High-resolution distribution map of apple orchards in China based on multisource data

Abstract

China is the world’s leading apple producer and vital to global fruit supply and food security. Apple production depends on yield and planting area, both strongly influenced by climate. Regional climate variations cause differing cultivation suitability, exacerbating production disparities. Although climate models can project impacts on apples, the lack of high-resolution distribution data limits reliable large-scale assessment. Mapping apple orchards is more challenging than field crops due to complex growing conditions. To tackle this challenge, we developed the first 30-m national-scale apple orchard distribution map of China (AOMC) for 2019–2021, covering 96% of China’s total apple orchard area. This dataset is generated based on an apple-phenology-driven algorithm framework, which integrates apple phenology and multi-source geographical information with the time-weighted dynamic time warping algorithm. Validations revealed that the dataset accurately represented orchard locations (Overall accuracy: 87%) and correlated strongly with municipal-level statistics (R²: 0.892–0.925). The AOMC supports the monitoring of spatiotemporal changes in China’s apple production patterns, serves as critical input for crop models, and advances the sustainable development and climate resilience of apple industry.

Background & Summary

Apple (Malus pumila) is a deciduous fruit crop widely distributed in temperate regions, and holds significant agricultural value as a globally prominent fruit species. China is the leading apple producer globally, with a total apple orchard area of 1.9425 million hectares and a production volume of 44.066 million tons, contributing significantly to the global fruit supply and food security. Apple growth and development are highly sensitive to climatic conditions, and production and quality are significantly affected by climate change. High-resolution spatial distribution data of fruit crops in China serve as the foundation for projecting fruit yield and quality, while also assessing food security^1–3. Alongside substantial achievements in the distribution mapping of staple crops⁴, the mapping of economic crops is now becoming a new research focus⁵. The existing CROPGRIDS dataset⁶ provides global apple harvested area data at a 10 km spatial resolution. Derived from interpolated statistical data, it is thus considered more appropriate for analyzing regional differences at broader scales⁷ rather than interpreting individual grid cells. Due to the complex and varied topography and climate in China, systematic identification of fruit cultivation areas at the national scale is constrained, with existing research largely limited to classification and identification within small regional areas. Yan et al. applied multi-temporal Sentinel-2 imagery and Random Forest classification to map apple orchards at the county level in Yantai, Shandong. Chen et al. developed an Orchard Mapping Index (OMI) based on the phenology of fruit trees and utilized Sentinel-2 imagery to map the distribution of various orchards across the Loess Plateau. However, the absence of high-resolution apple orchard spatial distributions limits the accuracy of large-scale yield and quality projections⁸, thereby impeding quantitative research on climate change’s impact on apple production⁹. Therefore, accurate and timely identification of apple orchard locations and areas is crucial for predicting apple yield and apple quality.

The integration of remote sensing interpretation with field survey samples has become an effective method for mapping crop cultivation distributions⁴. Compared with field surveys, remote sensing technologies offer more comprehensive data to monitor land use¹⁰. An increasing amount of satellite images has been extensively utilized in crop classification, with advancements in high-resolution imagery providing essential datasets to support identifying both crop and orchard lands¹¹. Numerous studies have identified planting maps for major food crops (wheat, rice, and maize) at the county, regional levels¹² and provincial scales¹³. These high-resolution crop thematic maps serve as the solid foundation for large-scale projections of crop yield and quality. Compared with research on food crops, research on orchard identification via remote sensing methods has been relatively limited, with most studies focused primarily on smaller regions, such as municipalities or provinces¹⁴. National-scale high-resolution distribution maps of apple orchards in China are unavailable, which hinders progress in related food security research⁹.

For crop and orchard mapping, both supervised and unsupervised classification methods combined with high-resolution remote sensing imagery have been extensively applied¹⁵. Machine learning dominates current classification methodologies¹⁶, which include decision trees¹⁷, random forest algorithm¹⁴, support vector machine algorithm¹⁸, and neural network classification methods¹⁹. However, these methods require great numbers of training samples at the state and national scales, and classification precision predominantly hinges on both the adequacy and reliability of field-collected reference data, which limits their performance at provincial and national levels²⁰.

Phenology-based classification algorithms are extensively utilized in crop identification²¹, functioning under the premise that individual crop species demonstrate distinct seasonal phenological patterns applicable to spatiotemporal classification²². Compared with machine learning methods, phenology-based classification methods require fewer training samples at the national scale, making them more suitable for large-scale mapping of crop and orchard distributions²³. Traditional phenology-based classification methods combine machine learning methods with phenological features from single or multiple time phases for crop classification¹⁵. Dynamic time warping (DTW) serves as a crop-cycle-driven vegetation mapping algorithm extensively employed that identifies crops by comparing the seasonal variations in vegetation indices among target crops and other classes²⁴, but accuracy may be affected by image quality and environmental factors shifting apple orchard phenology across regions, leading to misclassifications²⁴. The time-weighted dynamic time warping algorithm (TWDTW) offers an effective method for time series analysis²⁵. This algorithm incorporates time weights to address seasonal variations in crops, ensuring a balance between shape matching and phenological changes. It has shown excellent performance in identifying various crops and forests while requiring only a small amount of training data²³. Therefore, the TWDTW algorithm was employed in this study to ensure the reliability and national-scale consistency of the generated apple orchard dataset. Furthermore, the spatial distribution of apple orchards is also constrained by the regulation of policies and traditional agricultural practices, which makes the mapping process more complex. Integrating remote sensing, phenology-based methods, natural environmental factors, and socioeconomic conditions can increase the identification results accuracy.

