Smallholder Crop Type Mapping and Rotation Monitoring in Mountainous Areas with Sentinel-1/2 Imagery

: Accurate and timely crop type mapping and rotation monitoring play a critical role in crop yield estimation, soil management, and food supplies. To date, to our knowledge, accurate mapping of crop types remains challenging due to the intra-class variability of crops and labyrinthine natural conditions. The challenge is further complicated for smallholder farming systems in mountainous areas where ﬁeld sizes are small and crop types are very diverse. This bottleneck issue makes it difﬁcult and sometimes impossible to use remote sensing in monitoring crop rotation, a desired and required farm management policy in parts of China. This study integrated Sentinel-1 and Sentinel-2 images for crop type mapping and rotation monitoring in Inner Mongolia, China, with an extensive ﬁeld-based survey dataset. We accomplished this work on the Google Earth Engine (GEE) platform. The results indicated that most crop types were mapped fairly accurately with an F1-score around 0.9 and a clear separation of crop types from one another. Sentinel-1 polarization achieved a better performance in wheat and rapeseed classiﬁcation among different feature combinations, and Sentinel-2 spectral bands exhibited superiority in soybean and corn identiﬁcation. Using the accurate crop type classiﬁcation results, we identiﬁed crop ﬁelds, changed or unchanged, from 2017 to 2018. These ﬁndings suggest that the combination of Sentinel-1 and Sentinel-2 proved effective in crop type mapping and crop rotation monitoring of smallholder farms in labyrinthine mountain areas, allowing practical monitoring of crop rotations. types even in a complicated crop cultivation, smallholder farming system. Sentinel-1 is recommended for wheat and rapeseed classiﬁcation and Sentinel-2 for soybean and rapeseed classiﬁcation, particularly in mountainous areas. The crop rotations were monitored based on the reliable and high spatial resolution crop type maps. This study demonstrated the potential of combining Sentinel-1 and Sentinel-2 in crop type classiﬁcation and provided a reference for future crop rotation monitoring.


