Next Article in Journal
Hierarchical Feature Association and Global Correction Network for Change Detection
Previous Article in Journal
Comparison of FY-4A/AGRI SST with Himawari-8/AHI and In Situ SST
 
 
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 17
10.3390/rs15174140
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

Optimized Extraction Method of Fruit Planting Distribution Based on Spectral and Radar Data Fusion of Key Time Phase

by
Guobing Zhao
1,2,
Lei Wang
2,
Jianghua Zheng
1,3,*,
Nigela Tuerxun
1,
Wanqiang Han
1 and
Liang Liu
1
1
College of Geography and Remote Sensing Sciences, Xinjiang University, Urumqi 830046, China
2
Institute of Modern Forestry, Xinjiang Academy of Forestry Sciences, Urumqi 830063, China
3
Xinjiang Key Laboratory of Oasis Ecology, Xinjiang University, Urumqi 830046, China
*
Author to whom correspondence should be addressed.
Remote Sens. 2023, 15(17), 4140; https://doi.org/10.3390/rs15174140
Submission received: 4 July 2023 / Revised: 17 August 2023 / Accepted: 20 August 2023 / Published: 23 August 2023
(This article belongs to the Section Remote Sensing in Agriculture and Vegetation)
Browse Figures
Versions Notes

Abstract

:
With China’s fruit tree industry becoming the largest in the world, accurately understanding the spatial distribution of fruit tree growing areas is crucial for promoting socio-economic development and rural revitalization. Remote sensing offers unprecedented opportunities for fruit tree monitoring. However, previous research has mainly focused on UAV and near-ground remote sensing, with limited accuracy in obtaining fruit tree distribution information through satellite remote sensing. In this study, we utilized the Google Earth Engine (GEE) remote sensing cloud platform and integrated data from Sentinel-1, Sentinel-2, and SRTM sources. We constructed a feature space by extracting original band features, vegetation index features, polarization features, terrain features, and texture features. The sequential forward selection (SFS) algorithm was employed for feature optimization, and a combined machine learning and object-oriented classification model was used to accurately extract fruit tree crop distributions by comparing key temporal phases of fruit trees. The results revealed that the backscatter coefficient features from Sentinel-1 had the highest contribution to the classification, followed by the original band features and vegetation index features from Sentinel-2, while the terrain features had a relatively smaller contribution. The highest classification accuracy for jujube plantation areas was observed in November (99.1% for user accuracy and 96.6% for producer accuracy), whereas the lowest accuracy was found for pear tree plantation areas in the same month (93.4% for user accuracy and 89.0% for producer accuracy). Among the four different classification methods, the combined random forest and object-oriented (RF + OO) model exhibited the highest accuracy (OA = 0.94, Kappa = 0.92), while the support vector machine (SVM) classification method had the lowest accuracy (OA = 0.52, Kappa = 0.31). The total fruit tree plantation area in Aksu City in 2022 was estimated to be 64,000 hectares, with walnut, jujube, pear, and apple trees accounting for 42.5%, 20.6%, 19.3%, and 17.5% of the total fruit tree area, respectively (27,200 hectares, 13,200 hectares, 12,400 hectares, and 11,200 hectares, respectively). The SFS feature optimization and RF + OO-combined classification model algorithm selected in this study effectively mapped the fruit tree planting areas, enabling the estimation of fruit tree planting areas based on remote sensing satellite image data. This approach facilitates accurate fruit tree industry and real-time crop monitoring and provides valuable support for fruit tree planting management by the relevant departments.

1. Introduction

The fruit tree industry holds a significant position in the agricultural economy, and with China being the largest player in this industry, obtaining accurate information regarding the spatial distribution and planting areas of fruit tree orchards is crucial for promoting sustainable agroforestry development [1,2]. However, effectively monitoring and managing fruit tree planting areas presents challenges due to the intricate structure of orchards, the diverse range of area sizes, and the extended growth and development cycles [3,4].
Remote sensing has emerged as a fundamental tool for monitoring and mapping various land covers and land use types [5]. Advancements in remote sensing technologies, including high-resolution satellite images, hyperspectral sensors, LIDAR data, and the introduction of remote sensing cloud computing platforms like Google Earth Engine, have significantly improved the accuracy and efficiency of remote sensing interpretation [6,7]. Remote sensing has been used more frequently for the rapid acquisition of information on the growth status, yield, and quality of crops such as rice, corn, and wheat [8,9]. However, due to the perennial nature of fruit trees, including their variable growth characteristics, morphology, complex planting structures, and constraints imposed by row spacing and plant spacing, controlling the research scale and interpreting remote sensing images prove challenging [10]. In addition, the spatial resolution of traditional satellite remote sensing data is usually low, which cannot provide enough detailed and sufficient information to distinguish different tree species, which limits the ability to accurately identify and classify individual tree species, and thus there are many difficulties in the application of remote sensing in fruit tree planting information extraction. While the recent availability of high-resolution satellite data offers more detailed images, challenges such as purchase costs and extensive data processing operations persist. Consequently, much of the existing research on fruit tree planting information extraction remains confined to UAV remote sensing and near-ground remote sensing [11,12,13]. Traditional remote sensing classification methods are insufficient to meet the accuracy and real-time requirements of fruit tree planting. Thus, further research and exploration are warranted.
Accurate image timing selection is vital for crop extraction [14]. Different fruit tree species have distinct fertility periods, and the spectral characteristics of their leaves and fruits vary across seasons. Therefore, choosing a suitable time for remote sensing images is critical for enhancing classification accuracy [15]. Therefore, how to obtain “key temporal phase” satellite images and make full use of the spectral and spatial features to identify two or more crops with similar phenological calendars is a key issue to be studied in depth, and it is necessary to consider the temporal information of fruit trees to better distinguish tree species and improve classification accuracy.
Fruit trees are influenced by factors like species diversity, causing remote sensing pixels to contain signals from various tree species. The spatial mixing effect of vegetation further introduces interference and confusion into the spectral information. Different tree species exhibit variations in morphology, structure, and spectra [16]. Understanding and researching the effects of these differences on remote sensing classification is necessary to improve classification accuracy. Thorough species characterization and model development are pivotal in this endeavor. Machine learning holds promise for the remote sensing extraction of agricultural crops, including winter wheat and cotton [17,18]. By analyzing and modeling the study area by using machine learning algorithms, crop-related features, such as vegetation indices, texture features, and spectral features, can be extracted from remotely sensed images [19]. These features provide insights into a crop’s growth state, spatial distribution, and health [20]. While machine learning can train classification models to recognize features in remotely sensed images, fruit tree growing areas present unique challenges and limitations. Unlike other densely planted crops, the interpretation of fruit tree planting areas in remote sensing is influenced by the varying sizes of tree spacing and canopy, which depend on factors such as tree varieties, ages, and management practices. Therefore, machine learning classification models must exhibit adaptability to handle these variations effectively [21,22]. Object-oriented programming principles advocate breaking down problems into individual objects [23,24]. While machine learning and object-oriented programming belong to different fields, they can be effectively combined to extract fruit tree planting areas. Object-oriented design patterns facilitate the organization and management of complex machine learning systems, while machine learning algorithms enable data analysis and prediction. The algorithm’s robustness and generalization ability are crucial factors in improving the extraction of remote sensing planting areas of fruit trees, considering variations in the remote sensing data, fruit tree species, and growth stages [25]. Accordingly, a combined machine learning and object-oriented model was developed specifically for fruit tree planting areas, aiming to accurately identify and classify the planting areas.
This study aims to enhance the classification accuracy and monitoring capability of fruit tree planting areas by analyzing the key temporal phases and optimal features of characteristic planting areas. Additionally, it seeks to explore fast classification methods suitable for extracting fruit tree planting areas, estimate the area of fruit tree plantations, and determine their spatial distribution using the remote sensing cloud platform. The specific research objectives include the following: (1) Selecting Sentinel-1/2 and SRTM topographic data as the remote sensing image data and investigating the key temporal phases of remote sensing for identifying planting areas of common fruit tree species. (2) Analyzing the contribution of spectral, topographic, texture, polarization band, and other feature variables from multi-source data in identifying fruit tree planting areas. (3) Evaluating the effectiveness of various machine learning and object-oriented classification algorithms for extracting fruit tree planting areas, thus developing a comprehensive set of methods for the fruit tree classification and quantitative analysis of planting areas. (4) Extracting fruit planting information for 2022 from Aksu City, Xinjiang, to achieve accurate identification of the spatial distribution and quantitative estimation of fruit tree planting areas. The purpose of this study is to accurately capture the spatial distribution information of fruit tree planting areas and provide timely and precise digital and visual agricultural information related to fruit production for government management departments. This will contribute to the national and local socio-economic development strategies.

