Comparative Analysis of Supervised and Phenology-Based Approaches for Crop Mapping: A Case Study in South Korea

Abstract

This study aims to compare supervised classification methods with phenology-based approaches, specifically pixel-based and segment-based methods, for accurate crop mapping in agricultural landscapes. We utilized Sentinel-2A imagery, which provides multispectral data for accurate crop mapping. 31 normalized difference vegetation index (NDVI) images were calculated from the Sentinel-2A data. Next, we employed phenology-based approaches to extract valuable information from the NDVI time series. A set of 10 phenology metrics was extracted from the NDVI data. For the supervised classification, we employed the maximum likelihood (MaxLike) algorithm. For the phenology-based approaches, we implemented both pixel-based and segment-based methods. The results indicate that phenology-based approaches outperformed the MaxLike algorithm in regions with frequent rainfall and cloudy conditions. The segment-based phenology approach demonstrated the highest kappa coefficient of 0.85, indicating a high level of agreement with the ground truth data. The pixel-based phenology approach also achieved a commendable kappa coefficient of 0.81, indicating its effectiveness in accurately classifying the crop types. On the other hand, the supervised classification method (MaxLike) yielded a lower kappa coefficient of 0.74. Our study suggests that segment-based phenology mapping is a suitable approach for regions like South Korea, where continuous cloud-free satellite images are scarce. However, establishing precise classification thresholds remains challenging due to the lack of adequately sampled NDVI data. Despite this limitation, the phenology-based approach demonstrates its potential in crop classification, particularly in regions with varying weather patterns.

1. Introduction

Crop classification through remote sensing is crucial in ensuring food security and sustainable agriculture (Karthikeyan et al., 2020). However, obtaining accurate spatio temporal data on crop types remains a challenge, particularly in small holder farms (Wang et al., 2022a; Wang et al., 2022b). The development of efficient crop mapping algorithms is necessary to enable widespread application in large spatial domains (Ashourloo et al., 2022). Nevertheless, crop mapping presents significant difficulties with in the Land Use/Land Cover(LULC) classification community (Rahimi et al., 2022). One of the primary challenges is the limited availability of continuous in-situ data across vast areas, which is essential for conventional supervised classification (Zeng et al., 2020). Additionally, intra-class variability and inter-class similarity pose significant obstacles. Spectral-temporal characteristics of crops can differ widely across regions and years, while different crops may exhibit similar phenology and spectral features (Foerster et al., 2012).

There are multiple approaches available for deriving cropping patterns(Mahlayeye et al., 2022). One potential method involves obtaining a small number of highly detailed images during the crop-growing season. These images can be utilized to differentiate between various types of crops. Classical supervised land cover mapping techniques are well-suited for this task. However, the primary challenge lies in acquiring cloud-free images due to the combined impact of weather conditions and the limited revisit frequencies of most satellites with high-resolution capabilities (Hu et al., 2019). To address these challenges to classical supervised, techniques, phenology-based algorithms have been proposed, which analyze crop life cycles and develop temporal metrics(Liu et al., 2018; Qiu et al., 2015; Rahimi, 2024; Rahimi et al., 2021; Tian et al., 2019; Waldner et al., 2015). Vegetation phenology is a natural process that represents the annual cycle of plants, influenced by both biological and non-biological factors(Zhang et al., 2022).

Over the past few decades, various remote sensing approaches have been developed to map cropping intensity on large spatial and temporal scales. Traditionally, the normalized difference vegetation index (NDVI) or the enhanced vegetation index (EVI) has been calculated based on high-temporal-resolution datasets to create time series datasets that describe the crop’s life cycle (Zeng et al., 2020). Different methods have been proposed for mapping cropping intensity based on distinct phenological transitions, leading to two main groups of approaches. In the first group, annual cropping intensity is identified by applying a threshold to generate a binary crop phenology profile that indicates growing and non-growing periods (Liu et al., 2020a). The other group of studies focuses on identifying the peaks or valleys of crop phenology. Cropping intensity is then quantified by counting the number of peaks or valleys within a year. This approach provides an alternative method to assess cropping intensity based on distinctive phenological features (Liu et al., 2020b).

