Next Article in Journal
Evaluation and Applicability Analysis of GPM Satellite Precipitation over Mainland China
Next Article in Special Issue
Hybrid Deep Learning and S2S Model for Improved Sub-Seasonal Surface and Root-Zone Soil Moisture Forecasting
Previous Article in Journal
A Lightweight Dual-Branch Swin Transformer for Remote Sensing Scene Classification
Previous Article in Special Issue
Mapping the Age of Subtropical Secondary Forest Using Dense Landsat Time Series Data: An Ensemble Model
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 15
Issue 11
10.3390/rs15112867
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Mapping Insect-Proof Screened Citrus Orchards Using Sentinel-2 MSl Time-Series Images

by
Guobin Yu
1,
Li Zhang
1,2,*,
Lingxia Luo
1,
Guihua Liu
1,2,
Zongyi Chen
3 and
Shanshan Xiong
1
1
School of Geography and Environment, Jiangxi Normal University, Nanchang 330022, China
2
Key Laboratory of Poyang Lake Wetland and Watershed Research, Ministry of Education, Nanchang 330022, China
3
Hydrology and Water Resources Monitoring Center for Ganjiang River Upstream, Ganzhou 341000, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2023, 15(11), 2867; https://doi.org/10.3390/rs15112867
Submission received: 16 April 2023 / Revised: 26 May 2023 / Accepted: 29 May 2023 / Published: 31 May 2023
(This article belongs to the Special Issue GeoAI and EO Big Data Driven Advances in Earth Environmental Science)
Browse Figures
Review Reports Versions Notes

Abstract

:
Citrus is a crucial agricultural commodity of the hilly subtropical regions of southern China. Attempts in recent years to combat the destructive disease Huanglongbing (HLB) have led to citrus orchards being covered with insect-proof screens (IPS). Understanding which citrus orchards are covered by IPS is crucial for regional water and soil conservation, as well as control of plastic pollution. However, monitoring of orchards is complicated by IPS spectral interference in remotely sensed image classification. Here, an optimal feature combination scheme is developed and tested for mapping citrus orchards that use IPS. Seasonal Sentinel-2 images from 2021 were used to define indices for vegetation, plastic mulch, red edge, and texture. These were combined with topographic and land surface temperature using random forest classification to determine optimal feature discrimination combinations for orchards in Xunwu County, Jiangxi Province. Results show: (1) significantly higher visible light reflectance from IPS orchards ensures spectral discrimination between IPS covered and uncovered orchards. (2) After feature optimization, the seasonal spectral band has the highest accuracy (86%) in single feature classification. The addition of conventional indices and topographic-temperature features improves classification to 92%. (3) Xunwu County had 460 km2 of citrus orchard cover in 2021, with 88 km2 (19%) of that total being covered with IPS. Our method effectively and accurately maps citrus orchards with or without IPS coverage at 10 m resolution. The effective monitoring of large-scale IPS in other regions can now support the development of local and regional sustainable agricultural policies.

Graphical Abstract

1. Introduction

Citrus trees, which include oranges, lemons, grapefruits, pomelos, and limes, belong to the Rutaceae family. China is one of the earliest cultivators of citrus and continues to possess abundant resources and diverse varieties. In recent decades, the demand for citrus has led to a significant expansion of citrus orchards in China. As of 2021, China’s citrus planting area has reached 5648.55 km2 [1]. The subtropical region of southern China, including Jiangxi, Hubei, Sichuan, Guangdong, and neighboring provinces, is a primary production area for citrus. Ganzhou City in Jiangxi Province is a particularly important citrus production base, and it is also one of the most ecologically sensitive areas in the southern hilly region of China. Since 2013, due to the outbreak of citrus Huanglongbing (HLB), which is a devastating disease in citrus orchards [2], a large number of citrus trees have been cut down and burned, and the orchards have been reclaimed and renovated. These activities have resulted in increased soil erosion [3] and greenhouse gas emissions [4] in the southern hilly region of China. To prevent the spread of HLB, insect-proof screens (IPS) have been widely used since 2017. HLB is mainly transmitted by citrus psyllids, and IPS made of polyethylene mesh fabric can prevent insect proliferation. However, IPS are not biodegradable and difficult to recycle, leading to potential problems, such as plastic and soil pollution. Therefore, understanding the distribution of IPS usage in citrus orchards is crucial for quantifying its impact on the environment and planning the future development of the citrus industry.
Citrus planting is greatly affected by human factors and natural factors. Thus, citrus orchards have high heterogeneity in time and space [5]. On a larger scale, accurately estimating the citrus cultivation area is challenging. Fortunately, remote sensing image data can be employed to draw large-scale land use maps at low cost and high precision. In addition, land cover changes can be analyzed, allowing macroscopic monitoring of agricultural planting cycles and trends [6,7,8]. Medium or high-resolution optical satellite data, such as Landsat5 TM and Landsat8 OLI data [5,9], Sentinel-2 MSI data [10], CBERS-2B CCD data [11], and GF-2 PMS data [12], can be used to calculate vegetation indices, water indices [5,9,10], and texture features [13,14] of the image as classification features, which can then be combined with multivariate data, such as topographic factors [5,9]. Machine learning approaches, such as the support vector machine and random forest methods, can handle high-dimensional data features [15,16] and classify citrus orchard areas at a high classification accuracy. Among them, Xu et al. [5] used Landsat-8 OLI data for seasonal compositing and extracted various indices, including NDVI, SAVI, NBR, NDMI, and MNDWI. The random forest algorithm was applied to extract citrus orchards in the southern region of China for the years 2016 and 2018, demonstrating a certain level of reliability for healthy and open-air citrus orchards at a spatial resolution of 30 m. The recent outbreak of HLB has led to large areas of citrus orchards in southern Jiangxi being covered with IPS. Due to the white color of IPS, citrus orchards covered by IPS undergo a transition from a light green patch to a white or gray patch, thereby altering the original spectral features of the orchard. However, previous studies have not addressed or adequately acknowledged this issue, resulting in a gap in the literature regarding the extraction of orchards covered by IPS. Additionally, spectral interference from IPS may lead to inaccurate estimations of citrus orchard areas in earlier methods. To monitor the extent of IPS usage in southern Jiangxi and accurately estimate the area of citrus orchards, it is crucial to improve classification methods that can distinguish between orchards with and without IPS cover.
Through field observation and comparison of their materials, we found that IPS are similar to vegetable greenhouses and plastic mulch in physical characteristics. Medium-resolution optical satellite data, such as Landsat5 TM and Landsat8 OLI data [17,18,19] and Sentinel-2 MSI data [20], can be used to calculate spectral indices, such as the greenhouse vegetable land index (VI) [19], plastic-mulched land cover index (PMLI) [17], plastic greenhouse index (PGI) [21], and others. The extraction of vegetable greenhouses and plastic mulch has achieved satisfactory results, and those methods provided a foundation for us to better classify and discriminate IPS covered and uncovered citrus orchards. IPS coverage will reduce wind speed and solar radiation, thus changing the temperature and humidity inside the net [22,23]. In hilly and mountainous areas, the establishment of a large area of IPS cover will significantly reduce the solar radiation absorption rate in the sunny slope area, hinder the wind flow in the valley, and affect the local microclimate. Based on these observations, we also consider land surface temperature into our classification methods of IPS-covered orchards.
Vegetation characteristics can be used to better classify different vegetation types, and the seasonal composition of images can reflect the spectral characteristics of various vegetation types in different seasons. However, due to the cloudy and rainy climate in southern Jiangxi, it is often difficult for ordinary optical-band satellites to obtain clean images in this region [24]. Sentinel-2 has two complementary satellites and a revisit cycle of five days in middle and low latitudes, which facilitates the acquisition of seasonal cloudless images. In addition, the high spatial resolution of 10 m provides more possibilities for the classification of IPS covered or uncovered citrus orchards.
Multi-temporal image and multi-dimensional classification features require processing very large amounts of data, which makes it difficult to carry out classification calculation on local servers. Recent years have seen the rise in remote sensing computing cloud services, such as Google Earth Engine (GEE), Pixel Information Expert Engine (PIE), and other platforms, which provide speed and convenience in the processing of a large amount of remote sensing image data. The GEE has a vast archive of historical image datasets and integrates a wide range of remote sensing image processing algorithms, providing a strong platform for processing large-scale remote sensing data and image classification [25].
We, here, used seasonal Sentinel-2 MSI image data, and we analyzed the spectral differences between IPS covered and uncovered citrus orchards in the study area. We extracted more than 100 features that may contribute to classification using the GEE platform, using random forest and recursive feature elimination (RFE) methods for classification and feature optimization. This work can provide support for the sustainable development of local citrus cultivation and serve as a reference for remote sensing identification of IPS-covered orchards in other regions.

