1. Introduction
Crop monitoring by satellite remote sensing requires high spatial and temporal resolution image time series and ground campaigns to monitor the entire crop cycle with frequent ground acquisitions over extensive areas. With the ever-increasing number of satellites and the availability of free data, the integration of multisensor images in coherent time series offers new opportunities for land cover and crop type classification [
1]. In addition, satellites with a shorter revisit time (e.g., six days at the Equator for the Sentinel-1 constellation and five days at the Equator under cloud-free conditions for the Sentinel-2 constellation) and reconfigurable acquisitions (different viewing conditions can be applied for more frequent observation of a certain area) can be used to better identify the different growth cycle stages that are often imperceptible when using more sporadic data.
Many studies have combined optical and microwave images to improve mapping accuracy in agricultural scenarios [
2,
3,
4,
5,
6,
7]. SAR data are independent of solar illumination and depend on the wavelength and on the roughness, geometry and material contents of the targeted surface. In contrast, optical data are greatly influenced by cloud cover and represent the reflectance of solar energy from a target area. The potential of Earth Observation (EO) techniques for the management of land and water resources has also been widely acknowledged [
8,
9]. The repeatability of observations on a cyclic basis and the availability of high spatial resolution multispectral data are particularly suitable for cost-effectively mapping crops and irrigated areas with satisfactory accuracy [
10].
The amount of water required to meet the cropped field’s water loss through evapotranspiration is defined as the crop water requirement [
11,
12]. Water requirements vary from crop to crop and throughout the growing season of an individual crop. The FAO Penman–Monteith method is the standard method for the definition and computation of the reference evapotranspiration (
ETo) [
13,
14]. Crop evapotranspiration (
ETc) from crop surfaces under standard conditions is determined by crop coefficients (
Kc) that relate
ETc to
ETo. The dual crop coefficient approach separates
Kc into two separated coefficients, one for crop transpiration (
Kcb, basal crop coefficient) and another for soil evaporation (
Ke).
EO methodologies have been used to estimate
ETc due to the reflective properties of vegetation and to the relationship of
ETc with crop characteristics, such as the Leaf Area Index (LAI) and crop coefficient (
Kc).
ETc can be estimated from EO data using physics-based methods based on the surface energy balance [
15,
16,
17] or using empirical methods based on the use of vegetation indices [
10,
18,
19,
20]. Physics-based methods estimate latent heat flow through the surface energy balance, but the difficulties related to the measurement of its terms have led to a wider use of the empirical methods, in which the crop coefficient is obtained through the vegetation indices approach. The estimation of
ETc based on vegetation indices, usually the NDVI, is a modification of the crop coefficients method [
11,
12], in which
ETo is calculated based on meteorological data, and the crop coefficient introduces information related to the crop. In these methods, crop data are often obtained through tabulated values, which provide general data for several crops [
13,
14]. To improve the accuracy of crop water requirement estimation, the characterization of the
Kc curves must be improved and can be done using EO data, because crop characteristics are well correlated with the spectral reflectances [
21]. Thus, as proposed by several authors [
10,
18,
19,
20], the
Kc-NDVI approach establishes an empirical relationship between
Kc values obtained through field measurements and NDVI values retrieved from EO optical data. The equations used to estimate crop coefficient values based on vegetation indices, once calibrated and validated for a given location, can accurately estimate crop evapotranspiration [
10,
21].
This study aims to assess the potential of multitemporal and multisensor EO data for crop parameter (
Kcb values and the lengths of the crop’s development stages) retrieval and crop type classification at high spatial (10 m) and temporal (five days) resolution with a focus on irrigated agriculture. For this purpose, EO data (Sentinel-1A + SPOT-5 Take-5) are evaluated for irrigation requirement estimation based on a soil water balance model (IrrigRotation) [
22]. The main goals are to estimate crop parameters from NDVI and VV + VH backscattering time series and to calculate the crop irrigation requirements. On the other hand, the integration of Sentinel-1A VV + VH polarized data into the classification process is assessed and compared to the accuracies obtained only with spectral information from SPOT-5. This permits the determination of: (1) band combinations that provide better crop classification results; (2) the most critical dates for improved crop class discrimination; (3) and the number of observations required within a given growing season for a good classification accuracy.
