1. Introduction
Chestnut (
Castanea mollissima Blume) is a deciduous tree belonging to
Castanea of Fagaceae, a species that is originally from China. It is a famous nut tree and one of the earliest fruit trees domesticated and utilized in the country [
1], with high economic value and ecological benefits [
2]. China’s chestnut is of excellent quality and rich in nutrients, and its existing cultivation area and annual output rank first in the world [
3,
4]. Chestnut is mostly distributed in mountainous and hilly areas, and the chestnut industry has been leading the market in the area, increasing agricultural efficiency and farmers’ income in the mountainous regions of Beijing. Chestnut tree species have also demonstrated the effect of increasing surface coverage, improving soil, and reducing soil erosion in many soil and water conservation projects [
4]. However, the planting density of the chestnut economic forest is usually low, and carrying out weeding and other operations before the harvest of the fruit easily leads to long-term exposure and loosening of the undergrowth land, resulting in the degradation of vegetation and death of chestnut trees [
4,
5,
6]. Therefore, it is of great significance to establish an automatic remote sensing identification method for chestnut forest distribution to improve the efficient management of economic forests and help maintain China’s sustainable economic and social development.
Compared with traditional field monitoring, remote sensing image identification of tree species distribution has the advantages of large spatial range and flexible time-frequency, which can cover forest areas that are difficult for human beings to enter. Remote sensing imaging methods have been widely used in tree species monitoring [
7,
8,
9]. Optical remote sensing data are the mainstream data currently used in the classification of ground objects, and they can be divided into three categories according to spatial resolution: medium-resolution data (250–1000 m), high-resolution data (less than 10 m) and medium- and high-resolution data (10–250 m). For example, moderate resolution imaging spectrometer (MODIS) data have been widely used in regional-scale mapping, and the producer’s accuracy is usually low [
10,
11,
12] due to the serious mixed pixel effect. The details of high-resolution data are clearer, but the limited space–time coverage and high cost make it difficult to realize large-scale applications [
13,
14,
15,
16]. Among the medium- and high-resolution data, Sentinel-2 series images have high spatial resolution, rich spectral information, many selectable phases, and free access, which can support accurate and wide spatial coverage and long-term plant monitoring. Many studies have reported the successful implementation of classification mapping [
17,
18] with Sentinel-2 data on a regional scale. Based on Sentinel-2A multi-temporal images, Li et al. used an RF algorithm to identify and classify tree species in Dagujia Forest Farm [
19]. Gu et al. made full use of the rich information of Sentinel-2 data in the red edge band and used a variety of vegetation indices to combine them into a time series for crop classification [
20]. Jiang et al. put forward a remote sensing comprehensive classification scheme [
21] that is suitable for coniferous forests in the Tianshan mountains using a comprehensive analysis of the shadow distribution temporal characteristics, classification characteristics, and classifier selection based on two Sentinel-2 images with great differences in solar altitude and azimuth. However, a single-phase image can only extract limited spectral characteristics of vegetation, some of which are similar between different forests.
More and more researchers have begun to explore multiple-phase images to derive high-resolution and high-precision species mapping [
22,
23]. In our study, compared with the common classification of land cover types, it became more challenging to identify tree species. It is not enough to only use the differences in spectral and spatial texture in a single phase between the target tree species and other ground objects. Hence, we try to fully explore the differences in seasonal curves and phenological characteristics from multiple-phase images to better understand this distinction.
Regarding the requirement of large-scale and long-time series of remote sensing images in vegetation phenology analysis, the GEE (Google Earth Engine) Cloud Platform is considered to be a great software for this matter [
24,
25]. It stores publicly available remote sensing image data such as Sentinel/Landsat in Google’s disk array so that GEE users can conveniently extract, call, and analyze massive remote sensing big data resources [
26,
27]. Yang et al. automatically generated winter wheat training samples with the temporal and spectral characteristics of Sentinel-2 images on the GEE platform and used the OCSVM classification method to draw the winter wheat planting area map of Jiangsu Province in China in different seasons [
28]. Tian et al. proposed a pixel-based phenological feature composite method (Ppf-CM) to reduce the spatial variability of phenology but also enhance the spectral separation between
Spartina alterniflora and native species by superimposing two unique phenological features [
29].Ni developed enhanced Ppf-CM (Eppf-CM) in the GEE platform to make full use of several representative phenological stages of rice and obtained one of the rice maps with the highest spatial resolution of 10 m on the regional scale in Northeast China [
30]. Currently, just a few studies in the field are focused on remote sensing extraction of chestnut forests. Some domestic researchers have relied on high-resolution WorldView imagery to identify the distribution of chestnut trees scattered within agricultural and urban environments [
14,
15] and used the optimal temporal phase and classification features for chestnut forest extraction by integrating MODIS and Landsat data [
21]. However, there is still a lack of research concerning the utilization of Sentinel-2 imagery’s advantages for extracting vegetation phenology features to identify chestnut forests.