Previous studies have demonstrated the significant potential of crop phenology algorithms in agricultural remote sensing mapping (Rahimi, 2024). These algorithms have been successfully employed in various applications, including identifying the start of the season (SOS) and the end of the end of the season (EOS), detecting peak growth (PG) and peak drought (PD) periods, calculating the duration of the growth cycle, quantifying the growth rate and decline rate of vegetation indices (VIs) values and assessing cropping intensity by identifying the number of peaks and troughs in the VIs’ time series within a year (Araya et al., 2018; Filippa et al., 2016; Hufkens et al., 2018; Tan et al., 2010). These algorithms leverage remote sensing signals corresponding to key phenological stages to identify crops (Zeng et al., 2020). For example, a paddy rice mapping algorithm has been developed that utilizes unique features of paddy rice, such as transplanting or harvesting, and employs vegetation indices (Dong et al., 2016; Xiao et al., 2005). These findings highlight the valuable insights that can be gained by utilizing crop phenology in agricultural remote sensing and its potential for enhancing mapping and monitoring capabilities.

Phenology-based or thresholding methods require accurate threshold values to achieve precise classification (Tang et al., 2016). Supervised classification necessitates training data for each year, which becomes challenging for large-scale mapping. Phenology-based mapping relies on long and continuous time series data. Furthermore, all these methods exhibit limitations in accurately classifying areas with small-sized crop fields. In contrast, object-based image analysis (OBIA), as compared to traditional pixel-based classification, considers groups of pixels instead of individual pixels, contributing additional information such as spectral, textural, and geometric features (De Wit and Clevers, 2004). The OBIA approach comprises two main steps: image segmentation and classification. Segmentation involves partitioning the entire image into internally uniform and homogeneous regions or objects, generating additional spectral, textural, and geometric information.

In the classification process, each object is assigned to a specific class based on its spectral, textural, geometric, and customized properties (Singha et al., 2016). Previous studies have demonstrated the effectiveness of OBIA in various applications (De Castro et al., 2018; Qiu et al., 2021). Given the prevalence of small, fragmented rice fields in South Korea, the OBIA approach can be particularly advantageous as it focuses on objects or parcels rather than solely relying on single-pixel properties. Phenology, which encompasses the timing of plant growth events such as budburst, leafing, peak growth stage, flowering, and abscission, has been utilized in recent studies for crop classification. However, the potential of phenology within the OBIA framework for crop classification in regions like South Korea remains unexplored.

Mapping crops accurately in regions with frequent rain and cloud cover poses significant challenges, especially when dealing with small agricultural parcels. Traditional phenology-based approaches may face limitations in capturing the temporal dynamics of crop growth due to the inadequate availability of cloud-free images. Alternatively, supervised approaches such as the maximum likelihood (MaxLike) algorithm have shown promise in crop classification, relying on training data and spectral characteristics. In regions like South Korea with a high frequency of rainy and cloudy days, particularly during the peak vegetation growth period, and where agricultural parcels are predominantly small (less than 1 hectare), mapping crops using phenology-based approaches can be challenging. In this study, we aim to compare the effectiveness of a phenology-based approach and a supervised approach (Maximum likelihood classification termed MaxLike) algorithm, in crop classification.

2. Materials and Methods

2.1. Study Area

Our study area Angyo-ri (457312, 461839, 4048669, 40405646 UTM zone 52), with an area of 649.9 ha is in Andong City which is the capital of Gyeongbuk Province in South Korea. After conducting field surveys in the study area, several crop types were identified, including rice, pepper, potato, corn, and man-made structures such as greenhouses covered with plastic. These classes were deemed relevant for further analysis and were therefore considered in the subsequent stages of the study. Fig. 1 depicts the geographical position of Angyo-ri within South Korea. The satellite imagery displayed is from Sentinel-2A and was captured on August 27, 2022. The dark green areas depicted in this figure indicate rice crops, providing visual evidence that our study area is predominantly characterized by rice cultivation.

Fig. 1. Land use map of South Korea (Buchhorn et al., 2020) and the geographic location of the study area Angyo-ri in Andong City, South Korea.