2. Study Area and Data

2.1. Study Area

Xinjiang, China, is a region abundant in soil, water, light, and heat resources, making it highly suitable for the development of fruit plantation production [26,27]. The study area is situated in Aksu City (Figure 1) in southern Xinjiang. In addition, the study chose to ignore the desert area in the southeast direction. Aksu City is located at the northwestern edge of the Tarim Basin, which falls within a warm-temperate arid climate zone characterized by limited rainfall, high evaporation, and a dry climate [28]. The area’s flat terrain, fertile land, ample water supply, abundant sunlight, and extended frost-free period create favorable conditions for cultivating various crops. Spanning an area of 14,420 km², the region predominantly cultivates fruit trees, rice, and cotton. Notably, the fruit tree species grown here include apples, pears, jujubes, and walnuts. Furthermore, the study area is characterized by a combination of towns, deserts, and the Aksu River, which is one of the three major rivers in Xinjiang with international significance. Benefiting from its favorable climate and ample land resources, the Aksu region possesses a distinct advantage in the forest fruit industry. Notably, special fruit trees, such as jujubes, walnuts, apples, and pears, have emerged as some of the region’s prominent industries [29]. The development of the fruit tree industry holds significant importance in enhancing the income of local farmers, fostering rural economic growth, and promoting environmentally sustainable development in the region.

2.2. Satellite Data and Pre-Processing

Google Earth Engine (GEE) has revolutionized the processes of data downloading, storage, and processing. In recent years, GEE has gained significant popularity for its land cover classifications [7,30]. This study employed Sentinel-1’s synthetic aperture radar (SAR) ground-range detected images and Sentinel-2’s atmospheric bottom reflectance L2A-level images, obtained from GEE, for identifying fruit tree types [31,32]. During cloudy and rainy weather conditions, Sentinel-2’s optical images lack informative content. In contrast, the Sentinel-1 SAR satellite can capture valuable images, even in the presence of cloud cover. Additionally, the altitude of Aksu City gradually decreases from the northwest to the southeast regions, making the inclusion of topographic data essential for the classification process [33]. Thus, this study integrated Sentinel-1 SAR, Sentinel-2 optical, and topographic data to determine the optimal feature combination for fruit tree type extraction within the study area.

2.2.1. Sentinel-1 Data

Sentinel-1A’s IW mode offers dual polarization modes, namely vertical emission and vertical reception (VV) and vertical emission and horizontal reception (VH). The 10 m resolution images captured by the VV and VH polarizations are extensively employed for crop-type classification [34]. The Sentinel-1 toolbox in GEE was employed to conduct an orbital parameter calibration, thermal noise removal, radiometric calibration, and terrain correction for the Sentinel-1 data. To mitigate the impact of noise, the pre-processed SAR data underwent an element-wise calculation of the median within its time range. Subsequently, month-by-month 10 m spatial resolution image data of the VV and VH polarizations were obtained in 2022 for the purpose of feature space construction.

2.2.2. Sentinel-2 Data

Within GEE, the Sentinel-2 image collection comprises Level 2A and Level 2B data. Level 2A’s data underwent correction for bottom-of-atmosphere reflectance (BOA) orthorectification, resulting in improved color levels and reflectance accuracy [35]. Sentinel-2’s Level 2A images encompass 13 bands, comprising 4 bands at a 10 m spatial resolution (blue, green, red, and NIR), 6 bands at a 20 m spatial resolution (red edge 1, red edge 2, red edge 3, red edge 4, SWIR 1, and SWIR 2), and 3 bands at a 60 m spatial resolution (coastal aerosol, water vapor, and cirrus clouds). In this study, the bands with a 20 m and 60 m resolution were resampled to a 10 m resolution using the nearest neighbor method. The collection of Sentinel-2 images took place in 2022. The images with the least cloud cover were reconstructed, and the median synthesis method was employed to generate the final composite of monthly lunar images.

2.2.3. SRTM Data

The study area is characterized by prominent topographic features, and these terrain attributes directly or indirectly influence the growth and distribution of fruit trees, consequently impacting the spatial distribution of fruit tree cultivation areas [36,37]. The Shuttle Radar Topography Mission (SRTM) is a global topographic mapping project conducted by NASA (National Aeronautics and Space Administration) and international partners. SRTM employs synthetic aperture radar (SAR) sensors onboard the Space Shuttle to capture high-resolution topographic data across extensive geographical coverage. These data yield a digital elevation model (DEM), representing elevation information at various Earth surface locations [38]. In this study, the SRTM data were resampled to a 10 m resolution.

2.3. Training and Validation Sample Data

Field surveys were carried out by the researchers in the study area between 18 July and 7 August 2021 and between 8 July and 22 July 2022 (Figure 2). The identified fruit tree planting types included walnut, jujube, apple, and pear trees, while other ground objects, such as bare land, buildings, and wetlands, were not further categorized.
Through our field investigation, 4027 sample points were collected in the study area (Figure 1). To ensure the reliability and accuracy of the samples, a rigorous selection process was employed. This process involved using the data acquired from the field survey in conjunction with high-resolution images of various crops from Google Earth. The sample data was split into a 7:3 ratio, with 70% used for training and 30% for validation in the classification process. The details of the sample selection and classification are presented in Table 1.

3. Methods

Figure 3 illustrates the research framework, which consists of the following main components: (1) image data pre-processing and sample acquisition; (2) construction and optimization of the feature space; (3) classification of the study area using feature collaboration data with four methods, namely, the support vector machine (SVM), random forest (RF), object-oriented (OO), and RF + OO; (4) analysis of the typical characteristic variables by selecting the backscattering coefficient and normalized vegetation index (NDVI) for time series analysis to determine their contribution degree; (5) evaluation of the classification accuracy using different methods and analysis of variations in accuracy across different time phases of the four fruit tree plantation areas to estimate the fruit plantation areas’ extent and investigate their spatial distribution.

3.1. Feature Space Construction

The construction of the classification feature space is a crucial step in remote sensing image classification, and the selection of classification features holds particular significance. In this study, Sentinel-1 and Sentinel-2 data, along with SRTM terrain data, were processed. The feature space for classification was constructed by selecting S2 spectral bands, vegetation index features, terrain features, polarization features, and texture features.

3.1.1. Spectral Band and Vegetation Index Characteristics

Vegetation change and classification studies often investigate the correlation between vegetation and vegetation indices (VIs), with the NDVI (Normalized Difference Vegetation Index) being the most widely employed index [39]. In this paper, high spatial resolution Sentinel-2 data is utilized. From this data, the original 12 bands and four commonly used vegetation indices [40]—NDVI, RVI, EVI, and SAVI—were extracted, and their corresponding calculation expressions are provided in Table 2.

3.1.2. Polarization Features

Sentinel-1 radar data is a high-resolution imaging system that remains unaffected by clouds and rain, enabling continuous monitoring of the ground surface [41]. In this study, multi-temporal Sentinel-1 radar images were selected, with the VV and VH polarization bands being commonly employed for radar remote sensing. These bands provide valuable signal reflection information from different directions, playing a crucial role in the monitoring and classification of fruit tree crops. Pre-processed VV-polarized and VH-polarized time series images were obtained from the GEE platform to assess the impact of multi-temporal radar remote-sensing data on feature classification accuracy. Moreover, the VV- and VH-polarized bands are incorporated into the construction of the feature space.

3.1.3. Terrain Features

Terrain plays a significant role in fruit tree planting, with elevation being a primary factor. Different fruit tree species have specific elevation requirements, and certain planting areas may be constrained to specific elevation ranges, resulting in the division of fruit tree planting areas. Additionally, slope, which measures the degree of topographic relief, influences fruit tree suitability. Steep slopes are prone to soil erosion and root instability, rendering them unsuitable for fruit tree cultivation. Conversely, moderate slopes with favorable soil and water conservation conditions are more conducive to fruit tree growth. Moreover, aspect affects the sunlight and temperature distributions within fruit tree planting areas. Sunny slopes receive ample sunlight and higher temperatures, favoring light-loving fruit tree species, while shaded slopes experience reduced temperatures due to shadows, creating a suitable environment for shade-loving fruit tree species.
To integrate the aforementioned analysis, this study utilized the SRTML1_003 data provided by GEE, which was resampled to a 10 m resolution using the nearest neighbor method. The selected feature components for the feature space construction include elevation, slope, aspect, and hillshade, which are commonly used in terrain analyses.