Introduction
Since the advent of satellite remote sensing, land cover classification has been an essential and active topic in land use science and agriculture [1,2]. With the increasing earth observation capabilities, many regional or global land cover products have been produced, such as GLC2000 [3], GlobeLand30 [4], and FROM-GLC10/30 [1,5]. These products contain cropland class but lack specific information on crop types [6,7]. In agricultural applications, accurate and timely crop type classification is vital for global food security, farming policymaking, and international food trading [8]. It is also a prerequisite for crop rotation, an emerging farming policy in China to prevent long-term land degradation [9].
Accurate crop type mapping is challenging due to spectral similarities and intra-class variability induced by crop diversity, environmental conditions, and farm practices [9][10][11]. Numerous satellites with various temporal and spatial resolutions have been launched in the past decades, such as Landsat, Sentinel, and Moderate Resolution Imaging Spectroradiometer (MODIS). These satellite images could be categorized into low, medium, high, and very high spatial resolutions according to their pixel sizes in the range of large than 100 m, 10-100 m, 1-10 m, and less than 1 m, respectively [12,13]. The main crop types in the study area include wheat, rapeseed, soybean, and corn [57][58][59]. Based on multiple-year field surveys ( Figure 4) and historical studies [58,59], we found that wheat and rapeseed are mainly distributed in AEZ A, while soybean and corn are mostly in AEZ B. Limited rice fields could also be found in AEZ B. Therefore, in AEZ A, we focused on wheat and rapeseed crops, and in AZE B we focused on soybean, corn, and rice. From a management perspective, the local government has installed a crop rotation policy to improve crop production and land degradation prevention, which requires farmers to plant different crops each year [60]. Some farmers comply with the policy, but some do not because of various management and economic reasons [61].
The crop phenologies of crops vary across the study area ( Figure 2). Wheat and rapeseed are sown simultaneously (about April 15), but wheat's ripening and harvest stages are about two weeks earlier (August 15 for wheat and September 1 for rapeseed). Although corn, soybean, and rice are harvested in early October, corn is sown two weeks The elevation of the study area ranges from 150 to 1800 m along with the Greater Khingan Range (Figure 1c), and the average elevation in AEZ A is significantly higher than that in AEZ B, with the descending trend from the center. Most crops are distributed at the elevation of~600 m in AEZ A and~300 m in AEZ B, but their slopes are the same (<15 • ). The study area is characterized by long cold winter and short-cool summer. The annual average temperature is below 0°C, and the yearly cumulative temperature in AEZ A is about 500°C and little more in AEZ B. Impacted by the terrain, the precipitation in the two zones is also different: about 300-400 mm in AEZ A and more than 500 mm in AEZ B. According to the field boundary dataset from the local agricultural department, the crop fields of less than 16 ha account for approximately 90% of the total. In addition, more than Remote Sens. 2022, 14, 566 4 of 18 50% of the crop fields in the study area are smaller than 2.56 ha, which is the threshold of small farmland defined by Lesiv et al. [56] in their global field size studies.
The main crop types in the study area include wheat, rapeseed, soybean, and corn [57][58][59]. Based on multiple-year field surveys ( Figure 4) and historical studies [58,59], we found that wheat and rapeseed are mainly distributed in AEZ A, while soybean and corn are mostly in AEZ B. Limited rice fields could also be found in AEZ B. Therefore, in AEZ A, we focused on wheat and rapeseed crops, and in AZE B we focused on soybean, corn, and rice. From a management perspective, the local government has installed a crop rotation policy to improve crop production and land degradation prevention, which requires farmers to plant different crops each year [60]. Some farmers comply with the policy, but some do not because of various management and economic reasons [61].
The crop phenologies of crops vary across the study area ( Figure 2). Wheat and rapeseed are sown simultaneously (about April 15), but wheat's ripening and harvest stages are about two weeks earlier (August 15 for wheat and September 1 for rapeseed). Although corn, soybean, and rice are harvested in early October, corn is sown two weeks before soybean and rice (around May 5 for corn, May 20 for soybean and rice) [58,59]. In addition, wheat and rapeseed are grown a month earlier than are soybean and rice. The main crop types in the study area include wheat, rapeseed, soybean, and corn [57][58][59]. Based on multiple-year field surveys ( Figure 4) and historical studies [58,59], we found that wheat and rapeseed are mainly distributed in AEZ A, while soybean and corn are mostly in AEZ B. Limited rice fields could also be found in AEZ B. Therefore, in AEZ A, we focused on wheat and rapeseed crops, and in AZE B we focused on soybean, corn, and rice. From a management perspective, the local government has installed a crop rotation policy to improve crop production and land degradation prevention, which requires farmers to plant different crops each year [60]. Some farmers comply with the policy, but some do not because of various management and economic reasons [61].
The crop phenologies of crops vary across the study area ( Figure 2). Wheat and rapeseed are sown simultaneously (about April 15), but wheat's ripening and harvest stages are about two weeks earlier (August 15 for wheat and September 1 for rapeseed). Although corn, soybean, and rice are harvested in early October, corn is sown two weeks before soybean and rice (around May 5 for corn, May 20 for soybean and rice) [58,59]. In addition, wheat and rapeseed are grown a month earlier than are soybean and rice.

Sentinel-1 Imagery and Preprocessing
Sentinel-1 is a satellite with a C-band synthetic aperture radar (SAR)that provided a frequent coverage over the study area ( Figure 3 and Table 1). The Ground Range Detected (GRD) product was used in this study, with 10 m spatial resolution and Interferometric Wide Swath (IW) mode with two polarizations, vertical transmit/vertical receive (VV) and vertical transmit/horizontal receive (VH) in GEE ("COPERNICUS/S1_GRD"). Preprocessing algorithms were applied to the Sentinel-1 imagery available within GEE, including orbital file correction, GRD border noise removal, thermal noise removal, radiometric calibration, and terrain correction (https://developers.google.com/earth-engine/guides/sentinel1, accessed on: 16 December 2021). A total of 407 Sentinel-1 images in 2017 and 398 in 2018 (Figure 3a,b, Table 1) were acquired during crop growing seasons (April 15 to August 15 for AEZ A and May 20 to October 1 for AEZ B). Each year, a median composite algorithm was applied to all Sentinel-1 imagery to create a single mosaic image for the crop type classification [62] for the entire study area. . Each year, a median composite algorithm was applied to all Sentinel-1 imagery to create a single mosaic image for the crop type classification [62] for the entire study area.
Corresponding to the time window of Sentinel-1 images collected, we acquired the Sentinel-2 images in the GEE environment (Figure 3c,d). To obtain Sentinel-2 cloud-free imagery, we only chose images with less than 20% cloud cover. Then, the quality assessment band (QA60) containing the opaque and cirrus clouds information was applied to remove contaminated pixels [62]. The median function could eliminate the noise and abnormal pixels, so we generated a single high-quality Sentinel-2 imagery with a median composite algorithm. In addition to the original spectral bands, some vegetation indices were calculated to increase the variability of different crop types according to previous studies [24,57] (Table 2).