2. Materials and Methods

2.1. Study Area

Xunwu County is located on the southern border of Jiangxi Province between 24°30′40″–25°12′19″ north latitude and 115°21′22″–115°54′25″E east longitude and is under the jurisdiction of Ganzhou City. It is a mountainous agricultural county, which has a subtropical monsoon climate (Figure 1). The county is rich in sunlight, heat, rainfall, and water resources. It is blocked by high mountains in the east and south and is rarely affected by typhoons. The climate conditions are, thus, very suitable for the growth of citrus fruit trees. The terrain of Xunwu County is dominated by mountains and hills, accounting for 75% of the total area of the county. A large diurnal temperature difference in autumn can promote sugar accumulation in the expansion and ripening stages. The soil in this region is mainly red soil, accounting for 86% of the total soil area [26]. Red soil is rich in iron oxide, aluminum oxide, quartz, and other minerals, is acidic and sticky, and is an ideal soil for citrus growth. In recent years, national poverty alleviation projects have supported the rapid expansion of citrus cultivation in Xunwu County, reaching a peak of 301.73 km2 of planted area and 563,800 tons of output in 2013 [27]. However, in the same year, large areas of fruit trees in some villages and towns were infected with HLB disease, leading to the removal of diseased plants and a reduction of 103.8 km2 of orchard area from 2013 to 2017. Prevention and control measures were then implemented, resulting in an increase in citrus planting areas in 2017. Although HLB has been systematically removed, IPS use has continued for a new round of citrus breeding and protection of disease-free fruit trees. The orchard area covered by IPS has gradually increased since 2017. Consequently, two distinct types of orchard landscapes have emerged: healthy and open-air citrus orchards (Figure 2a), as well as citrus orchards covered with IPS (Figure 2c).

2.2. Data Sources

2.2.1. Sentinel-2

Sentinel 2 is a high-resolution multispectral imaging satellite composed of two satellites, 2A and 2B. Both satellites carry a Multi-Spectral Instrument (MSI). We, here, used the Level-2A product obtained from GEE (dataset ID: “COPERNICUS/S2_SR”) to extract spectral features and classification, which is an orthorectified and atmospherically corrected surface reflectance data set. We utilized 12 bands from the Sentinel-2 image and resampled all bands to a resolution of 10 m using the bilinear interpolation method. Different vegetation types have significant differences in phenology. We consider the climate characteristics in southern China to seasonally compile the images using a median method. The region is subject to frequent rainy conditions, so we use a threshold of less than 20% cloud cover to screen images and use the opaque clouds and cirrus information provided by the “QA60” band in the data for cloud removal. The images may have missing values after cloud removal and compositing, especially during the rainy season when it is cloudy, so the same period images of the previous or subsequent year were used to fill in the missing information. The images used in each season are shown in Table 1.

2.2.2. Landsat-8

Landsat 8 carries two sensors: the Operational Land Imager (OLI) and the Thermal Infrared Sensor (TIRS). We obtained data from GEE for a L2 Landsat product (dataset ID:” LANDSAT/LC08/C02/T1_L2”). OLI images are used to test the classification method of Xu et al. Band10, with a resolution of 100 m in TIRS, was used, here, to obtain the land surface temperature (LST), which is resampled to 10 m. Similarly, we used the “QA_PIXEL” band in the data for cloud removal and seasonal median compositing of images. Cloudy sky conditions often lead to missing values in LST data [28], so the same period images of the previous or subsequent year were used to fill the missing values. The number of images used after screening in each season is given in Table 2.

2.2.3. SRTM

The Shuttle Radar Topography Mission (SRTM) is an international research effort to obtain digital elevation models. The Vision3.0 (SRTM V3) obtained from GEE is used to generate topographic features (dataset ID is“ USGS/SRTMGL1_003”) with a resolution of 30 m.

2.2.4. Land Cover Reference Data

The land cover of Xunwu county was divided into eight categories: impervious, water, forest, cropland, citrus, insect-proof screen (IPS), bush, and bare land. We used Google Earth software to draw sample points on the high-resolution image by visual interpretation. All samples were drawn and sorted by one member of our team to ensure that all samples drawn follow the same standard. The drawing of sample points follows the principle of uniform distribution in Xunwu county, and the proportion of the number of various samples is roughly in line with the actual surface area ratio. A total of 2572 sample points in the study area in 2021 were collected (Table 3). We used 70% of the sample points for training and 30% for validation.

2.3. Methods

To obtain the optimal classification scheme for distinguishing citrus orchards and insect-proof nets, we designed an experimental plan for this study (Figure 3). Firstly, we used seasonal composites of Sentinel-2 images to account for vegetation changes and extracted four types of image features (spectral bands, conventional spectral index, red edge index, and texture). We also retrieved land surface temperature from the thermal infrared band B10 of Landsat 8 and combined it with topographic data to create “Topographic-temperature features.” Each feature is labeled with a number (1–4) to represent the four seasons (spring, summer, autumn, and winter). Overall, we obtained five sets of 39 features (138 when considering seasonal images), which are detailed in Table 4. To reduce redundant variables and identify the optimal classification features, we initially employed the recursive feature elimination algorithm on each set of five features to determine the best single classification feature. Subsequently, the best single feature was combined with other features groups, and the random forest method was employed for classification. The classification accuracy of all combination schemes was compared to determine the optimal classification scheme.