2.2. Data

In recent times, the launch of European Space Agency (ESA) twin Sentinel-2 satellites has brought about changes in data availability (Misra et al., 2020). These satellites provide 10-meter resolution data and revisit the same location every five days, which has significantly improved the acquisition of cloud-free images. Despite this progress, retrospective analyses using such dense high-resolution datasets are still not readily accessible. The availability of Sentinel-2 images provides a valuable resource for acquiring time series imagery. This unique combination of spatial and temporal resolution presents an unparalleled opportunity to capture detailed information on crop phenological dynamics (Misra et al., 2020). By leveraging these datasets, researchers can gain comprehensive insights into the temporal patterns and changes in crop growth stages overtime. Therefore, we used Sentinel-2A images from the year 2022 for both phenology-based and supervised classification approaches. Our selection process involved choosing images that were free from clouds and required minimal atmospheric correction. As a result, we identified 31 suitable images for NDVI extraction specifically for our study area. It is important to note that we focused on using images with a spatial resolution of 10 meters (Table 1).

Table 1. Distribution of Sentinel dataset across different months of the year

Numbers indicate the day of the month.

2.3. Phenology Metrics Calculation

For the phenology-based crop mapping, we initially computed 31 NDVI images for the study area. To smooth the NDVI values, we utilized the “sgolayfilt” function in the “signal” R package. This function applies a Savitzky-Golay (S-G) filter, which is a widely used method for smoothing noisy data (Araya et al., 2018). The filter helps to reduce fluctuations in the NDVI time series, providing a more continuous and representative pattern of vegetation growth overtime. When applying the Savitzky-Golay (S-G) filter, It is essential to exercise caution as it can potentially weaken the signals present in the data. Therefore, It is crucial to strike a balance between smoothing the data to reduce noise while preserving the underlying signals of interest (Tan et al., 2010). Subsequently, we calculated 10 phenology metrics for each pixel in the study area, as outlined in Table 2 (Filippa et al., 2016; Rahimi, 2024). The definition and description of each metric are provided in Table 2 to facilitate comprehension. Fig. 2 complements the understanding of these metrics by presenting a schematic representation of nine of them.

Table 2. Definition of the phonological indices calculated in the study (Filippa et al., 2016; Rahimi, 2024)

Fig. 2. The NDVI dynamics curve shows nine phonological metrics. DOY: day of the year.

In particular, the metrics “Min” and “Max” correspond to the minimum and maximum values of the NDVI trend line, respectively. The “Difference” metric is obtained by subtracting the minimum value from the maximum value. The metrics “TMin” and “TMax” represent the dates corresponding to the minimum and maximum NDVI values, respectively. The “SOS” (Start of Season) and “EOS” (End of Season) metrics indicate the dates when the NDVI value first time reaches 0.2 and last time reaches 0.2 on the opposite side of the NDVI trend line, respectively. By subtracting the SOS from the EOS, we obtain the “Timelength” metric, which represents the duration of the growing season. The “Count0.2” metric represents the number of dates in which the NDVI values are greater than 0.2. The “TafterMax” metric indicates the time it takes for the crop reflectance to reach 0.2 again after reaching the maximum NDVI value. The “AUC” (Area Under Curve) metric quantifies the area under the NDVI curve, which provides an overall measure of the vegetation growth intensity throughout the growing season and was calculated using the “trapz” function in R software. All other phenology metrics were also calculated using R software.

2.4. Pixel-Based and Segment-Based Phenology Mapping

We employed two techniques in the phenology mapping approach, namely pixel-based and segment-based. The reason was to compare the effectiveness of these two techniques. The pixel-based technique involved acquiring a set of NDVI values for each cell in the landscape throughout the year 2022. The phenology classification procedure involved several steps. Firstly, we used the “Raster to Point” function in ArcGIS software to convert the study area data into a point shapefile format. This point dataset was used to extract NDVI values for 31 time periods. The attribute table of the point shapefile was then exported to an Excel file for subsequent analysis in R software. This allowed for the calculation of 10 phenology metrics for each of the 64,985 points in the study area. Next, the “Point to Raster” function in ArcGIS was employed to convert the point data back into a raster format. This process assigned the calculated phenology metric values to their respective locations on the raster grid, resulting in phenology maps. These maps provided a spatial representation of the distribution of the phenology metrics across the study area.