3.1.4. Texture Features

Texture plays a vital role in remote sensing image classification, particularly in fruit tree planting areas where planting structures exhibit alternating arrangements of tree species. The technique employed to analyze an image’s texture is known as texture analysis, and the gray-level co-occurrence matrix (GLCM) method is widely used for its adaptability and robustness [42]. By computing textures, the GLCM can extract 18 texture metrics from 8-bit grayscale images. Among the numerous texture indicators computed with the GLCM, four indicators were selected for the classification stage (Table 3) due to their common usage in the field. These indicators are derived from normalized vegetation index images, which facilitate quantitative analysis and the determination of fruit tree growth status.

3.1.5. Feature Optimization

During the processing of remote sensing images for classification, it is essential to perform feature selection, as not all features significantly enhance accuracy. The objective is to identify and retain the features that contribute the most to crop classification [43]. Feature selection not only eliminates redundant features but also optimizes computational efficiency, leading to improved classification accuracy and model performance [44]. A total of 26 variables were chosen to construct the feature space, encompassing 12 spectral features, 2 polarization features, 4 topographic features, 4 texture features, and 4 vegetation index features. Employing the sequential forward selection (SFS) method, we successively added 26 features to the random forest classifier based on their contribution levels. Subsequently, we calculated the classification accuracy after each feature addition and compared the overall trend. The process continued until the accuracy reached a stable state or showed minimal improvement, indicating the identification of the best feature subset.

3.2. Classification Methods

To emphasize the advantages of remote sensing cloud platforms in image processing and classification, this study employed four classification methods that combine pixel-based and object-oriented approaches. The aim was to compare the accuracy of these algorithms in extracting typical fruit tree planting areas, namely random forest (RF), support vector machine (SVM), object-oriented (OO), and RF + OO. For the classification process, the sample data were initially divided into a 7:3 ratio for the training and validation sets. Subsequently, these five categories of features, which include original bands, polarization features, terrain features, texture features, and four vegetation indices, were inputted into the classifier to identify walnut, jujube, pear, and apple tree planting areas.

3.2.1. Random Forest

Random Forest (RF), introduced by Leo Breiman [45], is an ensemble learning classifier that combines decision trees. It utilizes the bootstrap resampling technique, wherein approximately two-thirds of the training samples are randomly selected with replacement to generate multiple decision trees [46]. The remaining one-third of the samples are used for internal cross-validation to assess the classification accuracy of the random forest. The GINI coefficient is employed to determine the splitting criteria at each node during the construction of the random forest [47].
The number of decision trees, denoted as T, plays a crucial role in RF classification. Different values of T yield varying classification effects. When T is small, RF classification tends to have a high out-of-bag error (OOB), which represents the proportion of incorrectly predicted data among the total. As T gradually increases, the OOB error gradually decreases. However, increasing T also leads to higher complexity of the random forest model, resulting in longer RF classification times. Therefore, selecting an appropriate value for T is essential. The RF classification achieves a stable and effective outcome only when T reaches a certain threshold. In this study, the optimal value of T was determined through multiple trials, with T = 145 yielding the best RF classification model (Figure 4).

3.2.2. Support Vector Machine

Support vector machine (SVM) was first proposed in 1964 as a supervised classification method based on a binary generalized linear classifier [48]. SVM is widely utilized in remote sensing image classification and fusion due to its excellent capabilities in data analysis and pattern recognition [49,50]. For parameter configuration, the kernel function is set to the radial basis (RBF), gamma is set to 0.5, and the cost is set to 10, while the remaining parameters are set to their default values.

3.2.3. Object-Oriented

Object-oriented (OO) classification methods are predominantly employed in remote sensing for segmenting and classifying high-resolution images and multispectral data, utilizing spatial, textural, and spectral information to produce precise classification results or vector outputs [51]. In the Google Earth Engine (GEE) platform, various image segmentation algorithms, such as GMeans, KMeans, and SNIC, can be utilized for object-oriented classification [52,53]. Among these algorithms, SNIC is chosen for image segmentation due to its superior preservation of detailed image information, making it suitable for scenarios requiring high-quality segmentation results compared to the other two algorithms [54]. During the object-oriented crop classification process using SNIC, several parameters need to be configured. After conducting multiple experiments, it was determined that setting the tightness to 0.1, connectivity to 8, neighborhood size to 256, and seed to 13 yields a segmentation unit that is more suitable for extracting fruit tree planting regions.

3.2.4. RF + OO

In the traditional random forest classification method, the minimum processing unit is a single image element, which often leads to the occurrence of “salt-and-pepper noise”. However, in the object-oriented classification method, the minimum processing unit is increased to a polygonal object with homogeneous characteristics. In the random forest combined with the object-oriented (RF + OO) classification method, the initial segmentation of the image is performed using the SNIC algorithm. By adjusting and optimizing the relevant parameters to achieve the best segmentation effect, a series of superpixels are generated, and a set of features is extracted from these superpixels. Finally, the RF classifier is utilized to train the extracted features and classify the entire image.

3.3. Evaluation of Feature Importance and Classification Accuracy

3.3.1. Feature Importance Assessment

The Gini coefficient is utilized to assess the impurity of a node, indicating the frequency at which a random instance is misclassified. Hence, the Gini coefficient serves as a measure to evaluate the importance of each variable [55,56]. In the RF model, multiple decision trees are constructed by randomly sampling both instances and features. Subsequently, the Gini coefficient of each feature is calculated for each tree, and the weighted average of the Gini coefficients across all trees is determined to obtain the importance score of each feature. As significant variables play a crucial role in reducing instance impurity, a higher Gini coefficient signifies greater importance of the corresponding variable [57]. In this study, the Gini coefficient is computed for each feature to evaluate its significance in the classification process. This approach facilitates the selection of the most informative splitting feature during decision tree construction, simplifying the model and enhancing efficiency and accuracy.

3.3.2. Accuracy Assessment

The classification results in this study were evaluated using the confusion matrix method. The confusion matrix is a comparison array that depicts the number of real pixels classified into specific classes and the number of pixels tested. It is represented by an N × N matrix [58]. Several metrics can be derived from the confusion matrix (Table 4), including producer accuracy (PA), user accuracy (UA), overall accuracy (OA), and the Kappa coefficient (KC) [59,60].