Hence, this study leverages the high temporal and spatial resolution of Sentinel-2 images to extract multiple classification features, including vegetation phenological characteristics. Consequently, we employed the Google Earth Engine (GEE) platform to establish and optimize a machine learning model, aiming to develop a high-precision method for automatically identifying planted chestnut forests and generating accurate distribution maps in northeastern Beijing. The implementation of this method not only offers crucial data support for soil erosion prevention and control studies in Beijing but also serves as a valuable reference for future research endeavors in this field.
3. Results
3.1. Phenological Characteristic Construction
Generally, the growth period of Chinese chestnut (that is, from germination to defoliation) is 245 days. It usually sprouts at the end of the first ten days, produces leaves in the middle ten days, and produces new shoots in the last ten days of March. Male flowers are exposed and gradually elongate in the first ten days, and female flowers appear in the second ten days of April. The flowers bloom in May and fade in early June, and the shell begins to expand and harden and their seed growth period is from July to August. The nuts mature in early September, and leaves fall in early November and then enter dormancy. The NDVI time series curve (
Figure 2) obtained in this study reflects the growth and development of Chinese chestnut in a year, as well as significant differences with other surface feature types. In the whole time series, the NDVI value of evergreen forest is relatively high, and the NDVI value of water bodies is mostly negative. Chinese chestnut has a relatively similar NDVI value to that of cultivated land and deciduous forest land. However, in the early and late growth stages, we observed a significantly different change compared to other ground objects.
Three key phenological periods with high discrimination were determined by analyzing the NDVI time series curve: ① the flowering period of Chinese chestnut (1 April to 31 June): During this time, the NDVI curve of cultivated land, Chinese chestnut forest, and deciduous forest exhibit distinct variations due to their different flowering characteristics. ② Chinese chestnut fruiting period (15 August to 31 October): At this time, the Chinese chestnut forest and deciduous forest are relatively dense. Most vegetation in the cultivated land has been picked in preparation for the Chinese chestnut harvest, but the deciduous and evergreen forests still have strong vegetation growth. ③ Dormancy period of Chinese chestnut (1 December to the end of February): The NDVI value of evergreen forest is significantly higher than that of other ground objects during this period, owing to its distinctive vegetation cover.
The reflectance curves of spectral bands of three phenological periods of Chinese chestnut were generated based on the Sentinel-2 median composite images of different phenological periods of flowering, fruiting, and dormancy of Chinese chestnut (
Figure 3). The reflectance of the red band (Band 4), the nearinfrared band (Band 8), and the shortwave infrared band 2 (Band 12) of different land cover types during the flowering period of
Castanea mollissima are highly distinguishable. The reflectance of the red band (Band 4) is mainly different in chestnut forest land, deciduous forest land, and cultivated land. The near-infrared band (Band 8) plays a major role in distinguishing non-vegetation areas, and the reflectance of the shortwave infrared band 2 (Band 12) is more distinguishable for six surface feature types. However, the reflectivity of other bands in the flowering period of Chinese chestnut has cross convergence, and there is no obvious difference among various ground features. The blue band (Band 2), green band (Band 3), and narrow and red edge band 4 (Band 8A) of the chestnut fruit stage are relatively distinct, which can be used as the classification spectral characteristics of the fruit stage. The reflectance of the blue and green bands in cultivated and construction land is higher than that of other surface features, while the reflectance of the red edge band 4 (Band 8A) in chestnut forest land is significantly higher than that of other land covers. In addition, during the dormancy period of Chinese chestnut, the red edge band 1 (Band 5), red edge band 2 (Band 6), red edge band 3 (Band 7), and shortwave infrared band 1 (Band 11) reflectance of the ground objects are significantly different.
We extracted the vegetation index in different phenological periods and outlined the box graph according to the Sentinel-2 images. Based on the median, mean, quarter, and three-quarter values of the box graph, we analyzed the distribution characteristics of the vegetation index in different phenological periods and selected the optimal vegetation index characteristics under different phenological periods to construct the phenological characteristics. The analysis results are shown in
Figure 4. The distribution characteristics of the EVI, NDVI, PSRI, and MTCI vegetation indices in the flowering period of Chinese chestnut are quite different from those of other land types. Furthermore, the distribution characteristics of the MNDWI, WDRVI, and REP vegetation indices in the fruiting period are also different from those of other land types. There are also considerable differences in the distribution characteristics of the GNDVI, NDBI, and SAVI vegetation indices in the dormancy period of other land types.
Based on the above analysis of spectral band reflectance and vegetation index values of each phenological period, we listed 20 characteristic factor combinations of three chestnut phenological periods to outline the phenological characteristics. The flowering period corresponded to seven key indicators, the fruiting period corresponded to six key indicators, and the dormancy period corresponded to seven key indicators, as shown in
Table 4.