Cropland Layer Mask
In the study area, the proportion of cropland is much less than that of the forest, grassland, and other land cover types. To avoid the interference of non-cropland cover types, we first masked out the non-cropland cover types before implementing crop type classification. The cropland extent product (https://search.earthdata.nasa.gov/, accessed on: 16 December 2021), Global Food Security-support Analysis Data at 30 m (GFSAD30), developed by Teluguntla et al. [41], was used to mask out croplands. The product obtained more than 94% overall accuracy over China with an independent ground truth data validation. Although it was produced by the Landsat-7/8 imagery around 2015, it could also be used for cropland masks in 2017 and 2018 as the change of cropland is negligible within several years [44,45] in the study area.

Crop Type Reference Data
We collected ground truth data during the crop growing seasons in 2017 and 2018. The location of each crop field was recorded with a geographic positioning system (GPS), and the corresponding field information was recorded manually. The field samples collected were mainly crops. Other land cover types such as forests, built-up, and water bodies were not gathered as they were not the focus of this study. We acquired a total of 2412 sample points in 2017 (soybean: 666, corn: 703, rice: 45, wheat: 534, rapeseed: 167, others: 297) and 2487 sample points in 2018 (soybean: 396, corn: 755, rice: 38, wheat: 368, rapeseed: 434, others: 496). The spatial distribution of sample points is shown in Figure 4. . The spatial distribution of sample points is shown in Figure 4. As field surveys in large areas are time-consuming and labor-intensive, not all samples collected are high-quality. Their quality was affected by two factors: (a) some of these samples were on the crop field edges rather than the interior of crop fields; (b) the spatial distribution of these samples was not well-proportioned. The uneven distribution in geographic space would cause potential errors in spatial representation and ultimately affect the classification accuracy [70]. To overcome these issues, we re-generated regions of interest (ROIs) by superimposing the collected field samples onto the high-resolution imagery (HRI) (e.g., Google Earth, Planet, or RapidEye) [46]. The high spatial resolution imagery was used to determine the extent, and field samples were used to label the ROIs as new ground-truthing data. The number of selected ROIs for each crop type is illustrated in Table 3. Although the number of ROIs was fewer than that of field samples, many more pixels were included in ROIs and thus provided sufficient and representative training data for crop type mapping. In the classification, 70% of the reference data were used for model training, and the remaining 30% was used for accuracy assessment.   As field surveys in large areas are time-consuming and labor-intensive, not all samples collected are high-quality. Their quality was affected by two factors: (a) some of these samples were on the crop field edges rather than the interior of crop fields; (b) the spatial Remote Sens. 2022, 14, 566 7 of 18 distribution of these samples was not well-proportioned. The uneven distribution in geographic space would cause potential errors in spatial representation and ultimately affect the classification accuracy [70]. To overcome these issues, we re-generated regions of interest (ROIs) by superimposing the collected field samples onto the high-resolution imagery (HRI) (e.g., Google Earth, Planet, or RapidEye) [46]. The high spatial resolution imagery was used to determine the extent, and field samples were used to label the ROIs as new ground-truthing data. The number of selected ROIs for each crop type is illustrated in Table 3. Although the number of ROIs was fewer than that of field samples, many more pixels were included in ROIs and thus provided sufficient and representative training data for crop type mapping. In the classification, 70% of the reference data were used for model training, and the remaining 30% was used for accuracy assessment.