4. Results

4.1. Feature Time-Series Analysis of Feature Variables

A total of 26 variables were chosen to construct the feature space, encompassing 12 spectral features, 2 polarization features, 4 topographic features, 4 texture features, and 4 vegetation index features. The heatmap of the classification features of crops (Figure 5a) reveals noticeable variations in the contribution level of each feature type across different months. The importance values of the classification features ranged from 25.0 to 235.0. During the germination stage of fruit tree crops in March and the mature stage in August and September, the features exhibited relatively high contributions to the classification process. However, as the fruit trees entered the dormant stage in November, the contributions of the features to the classification decreased.
Considering the differences in the importance of feature variables in each month, September was selected to generate a histogram of the importance of classification features (Figure 5b), and it can be seen that most of the spectral features can distinguish crops; however, only a few primitive bands have good discriminatory ability for crops. Considered together, the order of importance is polarization features > primitive bands > vegetation indices > texture features > topographic features. Among the 12 spectral bands, B11, B2, B5, B9, and B4 exhibited higher contributions to the classification process. They also demonstrated superior performances in the four additional vegetation indices, with the EVI having the highest level of contribution. Among the four texture feature indexes based on the NDVI, ASM, entropy, and contrast displayed higher classification contributions compared to correlation. Among the four topographic features, except elevation, the classification contribution of the other three topographic features is very low. This is because elevation, as one of the fundamental and significant topographic features, can indicate variations in the surface height and terrain relief, thereby influencing the growth and distribution of vegetation. On the other hand, other topographic features, such as slope and aspect, might be implicitly captured by elevation in certain cases, leading to their lower contribution to the classification results. Among the additional polarization features, the VV and VH polarization bands exhibited prominent classification contributions and yielded better results. The polarization characteristics of S1 are associated with crop growth conditions and other factors. The polarization characteristics of Sentinel-1’s radar data provide improved discrimination between various crops, serving as a vital basis for crop classification. Utilizing the sequence forward selection (SFS) method, the initial set of 26 candidate features was reduced to 20. As can be seen from the red points in Figure 5b, when the number of features involved in classification reaches 20, the overall accuracy of classification no longer increases substantially. That is, with the continuous addition of classification features, the overall accuracy reaches a stable state. This method enables the elimination of redundant features, enhances the classifier’s performance, and results in the identification of the optimal feature subset.
To gain deeper insights into the contribution level of feature variables in this classification process, the polarization features, VV and VH, along with the vegetation index feature, NDVI, were chosen. A time series approach was employed to analyze the response patterns of different fruit trees to these feature variables (Figure 6 and Figure 7). Among them, due to the effective smoothing and fidelity provided by the Harmonic Analysis of Time Series (HANTS) method, HANTS harmonic analysis was applied to fit the NDVI time series curves. This process facilitated noise removal and the reconstruction of the NDVI time series (Figure 6a,b). The results indicated that, except for the jujube tree, which consistently exhibited lower NDVI values throughout the year, the growth patterns of the other fruit trees did not display noticeable differences. These findings confirmed that relying solely on the NDVI as a classification criterion was insufficient and ineffective in distinguishing the four crops. Therefore, it was necessary to incorporate the remaining characteristic variables in the fruit tree crops’ classification process.
VV (vertical polarization) and VH (horizontal polarization) are the two polarization modes employed in synthetic aperture radar (SAR). These modes offer valuable insights into diverse scattering mechanisms that reflect a surface’s characteristics and exert a considerably greater influence than other variables in the classification process. The responses of VV and VH polarization bands vary with the season and growth stage of fruit crops throughout the phenological period (Figure 7). Initially, polarization reflectance is low in early crop growth due to factors like shorter vegetation, limited cover, and soil exposure. It gradually rises as a crop’s height and vegetation coverage increase. Upon maturity, polarization diminishes due to height saturation, stabilized cover, and increased soil exposure, leading to a decreased SAR signal reflection.

4.2. Accuracy Differences in Key Temporal Phases

Using the random forest combined with the object-oriented method, this study investigates the disparities in remote sensing accuracy for identifying fruit tree planting areas during different phases. The analysis considers the variations in phenology among different fruit tree planting areas throughout each month (Figure 8). Notably, the jujube planting area exhibited the most favorable remote sensing extraction effect, achieving a user accuracy of 99.18% and a producer accuracy of 96.62% in November. Conversely, the pear tree planting area displayed the least satisfactory remote sensing extraction effect, with a user accuracy peaking at 93.42% and a producer accuracy of 89.03% in November. Overall, the remote sensing extraction accuracy in different fruit tree planting areas can be ranked as follows: jujube tree > apple tree > walnut tree > pear tree. During the field survey conducted in November, jujube trees exhibited distinctive characteristics, including yellow leaves and red jujubes (Figure 9), which facilitated their differentiation. As a result, when performing single-crop extraction, jujube trees were found to be more readily identifiable based on the critical temporal phases of fruit trees compared to other fruit tree growing areas.

4.3. Comparison of Classification Models

The utilization of the cloud platform streamlines image pre-processing and enhances classification efficiency in this classification process. Furthermore, the random forest combined with the object-oriented (RF + OO) classification model demonstrates superior accuracy and results compared to standalone machine learning and object-oriented classification models.
Firstly, the overall accuracy (OA) and Kappa coefficient (KC) were evaluated for four models: random forest (RF), support vector machine (SVM), object-oriented (OO), and RF + OO (Figure 10b). SVM exhibited the lowest accuracy, with OA and KC values of 52.76% and 0.31, respectively, while the RF + OO approach achieved the highest accuracy, with OA and KC values of 94.68% and 0.92, respectively. Comparing it to the RF classification method alone, the RF + OO approach yielded an improvement of 4.52% in OA and 0.16 in KC. Consequently, the results conclusively demonstrate that the RF + OO classification method outperforms the other three approaches. The combination of the SNIC algorithm and the RF classifier proves effective in enhancing classification accuracy and handling large-scale remote sensing image data.
Subsequently, the classification outcomes for four representative fruit tree crops were evaluated based on their producer accuracy (PA) and user accuracy (UA) in the crop classification results (Figure 10a). The results reveal that apple trees achieved the highest producer accuracy at 97.32%, closely followed by pear trees at 97.13%. The walnut tree and jujube tree exhibited producer accuracies of 94.64% and 89.72%, respectively. Regarding user accuracy, apple trees also attained the highest value of 95.62%, trailed by jujube trees at 93.22%. Walnut trees and pear trees achieved user accuracies of 92.16% and 91.64%, respectively. In a comprehensive assessment, apple and pear trees demonstrate relatively high classification accuracy, whereas jujube trees and walnut trees exhibit relatively lower accuracy. To enhance the classification accuracy, further algorithm optimization and improvement in remote sensing data quality are necessary.

4.4. Crop Distribution Results

Based on the aforementioned research findings, this study establishes that September exhibits the highest importance in classification features, rendering it the most discriminative month for fruit tree classification. Consequently, September’s imagery was chosen to extract the planting area statistics and classification results due to its provision of valuable and distinctive information for accurately identifying fruit tree types. This involved generating the area of individual classified image elements using the ee.Image.pixelArea() function provided by GEE, followed by a summation of the image element areas using the ee.Reducer.sum() function. Consequently, employing the RF + OO classification method, the planting areas of different fruit tree crops and the percentage of the total fruit tree planting area in Aksu City were obtained (Table 5).
The overall classification results in Figure 11 (excluding the desert area in the southeast; Green, red, blue, yellow, and white correspond to walnuts, date trees, apples, pears, and others (roads, deserts, and residential areas)) reveal that the distribution of the four crops in the study area is mixed and blocky, with noticeable linear paths connecting different plots. Fruit tree planting exhibits distinct spatial distribution characteristics. Jujube trees are extensively cultivated in Aksu City, particularly in the first and second regiments of the Corps, as well as Keratale Town. The planting area of walnuts extends from the center of Aksu City to the surrounding areas, prominently present in Topruk Township, Baishtugman Township, Kumbashi Township, Aykule Town, and Karatale Town. Pear trees exhibit a similar spatial distribution as walnut trees, although with a smaller planting area. Apple trees are predominantly grown in large numbers in the northern part of Aksu City, specifically in the villages of Langan, Yukak Baridang, and near the experimental forest farm. Being a renowned hub for melons and fruits in Xinjiang, Aksu City has witnessed rapid development in the fruit industry. It has established a planting base with a specific focus on forest fruit species.

5. Discussion

5.1. Analysis of Feature Variables

This study assessed the effectiveness of integrating RF and object-oriented algorithms for extracting information on fruit tree planting areas. Sentinel-1 radar images, Sentinel-2 optical images, and SRTM data were utilized to map the fruit tree planting areas in Aksu City at a 10 m resolution in 2022. Moreover, this study explored the efficacy of feature optimization using the sequence forward selection algorithm with Gini coefficients. This approach aimed to reduce data dimensionality, enhance classification accuracy, and obtain more precise information on the distribution of fruit tree planting. The fruit tree industry in the Aksu area serves as a key sector for income generation and economic prosperity. The information on fruit-growing areas provided by this study holds great significance for the efficient management of the local fruit tree industry.
After feature optimization and employing a combined model for classification, the overall accuracy reached 94.68%, with a Kappa coefficient of 0.92. This study revealed that the contribution of different features to fruit tree classification varied. Notably, the Sentinel-1 backscatter coefficient feature exhibited significant variations in the feedback from different fruit tree planting areas and played a crucial role in the classification process. By incorporating VV and VH bands, the overall accuracy increased to 88.12%. However, the combination with Sentinel-2 data is still necessary due to the limited number of channels. These findings align with previous studies [61]. Nevertheless, this study utilized a combined classification model after feature optimization, achieving an overall accuracy of 94.68% when all optimization features were included. This study highlights the potential of radar data for fruit tree applications. Furthermore, future studies on fruit tree species’ classification and health parameter inversions can explore the generation of a novel remote sensing index specific to fruit trees based on the radar band time series. This index can enhance the accuracy of remote sensing parameter inversion for fruit trees. Moreover, in this study, the NDVI was employed to generate texture features using the grayscale co-occurrence matrix. These features held a moderate position in the classification feature ranking and effectively enhanced the classification accuracy. In a study by Nabil [62] et al. on horticultural tree identification, texture features were also incorporated into the classification; however, unlike our study, they generated texture features based on the Sentinel-1 radar band. Given the unique texture features in fruit tree planting areas resulting from row spacing and tree spacing and considering the proven high sensitivity of the Sentinel-1 radar band to fruit tree identification in this study, we propose utilizing the VV and VH polarization bands to generate texture indicators. This approach will facilitate a more comprehensive comparison and exploration in future research. Additionally, the study found that among the four classification features derived from the SRTM terrain data, the elevation variable had a significant contribution, while the slope and slope direction variables had the lowest contribution level. This disparity arises from certain fruit tree planting areas being constrained within specific elevation ranges, with elevation changes leading to divisions among these areas. Compared to features like vegetation indices, terrain features such as slope and slope direction exhibit greater short-term stability, resulting in a diminished ability to distinguish different fruit tree planting areas. These findings further validate the rationality and reliability of this study.