4. Discussion
In the NDVI time series in
Figure 6, the maize crop decreased from the end of mid-season until harvesting (senescence stage). This results from the fact that most of the maize parcels were sown in September/October 2014, while the first EO acquisition date was 26 March 2015 (DOY 85). The same trend was also observed in the VV + VH backscatter time series, indicating that both types of EO data are correlated. Although only half of a crop growth cycle is retrieved from the EO data, the results are promising because they demonstrate that this correlation is also possible for the late season period.
Bean parcels, sown in late February 2015, exhibited higher NDVI values in the beginning of the curve corresponding to the leaf development stage (
Figure 6). At the senescence stage, the bean crop presented the same trend as that of maize. Because soybean was sown in December 2014, the lower NDVI values in the beginning of the curve are explained by the crop growth stage (senescence) and by the typical short height of soybean (approximately 0.45 m). Pasture showed a smoother curve for the three time series, in agreement with the fact that pasture is always in the leaf development stage due to it being cut several times during the growing season.
From the NDVI single-plot curves, it is possible to distinguish the two different stage plots for maize (one leaf development stage plot and one flowering stage plot) from the remaining 26 senescence stage plots. For bean, there is also a clear distinction among the three different phenological stages (one leaf development plot, one fruit stage plot and three flowering stage plots). However, using the VV + VH backscatter curves, it is not possible to distinguish among different stage crop plots because, due to the lack of the northern part of the Sentinel-1A images for three epochs, 14 maize plots and two bean plots were not considered for the analysis (the excluded plots included the leaf development and the flowering stage plots for maize and the leaf development and fruit stage plots for bean).
In
Figure 7, a significant correlation is observed, especially between the VV backscatter and the NDVI values, demonstrating the consistency of both optical and microwave time series and that optical data affected by clouds can be replaced by microwave data. This overcomes one of the main limitations of these type of studies,
i.e., the reduced amount of EO data due to meteorological conditions. This result agrees with those of [
41], in which the potential of different TerraSAR-X incidence angles and polarizations for mapping sugarcane harvests, where a high correlation between the radar signal and NDVI index, calculated from SPOT-4/5, was observed. However, our results differ from those obtained during the AgriSAR campaign [
42], in which gamma VH backscatter correlated strongly with NDVI for canola and field pea throughout the entire growth cycle. In AgriSAR, it was verified that among cereal crops, the correlation with NDVI is much weaker when analyzed over the entire growth cycle, exhibiting only a strong correlation for the initial vegetative growth stages until booting/inflorescence.
According to the overall accuracy and kappa coefficient values in
Table 7, the NN classification results were more accurate than those of the SVM classifier (~8% for the overall accuracy and ~15% for the kappa coefficient). SVM applied in [
43] for land cover characterization using MODIS time series data was compared to two conventional non-parametric image classification algorithms: NN and classification and regression trees (CART). SVM generated an overall accuracy of 80% compared to 76% and 73% for NN and CART, respectively. However, some other studies reported that NN outperformed SVM and decision trees [
44,
45,
46].
The addition of the gamma backscattering bands to the SPOT-5 optical bands did not reveal any improvement in the results of the SVM and NN classifications (
Table 7). In a previous study [
47] of the Lower Tagus Valley, the use of a Sentinel-1A VH band together with the optical bands of a Landsat-8/OLI image enabled only a slight increase in the overall accuracy (1.6%) and in the kappa coefficient (2.5%) compared to the results of an SVM classification of the optical bands. Moreover, when adding only the Sentinel-1A VV band or both SAR bands to the optical bands, the overall accuracies and kappa coefficients were always lower than those obtained under the best band combination. Moreover, in [
48], the addition of backscatter intensity derived from Radarsat-2 images to the surface reflectances derived from Landsat-8/OLI images for crop classification in Ukraine slightly improved the overall classification accuracy from 1.5% to 4.0%.
