1. Introduction
Global-scale crop area mapping is important for tracking agricultural production and addressing issues relating to food security [
1,
2]. Conventional approaches for global crop mapping are based heavily on spaceborne approaches using multispectral optical sensors, such as the Moderate Resolution Imaging Spectroradiometer (MODIS), Landsat, Sentinel-2 and others [
2]. However, a majority of agricultural fields are just over two hectares in size, making moderate resolution platforms such as MODIS (250 m) unsuitable for mapping these smaller fields [
3]. Additionally, optical sensors such as Landsat and Sentinel-2 have a less frequent revisit than MODIS, and as such cloud cover can create large temporal gaps in the data record. Classifications can be less accurate when imaging opportunities are missed during critical crop growth periods [
4,
5].
Synthetic aperture radar (SAR) sensors offer unique abilities to assess agricultural landscapes due to their near-all-weather capabilities and sensitivity of microwave signals to the dielectric and structural characteristics of soils and crops (i.e., dielectric constant, roughness, orientation and density of canopy), in contrast to those derived from optical instruments [
6,
7]. Historically, space-based crop monitoring has relied primarily on the exploitation of optical data. The limited use of SAR technologies in operational mapping of soils and crops may be at least partially attributable to: (a) insufficient access to free and open SAR data; (b) lack of large-area acquisition strategies at appropriate scales; (c) poor quality digital elevation models required for processing; (d) complex data structures relative to optical data; and (e) the lack of standardized workflows for SAR data processing for land applications. The increase in the availability of SAR data from Sentinel-1A and 1B has created new opportunities for integrating moderate resolution SAR into operational crop mapping. These data are needed in cloud-prone regions, as well as during critical growth stages, to mitigate image gaps and establish robust monitoring programs [
8,
9,
10].
The optimal SAR frequency for crop identification is canopy dependent. For the sensor to be able to detect crop-related information, there must be a balance between: (1) sufficiently deep penetration of the microwave into the canopy as to permit scattering within the canopy elements; and (2) a sufficiently shallow penetration into the soil, to ensure that the scattering contribution from the soil is relatively small compared to the vegetation. Shorter C-band wavelengths are well suited to smaller biomass canopies as the scattering will originate primarily from within the canopy. However, when significant biomass accumulates due to crop type or growth stage, classification benefits from the data collected at longer L-band wavelengths [
11]. Classification accuracies are improved with the integration of SAR data acquired at multiple microwave frequencies [
7,
12]. Unfortunately, consistent standard coverages of L-band SAR data over global agricultural regions are more limited compared to C-band.
The NASA and Indian Space Research Organization (ISRO) SAR mission, NISAR, is an upcoming satellite mission that will collect data in the L- and S-bands. It is anticipated to launch in late 2022 or early 2023. NISAR will have an open data access policy that will allow for the wider use of L-band SAR for agricultural monitoring. NISAR will operate at a frequency of 1.26 GHz (25 cm wavelength) and have a mean revisit time of 12 days. NISAR science applications will cover a range of different domains. Of specific interest are data products relevant to agriculture such as crop area mapping, soil moisture retrievals and biomass estimation, all of which will be conducted in close collaboration with the United States Department of Agriculture (USDA), the Group on Earth Observations Global Agricultural Monitoring (GEOGLAM) through its underlying research and development group, the Joint Experiment for Crop Assessment and Monitoring (JECAM), and others [
13]. NISAR has a science requirement of being able to make crop area estimates at a 1 hectare resolution. This product will be validated over JECAM calibration-validation sites (
http://jecam.org/) with an intended classification accuracy exceeding 80% [
13].
The NISAR science algorithm for determining crop area consists of a temporal coefficient of variation (CV) approach that has already been applied and described in prior studies. These studies have included the use of SAR data acquired in the C- and L-bands, as well as simulated NISAR L-band data from airborne systems [
14,
15,
16,
17]. One of these studies compared biweekly airborne C- and L-band SAR data collected during the growing season (April–August) over an agricultural area in Germany, as part of the 2006 AgriSAR campaign [
14]. Using the CV approach, these researchers reported crop classification accuracies of 87% and 79% for the L-band and C-band data, respectively [
14]. This is the only study to date which has tested a dense C- and L-band time series for estimating crop area using the temporal CV approach. It is expected that crop area classification accuracies will vary by region of interest (ROI) depending on the relative prevalence of each crop and non-crop class within the ROI, as well as variations in the landscape due to climate, management practices, soils, and terrain. Backscatter can be expected to vary among crop types, and even from field to field planted with the same crop, depending on planting densities, growth stages, crop cultivars, row orientations, and how individual fields are managed (tilling, irrigation, fallowing) [
6,
18,
19].