5.2. Key Temporal Phase Analysis

Variations in classification accuracy across fruit tree growing areas and key time phases were observed. For example, jujube planting areas exhibited higher accuracy in July and November, corresponding to summer and autumn. In November, the jujube areas displayed distinctive yellow leaves and red dates, resulting in 99.18% user accuracy. This unique feature enables differentiation from other crops, enhancing overall accuracy. This study’s findings are a valuable reference for subsequent separate classifications, like jujube. Long-term changes were not examined, as fruit trees have prolonged growth periods, leading to stable planting. Statistical outcomes extend beyond the present, offering lasting significance and practical value for decision-making and management.

5.3. Classification Model Analysis

The extraction of fruit tree planting areas differs from the precise identification of individual tree species, which is commonly accomplished using UAV imagery integrated with deep learning methods [63]. These methods accurately detect the target tree species and extract the canopy area. However, they overlook the spacing between tree species due to their reliance on high-resolution UAV imagery. Additionally, UAV remote sensing is inadequate for large-scale crop identification due to storage limitations and economic inefficiencies. Consequently, utilizing satellite imagery with an appropriate resolution that covers the spacing of fruit tree planting is superior to the UAV remote sensing approach. This satellite-based method encompasses both fruit tree planting details and maintains accuracy under low-cost conditions. Research indicates that the SVM model exhibits the lowest classification accuracy, which is consistent with the findings from previous studies comparing it with traditional machine learning classification models [64,65]. Several factors contribute to the outcome that the SVM model was originally designed for binary classification problems, thereby complicating its application to large-scale multiclass problems. Moreover, the SVM model is sensitive to missing data and outliers, which greatly affects the determination of decision boundaries in the presence of noise and outliers, resulting in classification errors. In our research, the SVM + OO algorithm was also employed for classification purposes. However, its performance closely resembled that of utilizing SVM alone. As a result, we have not extensively elaborated on the SVM + OO approach within this paper. In contrast, the combined RF and object-oriented classification model employed in this study significantly improves the OA by 4.52% and the KC by 0.16 when compared to the RF classification method alone. Comparative analysis of the classification result images (Figure 11) demonstrates that the combined classification model effectively mitigates the noise generated by the pixel-based RF algorithm while preserving the plant spacing and plot characteristics of the fruit tree planting area. The combination model classification method investigated in this study for fruit tree planting areas serves as a valuable reference for subsequent remote sensing extraction over a wider range of fruit tree planting areas.

5.4. Uncertainty Analysis

Numerous studies have indicated that the crop area estimated through remote sensing does not always align precisely with the actual distribution area [66,67,68]. This divergence results from limited image resolution, challenging the detection of small areas. Interpretation encounters difficulties in assessing scattered small-scale fruit tree planting abundance. Classification methods and parameter selection affect accuracy. For precise large-scale fruit tree spatial data, nuanced judgments and specific factors are vital. Machine learning and object-oriented models enhance accuracy, effectively reducing the noise impact. This yields closely aligned results with the actual distribution, benefiting monitoring and interpretation.

6. Conclusions

This study focuses on the Aksu City area in Xinjiang, known for its distinctive fruit tree planting areas. Utilizing a remote sensing cloud computing platform, we employed multi-source remote sensing data, including Sentinel-1, Sentinel-2, and SRTM. These data were utilized to construct a classification feature space and analyze the critical temporal phases and optimal features associated with various fruit tree planting areas. Additionally, we compared multiple classification methods to investigate the suitability of RF and the combination of RF and object-oriented classification models for fruit tree planting areas, ultimately achieving the extraction of characteristic fruit tree planting structures in Aksu City in 2022. The statistical findings of this study hold substantial long-term significance and practical application value. They offer continuous support for decision-making and management within relevant departments, fostering the development of the fruit tree industry and rural revitalization. The main research findings are as follows: (1) Comparing the contribution of different categorical features, the backscatter coefficient feature of Sentinel-1 exhibited the most significant influence, followed by the original band feature and vegetation index of Sentinel-2. In contrast, the topographic feature made a relatively minor contribution. Furthermore, considering the comprehensive performance of the categorical feature variables throughout the year, their importance was highest in September compared to other months. (2) When comparing the classification accuracy of different fruit trees during each temporal phase, the jujube planting region exhibited the highest accuracy in November, achieving 99.18% user accuracy and 96.62% producer accuracy. The walnut planting region and apple planting region followed suit, while the pear planting region had the lowest classification accuracy, with 93.42% user accuracy and 89.03% producer accuracy. (3) We analyzed and validated the accuracies of the RF, SVM, OO, and RF + OO classification models. Among them, the RF + OO classification model achieved the highest classification accuracy (OA = 94.68%, Kappa = 0.92), while SVM exhibited the lowest classification accuracy (OA = 52.76%, Kappa = 0.31). (4) The total area of typical fruit tree plantations in Aksu City in 2022 amounts to 64,000 hectares. Among them, walnut tree plantations cover the largest area, with a statistical result of 27,200 hectares, accounting for 42.5% of the total fruit tree area. The second-largest plantations are dedicated to jujube and pear trees, occupying 13,200 hectares (20.6%) and 12,400 hectares (19.3%), respectively. Apple tree plantations cover the smallest area, with a statistical result of 11,200 hectares (17.5%). The classification outcome aligns with the actual crop distribution in the study area.

Author Contributions

Conceptualization, G.Z.; methodology, G.Z.; investigation, N.T. and G.Z.; writing—original draft preparation, G.Z.; writing—review and editing, J.Z., W.H. and L.L.; visualization, G.Z.; supervision, J.Z.; project administration, L.W. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by the Xinjiang Uyghur Autonomous Region Key R&D Special Project ‘Construction of a Forest and Fruit Resource Data System Based on “Space, Sky, and Earth” Multisource Remote Sensing Monitoring Technology’ (No. 20222101536).

Data Availability Statement

All data in this article can be obtained by reasonably contacting the corresponding author.