The lack of high-resolution spatial distribution map for apple orchards limits the accuracy of large-scale yield and quality projections, hindering quantitative research on climate variability’s impact on apple production and quality at the national scale. Therefore, the objective of this study is to automatically identify apple orchards spatial distribution in China covering 2019–2021 with 30 m spatial resolution. The development of the dataset involved the following steps: (1) Screening potential areas of apple orchard spatial distribution by integrating topographic, climatic, and existing planting patterns; (2) Fusing multi-source satellite imagery to generate gap-free vegetation index composites; (3) Collecting representative samples across provinces within the potential areas identified in Step 1 to define standard phenological curves; (4) implementing the TWDTW algorithm to map orchards by quantifying the similarity between standard phenological curves and pixel-level trajectories, with classification thresholds calibrated using province-level statistical yearbook data; and (5) validating the results through field-level assessment and statistical comparison. This provides basic data for monitoring the spatial distribution and area of apple orchards, further supporting refined simulations of apple yield and quality at the national scale and accurate division of apple production regions.

Methods

Study area

Apple orchards are widely distributed throughout China. This study identified areas of apple orchards harvested from 2019 to 2021 in 13 provinces in China covering an area of 1.975 × 10⁶ ha, including Shaanxi (SAX, 31.11%), Gansu (GS, 12.47%), Shandong (SD, 12.37%), Shanxi (SX, 7.24%), Liaoning (LN, 6.99%), Hebei (HB, 6.77%), Henan (HN, 5.90%), Xinjiang (XJ, 3.48%), Yunnan (YN, 2.74%), Sichuan (SC, 2.29%), Inner Mongolia (IM, 2.08%), Ningxia (NX, 1.43%), and Jiangsu (JS, 1.43%), which are the main apple producing provinces of China (Fig. 1) (https://data.stats.gov.cn/). These selected provinces constitute more than 96% of the total harvested areas of apple orchards. Agricultural production within the study region exhibits diversity. Staple crops are dominated by wheat and maize, while rice cultivation is primarily distributed in the southwestern areas. Among orchard types, apple orchards predominate. Peach, pear, and grape orchards are also common deciduous fruit tree types. Additionally, cultivation of evergreen fruit trees, such as citrus, occurs in the southwestern provinces.

Fig. 1.

Statistical areas for apple orchard identification in China, including 13 provinces. The red dots represent sample points obtained from the Google Earth Engine.

Data collection

Statistical data

Apple orchard area data at provincial and municipal levels for 2019 to 2021 were obtained from the National Bureau of Statistics of China (NBS, https://data.stats.gov.cn/index.htm, last access: 26 July 2025). It provides information on the harvested area and production for 13 major apple cultivation provinces, which were used to identify apple orchards and evaluate the accuracy.

Satellite data

Surface reflectance from Landsat-8 Operational Land Imager (OLI)²⁶ and Sentinel-2 Multispectral Instrument (MSI) (2018–2022)²⁷ was processed to generate 16-day 30 m soil-adjusted vegetation index (SAVI) composites using the Google Earth Engine (GEE). After atmospheric correction and cloud removal, cloud-free data from both satellites were combined. Maximum value composites were generated for each 16 days, followed by histogram matching to fuse Landsat-8 and Sentinel-2 SAVI at 30 m resolution. To reconstruct high-fidelity time series, we implemented a two-step pre-processing chain: (1) Linear interpolation was applied to fill data gaps not exceeding 32 days. (2) A subsequent dynamic-threshold filter²⁸ was then applied point-wise to detect and replace outliers caused by residual atmospheric noise. This filter flags a pixel as an anomaly if its SAVI value changes by more than 0.4 (|ΔSAVI| > 0.4) relative to adjacent valid observations and replaces it via interpolation, thereby preserving the natural phenological signal. The 16 d composited SAVI cloud-free frequencies from 2019 to 2021 for each pixel are shown in Fig. 2.

Fig. 2.

Good observation frequencies for the 16-day composited SAVI images in 2020 (a). Figures (b–d) illustrate the observation frequencies by province from 2019 to 2021.

The Sentinel-1 synthetic aperture radar (SAR) products²⁹ were acquired through GEE. This product was processed to acquire the backscatter coefficient for each image and convert it into decibels (dB). The refined-Lee filter, as outlined by Abramov, et al.³⁰, was applied for speckle noise correction. In this study, the vertical transmit and horizontal receive (VH) band was composited into 16-day maximum images (2018–2022) on GEE.

Climate data

To determine the suitable climatic conditions for apple cultivation, this study applied the CN05.1 dataset published by the Laboratory for Climate Studies of the China Meteorological Administration as observational data³¹. This dataset is generated by ground-based observational data from meteorological stations and essentially covers the entire land area of China (excluding statistical data for Taiwan Province).

Land use and land cover dataset

Digital earth model (DEM), forest canopy height data, China Land Cover Dataset (CLCD), and other geographic information products were used to screen potential areas of apple orchards in China. The datasets applied are summarized in Table 1.

Table 1.

