Identifying and monitoring the distribution of crops at high spatial resolution over large regions can help improve food security and achieve sustainable development (Vintrou et al., 2013; Inglada et al., 2015). Crop planting area maps can be used as a direct input to crop production prediction models (Fu et al., 2021) and can be potentially used to predict future planting maps (Zhang C. et al., 2019). Operational cropland mapping can also serve as an important input for modeling greenhouse gas emissions in agro-ecosystems, which plays an important role in determining regional carbon budgets (Crane-Droesch, 2018; Mohammadi et al., 2020). At present, agricultural statistics are collected by censuses and other sampling efforts to update the agricultural information on a regular basis but do not provide the detailed spatial patterns of croplands; moreover, the data collection process is time-consuming, labor-intensive, and expensive (Carletto et al., 2015). Therefore, more reliable and cost-effective methods are needed to identify crop planting area over regional scales.

The prevailing identification methods based on satellite data are machine learning methods such as supervised classifiers, including maximum likelihood classifier (MLC) (Arvor et al., 2011), support vector machine (SVM) (Löw et al., 2013), random forests (RF) (Chen H. et al., 2021) and deep learning (Chew et al., 2020; Xu et al., 2020); and unsupervised classifiers such as k-means (Hamada et al., 2019) and gaussian mixture model (GMM) (Wang et al., 2019). It is difficult for machine learning methods to extend classifier rules and parameters to regions outside of the areas for which they were trained (Rodriguez-Galiano et al., 2012) as all these methods are region- and phase-dependent due to period and region spectral variability. In addition, machine learning methods usually require a large amount of training data consisting of ground-truth observations obtained during the satellite overpass period (Millard and Richardson, 2015; Valero et al., 2016). For instance, the Cropland Data Layer (CDL) released by the United States Department of Agriculture (USDA) uses a decision tree classification method based on satellite observations to provide a 30-m resolution annual map of crop types in the United States (Boryan et al., 2011). The product relies on a large number of ground-truth samples collected each year during the growing season as training datasets (Boryan et al., 2011). In 2009, Nebraska alone used 251,016 Common Land Unit (CLU) polygon records for training (Boryan et al., 2011). Providing such large-scale data for crop mapping is expensive and time-consuming (Carletto et al., 2015). Also, the farm-level land use in the Farm Service Agency’s CLU dataset can only be accessed internally and is not publicly available. Therefore, machine learning algorithms that require as many training samples as possible are limited because of the lack of training datasets (Petitjean et al., 2012).

Another approach is to use a phenology-based classification algorithm, which usually extracts the growth calendar of crops and uses specific spectral features of the phenological period to distinguish crops (Fan et al., 2015; Ashourloo et al., 2019; Rad et al., 2019). Satellite-based vegetation index time-series data are commonly used to retrieve phenological indicators, such as season length, seasonal amplitude, and the number of effective peaks to identify crops (Geerken, 2009; Son et al., 2014; Liu J. et al., 2018; Guo et al., 2022). For example, Salehi Shahrabi et al. (2020) used the normalized difference vegetation index (NDVI) to define two parameters, i.e., the length of season and the slope from peak green to harvest, to identify maize. Some studies use different indexes to detect the unique phenological characteristics of crops within a certain time period (Dong et al., 2015). Several recent studies used the time-weighted dynamic time warping method (TWDTW) to quantify the differences of time series of satellite-based bands or indexes to distinguish various crops (Pan et al., 2021; Zheng et al., 2022a; Huang et al., 2022; Shen et al., 2022). Especially, this method requires a small volume of training samples, and it has shown better performance than commonly used machine learning algorithms (e.g., SVM and RF) for crop and forest classification (Belgiu and Csillik, 2018; Cheng and Wang, 2019).

However, there are large challenges for phenology-based classification algorithms to identify summer crops because of their common developmental patterns and similarity in growth calendars (Rao, 2008; Peña-Barragán et al., 2011; Vuolo et al., 2018). Among the summer crops, maize and rice dominate the cereals produced and consumed globally, and play an important role in ensuring food security. According to the Food and Agriculture Organization (FAO) of the UN, maize and rice accounted for 12% and 8% of the global production of primary crops in 2019, respectively (FAO, 2021). China is the largest producer of rice (FAO, 2021), the second largest producer of maize, and a major global importer of maize (FAO, 2017). Due to the very similar phenology of maize to many other summer crops, its spectral characteristics differ little from those of other crop types, making it difficult to ensure accurate classifications (Zhong et al., 2014; Tian et al., 2021). Skakun et al. (2016) found that maize and soybean share a common crop calendar and have highly similar spectral characteristics, which led to confusion in classification. Sibanda and Murwira (2012) used NDVI to distinguish summer crops and found there were no significant differences in the average NDVI among maize, cotton, and sorghum from the beginning of greening to late senescence. Moola et al. (2021) also found that maize showed low separability from vegetables such as peppers, tomatoes, and cucumbers, leading to classification confusion. In addition, maize has also been found to be indistinguishable from peanuts and sugar beets, and may need to be identified by subtle differences in its specific planting time and general growth patterns (Hoekman and Vissers, 2003; Qiu et al., 2021).