The use of multitemporal C-band Sentinel-1A images along with multitemporal SPOT-5 Take-5 images for crop classification in Wako-Kungo revealed an improvement of the overall classification accuracy when images are added cumulatively into the classification process. An overall accuracy of 91% (with a kappa coefficient of 81%) was obtained with a set of images from April 30 until June 4 (a total of 28 optical bands corresponding to the first 7 SPOT-5 Take-5 images of the time series), as expected, because they were acquired during and a few weeks after the field work when the crops were relatively vigorous. From that date on, classifications returned lower overall accuracies because there was a substantial decay in the NDVI and VV + VH backscatter values, indicating that the cultures are all in the last stage of senescence, having wilted and sometimes been already harvested.
In
Table 8, one can observe a significant misclassification between maize and pasture for the beginning of the time series, while a clear separation between soybean and bean seems to be evident, as expected from the analysis of
Figure 6. Pasture is the culture with the highest commission error (approximately 45%), meaning that the area occupied by this crop in the final map is significantly overestimated mainly due to misclassifications with maize, but also with soybean and bean. Bean has the highest omission error (approximately 23%), meaning that the area occupied by this crop in the final map is underestimated mainly due to misclassifications with maize.
ETo values present a small variability throughout the growing season, in agreement with the air temperature and solar radiation values shown in
Table 3, which present only small changes over time (
Figure 11).
ETc values decayed with the decreased
Kcb, consistent with the senescence stage shown by most crops. Only crop water requirement values are presented because during the late season, irrigation is interrupted for crop maturation. Additionally, in this study, the harvesting dates were unknown for all crop types. Thus, it was not possible to distinguish between the late season and the off season. Hence, without a through characterization of the complete growing season and of the cultural and irrigation practices in the region, it is difficult to accurately estimate the irrigation water requirements. From the analysis of the calculated volumes (
Figure 12), it is possible to verify the decrease in the values over time, again due to the senescence stage shown by most crops.
Better classification results would have been obtained if EO data were available for the entire crop growing season; thus, it would have been possible to estimate the crop irrigation requirements for the Wako-Kungo irrigation perimeter.
5. Conclusions
The purpose of this study was to assess the potential of multitemporal and multisensor EO data (Sentinel-1A and SPOT-5 Take-5) for crop parameter retrieval and crop type classification for agricultural water management in Angola.
The integration of microwave data (Sentinel-1A) with optical data (SPOT-5 Take-5) did not reveal any improvement in land cover mapping. From the two supervised classification methods used, NN had the highest accuracy of approximately 88% (with a kappa coefficient of approximately 73%) compared to that of the SVM classifier (85% and 68% for the overall accuracy and kappa coefficient, respectively). Higher classification accuracies were expected when using SAR data; however, the unavailability of images for more than half of the crop cycle (only the end of mid-season until harvesting was available) did not permit a better discrimination among crops. A multitemporal NN classification using just the first seven SPOT-5 Take-5 images produced a late mid-season map with an overall accuracy of 91% (with a kappa index of 81%).
The improved temporal resolution of the SPOT-5 Take-5 images, used in this study to simulate the ESA Sentinel-2 time series, is relevant for a better identification of the different crop growth cycle stages that are often imperceptible when using more sporadic data. Higher temporal resolution time series allow the retrieval of realistic values for
Kcb instead of the standard values proposed by FAO 56 [
13] and, consequently, a better estimation of the crop’s irrigation requirements. However, this aspect was not fully exploited for the same reason mentioned previously,
i.e., the lack of EO data for the complete growing season. Moreover, the consistency observed between the optical and microwave time series for all crop types enables the replacement of optical data affected by clouds with microwave data to increase the temporal resolution of the time series, providing a proxy measurement of crop development.