Geographic information products applied to screen potential areas of apple orchards in China.

Dataset	Spatial resolution	Year	Source
NASA SRTM1 V3.0	30 m	—	NASA46
CN05.1	0.25°	1961–2018	Wu and Gao31
NNGI_FCH_China V1.0	30 m	2019	Liu, et al.47
Annual China Land Cover Dataset	30 m	2020	Yang and Huang48
China wheat distribution map	30 m	2018	Shen, et al.35
China maize distribution map	30 m	2020	Dong, et al.20
Desert (sand) distribution data set in China	1 km	2019	Yan and Wang49
GWL_FCS30	30 m	2020	Liu, et al.50
Landsat 8 OLI	30 m	2018–2022	USGS26
Sentinel-2 MSI	10 m	2018–2022	ESA27
Sentinel-1 SAR	10 m	2018–2022	ESA29

Apple orchard classification method

This study developed a comprehensive approach to map apple orchard distribution by integrating potential cultivation areas, phenological characteristics, and satellite-based phenometry. The mapping algorithm encompassed five key components:

1. Screening potential areas of apple orchard spatial distribution. Based on the topographic and climatic requirements of apple growth as well as existing planting patterns, we combined statistical data, climate datasets, and geographic information products to identify suitable regions nationwide.

2. Pre-processing and compositing of satellite datasets. Landsat-8 OLI, Sentinel-2 MSI, and Sentinel-1 SAR images were filtered and pre-processed to remove cloud-affected pixels. The SAVI and VH were calculated using cloud-free satellite images, followed by 16-day maximum value compositing. Then the SAVI calculated from Landsat 8 OLI and Sentinel-2 MSI were histogram-matched and then fused to fill in the missing value.

3. Survey sample collection and phenological feature extraction. Representative samples of apple orchards and other land cover types were selected for each province based on field knowledge, website and high-resolution Google Earth imagery. These samples were combined with the processed satellite data to extract discriminative features and support accuracy assessment.

4. Apple orchard mapping. Within the potential areas identified in Step 1, pixel-scale SAVI time series were extracted from the 16-day composited data. The TWDTW method was employed to quantify the similarity between the standard phenological curves and each pixel’s trajectory (see Section “TWDTW algorithm” for more details). Province-level statistical yearbook data on orchard area were used to determine classification thresholds, enabling the generation of a national-scale apple orchard distribution map.

5. Accuracy assessment. The identification results were validated through two independent approaches: (i) field-level validation using a reserved set of survey samples, and (ii) comparison with apple orchard area statistics from statistical yearbooks. The complete workflow is presented in Fig. 3.

Fig. 3.

Flowchart of the apple orchard classification method.

All data processing and analysis were primarily conducted using Google Earth Engine (GEE) Code Editor, ArcGIS Desktop 10.6, and Python 3.8.15 within PyCharm Community Edition 2020.2.1.

Screening potential areas for apple orchards

The spatial distribution of apple orchards is shaped by natural environmental factors, policy constraints, and traditional farming practices, increasing the complexity of accurate mapping. To ensure that the classification results align more closely with reality, this study comprehensively considers factors such as current planting status, topography, climate, and land use when screening potential areas for apple cultivation (Table 2):

1. Existing apple areas. To ensure that the identified potential planting areas align with actual conditions, this study referred to statistical yearbooks from various provinces and cities. Counties with no apple production or where the apple harvest area constituted under 0.5% of the province’s overall cultivation area in the given province were identified separately to improve the identification accuracy⁵.

2. Topographic conditions. Based on previous studies, slope thresholds were applied to reflect topographic suitability. Apple orchards are predominantly located on slopes ≤ 25°, with optimal suitability below 15°¹⁴. Accordingly, potential planting areas were limited to slopes under 25°.

3. Climatic conditions. This study applied the apple climatic zoning indicators defined by Li, et al.³² and selected key climatic criteria to determine suitable planting areas.

4. Land use conditions. Potential areas were restricted to cropland and forestland. To comply with national policies protecting basic farmland for food security, protected zones were systematically excluded using high-resolution crop distribution data (wheat and maize). The desert (sand) and wetland distribution data are used to eliminate regions unsuitable for apple production. High-resolution vegetation canopy height datasets were employed to exclude pixels with excessively tall vegetation.

Table 2.

Screening Criteria for Potential Apple Orchard Areas.

Category	Dataset	Indicator	Metric
Topography	NASA SRTM1 V3.0	Slope14	≤25°
Existing apple area	Statistical Yearbooks	Area5	≥0.5%
Climate	CN05.1	Guarantee rate of <24 days of daily minimum temperature ≤−20 °C32	≥80%
		Mean temperature in summer32	<26 °C
		Annual precipitation32	≤950 mm
		Days of daily mean temperature ≥10 °C32	≥140 d
Land Use/Land Cover	China Land Cover Dataset	Land use type	Cropland, forestland
	China wheat distribution map	Crop type	Wheat
	China maize distribution map	Crop type	Maize
	Desert (sand) distribution dataset in China	Land use type	Desert
	GWL_FCS30	Land use type	Wetland
	NNGI_FCH_China V1.0	Canopy height	≤5 m

Regions that satisfy all four criteria were identified as potential distribution areas of apple orchards. The detailed screening criteria are summarized in Table 2.