2.3.1. Classification Feature Extraction

(1)
Conventional index features
The conventional index features are spectral indices obtained by combining visible light, near-infrared, and short-wave infrared bands of Sentinel-2 images, which can be calculated from common multispectral satellite images. Figure 4 shows the average spectral curve of various ground objects, drawn by the median composite image of Sentinel-2 MSI during 2021 and the sample points in Table 3. For ground features, such as forest, water, cropland, bare land, and impervious surfaces, the normalized difference vegetation index NDVI [29], enhanced vegetation index EVI [30], normalized difference moisture index NDMI [33], and modified normalized difference water index MNDWI [32] can improve classification accuracy due to their large spectral differences, especially in the red, NIR, and SWIR1 bands. However, other ground features, such as uncovered citrus, IPS covered citrus, and bush, have relatively small spectral differences and are more difficult to classify. For citrus orchards, there are discrete intervals between plants when planted reasonably close, which gives the orchards clear background soil color in an image. To accurately extract the citrus orchard area, the soil-adjusted vegetation index SAVI [31] can effectively reduce the influence of the soil background. We also used the plastic mulch index PMLI [17], which is a normalized operation performed on the red light B4 and short-wave infrared B11 bands, to better distinguish IPS-covered and -uncovered orchards based on their spectral curve comparisons. Furthermore, we considered seasonal changes by calculating the NDVI as a time series, using cloud-free images for the whole year in 2021. We then compiled the maximum and minimum of NDVI [5] to be added to the conventional index features.
(2)
Red-edge index features
We utilized the special red-edge bands B5, B6, B7, and B8A in Sentinel-2 data for combined calculation to reflect vegetation health information. This approach aided in expanding the spectral differences of different vegetation types. To investigate the impact of different red edge indices on the classification of various vegetation types in the study area, we computed seven red-edge indices using seasonal images.
(3)
Texture features
Principal Component Analysis (PCA) was used to analyze the spectral bands of Sentinel-2 seasonal images [38], and the first principal component was selected. We use the Grey-Level Co-Occurrence Matrix (GLCM) [39] calculated over 3 × 3 squares to compose eight types of feature groups.
(4)
Topographic-temperature features
We use seasonal Landsat-8 satellite images and the SMW (Statistical Mono-Window) algorithm to calculate land surface temperature (LST) data, and the contribution of LST data to the classification of IPS covered and uncovered regions was explored. The algorithm is based on an empirical relationship between top-of-atmosphere brightness temperatures in a single TIR channel and LST and utilizes simple linear regression and has simple and inexpensive implementation [40]. The SMW algorithm is described by Ermida, S.L et al. [41] and shared in GEE. In addition, we extract the elevation, slope, aspect, and hill shade of the study area using the GEE platform and SRTM data. To account for the impact of topographic fluctuations on surface temperature in the study area, we combined topographic factors and land surface temperature (LST) data to create a topographic-temperature feature group. The topographic features were selected based on their potential influence on surface temperature, and their combination with LST data allowed us to capture both the local and regional variability in temperature. The resulting feature group can help improve the accuracy of temperature-related analyses and modeling in the study area.

2.3.2. Random Forest

Random Forest (RF) is a machine learning algorithm that integrates multiple decision trees [42]. Random Forest first extracts N training samples from the training set with replacement each time to form a new training set, and then it uses the training set to generate several independent decision trees (which constitute the “random forest”), and the final classification result of each tree is determined by voting [43]. The random forest algorithm is well suited for processing high-dimensional classification features and has several advantages, including high precision, stable performance, and the ability to output feature importance. As a result, it has found widespread use in land use classification mapping [44,45,46]. We, here, use random forest for classification through GEE with the number of trees set to 100.

2.3.3. Feature Optimization

We employed the random forest algorithm to rank feature importance and utilized the recursive feature elimination (RFE) method to screen feature variables within each group and avoid data redundancy. We started by performing random forest classification on the initial feature set with K features, determining the importance of each feature, and calculating the classification accuracy. The feature with the lowest importance was removed from the current feature set, resulting in a new feature subset with K-1 features. This process was repeated until only one feature remained, and the classification accuracy was calculated for each feature subset. The feature subset with high classification accuracy and a small number of features was selected as the result of feature screening.
After screening each feature group, we combined the feature group with the highest accuracy with other feature groups and inputted them into the random forest classifier to calculate the accuracy of the optimal classification model. We compared the accuracy of various feature combinations to obtain the best classification model.

2.3.4. Accuracy Assessment

The verification sample is utilized to compute the confusion matrix and assess the accuracy of each scheme through selected indicators, including overall accuracy (OA), kappa coefficient, user accuracy (UA), and producer accuracy (PA) for both single and combined feature groups.

3. Results

3.1. Single Feature Group Optimization Analysis

We separately performed random forest classification for five extracted classification feature groups and used recursive feature elimination to create a classification accuracy line chart under gradual feature dimension reduction (Figure 5). The accuracy changes in features can be categorized into three stages: stable fluctuation, significant decline, and sharp decline. For instance, in seasonal spectral bands (Figure 5a), OA and Kappa fluctuate steadily during continuous elimination of minimum contribution classification features, with OA around 85% and Kappa around 83%. However, when the feature dimension reduces to 9, the classification accuracy begins to decline significantly, resulting in OA reduction to 82% and Kappa reduction to 79%. When the feature dimension reduces to 3, OA and Kappa decrease sharply. We choose the node before significant decline in OA as the optimal feature dimension of the feature group. For instance, for spectral band (Figure 5a) with feature dimension 9, we select 10 as the optimal feature dimension, retaining the first 10 features as the optimal feature variable after screening.
The same feature optimization is carried out for conventional index features, red edge index features, texture features, and topographic–temperature features. The final retained feature dimension and feature band are shown in Table 5. We obtained Figure 6 by counting the feature variables in Table 5 according to different seasons. The characteristic variables account for the largest proportion in winter, followed by summer and spring, and the smallest proportion occurs in autumn. We speculate that this is due to the following reasons. First, there are fewer cloudy and rainy days and more cloudless images in the study area in winter, so the seasonal image quality after composition is better. Second, citrus orchards also remain evergreen in winter, making them appear much different than yellow grass, cropland after harvest, and other easily confused features.
We optimized and classified the characteristics of five variable groups and evaluated their accuracy, determining the OA, Kappa, UA, and PA of IPS-covered and -uncovered citrus orchards. We added additional spectral features without considering seasonality, and obtained the classification accuracy of six feature groups, as shown in Figure 7. Seasonal spectral bands showed the highest accuracy in OA (86%), Kappa (84%), UA (87%), and PA (84%) of citrus. Non-seasonal spectral bands and conventional spectral indexes also performed reasonably well in various accuracies. However, the red edge index feature, texture feature, and topographic–temperature features had significantly lower accuracies. After comprehensive comparison, we concluded that seasonal spectral bands are the most advantageous for single-feature classification.

3.2. Combinatorial Feature Optimization Analysis

The seasonal spectral bands with the highest accuracy in the single features were combined with other feature groups by an exhaustive method, and a total of 15 different combination methods were obtained (Table 6). These combinations were sorted and coded according to the number of bands. Each combination was then classified by random forest and evaluated for accuracy. Figure 8 shows the accuracy evaluation result of 15 different combinations. The OA of combined feature classification was higher than that of single feature classification, and the OA of all schemes was above 86%. The spectral band combined with topographic–temperature features (C1) had the highest OA of 90% in pairwise feature combination. There was no significant improvement in the classification accuracy of citrus, but the UA and PA of IPS regions increased from 78% and 70% to 88% and 88%. These results suggest that topographic-temperature features have more contribution to the classification of IPS cover. Adding the conventional spectral index (C3) to the spectral band resulted in a significant improvement in UA for citrus, increasing from 87% to 94%. For IPS covered orchards, the UA and PA improved from 78% and 70% to 81% and 79%, respectively. On the other hand, there was no significant improvement in classification accuracy of IPS-covered and -uncovered orchards after adding the red edge index (C2) or texture (C4).
C6, C8, C11, C13, and C15 are high in accuracy indicators, around 85% or higher. These combinations all include conventional spectral index and topographic–temperature features, indicating their importance in classifying IPS-covered and -uncovered orchards. However, red edge index and texture did not contribute significantly. Therefore, we chose the spectral band + conventional spectral index + topographic-temperature (C6) as the optimal scheme due to its lower feature dimension (28), simpler classification model, higher computational efficiency, and key features.