It should be noticed that current studies mostly used a single satellite-based spectral band or index input to identify summer crops (Sun et al., 2019; Zhang S. et al., 2019; Zhong et al., 2019). A single spectral index may not be able to characterize complex crop development, making it difficult to distinguish differences in summer crops (Gella et al., 2021; Shen et al., 2022). For example, Belgiu et al. (2021) found similarities in the curves of NDVI for maize, potatoes, carrots, cereals, soybeans, and cauliflower, leading to a large misclassification. In recent years, the application of multi-source data has received wide attention (Liu W. et al., 2018; Wang S. et al., 2020; Chen H. et al., 2021). An increasing number of studies used multiple satellite-based spectral bands or indexes to identify jointly summer crops (Dong et al., 2015; Wang Y. et al., 2020). For example, Pan et al. (2021) identified paddy rice by detecting a decreased signal of SAR when paddy rice was irrigated and a decreased vegetation index during the harvest periods. However, these methods do not adequately account for differences between similar crops or natural vegetation types, and quantitative measures of phenological changes are limited (Silva Junior et al., 2020). To classify complex summer crops, it is necessary to combine multiple bands or indexes and find important spectral characteristics for identifying these crops.

To deal with the challenges mentioned above, here we developed a new phenology-based method to identify the planting areas of maize, one of the most important summer crops, by combining the phenological characteristics of multiple spectral bands and indexes. Specifically, we revised the TWDTW method by adding the rank maps of each band or index in different phenological periods to combine multiple bands or indexes for classifying. Therefore, the overall goals of this study are to: 1) select different spectral bands or indexes, and determine their own optimal phenological period for maize identification; 2) Based on the revised TWDTW algorithm, combine multiple bands or indexes to extract maize planting area, compare and select the best combination, and extend to other years. The optimal combination will enable timely mapping of maize planting areas with the help of a simple, robust, and automated TWDTW algorithm and is expected to be applied to the identification of other crops. At the same time, it also provides ideas to further explore the use of multi-band combinations for phenology-based classification.

First, we examined the performance of a single satellite-based spectral band or index for identifying maize. In order to detect the best phenological periods of each band and index, we used a running window to generate all possible phenological periods starting from DOY 137 and ending to DOY 289 (see Section 2.3), and we compared the performance of each band and index with all potential phenological periods by comparing the identified areas with statistical area at county level. The results showed that almost the performance of all spectral bands and indexes largely varied during the different periods (Figure 6). For example, REP containing the DOY 233–289 showed the best performance than those during other periods (Figure 6L). Therefore, we compared the performance of each band and index with the best phenological periods. There were large differences in the performance of various bands and indexes compared to the statistical area (Table 2). In all investigated spectral bands and indexes, REP showed the best performance, and the identified areas showed the highest correlation with statistical areas (R ² = .810) and the lowest RMAE (.343) (Table 2). In contrast, RE2 had the poorest performance with the R ² of .575 and the RMAE of .415 (Table 2).

We further examined if the combinations of two bands or indexes can improve the identification accuracy compared to one single band or index, and the phenological periods of each band and index were set as the best phenological periods shown in Table 2. There are a total of 66 combinations of two bands or indexes. The results showed large differences in the identification accuracy of various combinations, such as R ² ranging from .703 to .855 (Figure 7A). Most combinations outperformed the single band or index (Figures 7C, D). For example, the RMAE derived by the combination of NIR and SWIR2 decreased about 10% compared to the lower RMAE of one single band of these two bands (Figure 7D). On contrary, incorporating some bands or indexes will result in lower identification accuracy. It can be seen from Figure 7 that the combinations with LSWI showed worse accuracy. Most combinations with RE3, RE4, and NIR performed better (Figure 7A). The highest accuracy was achieved with the combination of SWIR2 and NIR, with R ² of .855 and RMAE of .258; this was followed by the combination of SWIR2 and RE4, with R ² of .853 and RMAE of .259 (Figures 7A, B).