3.2. Comparison of Identification Methods of Chestnut Forest
This study mainly involved the identification and comparative analysis of chestnut forests using various feature combinations and machine learning methods. Initially, four feature combinations were created based on whether phenological features and topographical features were considered, as outlined in
Table 5. Subsequently, a total of twelve methods were formulated by combining these feature combinations with three different machine learning models, i.e., SVM, DT, and RF. These twelve methods were then applied to identify land cover in the Miyun and Huairou districts using Sentinel-2 images, and the results are depicted in
Figure 5 for reference.
The accuracy results of each classification method (see
Figure 6) show that the method (i.e., the fourth combination of eigenvalues) based on spectral and vegetation characteristics considering topographic and phenological features has the best effect. In particular, considering phenological characteristics is evidently better than not considering phenological characteristics in these recognition methods. Compared with the third combination of eigenvalues, the overall accuracy and Kappa coefficient of the fourth combination of eigenvalues are improved by 4.1%, 0.73%, 1.62%, and 0.0541, 0.0096, and 0.0198, corresponding to the results of the SVM, DT, and RF algorithms, respectively. Among the three machine learning algorithms, the RF algorithm has the highest accuracy. Under the fourth eigenvalue combination, the overall accuracy and Kappa coefficient are increased to 98.52% and 0.9820, and the user’s accuracy and producer’s accuracy of chestnut forest land are increased to 97.25% and 99.46%, respectively.
Overall, the RF algorithm with spectral, vegetation, phenological, and terrain characteristics had the best performance, achieving the highest evaluation accuracy. Therefore, we selected this method to identify Chinese chestnut forests in the Miyun and Huairou Districts.
3.3. Identification Result and Accuracy Evaluation
The land cover map and the distribution map of Chinese chestnut forest land in the Huairou and Miyun areas of Beijing were obtained through the RF algorithm considering phenological and topographic characteristics. This map has a resolution of 10 m. The map is shown in
Figure 7.
By evaluating the identification results using field survey sample data, the confusion matrix for the RF model yielded the following values: TP (true positive) = 3664, FP (false positive) = 104, TN (true negative) = 8501, and FN (false negative) = 46. The overall accuracy of classification mapping is higher than 98%, and the Kappa coefficient is above 0.9. In the vegetated areas (chestnut forest, cultivated land, deciduous forest, and evergreen forest), the accuracy reaches 92.43%. Chestnut is wrongly classified into different tree species, mainly pear, walnut, hawthorn, and other fruit trees, as well as some terraced fields and cultivated land. In the non-vegetated area (including water area and construction land), the accuracy is nearly 100%.
Compared with the chestnut forest land area stated in the official statistical data (see
Table 6), the extracted area of chestnut in this study is 320.85 km
2, which accounts for 89.59% of the official statistical data’s 357.17 km
2. The distribution of chestnut in Huairou and Miyun districts is 47.56% and 52.44%, respectively, closely resembling the statistical data of 41.06% and 58.94%.
4. Discussion
Previous studies have demonstrated that the accuracy [
21] of Chinese chestnut forest identification can be effectively improved by utilizing medium- and high-resolution images. Our analysis results based on Sentinel-2 images, align with this finding and further highlight the advantages of classification mapping with Sentinel-2 images. These advantages include high spatial resolution, abundant spectral information, numerous selectable phases, and free accessibility. However, better recognition results would be achieved by separately identifying chestnut forests in flat terrain and mountainous areas.
Relying on the traditional consideration of images’ spectral and vegetation characteristics, we took into account the differences in the characteristics of vegetation in the process of growth and development. The introduction of vegetation phenological and topographic features in this study improved the quality of the chestnut forest distribution recognition map and guaranteed its recognition accuracy. It is consistent with the studies by Li et al. [
22] and Lei et al. [
23], who highlighted the importance of incorporating phenological characteristics and topographic features to enhance the classification accuracy in land cover mapping.
In order to avoid mistakes when classifying Chinese chestnut forests, early research introduced other auxiliary data to reduce the similarity and proved that machine learning algorithms are effective in improving the accuracy of ground object classification [
49]. Following the research hotspot in the field of machine learning, we selected three machine learning algorithms to identify the distribution of chestnut forests in Miyun District and Huairou District (Beijing). We found that the RF algorithm is significantly better than the SVM algorithm and the DT algorithm in terms of the producer’s accuracy, user’s accuracy, overall accuracy, and Kappa coefficient results. However, there were still some omissions or classification errors when the RF algorithm was selected and used in this study. We could identify some noise points in the classification results of chestnut forest land, so there is still room for improvement through improving the RF methods. For example, we could consider the interaction between feature indicators to enhance the accuracy.
Overall, although significant progress has been made through the integration of vegetation phenological characteristics and the RF algorithm classification, there are still certain limitations in this study. First, we did not separate the images into flat and mountainous areas for classification, which could potentially impact the accuracy of identification. Second, the inclusion of texture characteristics in image extraction for classification was not considered, thereby potentially limiting the performance improvement. Last, the interaction between feature indicators in the machine learning models was not taken into account.