3.3. Optimal Classification Maps of Citrus Orchards

The classification results obtained using the optimal feature combination are presented in Figure 9. We counted the IPS covered and uncovered citrus orchards according to their elevation (Figure 10). Based on the graphs, the orchards are mainly concentrated in low mountain and hilly regions, particularly at elevations ranging from 300–400 m. The central and northern areas of Xunwu County are characterized by low mountains and hills, giving rise to two concentrated regions of citrus orchards. These regions are characterized by favorable levels of accumulated temperature and light and are located at a safe distance from lowlands and plains, which makes them less prone to waterlogging during the rainy season. Furthermore, these regions are relatively close to major towns and enjoy better agricultural infrastructure and management than other areas.
The spatial distribution of IPS-covered citrus orchards was similar to those without IPS, which were both loosely concentrated in the central and northern hilly areas of Xunwu County. However, the distribution density of IPS-covered orchards in the northern hilly area was significantly higher than in the central hilly area. This suggests that there are spatial differences in the severity of disaster and response to HLB in Xunwu County. The northern part of Xunwu County has a trumpet-shaped opening to the north, with high mountains blocking the east and west sides and low terrain in the north and south. This type of environment is conducive to the north–south flow of near-surface air, which assists the transmission of HLB disease by the citrus psyllids and leads to orchards in the northern hills being more vulnerable. In contrast, the central region is surrounded by mountains on all sides, and there is only an outflow in the northeast corner, so IPS-covered regions are mainly seen near this gap, while the distribution is less in other areas. In addition, areas with high density of IPS-covered orchards were mostly near linear impervious areas, which are provincial roads. From this, we speculate that the cross-regional movement of humans may also be one of the potential factors for the spread of HLB.

4. Discussion

4.1. Recommendations for Classifying Citrus Orchards Based on Feature Optimization

Recursive feature elimination (RFE) is a useful technique for optimizing single feature groups by reducing the feature dimension while maintaining classification accuracy. This approach can also improve the computational efficiency and interpretability of the classification model by identifying the most essential variables. For example, according to the optimization results of conventional index features (Table 5), we can know that MNDWI is the most important feature, NDVI, SAVI, NDMI, PMLI also play an important role, but EVI does not seem to be necessary because it is eliminated in the optimization process. Among topographic-temperature features, elevation is the most important, followed by LST and aspect, and slope is essential, because once the slope is subtracted, the accuracy of this type of feature begins to drop significantly.
According to the analysis of the classification accuracy of single features (Figure 7), we recommend using the raw spectral bands of the image as the main feature for citrus orchard classification, with other features mentioned above as supplementary. It is also important to consider seasonal features as much as possible to increase the information content of the image. Our analysis of the optimal feature variables revealed that winter features accounted for the largest proportion, approximately 40% (Figure 6). Given that the number of images or computing resources may be limited, employing exclusively winter images for classification is a feasible approach.
Higher classification accuracy was produced by multiple feature combinations than single feature groups. The inclusion of conventional index features greatly improved the classification accuracy of image discrimination with both IPS-covered and IPS-uncovered citrus orchards. In addition, topographic–temperature factors played an important role in classification of IPS-covered orchards. However, the red-edge index and texture features had little effect on classification accuracy. The combination of seasonal spectral bands, conventional index features, and topographic-temperature features (C6) showed the best performance among all feature combinations, with accuracy scores above 85% for all metrics and fewer feature dimensions. Therefore, we recommend using this feature combination for the land cover classification of citrus orchards using insect-proof screens.

4.2. Discussion of Different Citrus Orchard Recognition Methods

To validate the superiority of the C6 method over the previous methods in identifying citrus orchards, we merged IPS sample points into the Citrus class to apply Xu’s method [5], named Lan_Xu, using Landsat 8 OLI data in 2021. Additionally, we applied Xu’s method to Sentinel-2 MSI data and named it Sen_Xu. The classification results of the three methods in two concentrated orchard areas are shown in Figure 11. The overall accuracy (OA) of Lan_Xu was 86%, with a calculated citrus orchard area of 507 km2. Sen_Xu achieved an OA of 88%, with a citrus orchard area of 505 km2. Sen_C6 showed the highest OA of 92%, with a citrus orchard area of 460 km2, including 88 km2 (19%) covered by IPS. The user’s and producer’s accuracy for each land cover class can be seen in Figure 12. It is obvious that, due to the 30 m resolution limitation, Lan_Xu mostly depicted citrus orchards as continuous patches with blurry boundaries. Many impervious surfaces, especially roads, and cropland in valleys, were misclassified as citrus orchards. Sen_Xu, benefiting from its higher spatial resolution of 10 m, had smaller mixed pixel area in the boundary areas between citrus and other land cover types, resulting in less confusion with impervious surfaces. However, this method still misclassified many bare lands and agricultural field boundaries as citrus orchards, as the available image features are insufficient to distinguish between them. Sen_C6 successfully identified both open-air citrus orchards and those covered by IPS, with clearer orchard boundaries. According to Figure 12, Sen_C6 exhibited certain advantages in the classification accuracy of citrus orchards, cropland, and bare lands. Therefore, we believe that Sen_C6 provides a more accurate estimation of citrus orchard area, while Lan_Xu and Sen_Xu tend to overestimate the citrus area due to misclassification of agricultural fields.

4.3. Importance and Application of Identify IPS

A spatial distribution map of the citrus orchards was provided in Xunwu County with a higher spatial resolution of 10 m and more refined classification with IPS-covered area, which additionally included more accurate estimates of the area of citrus orchards. Based on this method, a web-platform software was developed for the local governments to give more precise land use map and statistics along with comparison of multiple phases to investigate and monitor the development of citrus industry. It also provided a database for quantitative analysis of ecological problems caused by citrus orchards. Additionally, accurate knowledge of the distribution and area statistic data of IPS can assist local governments in tracing and controlling HLB. Since the service life of IPS is only three to five years, they can help the government deal with possible plastic and soil pollution in the future.

4.4. Limitations and Future Work

Of course, there are certain limitations in the classification result. The distribution of IPS-covered orchards cannot completely represent the distribution of the citrus orchards affected by HLB. The use of IPS is the prerogative of local farmers or orchard operators. Large-scale IPS require a significant economic cost, citrus cultivation under IPS cover will also increase labor costs, and IPS covers have noticeable negative effects on the pollination and fertilization of citrus. Therefore, it is normal for operators to use pesticides or other chemicals to prevent and control the affected citrus orchards, even if they have been affected by HLB. This may potentially result in an underestimation of the citrus orchard area affected by HLB in our findings. For citrus orchards affected by HLB and not covered by IPS, the utilization of UAV hyperspectral imagery proves to be an effective method for accurate identification [47,48,49,50].
There are also some limitations for the classification scheme. It is only suitable for hilly and mountainous areas with certain undulating terrain, such as the hilly areas in southern China. In this region, the topography–temperature features played an important role in extracting IPS, which can distinguish IPS from vegetable greenhouses located in low-lying areas or floodplains. Accordingly, this method is not suitable for citrus planting areas with flat terrain, which may lead to the misidentification of vegetable greenhouses as IPS-covered orchards.
Additionally, the addition of texture features did not significantly increase the explanatory power of other variables in our study. Fruit trees in citrus orchards are typically planted in fixed grids, showing obvious banding. When covered with IPS, these orchards lose their regularity, making them appear flatter and smoother. However, the texture differences between the two types of orchards are not adequately captured in the Sentinel-2 MSI imagery. In the future, research using higher resolution images with better adjustments to window and angle of texture extraction may provide more insight into the contribution of texture features to citrus orchard classification.

