1. Introduction
Accurate and timely information on the cropping area and crop type obtained from remote sensing data either or not in combination with ground surveys is key for estimating crop production. This information has significant environmental, policy, agricultural and economic implications for most national governments, since crop production figures are used for determining the amount of food to import or export at the end of the growing season [
1,
2]. The error introduced to crop production estimation from general agricultural land cover maps is minimized with accurate crop extent maps [
1,
3,
4,
5]. For remote sensing-based crop production estimates, the ideal approach would be to combine biomass proxies and crop maps. Biomass proxies have been available for decades and from different sensors at different spatial resolutions. However, creating crop-specific maps has remained a challenge. In general, cropland maps, regardless of crop type, have proved to improve crop production forecasting [
4].
Discriminating croplands from non-croplands and identifying different crop types can be achieved with remote sensing-based crop growth monitoring and in particular with indices that quantify the distinct green-up and senescence of the crop cycle [
5]. Since different crops show different spectral responses depending on their maturity stage, the temporal dimension of remote sensing data is most useful for identifying major crop types and their phenology [
1,
6,
7]. However, using remote sensing data in an operational context for crop area assessment requires a wide geographic coverage and high spatio-temporal resolution at a minimal cost [
6].
Each vegetated land cover class represents a distinctive phenology (i.e., green-up, maturity, senescence and dormancy). Different datasets have been used to monitor the crop signature in remote sensing. The temporal resolution of high spatial resolution data is too low to derive crop phenology directly, whereas medium/low resolution data do not have sufficient spatial resolution to capture the crop-specific signature [
8,
9]. Despite these limitations, several studies successfully used low to high spatial resolution data or a combination of different resolutions for arable crop identification in the Great Plains by using 1-km NOAA-AVHRR and 500-m MODIS time series [
6,
10], for paddy rice identification in Japan by using 500-m MODIS time series [
11] and in northeast China by using a Landsat-based phenology algorithm [
12]. The work in [
1] developed a method for combining high spatial resolution data (Landsat, 30 m) with high temporal resolution data (MODIS, 500 m) to achieve a superior classification of crops in the Mississippi River Basin. In the near future, crop mapping will be possible with Sentinel 2, which has a high spatial resolution (10 m) and a five-day revisiting time. Since our final goal is to use the outcome map for crop production estimates, we hypothesize that 100-m Proba-V data can fulfil the requirements of both revisiting time and spatial resolution for crop production mapping at the regional scale.
Different methods for discriminating cropland and mapping different crop types based on vegetation phenology exist. The work in [
6] investigated the class separability between specific crop types in time series vegetation index data using the Jeffries–Matusita distance. In another study, [
13] applied a cluster analysis and used the Euclidean distance to compute the temporal distance of enhanced vegetation index values among samples. Support vector machines were used to map abandoned agriculture at large scales with coarse-resolution MODIS imagery and phenology metrics calculated with TIMESAT [
14] (For the details of TIMESAT, please refer to [
15]). The work in [
16] used agro-meteorological data containing information on times of crop growth stages, which were utilized to obtain the average phenological pattern for each individual crop type.
Several studies were done on crop mapping using different remote sensing techniques and analyzing time series data at different resolutions. Machine learning algorithms, such as random forest, artificial neural network and support vector machine, perform significantly better compared to the traditional supervised classification methods [
17,
18]. Dynamic time warping (DTW) has emerged as a promising new technique for time series data mining applications that include land cover mapping [
19]. The major disadvantage of these methods is their computational complexity [
20]. Spectral matching techniques (SMTs, [
21]) are an innovative method of identifying and labelling information classes in historical time series data. Originally, SMTs were used in hyperspectral analysis of minerals [
21]. Time series data, however, can be treated in a similar manner as hyperspectral data where hundreds of bands stack a single instance of a hyperspectral image [
22]. According to the method, two time series are matched with the target ‘spectra’, which are acquired from ideal end-member classes known through census data, ground truth or maps of the study area. The work in [
21] tested the method using monthly AVHRR data in the Krishna River Basin in India and demonstrated that spectral similarity was the best method.
A number of factors affect the accuracy of crop classification. Differences in crop phenology [
23], agricultural field size [
24,
25] and the observation period length are all known to have a significant effect on accuracies. In addition, each crop has unique phenological features, which are affected by regional variations in climate and management practices [
6,
11,
26]. Furthermore, varying spectral responses from the soil can change the ability to discriminate the crop type throughout the growing season [
27]. Accordingly, image acquisition during those periods when crop separability is the highest is crucial to increase crop classification accuracy [
27,
28,
29].
The objective of this study is to develop a crop mapping approach applicable at a global level inspired by the SMT method [
21] on a seasonal basis using 100-m Proba-V NDVI data. Proba-V data at a 100-m spatial and five-day temporal resolution likely improve land monitoring studies compared to the 250-m spatial and eight-day temporal resolution of MODIS data, or the 300-m spatial and one-day temporal resolution of Proba-V, or the 10-km spatial and one-day temporal resolution of NOAA-AVHRR [
25]. Although the 100-m Proba-V data tend to be more advantageous, the time series data are currently limited, as they became available in May 2013. SMTs were used for seasonal crop area mapping with time series data by using different temporal windows throughout the growing season: from green-up to senescence, from green-up to dormancy and from minimum NDVI at the beginning of the growing season to minimum NDVI at the end of the growing season. The method aims to facilitate crop production estimates by developing crop-specific maps.