The potential areas for apple orchards in China are screened out through the consideration of multiple criteria (Fig. 4). The potential apple cultivation areas are distributed primarily in North China and the Shaanxi–Gansu–Ningxia region, which aligns with the orchard area distribution outlined in the statistical yearbooks.

Fig. 4.

Potential areas for apple orchards in China. Figures (a–d) and (f–h) show the potential apple areas under the intersection of multiple criteria. Figure (e) shows the Venn diagram of principles for screening potential areas for apple orchards. Figure (i) shows the final apple orchard map of China (AOMC) in 2020, resampled to a 10 km grid for visualization.

Generation of standard SAVI curves

To derive the standard phenological curves for apple orchards, we collected survey samples through visual interpretation of very high-resolution Google Earth imagery from 2020. This process integrated known field locations and ancillary information. Information on locally representative, stable, and mature commercial orchards was gathered through multiple sources, including literature, agricultural maps, local government websites, and social media platforms. Within each province, samples were collected according to the spatial pattern of the main production areas, covering different ecological planting zones. Cities with large apple orchard area in the statistical yearbook were prioritized for sampling. To comprehensively reflect the province’s cultivation conditions, a suitable number of samples were also selected from other prefecture-level cities where apple cultivation is present. Reference samples (forest land, cropland and other orchards) were selected nearby to maintain consistent image conditions. Apple orchard samples were selected to represent key agronomic and topographic variations, while samples from other land cover types included other orchards, staple crops, forests, and built-up areas. To ensure representativeness, sample points were randomly and uniformly distributed within each delineated patch. Furthermore, all sample points were validated using vegetation index time series to confirm their phenological consistency. This study collected 9142 samples, comprising 6162 samples for apple orchards and 2980 samples for other land use types (Table 3 and Fig. 1). All samples were randomly divided into training (70%) and validation (30%) sets. Training samples were used for feature extraction and analysis, while validation samples were used to evaluate classification accuracy.

Table 3.

Provincial sample totals by type in 2020.

Province	Apple orchard samples	Other samples	Total samples
Shaanxi (SAX)	1008	466	1474
Shandong (SD)	940	506	1446
Henan (HN)	612	310	922
Shanxi (SX)	620	300	920
Gansu (GS)	530	210	740
Hebei (HB)	466	264	730
Liaoning (LN)	450	216	666
Ningxia (NX)	320	144	464
Yunnan (YN)	310	106	416
Sichuan (SC)	236	160	396
Inner Mongolia (IM)	210	136	346
Xinjiang (XJ)	260	62	322
Jiangsu (JS)	200	100	300

Compared to the widely used normalized difference vegetation index (NDVI), the SAVI shows superior effectiveness in the identification and classification of apple orchards. The SAVI was computed using the near-infrared band (NIR) and red band (R) from surface reflectance images as follows:

$$\mathit{SAVI}=\frac{\left(1+L\right)\left(\mathit{NIR}-R\right)}{\mathit{NIR}+R+L},$$ 1

where NIR represents near-infrared reflectance, R denotes red band reflectance, and L is a soil adjustment factor to minimize the impact of soil surface reflectance, ranging from 0 to 1. In this study, L is 0.5, which is recommended in its original paper and has been widely adopted in crop identifications³³.

The SAVI time series was derived from 16-day composited images of Landsat 8 and Sentinel-2 to represent the seasonal variation in apple orchards. Owing to variations in climate and topographical conditions, apple orchards in each province exhibit unique phenological characteristics. For each province, apple orchard training samples were used to establish representative SAVI curves. Key phenological stages for each pixel were identified based on accumulated temperature³⁴. Each SAVI curve was temporally adjusted to align all provincial training samples to a consistent phenological timeline. Given the stability of apple phenology³⁴, we applied the 2020 standard SAVI time series to identify the spatial distribution of apple orchards for 2019–2021, thereby further examining the scalability of the method³⁵.

Adjusted SAVI curves were averaged to generate standard seasonal curves for each province. Apple orchards across all 13 provinces showed similar seasonal patterns (Fig. 5). Apple trees enter the early growth stage before Day of Year (DOY) 81 with low SAVI. Throughout the flowering stage (DOY 81–129), SAVI increases rapidly as the apple trees bloom and grow. From DOY 129 to DOY 257, during the fruit set stage, the trees reach peak growth and then gradually stabilize, with the SAVI following a similar pattern. The fruit expansion and colouring stage is characterized by stable SAVI values. After DOY 257, ripening and harvest occur, followed by leaf fall and dormancy, leading to a gradual decline and eventual stabilization of SAVI.

Fig. 5.

Standard curves of SAVI for apple orchards in 13 provinces of the study area.

To differentiate apple orchards from other land cover types, we conducted a detailed comparative analysis of standard SAVI curves (Fig. 6). Apple orchards showed distinctive traits in Google Earth imagery, displaying rougher canopies than crops, smaller canopies than pear trees, and more regular spatial patterns than both pear orchards and forests. Significant phenological differences were evident between apple and other fruit trees (Fig. 6). The flowering period was optimal for identification: pear orchards flowered slightly earlier but had shorter phenological cycles, and also exhibited earlier autumn leaf drop, leading to faster SAVI decline. Notable differences were also found between apple and peach orchards. Apple orchards typically flower from March to May, with SAVI stabilizing by June, whereas peach orchard SAVI continues rising to peak in August or September. Throughout this period, apple SAVI values remained higher than those of peach. These distinct flowering phenology patterns allow reliable differentiation of apple orchards from other fruit crops.