5. Conclusions

We, here, used a variety of spectral index features and texture features taken from Sentinel-2 images, combined with terrain and land surface temperature, to classify and extract citrus orchards in Xunwu County, with particular attention paid to the discrimination of IPS-covered and -uncovered orchards. The optimization and combination of various features allowed us to obtain a high-precision classification feature combination scheme. This scheme can effectively distinguish citrus orchards with or without IPS cover and provides guidance for remote sensing-based identification of citrus orchards with IPS cover. Our main conclusions are:
(1)
The reflectance of IPS-covered orchards in the visible band is significantly higher than that of uncovered orchards. Classification and extraction of different orchards in different regions are thus feasible when exploiting these spectral differences.
(2)
The seasonal spectral bands showed the most significant advantage in single feature classification after feature optimization, with an OA of 86%.
(3)
The combination of seasonal spectral bands + conventional spectral index + topographic–temperature factors showed the best performance of combined features. This combination can improve OA by 6% and increase the UA and PA of IPS by about 10% over single-feature approaches.
(4)
In 2021, the citrus orchard area in Xunwu County was 460 km2. Approximately 88 km2 of those citrus orchards are covered with IPS, accounting for 19% of the total.
(5)
The distribution of citrus planting areas in Xunwu County is predominantly concentrated in the central and northern hilly regions, with the density of IPS coverage being higher in the northern citrus orchards. The reason for this difference in distribution is likely due to a combination of topography and human activities.

Author Contributions

Conceptualization, G.Y. and L.Z.; Methodology, G.Y. and L.L.; Formal analysis, G.Y. and L.L.; Investigation, G.L. and S.X.; Resources, G.Y., G.L. and Z.C.; Writing—original draft preparation, G.Y.; Writing—review and editing, L.Z., L.L., Z.C. and S.X.; Visualization, G.Y. and L.L.; Supervision, L.Z.; Project administration, L.Z.; Funding acquisition, L.Z. and G.L. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the National Natural Science Foundation of China (Grant Nos. 41701514 and 41961004), the Natural Science Foundation of Jiangxi (No. 20224BAB202037 and 20224BAB203034), and the Key Laboratory of Poyang Lake Wetland and Watershed Research (Jiangxi Normal University), Ministry of Education (Nos. PK2021007 and PK2020003).

Data Availability Statement

The data presented in this study are available upon request from the corresponding author.

