0.1 to 0.4 range. In 2020, compared to 2019, the NDVI increase was concentrated within the 0.5 to 0.6 range. In 2021, compared to 2020, the increase in NDVI was concentrated within the 0.6 to 0.7 range. This means that the overall growth of walnuts in the current year is better than the previous year.

5. Conclusions

Based on the GEE cloud platform, this study used Sentinel-2 multi-temporal image data as the data source, extracted the walnut planting area of Ganquan Town from 2017 to 2021 after feature construction and optimization using the random forest classification algorithm, and applied the NDVI difference calculation method to the extracted walnut planting area in 2021 to realize research on walnut growth monitoring. Combined with climatic factors, this study monitored and analyzed walnut growth in Ganquan Town from April to mid-September 2021, and the results show that the method adopted in this study not only allowed for obtaining high-precision walnut planting distribution information, but also allowed for high-frequency growth monitoring. However, there are some shortcomings in the monitoring of walnut growth in this study: there are still errors in the data due to the inevitable correlation between the mixed pixels in the images and the original spectral bands.

Subsequent studies will perform feature operations on the raw spectra, combined with deep learning, to improve the accuracy of classification extraction. This study only focuses on the changes in walnut cultivation in Ganquan Town within the last five years, and subsequent studies will monitor the growth of walnut cultivation areas in the entire southern border region over a longer time period.