Fig. 6.

Examples of (a) colour and textures from Google Earth images and (b) frequency histograms of apple orchards and forests by VH during the apple blooming stage. Time series (c) of SAVI for various vegetation types.

Given the minimal variation in SAVI curves for apple orchards and forestlands, distinguishing between these land cover types using SAVI alone is inherently challenging. To overcome this limitation, Sentinel-1 SAR products were utilized for improved differentiation. According to the samples, the VH backscatter value in April, corresponding to the apple flowering period, was identified as a reliable discriminant. Specifically, the VH backscatter value for apple orchards in April is lower than −15 dB, whereas it is higher for forestlands (Fig. 6(b)). This discrepancy indicates that pixels with a VH value exceeding −15 dB are unlikely to correspond to apple orchards. By assigning greater dissimilarity to such pixels, the study effectively distinguished apple orchards from forestlands.

TWDTW algorithm

To map apple orchard spatial distributions in China, this study employed the TWDTW algorithm, an enhanced version of the DTW algorithm, widely recognized for its outstanding performance in crop classification and mapping applications^36,37. The TWDTW algorithm was applied to measure the similarity between the standard SAVI time-series curves of apple orchards and the SAVI time-series curves of unclassified pixels within potential areas for apple orchards.

The principle of the TWDTW algorithm operates as follows:

First, we assume there exists a reference curve U of length n and an unclassified curve V of length m. An n × m matrix $\psi$ is constructed from U and V, where each element ${\psi }_{i,j}$ represents the absolute difference between $U_i$ ∈ U (i = 1, …, n) and $V_j$ ∈ V (j = 1, …, m), augmented with a Logistic time cost ${\mathrm{\omega }}_{i,j}$ acting as a penalty term.

$${\psi }_{i,j}=\left|U_i-V_j\right|+{\mathrm{\omega }}_{i,j,}$$ 2

$${\mathrm{\omega }}_{i,j}=\frac{1}{1+e^{\left(-\alpha (g\left(t_i,t_j\right)-\beta \right)}},$$ 3

Where $g\left(t_i,t_j\right)$ represents the temporal interval (in days) between two corresponding dates. The midpoint $\beta$ and spread $\alpha$ were determined based on Maus et al.’s²⁵ methodology and the phenological stage durations of apple trees, with values set to $\alpha$ = 0.1 and $\beta$ = 50 days.

Using matrix $\psi$, we constructed a minimum-cost cumulative matrix D through recursive summation. The elements $d_{i,j}$ of this matrix are subject to the following boundary conditions:

$$d_{i,j}=\left\{\begin{array}{cc}{\psi }_{i,j},& i=1,j=1\\ {\sum }_{k=1}^i{\psi }_{k,j},& 1<i\le n,j=1\\ {\sum }_{k=1}^j{\psi }_{i,k},& i=1,1<j\le m\end{array}\right)$$ 4

For all other elements $d_{i,j}$, starting from the top-left corner of matrix D as the origin, the computation progresses recursively by accumulating the minimum ${\psi }_{i,j}$ value from three adjacent cells: left ($d_{i,j-1}$), upward ($d_{i-1,j}$), and diagonal upper-left ($d_{i-1,j-1}$). This recursive accumulation continues until reaching the bottom-right element $d_{n,m}$ of matrix D.

$$d_{i,j}={\psi }_{i,j}+\min (d_{i-1,j},d_{i-1,j-1},d_{i,j-1}),$$ 5

The TWDTW algorithm begins at the top-left cell $d_{1,1}$ and fills the entire matrix D as described above, resulting in a total accumulated cost at $d_{n,m}$, which is the TWDTW distance for the current alignment. Then, we shift the starting position of the unclassified curve V by one element and restart the computation (i.e., rebuild the matrix D for the new segment of V aligned with the entire reference curve U). Since the algorithm allows variable curve lengths, multiple segments of V are matched against U, producing multiple warping paths and thus multiple TWDTW distances. The minimal TWDTW distance among all segments is selected as the ‘optimal warping path’ and is defined as the ‘TWDTW difference’ between the curves U and V. A smaller TWDTW difference indicates higher similarity between V and U, implying a higher probability that the pixel belongs to the target feature.

In this study, we computed the TWDTW distance between the standard apple orchard SAVI curve U and each unclassified SAVI curve V within the potential areas. In order to capture the apple phenology more flexibly, we move the standard curve of each province forward and backward (the max time window less than 5-time intervals), calculate the distance between each pixel and each translated curve, and select the smallest distance as the final matching result of the pixel³⁵. Pixels were ranked by ascending TWDTW distance, and cumulative areas were calculated. Provincial thresholds were determined using statistical orchard area data: the TWDTW distance of the terminal pixel where the cumulative area matched the statistical value became the threshold. Pixels with TWDTW distance smaller than or equal to the provincial threshold were mapped as apple orchards.