Acknowledgments

We are grateful to Google Earth Engine for providing us with a free computing platform and to the Xinjiang Academy of Forestry Sciences for providing financial support for the field data.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Meng, X.; Rao, Y.; Tao, T.; Dong, S.; Jia, A.L.; Ma, H. A Review of Plant Breeders’ Rights Application and Granting for Fruit Trees in China from 2000 to 2019. Sci. Hortic. 2021, 276, 109749. [Google Scholar] [CrossRef]
  2. Xia, C.; Liu, Z.; Suo, X.; Cao, S. Quantifying the Net Benefit of Land Use of Fruit Trees in China. Land Use Policy 2020, 90, 104276. [Google Scholar] [CrossRef]
  3. Zhou, X.-X.; Li, Y.-Y.; Luo, Y.-K.; Sun, Y.-W.; Su, Y.-J.; Tan, C.-W.; Liu, Y.-J. Research on Remote Sensing Classification of Fruit Trees Based on Sentinel-2 Multi-Temporal Imageries. Sci. Rep. 2022, 12, 11549. [Google Scholar] [CrossRef]
  4. Gao, W.; Qiu, Q.; Yuan, C.; Shen, X.; Cao, F.; Wang, G.; Wang, G. Forestry Big Data: A Review and Bibliometric Analysis. Forests 2022, 13, 1549. [Google Scholar] [CrossRef]
  5. Phiri, D.; Morgenroth, J. Developments in Landsat Land Cover Classification Methods: A Review. Remote Sens. 2017, 9, 967. [Google Scholar] [CrossRef]
  6. Tamiminia, H.; Salehi, B.; Mahdianpari, M.; Quackenbush, L.; Adeli, S.; Brisco, B. Google Earth Engine for Geo-Big Data Applications: A Meta-Analysis and Systematic Review. ISPRS J. Photogramm. Remote Sens. 2020, 164, 152–170. [Google Scholar] [CrossRef]
  7. Zhao, Q.; Yu, L.; Li, X.; Peng, D.; Zhang, Y.; Gong, P. Progress and Trends in the Application of Google Earth and Google Earth Engine. Remote Sens. 2021, 13, 3778. [Google Scholar] [CrossRef]
  8. Huang, Y.; Chen, Z.; Yu, T.; Huang, X.; Gu, X. Agricultural Remote Sensing Big Data: Management and Applications. J. Integr. Agric. 2018, 17, 1915–1931. [Google Scholar] [CrossRef]
  9. Sishodia, R.P.; Ray, R.L.; Singh, S.K. Applications of Remote Sensing in Precision Agriculture: A Review. Remote Sens. 2020, 12, 3136. [Google Scholar] [CrossRef]
  10. Fassnacht, F.E.; Latifi, H.; Stereńczak, K.; Modzelewska, A.; Lefsky, M.; Waser, L.T.; Straub, C.; Ghosh, A. Review of Studies on Tree Species Classification from Remotely Sensed Data. Remote Sens. Environ. 2016, 186, 64–87. [Google Scholar] [CrossRef]
  11. Dainelli, R.; Toscano, P.; Di Gennaro, S.F.; Matese, A. Recent Advances in Unmanned Aerial Vehicles Forest Remote Sensing—A Systematic Review. Part II: Research Applications. Forests 2021, 12, 397. [Google Scholar] [CrossRef]
  12. Tang, L.; Shao, G. Drone Remote Sensing for Forestry Research and Practices. J. For. Res. 2015, 26, 791–797. [Google Scholar] [CrossRef]
  13. Guo, Q.; Zhang, J.; Guo, S.; Ye, Z.; Deng, H.; Hou, X.; Zhang, H. Urban Tree Classification Based on Object-Oriented Approach and Random Forest Algorithm Using Unmanned Aerial Vehicle (UAV) Multispectral Imagery. Remote Sens. 2022, 14, 3885. [Google Scholar] [CrossRef]
  14. Li, Y.; Chang, C.; Wang, Z.; Li, T.; Li, J.; Zhao, G. Identification of Cultivated Land Quality Grade Using Fused Multi-Source Data and Multi-Temporal Crop Remote Sensing Information. Remote Sens. 2022, 14, 2109. [Google Scholar] [CrossRef]
  15. Pan, L.; Xia, H.; Zhao, X.; Guo, Y.; Qin, Y. Mapping Winter Crops Using a Phenology Algorithm, Time-Series Sentinel-2 and Landsat-7/8 Images, and Google Earth Engine. Remote Sens. 2021, 13, 2510. [Google Scholar] [CrossRef]
  16. Chen, R.; Zhang, C.; Xu, B.; Zhu, Y.; Zhao, F.; Han, S.; Yang, G.; Yang, H. Predicting Individual Apple Tree Yield Using UAV Multi-Source Remote Sensing Data and Ensemble Learning. Comput. Electron. Agric. 2022, 201, 107275. [Google Scholar] [CrossRef]
  17. Talukdar, S.; Singha, P.; Mahato, S.; Shahfahad; Pal, S.; Liou, Y.-A.; Rahman, A. Land-Use Land-Cover Classification by Machine Learning Classifiers for Satellite Observations—A Review. Remote Sens. 2020, 12, 1135. [Google Scholar] [CrossRef]
  18. Wang, J.; Bretz, M.; Dewan, M.A.A.; Delavar, M.A. Machine Learning in Modelling Land-Use and Land Cover-Change (LULCC): Current Status, Challenges and Prospects. Sci. Total Environ. 2022, 822, 153559. [Google Scholar] [CrossRef]
  19. Cao, R.; Fang, L.; Lu, T.; He, N. Self-Attention-Based Deep Feature Fusion for Remote Sensing Scene Classification. IEEE Geosci. Remote Sens. Lett. 2021, 18, 43–47. [Google Scholar] [CrossRef]
  20. Zhao, Y.; Zhu, W.; Wei, P.; Fang, P.; Zhang, X.; Yan, N.; Liu, W.; Zhao, H.; Wu, Q. Classification of Zambian Grasslands Using Random Forest Feature Importance Selection during the Optimal Phenological Period. Ecol. Indic. 2022, 135, 108529. [Google Scholar] [CrossRef]
  21. Sheykhmousa, M.; Mahdianpari, M.; Ghanbari, H.; Mohammadimanesh, F.; Ghamisi, P.; Homayouni, S. Support Vector Machine Versus Random Forest for Remote Sensing Image Classification: A Meta-Analysis and Systematic Review. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 6308–6325. [Google Scholar] [CrossRef]
  22. Wessel, M.; Brandmeier, M.; Tiede, D. Evaluation of Different Machine Learning Algorithms for Scalable Classification of Tree Types and Tree Species Based on Sentinel-2 Data. Remote Sens. 2018, 10, 1419. [Google Scholar] [CrossRef]
  23. Shi, S.; Zhong, Y.; Zhao, J.; Lv, P.; Liu, Y.; Zhang, L. Land-Use/Land-Cover Change Detection Based on Class-Prior Object-Oriented Conditional Random Field Framework for High Spatial Resolution Remote Sensing Imagery. IEEE Trans. Geosci. Remote Sens. 2022, 60, 1–16. [Google Scholar] [CrossRef]
  24. Liu, T.; Yang, L.; Lunga, D. Change Detection Using Deep Learning Approach with Object-Based Image Analysis. Remote Sens. Environ. 2021, 256, 112308. [Google Scholar] [CrossRef]
  25. Pelletier, C.; Valero, S.; Inglada, J.; Champion, N.; Dedieu, G. Assessing the Robustness of Random Forests to Map Land Cover with High Resolution Satellite Image Time Series over Large Areas. Remote Sens. Environ. 2016, 187, 156–168. [Google Scholar] [CrossRef]
  26. Bao, J.; Li, J.; Wang, G.; Tang, Z.; Zhi, J. Branch Growth, Leaf Canopies and Photosynthetic Responses of Zizyphus Jujube Cv. “Huizao” to Nutrient Addition in the Arid Areas of Northwest China. Diversity 2022, 14, 914. [Google Scholar] [CrossRef]
  27. Yao, J.; Chen, Y.; Guan, X.; Zhao, Y.; Chen, J.; Mao, W. Recent Climate and Hydrological Changes in a Mountain–Basin System in Xinjiang, China. Earth-Sci. Rev. 2022, 226, 103957. [Google Scholar] [CrossRef]
  28. Xu, H.; Yang, J.; Xia, G.; Lin, T. Spatio-Temporal Differentiation of Coupling Coordination between Ecological Footprint and Ecosystem Service Functions in the Aksu Region, Xinjiang, China. Sustainability 2022, 14, 3483. [Google Scholar] [CrossRef]
  29. Li, H.; Pan, H.; Wang, D.; Liu, B.; Liu, J.; Zhang, J.; Lu, Y. Intercropping With Fruit Trees Increases Population Abundance and Alters Species Composition of Spider Mites on Cotton. Environ. Entomol. 2018, 47, 781–787. [Google Scholar] [CrossRef]
  30. Amani, M.; Ghorbanian, A.; Ahmadi, S.A.; Kakooei, M.; Moghimi, A.; Mirmazloumi, S.M.; Moghaddam, S.H.A.; Mahdavi, S.; Ghahremanloo, M.; Parsian, S.; et al. Google Earth Engine Cloud Computing Platform for Remote Sensing Big Data Applications: A Comprehensive Review. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 5326–5350. [Google Scholar] [CrossRef]
  31. Mullissa, A.; Vollrath, A.; Odongo-Braun, C.; Slagter, B.; Balling, J.; Gou, Y.; Gorelick, N.; Reiche, J. Sentinel-1 SAR Backscatter Analysis Ready Data Preparation in Google Earth Engine. Remote Sens. 2021, 13, 1954. [Google Scholar] [CrossRef]
  32. Duan, Q.; Tan, M.; Guo, Y.; Wang, X.; Xin, L. Understanding the Spatial Distribution of Urban Forests in China Using Sentinel-2 Images with Google Earth Engine. Forests 2019, 10, 729. [Google Scholar] [CrossRef]
  33. Sun, Z.; Xu, R.; Du, W.; Wang, L.; Lu, D. High-Resolution Urban Land Mapping in China from Sentinel 1A/2 Imagery Based on Google Earth Engine. Remote Sens. 2019, 11, 752. [Google Scholar] [CrossRef]
  34. Schlund, M.; Erasmi, S. Sentinel-1 Time Series Data for Monitoring the Phenology of Winter Wheat. Remote Sens. Environ. 2020, 246, 111814. [Google Scholar] [CrossRef]
  35. Yang, K.; Luo, Y.; Li, M.; Zhong, S.; Liu, Q.; Li, X. Reconstruction of Sentinel-2 Image Time Series Using Google Earth Engine. Remote Sens. 2022, 14, 4395. [Google Scholar] [CrossRef]
  36. Zakeri, F.; Mariethoz, G. A Review of Geostatistical Simulation Models Applied to Satellite Remote Sensing: Methods and Applications. Remote Sens. Environ. 2021, 259, 112381. [Google Scholar] [CrossRef]
  37. Tarolli, P. High-Resolution Topography for Understanding Earth Surface Processes: Opportunities and Challenges. Geomorphology 2014, 216, 295–312. [Google Scholar] [CrossRef]
  38. Su, Y.; Guo, Q. A Practical Method for SRTM DEM Correction over Vegetated Mountain Areas. ISPRS J. Photogramm. Remote Sens. 2014, 87, 216–228. [Google Scholar] [CrossRef]
  39. Wardlow, B.D.; Egbert, S.L. A Comparison of MODIS 250-m EVI and NDVI Data for Crop Mapping: A Case Study for Southwest Kansas. Int. J. Remote Sens. 2010, 31, 805–830. [Google Scholar] [CrossRef]
  40. Radočaj, D.; Šiljeg, A.; Marinović, R.; Jurišić, M. State of Major Vegetation Indices in Precision Agriculture Studies Indexed in Web of Science: A Review. Agriculture 2023, 13, 707. [Google Scholar] [CrossRef]
  41. Mandal, D.; Kumar, V.; Ratha, D.; Dey, S.; Bhattacharya, A.; Lopez-Sanchez, J.M.; McNairn, H.; Rao, Y.S. Dual Polarimetric Radar Vegetation Index for Crop Growth Monitoring Using Sentinel-1 SAR Data. Remote Sens. Environ. 2020, 247, 111954. [Google Scholar] [CrossRef]
  42. Tassi, A.; Vizzari, M. Object-Oriented LULC Classification in Google Earth Engine Combining SNIC, GLCM, and Machine Learning Algorithms. Remote Sens. 2020, 12, 3776. [Google Scholar] [CrossRef]
  43. Cheng, X.; Liu, W.; Zhou, J.; Wang, Z.; Zhang, S.; Liao, S. Extraction of Mountain Grasslands in Yunnan, China, from Sentinel-2 Data during the Optimal Phenological Period Using Feature Optimization. Agronomy 2022, 12, 1948. [Google Scholar] [CrossRef]
  44. Liu, Y.; Xiao, D.; Yang, W. An Algorithm for Early Rice Area Mapping from Satellite Remote Sensing Data in Southwestern Guangdong in China Based on Feature Optimization and Random Forest. Ecol. Inform. 2022, 72, 101853. [Google Scholar] [CrossRef]