Crop Type Classification
We chose the random forest (RF) classifier for crop type mapping. In the RF algorithm, the number of trees (n_tree) and the number of selected variables (n_var) are the two adjustable parameters that affect the performance of crop type mapping. However, many studies found that the classification performance was insensitive to these parameters. In this study, we first tested different parameters' classification performances (50, 100, 200, . . . ).
Although parameters vary, they demonstrated subtle differences in classification performance. Following the previous research [2,42], the n_tree was set to 100, and n_var was assigned to the square root of the number of variables, avoiding menial tasks to improve the classification efficiency.
To explore the most optimal features for crop type classification, we categorized the features derived from Sentinel-1 and Sentinel-2 into three groups (Table 4). We assessed the classification performance of each group and the combination of different groups. The RF classifier combined with Sentinel-1/2 images and training data was applied to the crop growing seasons of 2017 and 2018, respectively. If there was a crop type change from 2017 to 2018, we consider that crop rotation occurred. In general, crop rotation occurs if the crop type differs from one year to another.

Accuracy Assessment
As mentioned earlier, 30% of reference data was used to evaluate the classification accuracy of crop types. The four metrics, overall accuracy (OA), user's accuracy (UA), producer's accuracy (PA), and F1-score (Equation (1)), derived from the confusion matrix were applied.
where F1 − score i represents the F1-Score of class i. PA i and UA i represent the PA and UA of class i. Figure 5 demonstrates the differences among Sentinel-2 spectral bands, vegetation indices, and Sentinel-1 polarizations in crop type classification. As shown in Figure 5a, there are no significant differences in visible and NIR bands between wheat and rapeseed, but the reflectance of wheat is less than that of rapeseed in the SWIR band. In addition, the intra-class variability of wheat is larger than that of rapeseed in both NIR and SWIR bands. In contrast, wheat and rapeseed exhibit noticeable differences in VV and VH polarizations (Figure 5c), demonstrating an excellent potential for wheat and rapeseed mapping with Sentinel-1 satellite imagery.   Compared with wheat and rapeseed, soybean and corn were shown to have an opposite performance in feature discrimination. While soybean has higher NIR and SWIR values, there were no noticeable differences in VIs and Sentinel-1 polarizations. Soybean Remote Sens. 2022, 14, 566 9 of 18 and corn have higher values than wheat and rapeseed in all features except visible and SWIR bands. There was no significant difference in vegetation indices between wheat and rapeseed, and between soybean and corn (Figure 5b). It seems then that optical imagery has better performance in soybean and corn discrimination, and SAR features have better potential in wheat and rapeseed identification.  Combining all features can achieve a high OA (82.43% in 2017, 86.98% in 2018), but the combination of Sentinel-2 spectral and Sentinel-1 polarizations makes a better performance for both years (85.97% and 88.07%). This combination also achieved a high OA (92.91% and 89.36%) in AEZ B. The high accuracy (92.99% and 88.71%) obtained by the Sentinel-2 bands alone suggests spectral bands derived from optical satellites play an essential role in soybean and rapeseed (crops distributed in AEZ A) mapping. The mapping of soybean and corn (crops distributed in AEZ B) was more accurate than the wheat and rapeseed due to the higher classification accuracy demonstrated in AEZ B.

Crop Type Distribution and Rotation
The 10 m crop type maps and spatial details produced with the Sentinel-1 and Sentinel-2 imagery in the study areas in 2017 and 2018 are illustrated in Figure 8, and areas of crop types are demonstrated in Figure 9 Although the crop planting areas in the study area are small and scattered, numerous crop fields are concentrated on the right side of the study area. Further, some crop fields are distributed in the flanks of the ridge. Consistent with the distribution of field survey samples (Figure 4), the wheat and rapeseed were mainly distributed in AEZ A in both 2017 and 2018. Soybean, corn, and rice were distributed primarily in AEZ B (Figure 8a,d), suggesting the necessity and reasonability of zonal classification (Figure 1c). From close-up views benefiting from the high spatial resolution of Sentinel-1/2 imagery, the classification results exhibit the precise