Validation methods

The identification accuracy for apple orchards was evaluated in two aspects: (1) validation at the field level via a validation set of survey samples and (2) comparison with the apple orchard area statistics from the statistical yearbooks. The confusion matrix is a common method for evaluating classification accuracy, typically presented in an n-by-n format³⁸. It provides key accuracy metrics, including overall accuracy (OA), producer’s accuracy (PA), and user’s accuracy (UA), each reflecting different aspects of classification performance. In image accuracy assessment, the confusion matrix is used to compare identification results with ground truth data, visualizing the accuracy within the matrix³⁹. To further evaluate the effectiveness of the identification process, the coefficient of determination (R²), mean absolute error (MAE), and root mean absolute error (RMAE) were applied.

Data Records

The 30 m distribution maps of apple orchards in China (AOMC) from 2019–2021 are available online at figshare⁴⁰. The dataset is provided in Geotiff (*.tif) format. Data named “AOMC_30_<province>_<year>.tif” means apple orchard map of China for 2019–2021. The dataset is provided in the “Asia_North_Albers_Equal_Area_Conic” spatial reference system, and the spatial extent of the dataset ranges from 18°26′ N to 50°01′ N and 53°22′ E to 128°17′ E. The pixel values in the dataset are: (1: apple orchard; 0: non-apple orchard). These data can be opened and applied in geographical software and other statistical applications.

Technical Validation

To validate the reliability, the apple orchard spatial distribution was first identified at the provincial scale using the previously discussed methods and then validated with survey samples and statistical data. The identification accuracy for apple orchards was evaluated using two methods: (1) field-level validation with a validation set of survey samples and (2) comparison with apple orchard area data from yearbooks.

Spatial distribution consistency

Field-level validation was conducted on the validation set of survey samples from the 2020 ground-truth classification data. For all 13 provinces in this study, the OA varied from 80.27% to 95.65%, the UA for apple orchards varied between 88.06% and 97.16%, and the PA ranged from 73.60% to 97.17% (Table 4).

Table 4.

Accuracy assessment matrices for the apple orchard distribution map across 13 provinces in 2020.

Province	Reference	Sample Size (n)	Map		PA (%)	UA (%)	OA (%)
			Apple	Non-apple
Shaanxi	Apple	308	273	35	88.64	96.13	90.04
	Non-apple	154	11	143	92.86	80.34
Shandong	Apple	293	256	37	87.37	96.24	89.37
	Non-apple	149	10	139	93.29	78.98
Shanxi	Apple	211	166	45	78.67	95.40	83.07
	Non-apple	102	8	94	92.16	67.63
Henan	Apple	197	145	52	73.60	96.03	80.27
	Non-apple	97	6	91	93.81	63.64
Hebei	Apple	144	118	26	81.94	88.06	81.17
	Non-apple	79	16	63	79.75	70.79
Liaoning	Apple	141	108	33	76.60	94.74	80.50
	Non-apple	59	6	53	89.83	61.63
Xinjiang	Apple	87	70	17	80.46	100.00	83.50
	Non-apple	16	0	16	100.00	48.48
Gansu	Apple	141	137	4	97.16	96.48	95.65
	Non-apple	66	5	61	92.42	93.85
Ningxia	Apple	97	90	7	92.78	97.83	93.75
	Non-apple	47	2	45	95.74	86.54
Sichuan	Apple	69	62	7	89.86	96.88	92.31
	Non-apple	48	2	46	95.83	86.79
Yunnan	Apple	93	83	10	89.25	96.51	89.68
	Non-apple	33	3	30	90.91	75.00
Inner Mongolia	Apple	56	45	11	80.36	91.84	84.21
	Non-apple	39	4	35	89.74	76.09
Jiangsu	Apple	76	69	7	90.79	94.52	89.52
	Non-apple	29	4	25	86.21	78.12

The performance of the proposed method varies by province. Shaanxi Province, which accounts for more than 32% of China’s apple orchard area, achieved a UA of 96.13%, a PA of 88.64%, and an OA of 90.04%. The OA in all surveyed provinces exceeded 80% (Table 4). The validation outcomes showed that the proposed algorithm effectively identified the distribution of apple orchard samples across China in 2020. Therefore, the proposed method was employed at the provincial scale. Validation from 2020 revealed consistent performance across the 13 major apple-producing provinces in China, with an overall accuracy of 87.00%.

Statistical comparison

Compared with agricultural statistics from various provinces from 2019–2021, the method provided consistent apple orchard area estimates for China. The estimated apple orchard area for 2020 demonstrated strong agreement with the agricultural statistical data obtained at the municipal level from statistical yearbooks (Fig. 7), with an R² of 0.912 (N = 179), a corresponding MAE of 3.399 Kha, and an RMAE of 31.7%. Similarly, the identified apple orchard areas for 2019 (R² = 0.925) and 2021 (R² = 0.892) also strongly agreed with the municipal-level agricultural statistics, respectively.

Fig. 7.

Comparison of agricultural statistics and identified apple orchard areas across municipalities in China (2019–2021) (a–c). The dashed line represents the 1:1 reference, while the red line indicates regression. N is the number of municipalities in the study area, and the unit of MAE is 1000 ha.