After identifying the crop classes and obtaining the necessary thresholds, the Structured Query Language (SQL) language in ArcGIS was utilized to assign the appropriate crop type to each polygon. Crop phenology thresholds were identified by analyzing phenology maps and employing visual inspections, aided by phenology NDVI curves. This approach allowed us to discern distinct patterns in the data and accurately determine the thresholds necessary for land cover classification. The specific equations for each crop are as follows:

Rice Equation: Rice = (Max ≥ 0.52) ˄ (TMax = 234) ˄ ¬ ((Max ≤ 0.2) ˄ (TMax ≠ 184)) ˄ ¬ (TMin = 169 ˅ TMin = 74)

Man-made Equation: (Max ≤ 0.2) ˄ (TMax ≠ 184)

Corn Equation: (TMax = 184)

Potato Equation: ((TMax = 304) ˅ (TMax = 324) ˅ (TMax = 154)) ˄ ¬ ((Max ≤ 0.2) ˄ (TMax ≠ 184))

Pepper Equation: (SOS = 144) ˄ (TMax = 239)

˄ (Logical AND): The symbol ˄ represents the logical AND operator, which signifies that both conditions on either side must be true for the overall expression to be true. For instance, in the expression “A ˄ B”, both A and B must hold true for the entire statement to be true.

¬ (Logical NOT): The symbol ¬ denotes the logical NOT operator, which negates or reverses the truth value of the following expression. When applied to a statement, it indicates the opposite of that statement’s truth value. For example, in “¬A”, if A is true, ¬A is false, and vice versa.

≠ (Not Equal To): This symbol denotes the not equal to operator, signifying that the values or conditions on either side of the symbol are not equal. In an expression like “A ≠ B”, it indicates that A and B are different from each other.

˅ (Logical OR): The symbol ˅ represents the logical OR operator, which signifies that at least one of the conditions on either side must be true for the entire expression to be true. For instance, in “A ˅ B”, if either A or B is true (or both are true), the entire statement is considered true.

On the other hand, the segment-based technique employed a different approach. Initially, the target image obtained on August 27, 2022, was segmented using eCognition software (Nussbaum et al., 2008). The segmentation analysis was performed using a scale parameter of 20 and a color threshold of 0.9. This process aimed to group similar pixels based on their spectral characteristics and spatial relationships. As a result, an output shapefile consisting of 4,699 segments was generated. Each segment represented a distinct region within the study area, characterized by homogeneous attributes such as color and texture. Average NDVI values were then calculated using ENVI software for each segment. The subsequent steps followed a similar methodology to the pixel-based approach, where phenology indices were derived based on the average values. By utilizing these two techniques, we aimed to map crops based on their phenological characteristics. The comparison between the pixel-based and segment-based approaches allowed us to assess their respective abilities and determine which technique performed better in this study.

2.5. Pixel-Based Supervised Maximum Likelihood Classification

In the supervised classification step, we employed the MaxLike algorithm at the pixel level to compare the results with the phenology-based approach. To perform this step, we conducted a field survey of the study area in August 2022 to visually identify and locate the different crop types present. Given that the study area was primarily dominated by rice crops, we focused on recording the locations of other crops using GPS devices. These recorded locations served as training samples in the ENVI software, which allowed us to train the MaxLike algorithm for the accurate classification of crop types in the study area.

2.6. Accuracy Assessment

To calculate the overall accuracy and kappa coefficients for each classified map, a total of 400 random points were generated across the 4 study areas. These points were then cross-referenced with Google Earth imagery and the previously collected field data to determine the actual crop type for each point. By comparing the assigned crop type from the classified map with the actual crop type, and confusion matrix, the overall accuracy and kappa coefficients were calculated as evaluation metrics for the classification accuracy.

3. Results

Fig. 3 shows some examples of phenological maps calculated for the study area. The map representing the “Min” metric (Fig. 3a) reveals variations in the minimum values across different pixels, providing some potential for crop separation. On the other hand, the map representing the “Max” metric (Fig. 3b) shows distinct agricultural parcels. Interestingly, when comparing this map with the satellite image presented in Fig. 1(a) remarkable similarity can be observed, indicating the usefulness of the “Max” metric in crop classification. The maps representing the “SOS” and “EOS” metrics (Figs. 3c, d) also exhibit notable patterns, highlighting certain farms. The “Timelength” metric (Fig. 3e) map shows similarities with the “SOS” metric, both of which can assist in differentiating farms. The “TMin” map (Fig. 3f), with its pink color representing rice crops, effectively highlights several small farms throughout the study area. The “TMax” map (Fig. 3g) also displays some potential for crop separation. Additionally, the “AUC” map (Fig. 3h) proves to be helpful, as it highlights specific areas like the “Max” metric map.