Acknowledgments

The authors sincerely thank all anonymous reviewers who provided detailed and valuable comments or suggestions to improve this manuscript.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. FAQ. Food and Agriculture Organization. Available online: https://www.fao.org/faostat/en/#data/QCL (accessed on 13 February 2023).
  2. Bove, J.M. Huanglongbing: A destructive, newly-emerging, century-old disease of citrus. J. Plant Pathol. 2006, 88, 7–37. [Google Scholar]
  3. Xiong, M.Q.; Sun, R.H.; Chen, L.D. A global comparison of soil erosion associated with land use and climate type. Geoderma 2019, 343, 31–39. [Google Scholar]
  4. Guo, L.B.; Gifford, R.M. Soil carbon stocks and land use change: A meta analysis. Glob. Change Biol. 2002, 8, 345–360. [Google Scholar]
  5. Xu, H.; Qi, S.; Li, X.; Gao, C.; Wei, Y.; Liu, C. Monitoring three-decade dynamics of citrus planting in Southeastern China using dense Landsat records. Int. J. Appl. Earth Obs. Geoinf. 2021, 103, 102518. [Google Scholar] [CrossRef]
  6. Ye, S.; Rogan, J.; Sangermano, F. Monitoring rubber plantation expansion using Landsat data time series and a Shapelet-based approach. ISPRS J. Photogramm. Remote Sens. 2018, 136, 134–143. [Google Scholar]
  7. Battude, M.; Al Bitar, A.; Morin, D.; Cros, J.; Huc, M.; Sicre, C.M.; Le Dantec, V.; Demarez, V. Estimating maize biomass and yield over large areas using high spatial and temporal resolution Sentinel-2 like remote sensing data. Remote Sens. Environ. 2016, 184, 668–681. [Google Scholar] [CrossRef]
  8. Li, H.; Hong, L. Spatio-temporal land use/land cover dynamics and its driving forces in the Mekong Basin using Landsat imageries from 1988 to 2017. Geocarto Int. 2022, 37, 14676–14698. [Google Scholar]
  9. Xu, H.; Liu, C.; Wang, J.; Qi, S. Study on Extraction of Citrus Orchard in Gannan Region Based on Google Earth Engine Platform. Geo. Inf. Sci. 2018, 20, 396–404. [Google Scholar]
  10. Morell-Monzó, S.; Estornell, J.; Sebastiá-Frasquet, M.-T. Comparison of Sentinel-2 and High-Resolution Imagery for Mapping Land Abandonment in Fragmented Areas. Remote Sens. 2020, 12, 2062. [Google Scholar]
  11. Silva, A.F.; Barbosa, A.P.; Zimback, C.R.L.; Landim, P.M.B. Geostatistics and remote sensing methods in the classification of images of areas cultivated with citrus. Eng. Agric. 2013, 33, 1245–1256. [Google Scholar] [CrossRef]
  12. Wang, S.; Chen, Y.L. The information extraction of Gannan citrus orchard based on the GF-1 remote sensing image. IOP Conf. Ser. Earth Environ. Sci. 2017, 57, 012001. [Google Scholar] [CrossRef]
  13. Morell-Monzó, S.; Sebastiá-Frasquet, M.-T.; Estornell, J. Land Use Classification of VHR Images for Mapping Small-Sized Abandoned Citrus Plots by Using Spectral and Textural Information. Remote Sens. 2021, 13, 681. [Google Scholar] [CrossRef]
  14. Reis, S.; Tasdemir, K. Identification of hazelnut fields using spectral and Gabor textural features. ISPRS J. Photogramm. Remote Sens. 2011, 66, 652–661. [Google Scholar] [CrossRef]
  15. Chang, N.-B.; Bai, K. Multisensor Data Fusion and Machine Learning for Environmental Remote Sensing; CRC Press: Boca Raton, FL, USA, 2018. [Google Scholar]
  16. Li, M.; Ma, L.; Blaschke, T.; Cheng, L.; Tiede, D. A systematic comparison of different object-based classification techniques using high spatial resolution imagery in agricultural environments. Int. J. Appl. Earth Obs. Geoinf. 2016, 49, 87–98. [Google Scholar] [CrossRef]
  17. Lu, L.; Di, L.; Ye, Y. A Decision-Tree Classifier for Extracting Transparent Plastic-Mulched Landcover from Landsat-5 TM Images. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2014, 7, 4548–4558. [Google Scholar] [CrossRef]
  18. Novelli, A.; Tarantino, E. Combining ad hoc spectral indices based on LANDSAT-8 OLI/TIRS sensor data for the detection of plastic cover vineyard. Remote Sens. Lett. 2015, 6, 933–941. [Google Scholar] [CrossRef]
  19. Zhao, G.X.; Li, J.; Li, T.; Yue, Y.D.; Warner, T. Utilizing landsat TM imagery to map greenhouses in Qingzhou, Shandong Province, China. Pedosphere 2004, 14, 363–369. [Google Scholar]
  20. Zhang, P.; Du, P.; Guo, S.; Zhang, W.; Tang, P.; Chen, J.; Zheng, H. A novel index for robust and large-scale mapping of plastic greenhouse from Sentinel-2 images. Remote Sens. Environ. 2022, 276, 113042. [Google Scholar] [CrossRef]
  21. Yang, D.; Chen, J.; Zhou, Y.; Chen, X.; Chen, X.; Cao, X. Mapping plastic greenhouse with medium spatial resolution satellite data: Development of a new spectral index. ISPRS J. Photogramm. Remote Sens. 2017, 128, 47–60. [Google Scholar] [CrossRef]
  22. Perillo, A.C.; Kucharik, C.J.; Meehan, T.D.; Serbin, S.P.; Singh, A.; Townsend, P.A.; Whitney, K.S.; Gratton, C. Use of insect exclusion cages in soybean creates an altered microclimate and differential crop response. Agric. For. Meteorol. 2015, 208, 50–61. [Google Scholar] [CrossRef]
  23. Yang, G.; Guo, Z.; Ji, H.; Sheng, J.; Chen, L.; Zhao, Y. Application of insect-proof nets in pesticide-free rice creates an altered microclimate and differential agronomic performance. PeerJ 2018, 6, e6135. [Google Scholar] [CrossRef]
  24. Wu, L.; Sun, G.; Miao, Z.; Zhang, A.; Feng, H.; Hu, J.; Yang, Z.; Wang, W.; Chen, B.; Tang, Y. On subtropical remote sensing in China:Research status, key tasks and innovative development approaches. J. Remote Sens. 2022, 26, 1483–1503. [Google Scholar]
  25. Shelestov, A.; Lavreniuk, M.; Kussul, N.; Novikov, A.; Skakun, S. Exploring Google Earth Engine Platform for Big Data Processing: Classification of Multi-Temporal Satellite Imagery for Crop Mapping. Front. Earth Sci. 2017, 5, 1–10. [Google Scholar] [CrossRef]
  26. Li, S.; Zhou, X.; Yin, J.; Xiao, J. Comprehensive Division of Navel Orange Planting Region in Xunwu County by Climatic, Soil and Landform Conditions. Acta Agric. Jiangxi 2007, 5, 40–43. [Google Scholar]
  27. Xunwu County Government. Xunwu County Government Information Disclosure Website. Available online: http://www.xunwu.gov.cn/xwxxxgk/zfxxgkzn/xxgk_tt.shtml (accessed on 20 May 2023).
  28. Ding, L.R.; Zhou, J.; Li, Z.L.; Ma, J.; Shi, C.X.; Sun, S.; Wang, Z.W. Reconstruction of Hourly All-Weather Land Surface Temperature by Integrating Reanalysis Data and Thermal Infrared Data From Geostationary Satellites (RTG). IEEE Trans. Geosci. Remote Sens. 2022, 60, 1–17. [Google Scholar]
  29. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef]
  30. Huete, A.; Justice, C.; Liu, H. Development of vegetation and soil indices for MODIS-EOS. Remote Sens. Environ. 1994, 49, 224–234. [Google Scholar] [CrossRef]
  31. Huete, A.R. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 1988, 25, 295–309. [Google Scholar] [CrossRef]
  32. Xu, H. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery. Int. J. Remote Sens. 2006, 27, 3025–3033. [Google Scholar]
  33. Wilson, E.H.; Sader, S.A. Detection of forest harvest type using multiple dates of Landsat TM imagery. Remote Sens. Environ. 2002, 80, 385–396. [Google Scholar] [CrossRef]
  34. Gitelson, A.A.; Gritz, Y.; Merzlyak, M.N. Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves. J. Plant Physiol. 2003, 160, 271–282. [Google Scholar]
  35. Dash, J.; Curran, P.J. Evaluation of the MERIS terrestrial chlorophyll index (MTCI). Adv. Space Res. 2007, 39, 100–104. [Google Scholar] [CrossRef]
  36. Navarro, G.; Caballero, I.; Silva, G.; Parra, P.-C.; Vazquez, A.; Caldeira, R. Evaluation of forest fire on Madeira Island using Sentinel-2A MSI imagery. Int. J. Appl. Earth Obs. Geoinf. 2017, 58, 97–106. [Google Scholar] [CrossRef]
  37. Gitelson, A.; Merzlyak, M.N. Spectral Reflectance Changes Associated with Autumn Senescence of Aesculus hippocastanum L. and Acer platanoides L. Leaves. Spectral Features and Relation to Chlorophyll Estimation. J. Plant Physiol. 1994, 143, 286–292. [Google Scholar]