  45. Breiman, L. Random Forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  46. Liang, B.; Zhao, R.; Tan, J.; Xia, L.; Cao, H.; Wu, S.; Yang, P. The Application of Compact Polarization Decomposition in the Construction of a Dual-Polarization Radar Index and the Effect Evaluation of Rape Extraction. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2023, 16, 5315–5330. [Google Scholar] [CrossRef]
  47. Behnamian, A.; Millard, K.; Banks, S.N.; White, L.; Richardson, M.; Pasher, J. A Systematic Approach for Variable Selection With Random Forests: Achieving Stable Variable Importance Values. IEEE Geosci. Remote Sens. Lett. 2017, 14, 1988–1992. [Google Scholar] [CrossRef]
  48. Mountrakis, G.; Im, J.; Ogole, C. Support Vector Machines in Remote Sensing: A Review. ISPRS J. Photogramm. Remote Sens. 2011, 66, 247–259. [Google Scholar] [CrossRef]
  49. Liu, P.; Choo, K.-K.R.; Wang, L.; Huang, F. SVM or Deep Learning? A Comparative Study on Remote Sensing Image Classification. Soft Comput. 2017, 21, 7053–7065. [Google Scholar] [CrossRef]
  50. Razaque, A.; Ben Haj Frej, M.; Almi’ani, M.; Alotaibi, M.; Alotaibi, B. Improved Support Vector Machine Enabled Radial Basis Function and Linear Variants for Remote Sensing Image Classification. Sensors 2021, 21, 4431. [Google Scholar] [CrossRef]
  51. Liu, Y.; Li, M.; Mao, L.; Xu, F.; Huang, S. Review of Remotely Sensed Imagery Classification Patterns Based on Object-Oriented Image Analysis. Chin. Geogr. Sci. 2006, 16, 282–288. [Google Scholar] [CrossRef]
  52. Luo, C.; Qi, B.; Liu, H.; Guo, D.; Lu, L.; Fu, Q.; Shao, Y. Using Time Series Sentinel-1 Images for Object-Oriented Crop Classification in Google Earth Engine. Remote Sens. 2021, 13, 561. [Google Scholar] [CrossRef]
  53. Yang, L.; Wang, L.; Abubakar, G.A.; Huang, J. High-Resolution Rice Mapping Based on SNIC Segmentation and Multi-Source Remote Sensing Images. Remote Sens. 2021, 13, 1148. [Google Scholar] [CrossRef]
  54. Tu, Y.; Chen, B.; Zhang, T.; Xu, B. Regional Mapping of Essential Urban Land Use Categories in China: A Segmentation-Based Approach. Remote Sens. 2020, 12, 1058. [Google Scholar] [CrossRef]
  55. Menze, B.H.; Kelm, B.M.; Masuch, R.; Himmelreich, U.; Bachert, P.; Petrich, W.; Hamprecht, F.A. A Comparison of Random Forest and Its Gini Importance with Standard Chemometric Methods for the Feature Selection and Classification of Spectral Data. BMC Bioinform. 2009, 10, 213. [Google Scholar] [CrossRef] [PubMed]
  56. Cheng, K.; Wang, J. Forest Type Classification Based on Integrated Spectral-Spatial-Temporal Features and Random Forest Algorithm—A Case Study in the Qinling Mountains. Forests 2019, 10, 559. [Google Scholar] [CrossRef]
  57. Gromski, P.S.; Xu, Y.; Correa, E.; Ellis, D.I.; Turner, M.L.; Goodacre, R. A Comparative Investigation of Modern Feature Selection and Classification Approaches for the Analysis of Mass Spectrometry Data. Anal. Chim. Acta 2014, 829, 1–8. [Google Scholar] [CrossRef] [PubMed]
  58. Cai, L.; Shi, W.; Miao, Z.; Hao, M. Accuracy Assessment Measures for Object Extraction from Remote Sensing Images. Remote Sens. 2018, 10, 303. [Google Scholar] [CrossRef]
  59. Foody, G.M. Explaining the Unsuitability of the Kappa Coefficient in the Assessment and Comparison of the Accuracy of Thematic Maps Obtained by Image Classification. Remote Sens. Environ. 2020, 239, 111630. [Google Scholar] [CrossRef]
  60. Allouche, O.; Tsoar, A.; Kadmon, R. Assessing the Accuracy of Species Distribution Models: Prevalence, Kappa and the True Skill Statistic (TSS): Assessing the Accuracy of Distribution Models. J. Appl. Ecol. 2006, 43, 1223–1232. [Google Scholar] [CrossRef]
  61. Chabalala, Y.; Adam, E.; Ali, K.A. Machine Learning Classification of Fused Sentinel-1 and Sentinel-2 Image Data towards Mapping Fruit Plantations in Highly Heterogenous Landscapes. Remote Sens. 2022, 14, 2621. [Google Scholar] [CrossRef]
  62. Nabil, M.; Farg, E.; Arafat, S.M.; Aboelghar, M.; Afify, N.M.; Elsharkawy, M.M. Tree-Fruits Crop Type Mapping from Sentinel-1 and Sentinel-2 Data Integration in Egypt’s New Delta Project. Remote Sens. Appl. Soc. Environ. 2022, 27, 100776. [Google Scholar] [CrossRef]
  63. Tian, H.; Fang, X.; Lan, Y.; Ma, C.; Huang, H.; Lu, X.; Zhao, D.; Liu, H.; Zhang, Y. Extraction of Citrus Trees from UAV Remote Sensing Imagery Using YOLOv5s and Coordinate Transformation. Remote Sens. 2022, 14, 4208. [Google Scholar] [CrossRef]
  64. Adugna, T.; Xu, W.; Fan, J. Comparison of Random Forest and Support Vector Machine Classifiers for Regional Land Cover Mapping Using Coarse Resolution FY-3C Images. Remote Sens. 2022, 14, 574. [Google Scholar] [CrossRef]
  65. Zafari, A.; Zurita-Milla, R.; Izquierdo-Verdiguier, E. Evaluating the Performance of a Random Forest Kernel for Land Cover Classification. Remote Sens. 2019, 11, 575. [Google Scholar] [CrossRef]
  66. Goodwin, N.R.; Coops, N.C.; Tooke, T.R.; Christen, A.; Voogt, J.A. Characterizing Urban Surface Cover and Structure with Airborne Lidar Technology. Can. J. Remote Sens. 2009, 35, 297–309. [Google Scholar] [CrossRef]
  67. Potapov, P.V.; Turubanova, S.A.; Tyukavina, A.; Krylov, A.M.; McCarty, J.L.; Radeloff, V.C.; Hansen, M.C. Eastern Europe’s Forest Cover Dynamics from 1985 to 2012 Quantified from the Full Landsat Archive. Remote Sens. Environ. 2015, 159, 28–43. [Google Scholar] [CrossRef]
  68. Fritz, S.; See, L.; McCallum, I.; You, L.; Bun, A.; Moltchanova, E.; Duerauer, M.; Albrecht, F.; Schill, C.; Perger, C.; et al. Mapping Global Cropland and Field Size. Glob. Chang. Biol. 2015, 21, 1980–1992. [Google Scholar] [CrossRef]
Remotesensing 15 04140 g001
Figure 1. Overview of the study area.
Figure 1. Overview of the study area.
Remotesensing 15 04140 g001
Remotesensing 15 04140 g002
Figure 2. Field pictures of fruit trees: (a) walnuts; (b) jujubes; (c) apples; (d) pears.
Figure 2. Field pictures of fruit trees: (a) walnuts; (b) jujubes; (c) apples; (d) pears.
Remotesensing 15 04140 g002
Remotesensing 15 04140 g003
Figure 3. Technology roadmap.
Figure 3. Technology roadmap.
Remotesensing 15 04140 g003
Remotesensing 15 04140 g004
Figure 4. T-value optimization diagram.
Figure 4. T-value optimization diagram.
Remotesensing 15 04140 g004
Remotesensing 15 04140 g005
Figure 5. Classification feature heat map and feature importance map: (a) classification feature heat map in different months; (b) feature importance statistical map in September.
Figure 5. Classification feature heat map and feature importance map: (a) classification feature heat map in different months; (b) feature importance statistical map in September.
Remotesensing 15 04140 g005
Remotesensing 15 04140 g006
Figure 6. Time series curves of four crops: (a) NDVI; (b) HANTS harmonic analysis.
Figure 6. Time series curves of four crops: (a) NDVI; (b) HANTS harmonic analysis.
Remotesensing 15 04140 g006
Remotesensing 15 04140 g007
Figure 7. Sentinel-1 SAR-band time series: (a) VV polarization characteristic curve; (b) VH polarization characteristic curve.
Figure 7. Sentinel-1 SAR-band time series: (a) VV polarization characteristic curve; (b) VH polarization characteristic curve.
Remotesensing 15 04140 g007
Remotesensing 15 04140 g008
Figure 8. Identification accuracy of different planting areas of fruit trees in each month: (a) walnut; (b) jujube; (c) apple; (d) pear.
Figure 8. Identification accuracy of different planting areas of fruit trees in each month: (a) walnut; (b) jujube; (c) apple; (d) pear.
Remotesensing 15 04140 g008
Remotesensing 15 04140 g009
Figure 9. Morphology of jujube tree in November.
Figure 9. Morphology of jujube tree in November.
Remotesensing 15 04140 g009
Remotesensing 15 04140 g010
Figure 10. Classification accuracy of different classification methods: (a) producer accuracy and user accuracy, (b) overall accuracy and Kappa coefficient.
Figure 10. Classification accuracy of different classification methods: (a) producer accuracy and user accuracy, (b) overall accuracy and Kappa coefficient.
Remotesensing 15 04140 g010
Remotesensing 15 04140 g011
Figure 11. Fruit tree distribution map: (a) Sentinel-2 image; (b) local image of Sentinel-2 with real samples; (c) GLCM; (d) SVM; (e) RF; (f) OO; (g) RF + OO.
Figure 11. Fruit tree distribution map: (a) Sentinel-2 image; (b) local image of Sentinel-2 with real samples; (c) GLCM; (d) SVM; (e) RF; (f) OO; (g) RF + OO.
Remotesensing 15 04140 g011
Table 1. Category and number of sample points.
Table 1. Category and number of sample points.
GeocodeSpeciesSample Size
0Walnut736
1Jujube727
2Apple1412
3Pear652
4Other500
Total 4027
Table 2. Calculate vegetation index.
Table 2. Calculate vegetation index.
Vegetation IndexAbbreviationsBased on S2 Expressions
Normalized Difference Vegetation IndexNDVI(B8 − B4)/(B8 + B4)
Ratio Vegetation IndexRVIB8/B4
Enhanced Vegetation IndexEVI2.5 × ((B8 − B4)/(B8 + 6 × B4 − 7.5 × B2 + 1))
Soil Adjusted Vegetation IndexSAVI(B8 − B4) × (1 + 0.5)/(B8 + B4 + 0.5)
Table 3. Texture features.
Table 3. Texture features.
FeaturesDescription
Angular Second Moment (ASM)Measure the uniformity or energy of the image grayscale distribution
ContrastContrast measurement based on local grayscale changes
CorrelationMeasure linear correlation of adjacent pixel grayscale
EntropyMeasure the degree of confusion between pixels in an image
Table 4. Accuracy evaluation index.
Table 4. Accuracy evaluation index.
Evaluation MetricsDescription
PARatio of the number of pixels correctly classified as that category to the actual total number of pixels referenced in that category
UARatio of the total number of pixels correctly classified as that category to the total number of pixels classified as that category
OAComprehensive evaluation of the quality of classification results
KCMetrics to check whether the model prediction results are consistent with the actual classification results
Table 5. Planting area of fruit trees.
Table 5. Planting area of fruit trees.
SpeciesArea/HectarePercentage
Walnut27,20042.5%
Jujube12,40019.3%
Apple11,20017.5%
Pear13,20020.6%
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