Crop Type Distribution and Rotation
The 10 m crop type maps and spatial details produced with the Sentinel-1 and Sentinel-2 imagery in the study areas in 2017 and 2018 are illustrated in Figure 8, and areas of crop types are demonstrated in Figure 9 Although the crop planting areas in the study area are small and scattered, numerous crop fields are concentrated on the right side of the study area. Further, some crop fields are distributed in the flanks of the ridge. Consistent with the distribution of field survey samples (Figure 4), the wheat and rapeseed were mainly distributed in AEZ A in both 2017 and 2018. Soybean, corn, and rice were distributed primarily in AEZ B (Figure 8a,d), suggesting the necessity and reasonability of zonal classification (Figure 1c). From close-up views benefiting from the high spatial resolution of Sentinel-1/2 imagery, the classification results exhibit the precise boundaries of crop fields (Figure 8b,c,e,f). Although a pixel-based approach was applied in this study, the classification results were similar to those from the object-based method. Each crop demonstrated a clear boundary with other crops. In addition, salt-and-pepper noise of non-crop types, which often seriously affects the classification performance in pixel-based methods, is minor in this study with the operation of non-cropland layer mask and reliable classification approaches. boundaries of crop fields (Figure 8b,c,e,f). Although a pixel-based approach was applied in this study, the classification results were similar to those from the object-based method. Each crop demonstrated a clear boundary with other crops. In addition, salt-and-pepper noise of non-crop types, which often seriously affects the classification performance in pixel-based methods, is minor in this study with the operation of non-cropland layer mask and reliable classification approaches.    (Figures 8c,f and 10f). Furthermore, except for a few crop fields that re unchanged across years, the crop rotation exhibited in Figure 10 was widespread in AEZ A and B (Figure 11). These classification results showed that the crop types were not always consistent in the same crop field across different years, suggesting that crop rotation is exercised in the study area. The crop type rotation areas and spatial details are shown in Figures 10 and 11. While the rice planting areas are consistent between 2017 and 2018 (Figure 8c,f), the rotation pattern mainly exists in wheat-rapeseed (Figures 8b,e and 10e) and soybean-corn (Figures 8c,f and 10f). Furthermore, except for a few crop fields that remain unchanged across years, the crop rotation exhibited in Figure 10 was widespread in both AEZ A and B ( Figure 11).

Discussion
This study demonstrated that the accuracy of crop type mapping was mainly affected by the frequency and quality of remotely sensed imagery, number and representativeness of training data and classification features [71][72][73]. As shown in previous research, this