According to the above comparisons on combinations of two bands shown in Figure 7, the combinations with RE3, RE4, and NIR performed better than other combinations. Therefore, we further selected all combinations with these three bands to generate the combinations of three bands or indexes by integrating another band or index. The classification accuracy of most three bands or indexes combinations was not much improved compared to the two bands or indexes combinations (Figures 8C, D). The highest accuracy was achieved by the combination of NIR, SWIR2, and green, with R ² of .856 and RMAE of .266; the second was with the combination of NIR, SWIR2, and red, whose accuracy is very close to the highest (R ² = .856, RMAE = .273) (Figures 8A, B). Similarly, there was little improvement in the classification accuracy of the combinations of four bands or indexes, with a decrease in R ² for almost all combinations, and an increase in RMAE for most combinations (Figures 9C, D). The above results indicated that as the number of bands or indexes in the combination increases to three or more, the accuracy (R ² and RMAE) no longer increases.

To better verify the classification accuracy of each combination, we calculated the UA, PA, OA, and kappa using the field survey data in 2019. The accuracies of the maize maps were not identical using different bands or indexes. SWIR2 had the highest accuracy (OA 88.69%), followed by NDVI (OA 86.22%); while red (OA 79.68%) and REP (OA 75.24%) had much lower accuracies (Table 3). The combinations obtained significant improvements compared to the basic single band or index. For example, the combination of RE4 and SWIR2 increased OA up to 10.90% compared to RE4 and 3.08% compared to SWIR2. Besides, with the increase of the number of bands or indexes in the combination, the accuracy was also improved. The average OA of the two, three, and four bands or indexes combinations were 89.94%, 90.75%, and 92.07%, respectively. Finally, a comprehensive evaluation of R ², RMAE and OA showed that the combination of RE4 and SWIR2 was the best among the 15 combinations tested in the study.

The combination of RE4 and SWIR2 was used for maize mapping in 2017–2020. At the county level, the identified maize area matched very well with the statistical area (Figure 10). The scatters were close to the 1:1 line, and the R ² reached above .83 with RMAE lower than .27. The comparison of the maize map based on the combination of RE4 and SWIR2 with the very high-resolution images of maize from Google Earth showed high spatial consistency in the two regions (Figure 11) and was able to exclude buildings, roads and other crops, and only slightly misclassified some small areas.

It has been a significant challenge to identify summer crops because of similar phenology characteristics (de Souza et al., 2015; Wang et al., 2022). Multitemporal scenes of one single spectral band or index are commonly used for crop identification (Kussul et al., 2017; Ienco et al., 2019). Our results also showed the similar phenological characteristics among the four summer crops including maize, peanut, soybean, and rice (Figure 4). Therefore, it is quite difficult to map maize relying on a single band or index due to spectral confusion in summer crops. Even when two or more bands and indexes are input for training, phenological characteristics cannot be extracted effectively (Abubakar et al., 2020; You and Dong, 2020; Chen Y. et al., 2021). Previous methods usually used machine learning methods, and limited the application capability to the other regions (Rodriguez-Galiano et al., 2012).

In this study, a multi-band recognition attempt was conducted based on the TWDTW algorithm. The results showed that the combinations of two and more bands and indexes can effectively improve the identification performance compared to one single band or index (Figures 7–9). One of the most important theory basics is large sensitivity differences of various crops to different spectral bands or indexes (Yin et al., 2020; Zhao et al., 2020). Studies have found that the reflectance in the red edge region is mainly affected by leaf absorption and canopy scattering (Baret et al., 1994; Yang and van der Tol, 2018). For the same total leaf chlorophyll content, absorption of a maize leaf in the red edge region is the same as that of a soybean leaf (Peng and Gitelson, 2012). But the scattering of a spherical canopy (soybean) is higher than that of a heliotropic canopy (maize) (Nguy-Robertson et al., 2012). Therefore, the RE4 reflectance of maize is lower than that of soybean (Peng et al., 2011). And the leaf reflectance of peanut in the red edge regions is very close to that of soybean (Buchaillot et al., 2022). SWIR2, another member of the best combination, is very sensitive to changes in soil moisture (Khanna et al., 2007). The differences between maize and paddy rice are striking during the irrigated period of paddy rice, with a much lower value of SWIR2 in paddy rice than in maize (Figure 4).

The new method also highlighted the importance of selecting phenological periods for spectral bands. In general, the classification accuracy of the TWDTW classifiers depends on the distinguishability of the temporal profiles of different crops (Gella et al., 2021). This study examined the difference in the phenological periods of each band or index for identifying maize, and only one band (RE4) showed the best performance using the temporal variations during the entire growing periods, whereas the other bands or indexes are better in certain phenological stages. For example, LSWI, SWIR1 and SWIR2 performed the best during DOY 161–225 (early and middle growth stage) (Table 2), adding data acquired in the late period (DOY 233–289) only brings information redundancy, which has little effect on improving classification accuracy. This phenomenon was also observed by Jia et al. (2012) when they investigated the ability of SAR data in the North China Plain to classify crops. They found that the information contained in two temporal SAR datasets acquired in late jointing and flowering periods is enough for crop classification; Sun et al. (2019) extracted different temporal features to identify crops, and found that the classification accuracy varies widely. In May (flowering period), the crops obtained the best OA and kappa. Research showed that rather than using the entire phenological period, images of optimal time periods can achieve higher classification accuracy (Murakami et al., 2001; Hao et al., 2015).