Fig. 3. Examples of phenological maps calculated for the study area: (a) Min, (b) Max, (c) SOS, (d) EOS, (e) Timelength, (f) TMin, (g) TMax, and (h) AUC.

Fig. 4 displays the NDVI dynamics curves of different crops after applying the S-G filter. The graph reveals distinct patterns in certain segments of the trend lines, indicating the potential for crop differentiation. For instance, the trend line for man-made structures consistently exhibits NDVI values below 0.2 throughout the year, suggesting that these structures can be easily classified. Similarly, other crops demonstrate unique patterns in the maximum NDVI values, allowing for differentiation based on the “Max” metric. It is important to note, that these trend lines are drawn based on pure samples, and the actual classification of different crops may pose challenges in real-world scenarios.

Fig. 4. The NDVI dynamics curve showing different crops.

3.1. Visual Assessment of Crop Mapping

Fig. 5 displays the classified maps of the study area. Upon visual interpretation, no significant differences are observed among these classified maps. However, one notable distinction lies in their ability to capture and classify marginal cells associated with small farms. The presence of scattered pixels, often referred to as a “salt and pepper” pattern, is more prominent in the MaxLike (Fig. 5a) and pixel-based (Fig. 5b) classification maps compared to the segment-based map (Fig. 5c). This suggests that the segment-based approach may offer improved delineation and smoother boundaries for agricultural parcels.

Fig. 5. Crop classified maps of the study area: (a) MaxLike, (b) pixel-based phenology, and (c) segment-based phenology.

3.2. Accuracy Assessment of Classified Maps

Table 3 provides an overview of the classification accuracy of the crop maps based on 400 random points, categorized according to three classification methods: MaxLike, Pixel-based, and Segment-based. Each method’s accuracy is evaluated for various crop classes, including Rice, Man-made, Pepper, Potato, and Corn. Under the MaxLike classification, the overall accuracy ranges from 0.89 to 0.93, with corresponding kappa coefficients between 0.74 and 0.85. The kappa coefficient is a measure of agreement between the classified maps and the ground truth data, with values closer to 1 indicating a higher level of agreement. The range of kappa coefficients suggests that different classification approaches or methods used in the study have yielded varying levels of agreement with the ground truth.

Table 3. The classification accuracy of the crop maps (based on 400 random points)

In particular, the MaxLike method demonstrates high accuracy in classifying Rice, with values ranging from 0.89 to 0.93. However, it shows limitations in accurately classifying other crops, such as Man-made and Pepper, where the classification accuracy is lower. Conversely, the Pixel-based and Segment-based classifications exhibit higher overall accuracy, ranging from 0.92 to 0.93, and kappa coefficients ranging from 0.81 to 0.85. These methods demonstrate improved performance in classifying all crop types, including Man-made, Pepper, and Potato, with higher accuracy compared to the MaxLike method. Overall, the Segment-based classification shows greater consistency and accuracy across all crop classes, suggesting their effectiveness in accurately mapping land cover types within the study area.

4. Discussion

This study compared the application of phenology-based approaches and supervised classification techniques for crop mapping in a specific study area. By utilizing remote sensing data and analyzing various phenology metrics, the study aimed to understand the effectiveness of these methods in accurately identifying and mapping different crops. The phenology-based approach relied on calculating NDVI values and extracting phenology metrics to characterize crop growth patterns. This method provided valuable insights into the temporal dynamics of crops and their phenological transitions. On the other hand, the supervised classification technique, specifically the MaxLike algorithm, utilized training samples and statistical analysis to classify crops based on spectral characteristics. The comparison of these two approaches allowed for a comprehensive evaluation of their performance in the study area. Additionally, the study employed field surveys and ground truth data to assess the accuracy of the classified maps, utilizing metrics such as overall accuracy and the kappa coefficient. The results of the study demonstrated that the phenology-based approach, particularly the segment-based method, achieved acceptable classification accuracies, indicating its effectiveness in differentiating various crop types.