  38. Henebry, G.M.; Rieck, D.R. Applying principal components analysis to image time series: Effects on scene segmentation and spatial structure. Int. Geosci. Remote Sens. Symp. 1996, 1, 448–450. [Google Scholar]
  39. Haralick, R.M.; Shanmugam, K.; Dinstein, I.H. Texture Features for Image Classification; IEEE: New York, NY, USA, 1975. [Google Scholar]
  40. Duguay-Tetzlaff, A.; Bento, V.A.; Goettsche, F.M.; Stoeckli, R.; Martins, J.P.A.; Trigo, I.; Olesen, F.; Bojanowski, J.S.; da Camara, C.; Kunz, H. Meteosat Land Surface Temperature Climate Data Record: Achievable Accuracy and Potential Uncertainties. Remote Sens. 2015, 7, 13139–13156. [Google Scholar]
  41. Ermida, S.L.; Soares, P.; Mantas, V.; Göttsche, F.-M.; Trigo, I.F. Google Earth Engine Open-Source Code for Land Surface Temperature Estimation from the Landsat Series. Remote Sens. 2020, 12, 1471. [Google Scholar] [CrossRef]
  42. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  43. Svetnik, V.; Liaw, A.; Tong, C.; Culberson, J.C.; Sheridan, R.P.; Feuston, B.P. Random forest: A classification and regression tool for compound classification and QSAR modeling. J. Chem. Inf. Comput. Sci. 2003, 43, 1947–1958. [Google Scholar] [CrossRef]
  44. Smith, A. Image segmentation scale parameter optimization and land cover classification using the Random Forest algorithm. J. Spat. Sci. 2010, 55, 69–79. [Google Scholar]
  45. Ma, Y.; Jiang, Q.; Meng, Z.; Li, Y.; Wang, D.; Liu, H. Classification of land use in farming area based on random forest algorithm. Trans. Chin. Soc. Agric. Mach. 2016, 47, 297–303. [Google Scholar]
  46. Pal, M. Random forest classifier for remote sensing classification. Int. J. Remote Sens. 2005, 26, 217–222. [Google Scholar] [CrossRef]
  47. Fan, D.; Wang, F.; Yang, D.; Lin, S.; Chen, X.; Lan, Y.; Deng, X. Citrus Huanglongbing Detection Based on Multi-Modal Feature Fusion Learning. Front. Plant Sci. 2021, 12, 809506. [Google Scholar]
  48. Valdés, R.A.; Ortiz, J.C.D.; Beache, M.B.; Cabello, J.A.; Chávez, E.C.; Pagaza, Y.R.; Fuentes, Y.M.O. A review of techniques for detecting Huanglongbing (greening) in citrus. Can. J. Microbiol. 2016, 62, 803–811. [Google Scholar] [CrossRef]
  49. Moriya, É.A.S.; Imai, N.N.; Tommaselli, A.M.G.; Berveglieri, A.; Honkavaara, E.; Soares, M.A.; Marino, M. Detecting citrus huanglongbing in brazilian orchards using hyperspectral aerial images. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2019, 42, 1881–1886. [Google Scholar] [CrossRef]
  50. Deng, X.; Zhu, Z.; Yang, J.; Zheng, Z.; Huang, Z.; Yin, X.; Wei, S.; Lan, Y. Detection of Citrus Huanglongbing Based on Multi-Input Neural Network Model of UAV Hyperspectral Remote Sensing. Remote Sens. 2020, 12, 2678. [Google Scholar] [CrossRef]
Remotesensing 15 02867 g001 550
Figure 1. Study area. (a) The location of study area in China. (b) Sentinel-2 true-color image of Xunwu County in 2021 (true-color image defined by combinations of B4 (red), B3 (green), and B2 (blue) bands).
Figure 1. Study area. (a) The location of study area in China. (b) Sentinel-2 true-color image of Xunwu County in 2021 (true-color image defined by combinations of B4 (red), B3 (green), and B2 (blue) bands).
Remotesensing 15 02867 g001
Remotesensing 15 02867 g002 550
Figure 2. Comparison of citrus orchards with or without IPS coverage. (a) Unmanned aerial vehicle (UAV) images of healthy citrus orchards without IPS cover. (b) Google Earth images of healthy citrus orchards without IPS cover. (c) UAV images of citrus orchards covered with IPS after suffering HLB. (d) Google Earth images of citrus orchards with IPS cover after HLB.
Figure 2. Comparison of citrus orchards with or without IPS coverage. (a) Unmanned aerial vehicle (UAV) images of healthy citrus orchards without IPS cover. (b) Google Earth images of healthy citrus orchards without IPS cover. (c) UAV images of citrus orchards covered with IPS after suffering HLB. (d) Google Earth images of citrus orchards with IPS cover after HLB.
Remotesensing 15 02867 g002
Remotesensing 15 02867 g003 550
Figure 3. Flowchart of obtaining the optimal classification scheme experiment.
Figure 3. Flowchart of obtaining the optimal classification scheme experiment.
Remotesensing 15 02867 g003
Remotesensing 15 02867 g004 550
Figure 4. Spectral curves of different land cover types in the study area are determined from Sentinel-2 MSI images.
Figure 4. Spectral curves of different land cover types in the study area are determined from Sentinel-2 MSI images.
Remotesensing 15 02867 g004
Remotesensing 15 02867 g005 550
Figure 5. Recursive feature elimination results. (a) Spectral bands. (b) Conventional spectral index. (c) Red-edge index. (d) Textural features. (e) Topographic-temperature features. The red circles indicate the feature dimension where OA begins to show a significant decrease.
Figure 5. Recursive feature elimination results. (a) Spectral bands. (b) Conventional spectral index. (c) Red-edge index. (d) Textural features. (e) Topographic-temperature features. The red circles indicate the feature dimension where OA begins to show a significant decrease.
Remotesensing 15 02867 g005
Remotesensing 15 02867 g006 550
Figure 6. Seasonal distribution of filtered characteristic variables.
Figure 6. Seasonal distribution of filtered characteristic variables.
Remotesensing 15 02867 g006
Remotesensing 15 02867 g007 550
Figure 7. Classification accuracy of single features after screening.
Figure 7. Classification accuracy of single features after screening.
Remotesensing 15 02867 g007
Remotesensing 15 02867 g008 550
Figure 8. Classification accuracy of combined features.
Figure 8. Classification accuracy of combined features.
Remotesensing 15 02867 g008
Remotesensing 15 02867 g009 550
Figure 9. Classification map produced by optimal feature combination.
Figure 9. Classification map produced by optimal feature combination.
Remotesensing 15 02867 g009
Remotesensing 15 02867 g010 550
Figure 10. Elevation distribution of citrus orchards and IPS. The outer ring is the citrus orchard without IPS coverage, and the inner ring is IPS with coverage.
Figure 10. Elevation distribution of citrus orchards and IPS. The outer ring is the citrus orchard without IPS coverage, and the inner ring is IPS with coverage.
Remotesensing 15 02867 g010
Remotesensing 15 02867 g011 550
Figure 11. Comparison of the classification results using three methods in two concentrated citrus regions. (a,e) using Lan_Xu. (b,f) using Sen_Xu. (c,g) using Sen_C6. (d,h) are high-resolution Google Earth images of the two areas.
Figure 11. Comparison of the classification results using three methods in two concentrated citrus regions. (a,e) using Lan_Xu. (b,f) using Sen_Xu. (c,g) using Sen_C6. (d,h) are high-resolution Google Earth images of the two areas.
Remotesensing 15 02867 g011
Remotesensing 15 02867 g012 550
Figure 12. Comparison of accuracy of three methods and extraction area of citrus orchards.
Figure 12. Comparison of accuracy of three methods and extraction area of citrus orchards.
Remotesensing 15 02867 g012
Table
Table 1. Sentinel-2 MSI images used in different seasons in the study area.
Table 1. Sentinel-2 MSI images used in different seasons in the study area.
SeasonNumber of ImagesPath/RowImage Acquisition Date
Winter14T50RLP22 December 2020, 27 December 2020, 1 January 2021, 21 January 2021, 31 January 2021, 15 February 2021, 20 February 2021
T50RLN22 December 2020, 27 December 2020, 1 January 2021, 31 January 2021, 5 February 2021, 15 February 2021, 20 February 2021
Spring2T50RLP16 April 2020
T50RLN16 April 2020
3T50RLP17 March 2021