With the impending launch of NISAR it is imperative to conduct more extensive evaluations of the NISAR Level 2 Cropland Area science algorithm, the temporal CV approach. The objectives of this study are to conduct further testing of the temporal CV approach, specifically with respect to: (1) providing first estimates at the NISAR calibration-validation site in southwestern Canada called Carman; (2) making comparisons of the temporal CV approach between Sentinel-1 C-band and PALSAR-2 L-band SAR data, to help elucidate the differences and synergies between the two frequencies; and (3) to make comparisons between in situ biomass data to the CV values obtained using Sentinel-1 and PALSAR-2 data.
Differences in the performance of L- and C-band SAR are expected among ROIs given differences in crop and non-crops mixes, variations in management practices and different cropping seasons. For example, L-band frequencies may tend to underestimate crop acreages because the CV over low-biomass fields may be small, leading to misclassifications of crop to non-crop. On the other hand, while the C-band would be better at detecting low biomass crops, this higher frequency signal may saturate for high biomass crops, yielding smaller CV values, and leading to classifications as a non-crop. This study examines backscatter response for a single growing season, and as such, CV values may also provide information on crop biomass. For example, CV values are expected to be larger for crops where biomass accumulation over the season is more substantial. Consequently, the magnitude of the CV may inform field-to-field differences in crop biomass.
The innovations of this study are specific to the temporal CV approach and lie in: (1) assessing whether NISAR’s accuracy requirement of 80% can be met at Carman, when using PALSAR-2 L-band data as proxy; (2) determining how classification accuracies vary by crop and frequency; and (3) to use in situ biomass data to examine whether temporal CV values are consistent with crop biomass, using the hypothesis that substantially lower (higher) biomass crops will have substantially lower (higher) temporal CV values.
5. Conclusions
This study presents first results of the temporal CV approach, an algorithm to be used for NISAR’s Cropland Area product, over a NISAR calibration-validation site located near Carman in Manitoba, Canada. We employed both C- and L-band SAR data from the Sentinel-1B and ALOS-2 PALSAR-2 satellites respectively to generate a crop area estimate. Each pixel was classified as crop if its temporal CV value exceeded a threshold value. The optimal threshold value used in this study was determined using a receiver operating curve approach, which is robust and yielded close to the maximum possible accuracy when using CV values. Evaluations were performed against the Annual Cropland Inventory (ACI), which contains detailed land cover classification on 30 m × 30 m pixels and includes a wide range of crop and non-crop classes. Comparisons show that crop area estimates were considerably better when using the C-band (84%) compared to the L-band over Carman (74%). A more detailed look at the classifications by ACI class revealed that the L-band classifications performed poorly (< 60%) due to classifying many soybean fields as non-crop, and many of the major non-crop classes (urban, grassland, and pasture) as crop. Whereas limiting factors for Sentinel-1 accuracy were relatively poor performances over urban (53%), pasture (65%), and corn (73%). Thus, both frequencies are useful for cropland classifications, and performance over a given region will mainly depend on the crop and non-crop types and their relative prevalence within the ROI. Lastly, because we noted that CV values showed large variation by fields when using PALSAR-2 data, we also sought out available in-situ biomass data to provide further context. Comparisons of CV values to in-situ biomass data collected a eight different times and in nine fields (five canola, four soybeans) revealed that the crop with substantially lower biomass (soybean) also had substantially lower CV values in both the C- and L-bands. This is an interesting result and speaks to the potential opportunity in also using the CV approach for making biomass estimates in addition to computing crop area.
This work, like others but using NISAR’s Level 2 Cropland Area science algorithm (temporal CV), demonstrates the added value of using both C- and L-band SAR data over agricultural areas. L-band retrievals provided added value compared to C-band over corn (97% vs. 73%), whereas C-band data provided added value over soybeans (86% vs. 57%). We also showed that this approach is acceptable for making crop and non-crop classifications with L- and C-band data, but not for crop classifications due to many land cover types having comparable CV values. This and other studies had already indicated that the temporal CV approach appears to be robust and fairly accurate for crop and non-crop classifications. Future work should build on the initial results reported here that CV values might also be useful for making estimates of agricultural biomass and, eventually, crop yields. Such studies will be of particular relevance as open source L-band SAR data becomes available through NISAR in 2022/2023, which will enable frequent and global scale, multi-frequency mapping of agricultural parameters such as crop area, but potentially also crop yields and soil moisture at sub-field scales.