Earlier studies on crop mapping techniques using optical images can be divided into different groups. One group focuses on individual cloud-free images and employs approaches based on image statistics. Examples include unsupervised classifiers like the self-organizing data analysis technique, as well as supervised classifiers such as maximum likelihood and support vector machine. For instance, Nitze et al. (2012) found that support vector machine classifiers with specific kernels were particularly successful in classifying crops in their dataset. They also observed that incorporating a multi-temporal approach enhanced the accuracy for specific crop types. The researchers then shifted towards using object-based methods after initially employing simple pixel-based classification. The comparison revealed that object-based methods exhibited higher ability and performance compared to pixel-based methods for land-cover classification (Ma et al., 2017). However, supervised classification in remote sensing does have some important weaknesses. Some of these weaknesses include training data, bias, generalization issues, sensitivity to training set size, spectral variability, and difficulty in incorporating spatial information (Ozdogan et al., 2010). Then, the trend of crop classification shifted towards pixel-based (Han et al., 2022) and segment-based phenology mapping methods (De Castro et al., 2018; Singha et al., 2016).

In our study, we aimed to compare supervised classification with phenology-based approaches of pixel-based and segment-based methods, in regions with challenges such as rainy days and limited availability of cloud-free images, such as South Korea. Despite the advantage of supervised classification in terms of shorter processing time, we found that the MaxLike method yielded lower classification accuracy compared to phenology-based approaches. As a result, our study suggests using segment-based phenology mapping as a more suitable approach for regions like South Korea, where continuous cloud-free satellite images are less likely to be available. This highlights the potential of phenology-based methods in overcoming the limitations posed by adverse weather conditions and limited data availability for accurate crop classification in such regions.

However, we encountered a limitation in using phenology-based approaches to determine suitable thresholds for differentiating crops. Due to a lack of adequately sampled NDVI data at critical time points, there was a possibility of missing valuable information, making it challenging to establish precise classification thresholds. This highlights the significance of having comprehensive and well-distributed temporal data for reliable analysis using phenology-based methods. Despite this limitation, the study findings indicated that phenology-based approaches still outperformed supervised methods in regions with frequent rainfall and cloudy conditions, such as South Korea. This suggests that the inherent temporal dynamics captured by phenological metrics can compensate for the difficulties posed by adverse weather conditions, leading to improved accuracy in classifying crops compared to traditional supervised methods. These results underscore the potential of phenology-based approaches in regions with diverse weather patterns, where relying solely on supervised methods may be less effective due to factors like cloud cover and atmospheric variations. However, it is crucial to acknowledge the limitations of the phenology-based approach, particularly the need for comprehensive and well-distributed temporal data to ensure accurate determination of thresholds and reliable crop classification. Future studies could focus on enhancing data collection strategies and developing robust techniques for threshold determination to further enhance the performance of phenology-based approaches in challenging environmental conditions.

5. Conclusions

This study revealed that among the methods evaluated, the phenology-based approach demonstrated superior performance compared to the classical supervised. The phenology-based approach, utilizing phenological metrics derived from remote sensing data, showcased a higher ability to accurately classify different crops. In contrast, the pixel-based method exhibited lower performance in crop classification than segment-based. The findings suggest that incorporating phenological information and temporal dynamics into the classification process can enhance the accuracy and effectiveness of crop mapping. The phenology-based approach leverages the patterns and trends in vegetation indices over time to differentiate between different crop types, leading to improved classification results. By considering spatial relationships and aggregating pixels into segments, the segment-based approach was able to capture more coherent and meaningful agricultural parcels, resulting in enhanced classification accuracy and smoother boundaries. These results highlight the importance of selecting appropriate classification methods that leverage temporal and spatial information for accurate crop mapping. We proposed phenology-based classification as a potential approach to mitigate the challenge of the limited availability of cloud-free images. By incorporating phenological information from satellite imagery, we can enhance the accuracy and robustness of crop classification models, thereby overcoming the limitations posed by cloud cover and ensuring the reliability of land cover mapping efforts.