Discussion
This study demonstrated that the accuracy of crop type mapping was mainly affected by the frequency and quality of remotely sensed imagery, number and representativeness of training data and classification features [71][72][73]. As shown in previous research, this study demonstrated that multi-sensor compositing enhanced data availability and improved crop type classification accuracy [57,74]. The combination of Sentinel-1 and Sentinel-2 images provided more observations (Figure 3) and, more importantly, distinguishable features for the classes with similar spectral properties ( Figure 5). Although compositing Landsat (Landsat-liked) and Sentinel-2 could also improve data availability, the 30 m spatial resolution is inadequate in most developing countries with smallholder farming systems [22,23]. LiDAR data or high spatial resolution images were also recommended to improve crop type mapping in smallholder regions [75,76], but the data scarcity limited its developments. In addition, this study has revealed that the freely available images of Sentinel-1 and Sentinel-2 could achieve high accuracy for smallholder crop type mapping over mountainous areas.
Sentinel-1 polarizations had better performance than Sentinel-2 bands in wheat and rapeseed classification, which may result from the flowering stage of rapeseed. The flowering stage was frequently used for crop type classification as apparent differences in spectral reflectance exist over this growth period for many crops [77,78]. Wheat and rapeseed have distinctive spectral features during the flowering stage. While the rapeseed flowers are yellow, wheat is inconspicuous [63], and the differences could also be found in the green and red bands. Compared with Sentinel-2 bands, the differences exhibited in Sentinel-1 polarizations are more prominent. Using these features, Dong et al. (2021) directly levered a robust threshold of Sentinel-1 VH polarization to distinguish rapeseed from wheat in winter wheat mapping across entire China [30]. Mercier et al. (2020) applied the Sentinel-1 polarizations to predict wheat and rapeseed phenological stages [79].
The canopy structures and biophysical properties of soybean and corn are utterly different. Since previous studies have demonstrated the sensitivity of SAR image to crop canopy structures and biophysical parameters [80,81], we assumed that Sentinel-1 polarizations have better performance than Sentinel-2 spectral bands in soybean and corn discrimination. However, the differences between soybean and corn in VV and VH polarizations were not significant, similar to the finding by other studies [24] in the same regions of China. A similar feature for VV and VH polarizations existed for the growing season. It may be because the study area is in mountainous areas, and differences in topography dwarf the differences in crop height. Among Sentinel-2 features, visible bands also exhibited a similar performance because visible bands had notably strong scattering by the atmosphere [82]. Compared with visible bands of Sentinel-2, SWIR bands demonstrated a better distinction between soybean and corn [8,83]. The effect of red-edge bands in crop classification has been shown in many crop type classification research [25,57].
The consistent classification of crop type in 2017 and 2018 revealed the significant inter-annual changes in crop distribution ( Figure 10). It is essential for developing countries like China to balance environmental sustainability, agriculture, and food security with limited cropland resources and large populations. To this end, the crop planting area and structural adjustment policies were conducted in northeast China in recent years. The policy incentivized frequent crop rotations of different crops [57,60,61]. It was the main reason we could observe the widespread rotations in corn-soybean or wheat-rapeseed over the study area. Limited by field surveys, only two years of crop type were mapped, but the consistent observations have the potential for multiple-year crop rotation assessments [19,60].
Although the approach used in this study achieved high accuracies for crop classification, some uncertainties still exist. First, to avoid non-crop interference, we used the 30 m cropland product to mask out the non-crop areas. The classification errors in this product may have impacts on the accuracy of the resultant crop type maps [44]. In addition, mixed pixels derived from the 30m spatial resolution product could also cause errors in crop type classification and rotation monitoring with 10 m Sentinel-1/2 images [84]. The narrow lines seen on the crop rotation map ( Figure 10) were resulted from the spatial resolution differences between Sentinel imagery and the GFSAD30 product. If 10 m cropland products are available, more accurate classification and rotation results will be generated. Second, we applied Sentinel-2 TOA products in this study as the SR data was unavailable in the GEE platform during the study period. Although poor observations and contaminated pixels, including snow/ice, opaque, and cirrus, were processed based on the quality assessment band, the abnormal values also generated uncertainties in classification [7]. Finally, the uneven image numbers across different years (Table 1) may result in inconsistent observations between 2017 and 2018.
While the salt and pepper noises in classification results were minimal, the objectbased classification approaches might also be appropriate in mapping crop types, according to previous studies [62,85,86]. Vegetation indices derived from Sentinel-2 imagery showed little difference in crop type classification. More vegetation indices should be explored in future studies to improve classification accuracy. Limited by the ground truth data availability, only two years of crop type classification and rotation monitoring were assessed in this study. We expected that long-term field surveys or new techniques for training data collections would be available in the future so that crop type rotation monitoring in multiple years is feasible.

Conclusions
Sentinel-1/2 images, combined with random forest classifier and Google Earth Engine platform, were used to map crop types and assess crop rotation patterns in a typical smallholder system located in mountainous areas. We combined different approaches to explore the optimal features for crop type classification in smallholder farms. The results demonstrated that the Sentinel-1 polarizations have better performance for wheat and rapeseed discrimination, and the Sentinel-2 bands exhibit obvious advantages in soybean and corn classification. With the optimal features identified (Sentinel-1 VV and VH bands for wheat and rapeseed; Sentinel-2 red edge and SWIR bands for soybean and corn), we achieved a high overall accuracy of~90% for most crops in the study area and showed clear separations from other crop types even in a complicated crop cultivation, smallholder farming system. Sentinel-1 is recommended for wheat and rapeseed classification and Sentinel-2 for soybean and rapeseed classification, particularly in mountainous areas. The crop rotations were monitored based on the reliable and high spatial resolution crop type maps. This study demonstrated the potential of combining Sentinel-1 and Sentinel-2 in crop type classification and provided a reference for future crop rotation monitoring.