The correlation between the statistics data and identified harvest areas varied by province (Fig. 8). At the municipal level, R² values across provinces were between 0.770 and 0.968, while RMAE ranged from 13.1% to 48.6%. Validation in all provinces showed strong performance, with R² values and slopes near 1 (Figs. 7, 8). The R² values for all provinces exceeded 0.77, with an average R² of 0.912. Shandong and Shanxi provinces had higher R² values (0.969 and 0.955) and lower RMAE values (0.282 and 0.272), and provinces in northern China displayed higher R² values and lower RMAE values than southern provinces did. Slopes of the regression equations comparing identified and statistical data exceeded 0.76 (0.76–1.28). In general, this method can capture the distribution of apple orchards in China well and performs better in northern China than in southern China, which is related to the quality of the images. The standard SAVI time series acquired from a single year (2020) was applied to other years (2019 and 2021) to evaluate the capability of this method. The results revealed that R², slope, and RMAE exhibited little difference in most states from 2019–2021 (Fig. 9).

Fig. 8.

Comparison between the agricultural statistical area and the estimated area of apple orchards at the municipality level for each province (a–m) and China (n) for 2020. The black dot represents the municipal-level apple orchards area summed from pixel level. The dashed line denotes the 1:1 line, and the red line denotes the regression line. The shaded areas represent the 95% confidence intervals (CIs). The unit of MAE is 1,000 ha.

Fig. 9.

Comparison of agricultural statistics and identified apple orchard areas (RMAE, R², and slope) at the municipal level for each province from 2019 to 2021 (a–c), with violin plots displaying annual distributions (d–f).

The validation results demonstrate that the proposed method is highly effective in accurately delineating the spatial distribution of apple orchards, thereby affirming its reliability and robustness.

Spatial distribution

This study produced the first national-scale distribution map of apple orchards in China (AOMC) at a 30 m spatial resolution using the TWDTW algorithm, covering 13 provinces for the period 2019–2021 (Fig. 10). The validation results showed that the proposed identification method can effectively identify the distribution of apple orchards across China in 2020. Apple orchards in China are located mainly in North China, Northeast China, the Shaanxi–Gansu–Ningxia region, and Xinjiang Province. Representative areas across various provinces are selected to illustrate the detailed information captured by the generated apple orchard maps, including Xinjiang (XJ, b1), Liaoning (LN, b2), Shandong (SD, b3), Shaanxi (SAX, b4) and Gansu (GS, b5) Provinces (Fig. 10). SAX, GS and SD are the top three provinces in terms of apple orchard area in China, which accounting for 55.95% of the apple orchard area.

Fig. 10.

Final identification map of apple orchards in China in 2020 (a). The left column in Figures (c1-c5) shows zoomed-in high-resolution images of the orange dots in representative areas (b1-b5), and the right column shows the presence (green) and absence (yellow) of the identified apple orchard map (a). The orange lines in Figures (c1-c5) are the boundaries of real apple orchards.

These representative areas across various provinces were selected to illustrate the detailed information captured by the generated apple orchard maps, and the apple orchard map was compared with high-resolution Google Earth images (Fig. 10). The maps captured the spatial distribution of apple orchards and included detailed features like clear boundaries. SAX is the largest apple production province in China, with apple orchards mainly concentrated in the Weibei dry plateau area, particularly in Yanan, Xianyang, and Weinan. The terrain in these orchard areas is characterized by mountains and dry plateaus. On plateaus, apple orchards are widely and continuously distributed, with well-defined and structured boundaries. Similar characteristics are also observed in apple orchards in southeastern GS. SD and LN are major apple production regions in northern China, with orchards primarily distributed across hilly and coastal areas, particularly on the Jiaodong Peninsula and Liaodong Peninsula. In hilly areas, due to the influence of complex terrain, apple orchards tend to be spatially fragmented and scattered, with small orchard sizes and indistinct boundaries. Aksu is one of XJ’s renowned apple cultivation areas and is located on the northern edge of the Tarim Basin and the southern slopes of the Tianshan Mountains. The apple orchards in Aksu are distributed primarily across oases and alluvial plains and are characterized by relatively flat terrain. These orchards are discreetly distributed, with well-defined boundaries and clear structural organization.

To further validate the map’s performance, multiple spatial comparison methods were employed (Fig. 10). In Aksu, XJ, and Yantai, SD, comparisons between the identification results and sample points demonstrated the map’s satisfactory performance. Representative apple orchards from different provinces were also selected for validation. In Dalian, LN, the Tuofeng Apple Production Park was chosen. Owing to the hilly terrain in this region, apple orchards are distributed over large areas but are fragmented, with unclear boundaries. In Yan’an, SAX, the Luochuan Apple Experimental Station of Northwest Agriculture and Forestry University, and the Touzhai Production Base of the Ganfu Fruit Industry in Baiyin, GS were selected. These two apple orchards exhibit clear boundaries and noticeable texture differences compared with the surrounding environment in Google Earth imagery. Further comparisons of the extracted results with Google Earth imagery confirmed the map’s satisfactory performance.

Comparison with CROPGRIDS datasets

The AOMC was resampled to a 10 km resolution and compared with the CROPGRIDS dataset (Fig. 11). The results show that the distribution map from this study aligns closely with CROPGRIDS in terms of spatial coverage, effectively capturing the locations of major apple-producing regions and revealing more detailed spatial distribution characteristics.

Fig. 11.

Comparisons between the (a) 10 km CROPGRIDS datasets and the (b) apple orchard map of China (AOMC) for 2020.