5. Discussion
This study demonstrated the suitability of spectral matching techniques (SMTs) for mapping crop types using 100-m Proba-V data for the 2014–2015 season. The methodology integrated multi-temporal satellite imagery and parcel boundaries retrieved from both the SIGMA project and ‘GDI-Flanders’ databases. The SMTs were ideal for analyzing remote sensing time series data during the crop growth period. We calculated spectral similarity values (SSV), which are measures of the shape and magnitude similarities of the time series spectra and found the most useful SMTs, similar to [
21]. Subsequently, SMTs were applied to match the ideal spectra, i.e., the reference NDVI profiles, to the class spectra, i.e., the individual pure pixel NDVI profiles.
The methodology demonstrated that 100-m Proba-V has the potential to be used in crop area mapping across different regions in the world. Proba-V is a relatively new satellite, and therefore, there are limited studies available for crop mapping. The work in [
48] reported crop identification accuracies in the range of 72.4%–86.2% for 100-m Proba-V data for mapping summer and winter crops in Bulgaria. In another study, [
49] achieved an overall accuracy of 84% using the 100-m Proba-V sensor for cropland mapping of Sahelian and Sudanian agro-ecosystems. These reported ranges are in line with our results. When using post-classification, the overall accuracy (%) ranged between 65 and 86, and the kappa coefficient changed from 0.43–0.84. In general, post-classification improved the overall accuracy results around an additional 6%–7% for Ukraine and Brazil and 11% for Russia compared to the initial classification results. For Belgium, the post-classification technique did not improve the classification results (see
Table 3 and
Table A1). Our results are best in Sria, Russia, followed by Kyiv, Ukraine, Flanders-Belgium and Sao Paulo, Brazil. A couple of reasons could explain the differences between accuracies across the different study areas. Firstly, better results were observed in the areas where the crop phenological development was not spread over a long time period. For instance, the planting time for maize in Brazil stretched from August–December with a period of highest activity in October and November. This prevented extracting the distinctive characteristic of the reference NDVI profiles. Secondly, the parcel sizes played an important role. When parcels covered a small number of satellite pixels, the results were less accurate, as was the case for Belgium. Thirdly, classification errors of crop types increased when the time window covered only part of the cropping period. Another reason behind the classification errors is related to the number of ground-truth parcels available from the study site, as is the case for the winter barley fields in Kyiv, Ukraine, compared to other crop types in the same site. Finally, crops with similar growing periods might cause classification errors, such as sunflower and maize in Sria, Russia. In addition, the extent of the study area played a role. Accuracies potentially improved when region specific NDVI reference profiles were included from different agro-ecological regions. Based on these results, crop area mapping was challenging, but the use of 100_m Proba-V proved a valid option even when mapping at the field level.
Our results were in close agreement with other studies that used different classification methods and/or other higher resolution satellite images. We used a multi-temporal sequence of 100-m Proba-V images covering one to two growing seasons. The work in [
8] reported an overall accuracy of 63% for vegetation mapping in southern Norway using 25-m resolution Landsat images. Another similar study reported an overall accuracy of 62.7% using the NDVI temporal profiles approach and 72.8% using a maximum likelihood classifier in the northeast of Germany with phenological information and spectral-temporal profiles from Landsat TM/ETM [
16]. The use of multiple sensors seemed to increase the accuracy. For instance, [
40] updated the crop classification in the land cover database of The Netherlands by combining Landsat TM, IRS-LISS3 (Indian Remote Sensing Satellite—Linear Imaging Self Scanner) and ERS2-SAR (European remote sensing satellite 2—synthetic aperture radar) and reported an overall accuracy value of 90%. Almost one million pixels were used at the national level covering not only the different types of cereals, but also grassland and flower bulbs. The use of homogeneous pixels improved the classification accuracy. The overall accuracy ranged from 73% for very heterogeneous pixels to 89% for homogeneous pixels in North Carolina and Virginia with 250-m MODIS NDVI [
50]. The number of homogenous pixels used in their study was 1014, which included 475 pixels for agriculture. We presented specific crop mapping results per-field. In another study, both per-field and per-area results were presented. The work in [
51] reported a maximum overall accuracy of 66% and a kappa coefficient of 0.60 per field and a maximum overall accuracy of 70% and a kappa coefficient of 0.64 per area for mapping specific crop types in Central Valley of California based on the time series of Landsat TM/ETM+.
Although our method showed promising results in crop area mapping, we identified a number of limitations. The reference NDVI profiles for the growing season of each crop type had to be defined in advance, either based on ground data, on user knowledge of the field or on a literature review. Another limiting factor occurred when the parcel size was smaller than the pixel size. Having larger parcel sizes than pixel sizes was an advantage, particularly because pure pixels tremendously improved the classification results.
The maps based on our methodology could be extended to regional or national-level crop production estimations and all crop types of interest. We showed that the within-field spectral variability could be reduced with accurate field boundaries. These boundaries eliminated classification errors due to mixed pixels [
40]. Object-based image analysis could enable the detection of field boundaries in regions without parcel information. To this extent, [
51] used image segmentation to delineate the field borders prior to classification.