According to the validation using field surveys and statistical area at county level, this study showed the good performance of the new method for identifying maize. In addition, we also investigated the temporal and spatial patterns of identified maize, which are another important information for judging the method reliability (Pan et al., 2021; Zheng et al., 2022b). The study area (i.e., Henan Province) is one of the largest contributors of maize production in China (Shen et al., 2022), and there are large areas to continuously plant maize (Zhang et al., 2018). The results of the combination of RE4 and SWIR2 showed that more than 95% of fields have been planted maize for four continuous years (Figure 12C). However, single-band RE4 and SWIR2 could not reflect the characteristics of large-scale continuous maize cultivation in Henan Province (Figures 12A, B). In addition, previous studies highlighted that satellite-based classification may result in “salt-and-pepper” noise, i.e., large amounts of isolated single or small area patches consisted of identified maize pixels (Zheng et al., 2022a). Our results indicated a high proportion of isolated large patches of identified maize (Figure 12D). Especially, compared to the classification map based on one single band or index, there are lower proportion of small area patches.

Nowadays, the world is in a rapid stage of agricultural modernization, but food security remains a top priority. Maize is one of the most widely produced cereals in the world and can be used for human food, livestock feed, and bioenergy (Ranum et al., 2014). Accurate and timely information of maize acreage is critical for regional and global food security as well as international food trade. The high-resolution crop distribution map can not only be used as a base map to predict crop production and improve the accuracy of large-scale crop yield simulations (Becker-Reshef et al., 2010; Franch et al., 2015; Wang S. et al., 2020), but also be used as a reference to predict its future planting distribution (Chu et al., 2021). Furthermore, agriculture has the unique potential to provide a beneficial contribution to the global carbon budget. For example, agriculture produces large amounts of nitrogen, one of the long-lived greenhouse gases, due to the use of fertilizers such as nitrogen fertilizers (Northrup et al., 2021; Pan et al., 2022). To better understand food security and greenhouse gas emissions, high spatial-resolution cropland area monitoring is necessary. Our improved method, the TWDTW method that combines multi-source data and phenological information, is suitable for the identification of common summer crops and can be extended to other regions. The produced maps can help to dynamically understand the planting distribution of summer crops, and can greatly help policy decisions related to agriculture and emission reduction.

Although our results showed robust classification by the new method, there are still some uncertainties that need to be resolved in the future. First, optical satellite remote sensing data are affected by clouds and rain, and there were few effective images in central Henan Province, especially in 2017 (Figure 13). While our algorithm screens for good-quality observations, data gaps due to persistent cloud cover create difficulties identifying crop cycles. Some fields may be mistakenly classified as non-maize. To address this, we need more remote sensing data with higher spatial and temporal resolution; another approach is to rely on the fusion of multi-source remote sensing data to produce products with high spatial and temporal resolution (Li et al., 2017; Zhu et al., 2017). In addition, another limitation comes from the impact of composite planting structures (Ghosh et al., 2006; Maitra et al., 2021). The regular maize-soybean strip intercropping, originally popularized in northern China, and maize-soybean relay-strip intercropping extended in southwestern China (Du et al., 2018), generating more complex phenological characteristics. A flexible utilization of the phenological period with TWDTW may be able to better deal with these issues.

This study developed a new phenological-based method to identify the summer crop based on multiple bands with their specific phenological periods. As an example, this study examined the performance of this new method for identifying maize in Henan Province of China. Ten spectral bands and three synthetic indexes derived from the Sentinel-2 dataset were used based on the revised TWDTW method. Time sliding was first performed on all bands and indexes with different time window sizes on 137–289 days of year to obtain the best identification phenological periods. On this basis, the performance of the multi-band combination TWDTW method in extracting maize planting area was evaluated. The combination of RE4 and SWIR2 had the best performance among all combinations, with overall classification accuracy reaching 91.77%. A total of more than 100 counties were selected for data accuracy assessment from 2017 to 2020, showing that the maize planting area estimated by this combination correlated well with the statistical data, with R ² greater than .83 and RMAE lower than .27. The combined effect was more accurate than for all single bands; with the increase in the number of bands or indexes in the combination, the overall classification accuracy improved with the use of up to three bands. The results indicated a robust potential for combining multiple bands or indexes and crop phenological information using the TWDTW method in the application of maize planting area monitoring.