Although CROPGRIDS has derived the global apple harvest area at a 10 km resolution based on interpolated statistical data. However, this method does not account for the actual growing conditions of apples, and the data merely represent the average harvest proportion per grid unit from 2000 to 2020, with a relatively low spatial resolution. Therefore, it is more suitable for analysing regional differences at larger scales rather than explaining individual grid units⁷.The AOMC generated in this study exhibits more realistic distribution patterns of apple orchards in China owing to the consideration of the suitable environment conditions for apple growth.

Uncertainties and prospects

While our method can well identify apple harvest areas in China, some uncertainties still exist.

First, due to the differences in sensor parameters and observation limitations, the quality of remote sensing images is uneven. Although the Sentinel-2 data of the same period is selected for image fusion and filling, there still exist some potential errors. In addition, the SAVI dataset has a time interval of 16 days, which may introduce bias when accurately capturing the phenological characteristics of apple orchards.

Second, to differentiate apple orchards from other orchards and forested areas, this study analysed standardized SAVI curves across vegetation types (Fig. 4). Phenological characteristics enabled discrimination among crops; however, class separability remains significantly compromised by spectral resemblance between apple and pear orchards during critical phenological phases. Notably, this similarity persists even though pears demonstrate an abbreviated phenological cycle with earlier phenophase onset compared to apples. This finding aligns with Chen, et al.⁴¹ empirical research in Linyi, Shandong, which documented spectral confusion between these phenologically analogous temperate pome crops in medium-resolution imagery. Future enhancements could integrate higher spatiotemporal resolution data with: (1) multidimensional spectral-temporal metrics, (2) red-edge band analysis¹⁴, and (3) spatial texture features⁴². The development of AI-driven spectral-phenological-topographic decoupling models shows particular promise for improving inter-crop discrimination^43,44.

Third, the identification accuracy for Xinjiang was lower than that for other provinces. Apple cultivation in Xinjiang is mainly concentrated plantings in oasis cores (Aksu⁴⁵, Xinjiang) alongside scattered, small-scale orchards in other regions. This disparity results in a provincial phenological curve that is less effective in characterizing marginal orchards, especially those in counties contributing under 0.5% to the provincial apple area. This can introduce local confusion between small orchards and other sparse vegetation in similar settings. Since Xinjiang contributes only about 3.48% to the national apple harvested area (Fig. 1) and the method performs reliably in its core production regions such as Aksu (Fig. 10(c1)), the impact of this provincial uncertainty on the national-scale dataset can be regarded as limited.

Future studies could incorporate object-based recognition approaches alongside satellite imagery with enhanced spatial and temporal resolutions to potentially achieve more precise mapping outcomes. Additionally, classification accuracy may be influenced by the reliability of meteorological datasets, agricultural statistics, and land-use products. Therefore, as high-precision basic farmland datasets and updated land-use inventories become increasingly accessible, they would likely provide substantial support for improving orchard identification accuracy. The typical uncertainties discussed above and their corresponding potential improvements are summarized in Table 5.

Table 5.

Uncertainties and potential improvements.

Uncertainty source	Primary Cause	Future improvements
Data Source & pre-processing	Noise and spectral distortion arising from the fusion of multi-source imagery. The 16-day composite cycle may miss rapid phenological changes.	Employ higher spatiotemporal resolution imagery and advanced fusion algorithms.
Phenological confusion	Similarity in phenological curves between apple and pear trees.	Integrate multi source classification features with machine learning14 and explore biochemical markers using hyperspectral data51.
Data quality & availability	Data quality of data used to screen potential planting areas.	Adopt higher-precision thematic data (basic farmland) and alternative datasets.
Salt-and-pepper noise effect	Pixel-based analysis of SAVI series leads to salt-and-pepper noise and fragmented outputs.	Improve patch integrity by adopting object-based analysis (OBIA) over pixel-based methods36.

Usage Notes

China is the leading apple producer globally, accounting for about 47% of global apple production. Using available Landsat-8 and Sentinel datasets, along with multisource data, including geographic information, remote sensing products, and meteorological data, we integrated various phenological conditions of apples in China to generate the first 30 m national-scale apple orchard distribution map in China (AOMC). This study mapped the distribution of apple orchards across 13 provinces in China (2019–2021), covering 96% of the nation’s apple cultivation area. The method effectively identifies apple orchard distribution areas with minimal training sample data. Validation results based on 2,980 samples obtained from Google Earth indicated that the UA, PA, and OA in 2020 across the 13 major apple-producing provinces in China were 84.79%, 95.47%, and 87.00%, respectively. The distribution map represents the locations of apple orchards within these 13 provinces. These maps help monitor the distribution and extent of apple orchards, providing essential data for national-scale, refined simulations of apple production.

Author contributions

Wenqiang Xie: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Software, Validation, Visualization, Writing - original draft, Writing - review and editing. Yunkun Song: Formal analysis, Investigation, Writing, review, and editing. Xuefeng Cui: Formal analysis, Funding acquisition, Supervision, Writing - review and editing. Xiaodong Yan: Conceptualization, Formal analysis, Project administration, Supervision, Writing - review and editing. Shuaifeng Song: Writing - review and editing.

Data availability

The 30 m distribution maps of apple orchards in China (AOMC) from 2019–2021 are available online at figshare⁴⁰.

Code availability

The code is available at: https://code.earthengine.google.com/ec4c709d309507b3d168616abab14419?noload=true.

Competing interests

The authors declare no competing interests.