Zhao, G.; Wang, L.; Zheng, J.; Tuerxun, N.; Han, W.; Liu, L. Optimized Extraction Method of Fruit Planting Distribution Based on Spectral and Radar Data Fusion of Key Time Phase. Remote Sens. 2023, 15, 4140. https://doi.org/10.3390/rs15174140

AMA Style

Zhao G, Wang L, Zheng J, Tuerxun N, Han W, Liu L. Optimized Extraction Method of Fruit Planting Distribution Based on Spectral and Radar Data Fusion of Key Time Phase. Remote Sensing. 2023; 15(17):4140. https://doi.org/10.3390/rs15174140

Chicago/Turabian Style

Zhao, Guobing, Lei Wang, Jianghua Zheng, Nigela Tuerxun, Wanqiang Han, and Liang Liu. 2023. "Optimized Extraction Method of Fruit Planting Distribution Based on Spectral and Radar Data Fusion of Key Time Phase" Remote Sensing 15, no. 17: 4140. https://doi.org/10.3390/rs15174140

APA Style

Zhao, G., Wang, L., Zheng, J., Tuerxun, N., Han, W., & Liu, L. (2023). Optimized Extraction Method of Fruit Planting Distribution Based on Spectral and Radar Data Fusion of Key Time Phase. Remote Sensing, 15(17), 4140. https://doi.org/10.3390/rs15174140

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