T50RLN1 April 2021, 21 April 2021
Summer2T50RLP29 August 2020
T50RLN29 August 2020
5T50RLP25 July 2022, 30 July 2022, 24 August 2022
T50RLN30 July 2022, 24 August 2022
Autumn6T50RLP18 September 2021, 28 September 2021, 3 October 2021, 27 November 2021
T50RLN12 November 2021, 27 November 2021
Table
Table 2. Landsat-8 OLI and TIRS images were used in different seasons in the study area.
Table 2. Landsat-8 OLI and TIRS images were used in different seasons in the study area.
SeasonNumber of ImagesPath/RowImage Acquisition Date
Winter7121,0429 December 2019, 10 January 2020, 27 February 2020,
121,0439 December 2019, 25 December 2019, 10 January 2020, 27 February 2020
12121,04211 December 2020, 27 December 2020, 12 January 2021, 13 February 2021, 14 December 2021
121,04311 December 2020, 27 December 2020, 12 January 2021, 28 January 2021,
13 February 2021, 14 December 2021, 30 December 2021
Spring8121,04214 March 2020, 15 April 2020, 1 May2020, 17 May 2020
121,04314 March 2020, 15 April 2020, 1 May2020, 17 May2020
6121,0421 March 2021, 17 March 2021, 2 April 2021
121,0431 March 2021, 17 March 2021, 2 April 2021
Summer7121,0424 July 2020, 20 July 2020, 5 August 2020, 21 August 2020
121,0432 June 2020, 18 June 2020, 21 August 2020
7121,04223 July 2021, 8 August 2021, 24 August 2021
121,0435 June 2021, 23 July 2021, 8 August 2021, 24 August 2021
Autumn5121,04224 October 2020
121,0436 September 2020, 22 September 2020, 24 October 2020, 25 November 2020
12121,0429 September 2021, 25 September 2021, 11 October 2021, 27 October 2021,
12 November 2021, 28 November 2021
121,0439 September 2021, 25 September 2021, 11 October 2021, 27 October 2021,
12 November 2021, 28 November 2021
Table
Table 3. Ground truth sample points.
Table 3. Ground truth sample points.
ImperviousForestWaterCroplandCitrusBushBare landIPSTotal
Number1757871314523572831372502572
Table
Table 4. All image features for classification.
Table 4. All image features for classification.
Feature TypeFeature NameAbbreviationFormulaReferences
Spectral bandsAerosolsB1
BlueB2
GreenB3
RedB4
Red-edge 1B5
Red-edge 2B6
Red-edge 3B7
NIRB8
Water vaporB9
SWIR1B11
SWIR2B12
Conventional spectral indexNormalized difference vegetation indexNDVI NIR   -   Red NIR   +   Red [29]
Enhanced Vegetation IndexEVI 2.5   ×   ( NIR   -   Red ) NIR   +   6   ×   Red   -   7.5   ×   Blue   +   1 [30]
Soil-adjusted
vegetation index		SAVI ( 1   +   L )   ×   ( NIR   -   Red ) NIR   +   Red   +   L ,   L = 0 . 5 [31]
Modified normalized difference water indexMNDWI Green   -   SWIR 1 Green   +   SWIR 1 [32]
Normalized difference moisture indexNDMI NIR   -   SWIR 1 NIR   +   SWIR 1 [33]
Plastic-mulched landcover
index		PMLI SWIR 1   -   Red SWIR 1   +   Red [17]
Maximum NDVI compositeNDVImax max   { NDVI ( i , j ) 1 ,   NDVI i , j 2   , …   ,
NDVI i , j n   } 		[5]
Minimum NDVI compositeNDVImin min   { NDVI ( i , j ) 1 ,   NDVI i , j 2   , …   ,  
NDVI i , j n } 		[5]
Red-edge indexChlorophyll Index Red-EdgeCIre Red - edge   3   Red - edge   1   -   1 [34]
Normalized Difference
red edge 1		NDre1 Red - edge   2   -   Red - edge   1   Red - edge   2   +   Red - edge   1   [34]
Normalized Difference
red edge 2		NDre2 Red - edge   3   -   Red - edge   1   Red - edge   3   +   Red - edge   1   [34]
Meris terrestrial chlorophyll
index		MTCI Red - edge   2   -   Red - edge   1   Red - edge   1   -   Red [35]
Normalized Difference Vegetation Index red-edge 1 narrowNDVIre1 NIR   -   Red - edge   1   NIR   +   Red - edge   1   [36,37]
Normalized Difference Vegetation Index red-edge 1 narrowNDVIre1 NIR   -   Red - edge   2   NIR   +   Red - edge   2   [36,37]
Normalized Difference Vegetation Index red-edge 1 narrowNDVIre1 NIR   -   Red - edge   3   NIR   +   Red - edge   3   [36,37]
Textural featuresSum AverageSAVG Σ i ,   j = 0   N - 1 ip = u
VarianceVAR   Σ i , j = 0 N - 1 p i , j   i - u 2
Inverse Difference MomentIDM   Σ i , j = 0 N - 1 p   i ,   j 1 +   i - j 2
ContrastCON   Σ i , j = 0 N - 1 p i , j   ( i - j ) 2
DissimilarityDISS   Σ i , j = 0 N - 1 p i , j   | i - j |
EntropyENT Σ i , j = 0 N - 1 - p i , j log ( p ( i , j ) )
Angular Second MomentASM   Σ i , j = 0 N - 1 p   i , j 2
CorrelationCORR Σ i , j = 0 N - 1 ijp   i , j - μ x μ y σ x σ y
Topographic-temperature featuresDEMDEM
SLOPESLOPE
AspectAspect
Hill shadeHillshade
Land surface temperatureLST
Note: NDVI i , j n refers to the pixel value of the nth NDVI in row i and column j. In the texture feature, N indicates the gray level, p (i, j) represents the frequency of pixel pairs in a specific position relation, and μ x , μ y , σ x and σ y   are, respectively, the mean and variance of px (i) = ∑kpx (i, k) and px (j) = ∑kpy (k, j).
Table
Table 5. Results after feature screening.
Table 5. Results after feature screening.
Feature TypeBand NameNumber of Bands
Spectral bandsB1_4, B12_2, B2_4, B12_1, B4_4, B11_4, B8_4, B12_4, B7_1, B7_410
Conventional spectral indexMNDWI_3, MNDWI_2, NDVI_MIN, MNDWI_1, NDVI_MAX, SAVI_4, NDMI_1, PMLI_4, NDVI_4, NDMI_4, PMLI_111
Red-edge indexMTCI_3, MTCI_2, NDVIRE2_2, NDRE1_4, NDRE2_4, MTCI_4, NDVIRE1_4, CIRE_4, NDVIRE1_29
Textural featuresSAVG_2, SAVG_1, IDM_1, CON_1, CORR_2, CON_2, IDM_2, CORR_1, DISS_1, DISS_2, VAR_2, VAR_1, ASM_2, SAVG_314
Topographic-temperature featuresElevation, LST_4, LST_1, LST_2, LST_3, Aspect, Slope7
Note: The numbers 1–4 in the characteristic variables represent four different seasons: spring, summer, autumn, and winter. Band names are sorted by importance score from highest to lowest.
Table
Table 6. Combination schemes obtained by an exhaustive method. Sorted by feature dimension.
Table 6. Combination schemes obtained by an exhaustive method. Sorted by feature dimension.
Combination SchemeNumber of BandsCode
Spectral bands + Topographic-temperature17C1
Spectral bands + Red-edge index19C2
Spectral bands + Conventional spectral index21C3
Spectral bands + Texture24C4
Spectral bands + Red-edge index + Topographic-temperature26C5
Spectral bands + Conventional spectral index + Topographic-temperature28C6
Spectral bands +Conventional spectral index + Red-edge index30C7
Spectral bands + Topographic-temperature + Texture31C8
Spectral bands + Red-edge index + Texture33C9
Spectral bands + Conventional spectral index + Texture35C10
Spectral bands + Conventional spectral index + Red-edge index + Topographic-temperature37C11
Spectral bands + Red-edge index + Topographic-temperature + Texture40C12
Spectral bands + Conventional spectral index + Topographic-temperature + Texture42C13
Spectral bands + Conventional spectral index + Red-edge index + Texture44C14
Spectral bands + Conventional spectral index + Red-edge index + Topographic-temperature + Texture51C15
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Yu, G.; Zhang, L.; Luo, L.; Liu, G.; Chen, Z.; Xiong, S. Mapping Insect-Proof Screened Citrus Orchards Using Sentinel-2 MSl Time-Series Images. Remote Sens. 2023, 15, 2867. https://doi.org/10.3390/rs15112867

AMA Style

Yu G, Zhang L, Luo L, Liu G, Chen Z, Xiong S. Mapping Insect-Proof Screened Citrus Orchards Using Sentinel-2 MSl Time-Series Images. Remote Sensing. 2023; 15(11):2867. https://doi.org/10.3390/rs15112867

Chicago/Turabian Style

Yu, Guobin, Li Zhang, Lingxia Luo, Guihua Liu, Zongyi Chen, and Shanshan Xiong. 2023. "Mapping Insect-Proof Screened Citrus Orchards Using Sentinel-2 MSl Time-Series Images" Remote Sensing 15, no. 11: 2867. https://doi.org/10.3390/rs15112867

APA Style

Yu, G., Zhang, L., Luo, L., Liu, G., Chen, Z., & Xiong, S. (2023). Mapping Insect-Proof Screened Citrus Orchards Using Sentinel-2 MSl Time-Series Images. Remote Sensing, 15(11), 2867. https://doi.org/10.3390/rs15112867

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop