1.1.

Rationale for This Research: Advances Over Existing State of Knowledge

Existing cropland mapping methods, including the previously developed ACCA, operate on an annual basis. For example, ACCA reported by Thenkabail and Wu¹⁴ computed irrigated and rainfed croplands extents and areas at the end of the year based on remote sensing data such as Moderate Resolution Imaging Spectroradiometer (MODIS) and Landsat and various secondary data. Yet, agricultural food security policies often call for cropland products during various stages of the crop growing season. For instance, during the recent severe drought across the western U.S., there was a constant demand for cropland data throughout the growing season from policy makers, resource managers, and researchers.¹⁵ Increasing climate variability poses further uncertainty in agriculture productivity and seasonality in different regions across the world, and therefore the need for more frequent cropland products throughout the growing season from various end-users cannot be overemphasized. Thereby, the first advance in this research is to develop ACCA to produce seasonal (e.g., month by month during the entire growing season) cropland areas and extents.

Further, even though numerous studies have focused on mapping cultivated cropland (referred to as “cropland” hereafter) across spatial and temporal scales,^{6–8,16–21} few have attempted to delineate the extent of fallow croplands (referred to as “fallowland” hereafter). While fallowland statistics are available for some places, it is widely accepted that they have very high uncertainty associated with them, since these fallowland extents/areas have not been systematically studied. The fallowlands consist of complex spectral signatures (e.g., barren, sparse grass/shrub cover, weed are present), have widely varying temporal dynamics, and lack field observations and reference data. Yet, fallowlands have great implications on water use, food production, nutrient cycling, and carbon sequestration and therefore climate change.²² The agricultural practices on fallowlands have large impacts on soil organic carbon and ${\mathrm{CO}}_2$ emissions, economic subsidies, and potential environmental impacts, such as erosion, ground water contamination, and trace greenhouse gas productions.^23–25 Mapping of fallowlands and capturing its temporal dynamics throughout the growing season can contribute to an improved understanding of the potential for different management practices to mitigate the negative effects of drought-related cropland fallowing. For example, shortage of water for irrigation and crop production is one of the major impacts of drought in the heavily cultivated Central Valley in California, and therefore timely and accurate information on fallowland acreage is extremely useful in identifying the extent of changes in fallowed acreage due to water shortage during drought and guiding decision making with respect to requests for local water transfers, county drought designations, or state emergency proclamations. Therefore, routine and precise mapping of cropland and fallowland distributions in California has great implications on water use and crop water productivity, which can be achieved in an automated fashion especially under the projected more severe drought for the western U.S.⁴ Thereby, the second advance in this research is to build production of fallow cropland extents and areas in the ACCA.

The overarching goal of this study was to develop a seasonal ACCA using MODIS 250-m remote sensing data to produce two unique products: (1) seasonal cropland extents and areas and (2) seasonal fallowland extents and areas. Even though earlier ACCA developed by Thenkabail and Wu¹⁴ computed cropland extents and areas, it was limited to annual products and did not produce fallowlands; both of which are serious limitations. The seasonal ACCA is developed to have automated, rapid, and accurate reproduction capability. State of California was chosen as a pilot region for reasons discussed earlier. Preliminary stakeholder requirements for fallowland acreage monitoring in California indicated that the uncertainty of $\pm 25\%$ may be tolerable (NIDIS California Pilot web conference, 2012), given the current prevailing lack of fallowland monitoring capability. Meanwhile, user requirements for cropland accuracy are vague or are still under development.

Seasonal ACCAs were developed for 5 months throughout the growing season, including June, August, September, October, and December 2012, and tested using independent datasets from June, August, September, and December 2011. Cropland and fallowland mapping accuracies were also evaluated based on field surveys in April, May, June, September, and October 2012. Cropland areas of the 58 counties of California that were derived from the ACCA were also evaluated using independent data sources from farmer-reported surveys [e.g., U.S. Department of Agriculture Farm Service Agency (USDA FSA)]. Given the increasing uncertainty of cropland dynamics in a changing climate, mapping of cropland extent and area on a monthly basis throughout the growing season will be of great value in assessing drought impacts on agricultural production as well as patterns in short- and long-term fallowing of farmland.

Historically, though the use of sensors such as those carried on the Landsat missions can be problematic in areas with high-cloud cover during the growing season, production of fine-resolution cropland maps relied on 1- to 30-m resolution remote sensing imageries. Recent work on cropland classification methods has utilized data from the MODIS instrument, which provides an alternative for monitoring cropland areas ^{8,17,20,21,26–34} across multiple temporal and spatial scales due to its high-temporal coverage, moderate spatial resolution, and high-quality time-composite products that largely resolve the cloud issues. Thus, in this research, MODIS 250-m data were used to produce seasonal cropland and fallowland extents and areas.

2.1.

Study Area

The state of California, USA (32°N to 42°N and 114°W to 124°W) was used as the study area to develop, test, and implement a seasonal ACCA. California is one of the most productive agricultural regions in the world, where its agricultural exports play a major role in the state’s economy.³⁵ California produces a variety of crops, some of which are unique commodities grown only in California. According to the USDA National Agricultural Statistics Service (NASS) California Field Office, in 2011, the state’s 81,500 farms and ranches generated record $43.5 billion products revenue, up from the $38 billion reached during 2010, and therefore California remained the top state in cash farm receipts as it comprised 11.6% of the U.S. total.³⁶ According to the 2012 NASS CDL for California, about $\mathrm{40,000}{\mathrm{km}}^2$ (9.4% of California’s total area of $\mathrm{424,000}{\mathrm{km}}^2$) was cropland. Farming accounts for about 84% of all human water use in California with the majority of California’s cropland being irrigated.³⁷ The climate is characterized by winter/spring precipitation followed by summer drought. Thus, fallowlands may support weeds or other rainfed volunteer vegetation earlier in the year, but during summer, they are typically very low green vegetation fractions, or barrens, due to the lack of summer rainfall and irrigation.

2.2.

Remote Sensing Data

An ACCA for California was developed using MODIS 250-m data for the growing season of 2012. Unlike a previous ACCA in Tajikistan, where a combination of MODIS and Landsat data were used,¹⁴ only MODIS time series data were used here, because Landsat Thematic Mapper 5 (TM5) data were discontinued in 2012 and scanline issues associated with the Landsat 7 Enhanced TM (ETM) made the data difficult to use for this application. MODIS normalized difference vegetation index (NDVI) time series have been used widely and proven to be efficient in mapping cropland.^{8,17,20,21,27,29–31,34} In this study, MODIS data were solely used to develop the ACCA, and the approach leveraged the frequent temporal coverage provided by MODIS for the seasonal mapping of cropland. In addition, the MODIS-based case study of California in this research can have great implications for implementation of such automated cropland mapping methods to other areas around the world, where higher temporal resolution of MODIS compared with Landsat largely increases the probability of acquiring frequent cloud-free images.

MODIS Terra surface reflectance 8-day composite level 3 (L3) Global 250-m data were obtained through the National Aeronautics and Space Administration (NASA) Earth Observing System Data and Information System (EOSDIS, http://earthdata.nasa.gov/) for the years 2011 and 2012. This dataset is a L3 composite with each pixel containing the best possible L2 observation during an 8-day period, as selected on the basis of high-observation coverage, low-view angle, the absence of clouds or cloud shadow, and aerosol loading.³⁸ The L3 products have been atmospherically corrected for atmospheric scattering and absorption from atmospheric gases and aerosols, and therefore the resulting surface reflectance is as it would have been measured at ground level. In addition, the L3 products have also been geometrically corrected. All L3 data have been geolocated into a specific map projection, and the gridding process converted the input observation space to the output geometrically correct level.³⁹ We obtained all the available 8-day composite data from Jan. 1, 2011, to Dec. 31, 2012. Two bands (band 1 and band 2) of the MODIS surface reflectance data were used to calculate NDVI by using the equation: $\mathrm{NDVI}=(\text{band}2-\text{band}1)/(\text{band}2+\text{band}1)$, where bands 1 and 2 are red and near-infrared bands, respectively. The NDVI values, which ranged from $-1$ to $+1$, were then converted to 8-bit scaled NDVI ranging from 0 to 255 by using the equation: Scaled $\mathrm{NDVI}=\text{Calculated}\mathrm{NDVI}\times 127.5+127.5$. Upon generating the scaled NDVI for all MODIS 250-m 8-day composites, a monthly maximum value composite was applied by taking the maximum NDVI value for each pixel from all 8-day composites within the same month, resulting in one maximum MODIS NDVI composite for each month. All monthly maximum MODIS NDVI composites within the same year were then stacked together into a data cube for the ACCA development.¹⁴

2.3.

Automated Cropland Classification Algorithm Development

The process of the ACCA development first involves the generation of sets of rules using MODIS NDVI data cube to derive a cropland layer that matches a reference map. For California and the U.S., the USDA annual CDL provides an ideal reference that includes accuracy statistics for all crop classes included in the map. Contrary to the previous ACCA developed in Tajikistan,¹⁴ where the reference cropland layer had to be first created using laborious classification methods, in this study, routinely available CDLs with high accuracies ($\sim 80\%$ averaged across crops) were available from NASS for the state of California each year since 2007.

The total cropland areas were comprised of areas cultivated in a given year and areas left fallow (referred to as “fallowland” and looked similar to bare soil or very low-cover vegetation) in the same year. We obtained the year-end CDLs from the NASS CropScape portal ( http://nassgeodata.gmu.edu/CropScape/) from 2007 through 2011 to create a 5-year cropland mask including both cultivated and fallow areas, representing the maximum potential cultivated and fallow cropland areas occurred in California within the 5 years, where ACCA can be applied to delineate cropland from noncropland as well as cultivated versus fallow cropland. Within the 5 years, about 20% of the croplands were left fallow in any given year, although the spatial locations of these fallowlands, typically, changed from year to year. We also obtained monthly fallowland data products throughout the growing season from June to December of 2012 from NASS, from which monthly simulated cropland layers (SCLs) were generated as the remaining area from the 5-year cropland mask. Upon resampling the MODIS data cube to 30-m resolution as the SCL, the thresholds of ACCA rules were determined by trial and error, where a threshold of a rule was set when $\le 10\%$ error of omission was achieved comparing with the reference SCL. During the ACCA development process, one single rule can only map a certain portion of the total cropland area compared with the reference cropland layer, and additional rules were then written to delineate the unresolved cropland left over from the previous rule sets; therefore, no overlap of cropland area occurred among multiple rule sets. Eventually, all rule sets were compiled together into one single algorithm when over 90% of the total cropland from the ACCA-derived cropland layer matched with the reference cropland layer pixel-by-pixel. The seasonal ACCA delineated actively cultivated areas and left aside the fallowlands within the total cropland areas. Thus, both cropland and fallowlands were established throughout the growing season. All algorithms coding and testing were conducted in the ERDAS Imagine 2011 Modeler.

2.4.

Accuracy Assessment

In this research, three unique approaches of accuracy assessments were adopted to ensure the strength of the seasonal ACCA.

2.4.3.

Ground-based field surveys: field data validating remote sensing products

In the early growing season of 2012, 10 east-west field transects were established in California, spanning horizontally across the major cropland areas in California (Fig. 1). The transects ranged in distance from 20 to 54 km across the Central Valley. Data were collected for every field along each transect based on visual inspection of each field at a location on the edge of the field adjacent to an access road. Digital photos were taken for each field, and descriptive information was recorded including geographic location (latitude and longitude), initial subjective classification (e.g., crop and idle at present time), bare soil condition (e.g., present or not, tilled, beds shaped, irrigated wet soil, and flooded), weed condition (e.g., present or not, color, fractional cover, height, and type), cover crop condition (e.g., present or not, color, fractional cover, height, and type), and crop condition (e.g., present or not, type, structure, color, and condition, including emergent, growing, senescent, residue, recently harvested, fractional cover, height, and irrigation system type). Data were entered into a geographic information system database that linked geo-registered field locations and associated attributes collected during the field surveys. Two early season field surveys were conducted for each transect (April/May and May/June) and sampled all fields along each transect route. Since the original surveys contained a very small percentage of fields that did not have a clearly established crop, two late season field surveys were conducted in September and October. Due to logistical constraints, these surveys focused on a subset of fields identified as fallow by either the NASS or ACCA fallowland maps for August 2012. A general principle of ground data collection for accuracy assessment was to collect 100 samples per classification category (e.g., cultivated or fallow cropland) when dealing with a large area.⁴¹ In order to meet such sampling criteria for accuracy assessment, we sampled over 1500 fields across the 10 transects throughout the growing season of 2012. We used the early season field surveys to assess the accuracy of ACCA-derived cropland map, and late season field surveys to validate ACCA-derived fallowland map of 2012.

Fig. 1

Early seasons field transects for mapping cropland and fallowland in California. The cropland and fallowland maps are from the NASS Cropland Data Layer (CDL) of 2012. Late season sampling was concentrated in the fallowland region near 35N/120W.

3.1.

Seasonal ACCA for Cropland of California

Seasonal ACCAs were developed throughout the growing season of 2012 for the months of June, August, September, October, and December based on the reference SCLs for the corresponding months (e.g., Fig. 2 illustrates the seasonal ACCA for the month of August 2012). The algorithm was designed to detect the presence of crops, as indicated by elevated NDVI or standalone near-infrared reflectance. The algorithm development process started with using the yearly total NDVI to delineate cropland, followed by using NDVI from critical months out of the year to map seasonal croplands (e.g., Fig. 2). The algorithm for each month followed a similar structure, but the thresholds of the rule sets varied as the growing season progressed and more monthly MODIS NDVI data layers were incorporated into the dataset. The ACCA for the month of August (Fig. 2), for example, involved MODIS NDVI data for the months of January through July. As the growing season progressed, more monthly MODIS NDVI data layers were used in the algorithm in the subsequent months to more rigorously map cropland. In addition, the band 2 (near-infrared) surface reflectance during July was also used in the algorithm (except for June ACCA). During the algorithm development process, the cumulative monthly NDVI was the most effective in delineating the largest amount of cultivated area and set as Rule 1 (Fig. 2). In the August ACCA (Fig. 2), for example, Rule 1 using monthly total NDVI from January to July delineated 56% of the total cropland area. Rules 2 and 3 used monthly total NDVI as well as individual monthly NDVI of critical months during the growing season to delineate additional cropland area that was not captured by Rule 1 (Fig. 2). Rules 4 and 5 used seasonal (e.g., April to July and January to April) cumulative NDVI as well as MODIS band 2 surface reflectance of July to delineate the remaining cropland (Fig. 2).

Fig. 2

Sample automated cropland classification algorithm (ACCA) of August 2012 for the state of California. Total NDVI is the additive sum of monthly NDVI, and MODIS surface reflectance has a scale factor of 0.0001.

3.2.

ACCA-Derived Cropland Maps

We applied the seasonal ACCAs (June, August, September, October, and December) to the MODIS monthly NDVI data cube of 2012 (Sec. 2.2) and generated cropland layers for the corresponding months. To avoid repetition, only cropland output for August was illustrated with its algorithm shown in Fig. 2. The ACCA-derived cropland map for August 2012 agreed well with the August SCL, demonstrated by an overwhelmingly overlapped cropland area captured both in the ACCA-derived cropland map and SCL [Fig. 3(a)]. There were some fractional areas that were captured only by ACCA in the northern Central Valley, and some other isolated areas only present in SCL in the southern Central Valley [Fig. 3(a)]. The August ACCA (Fig. 2) was then applied on an independent MODIS monthly NDVI data cube from 2011, and the output was compared with the August SCL of 2011 [Fig. 3(b)]. Similar to the one in 2012 [Fig. 3(a)], the ACCA-derived cropland map of August 2011 showed substantial agreement with SCL of August 2011 [Fig. 3(b)]. Compared with the results from August 2012 [Fig. 3(a)], there were slightly more cropland areas that were only present in the ACCA-derived cropland map in the northern Central Valley and very minimal cropland areas that were solely in the SCL throughout the state in the ACCA-derived cropland map of August 2011[Fig. 3(b)].

Fig. 3

Comparison of ACCA-derived cropland maps and SCLs of California for (a) August 2012 and (b) August 2011. Common cropland areas in both ACCA-derived croplands and SCLs, areas only in ACCA-derived croplands, and areas only in SCLs were shown here.

3.3.

Accuracy Assessment for ACCA-Derived Cropland Maps

3.3.2.

Area-based comparison with ground truth data at the county level

USDA FSA has produced county-level crop acreage data on a monthly basis throughout the growing season since August 2011. County-level cropland area was compared between the ACCA-derived cropland map and the FSA county crop acreage data. We illustrated here the linear regression between ACCA-derived cropland area and crop acreage data for December 2012 for the 58 counties of California with a slope of 1 [Fig. 4(a)]. Linear regressions showed that the ACCA agreed well with the crop acreage data over the growing season (Table 4), although ACCA tended to overestimate the cultivated area for September and October 2012 (Table 4), since not all farmers participate in the crop acreage data collection program throughout the growing season. We also compared ACCA-derived fallowland areas to the farmers reported fallow acreage at the county level in 2012. The vast majority of the fallowland of California was located within the Central Valley, where fallowland acreage was consistently reported in the FSA dataset. $R$-square values of linear regressions between ACCA-derived fallowland areas and FSA crop acreage data fallowland areas during the growing season of 2012 were all above 0.72 and the slopes were above 1, indicating that the ACCA overestimated the fallowland area consistently (Table 4).

Fig. 4

The ACCA-derived county-level cropland areas of (a) December 2012 and (b) December 2011 for California, in comparison with cropland areas derived from crop acreage data from the Farm Service Agency (FSA) for the corresponding months.

Table 4

Summary of linear regressions of ACCA-derived county-level cropland and fallowland (Central Valley counties) areas against the crop acreage data from the FSA throughout the growing season of 2012 (year during which ACCA was developed) and 2011 (independent year).

	Cropland 2012		Cropland 2011		Fallowland 2012		Fallowland 2011
Month	Slope	$R$-square	Slope	$R$-square	Slope	$R$-square	Slope	$R$-square
August	1.07	0.70	1.35	0.75	1.24	0.72	3.37	0.85
September	1.26	0.70	1.24	0.76	2.38	0.77	3.15	0.77
October	1.46	0.72	1.22	0.77	2.54	0.79	3.07	0.79
December	1.00	0.72	1.19	0.77	2.62	0.75	3.39	0.80

To test the repeatability of ACCA-generated cropland area statistics, we applied the ACCA of December 2012 to the MODIS data cube of 2011 to derive a cropland map of December 2011 and compared the county-level statistics of ACCA-derived cropland area of December 2011 versus the crop acreage data of December 2011 from the FSA. The linear regression showed an $R$-square of 0.77, and ACCA-derived cropland area tended to overestimate the reported county-level cropland area from the farmers [Fig. 4(b)]. Throughout the growing season of 2011, the county-level cropland area statistics agreed well with the crop acreage data with the $R$-square values of the linear regressions above 0.75 for all months (Table 4). The county-level fallowland areas showed a good agreement with the FSA crop acreage data fallow area with $R$-square values of linear regressions over 0.77 across all months. Similar to the results from 2012, ACCA tended to overestimate the cultivated and fallow area with the slopes above 1 (Table 4), which could be due to the data collection process of the crop acreage data, where not all farmers were included. For cropland, as the growing season progressed from August to December in 2011, ACCA more accurately captured the cropland area from the crop acreage data, demonstrated by a progressively smaller slope (closer to 1) and larger $R$-square value (Table 4).

3.3.3.

Field survey-based accuracy assessment

Two groups of field surveys were conducted throughout the growing season of 2012 including the early season (April to June) and late season (September to October). We used the early season “crop-present” survey fields to assess the accuracy of the ACCA-derived cropland map of August 2012, and late season “fallow” fields to assess the accuracy of the ACCA-derived fallowland map of October 2012. The early season field surveys covered the heavily cultivated Central Valley area including all major crops [Fig. 5(a)]. The ACCA-derived cropland map for August 2012 showed a very high producer’s accuracy of 95% and user’s accuracy of 96% (Table 5). The ACCA-derived fallowland map for October 2012 agreed well with the field survey with a producer’s accuracy of 84% and user’s accuracy of 76% (Table 5) based on 210 surveyed fields, where higher concentrations of fallowland was located in California [Fig. 5(b)]. Greater uncertainty in fallowland extent compared with the cultivated area was expected, given its complex spectral composition and temporal dynamics. Some level of class confusion between cultivated and fallow croplands resulted from spectral insensitivity. For example, newly established plantations on a previously fallowland [see the bottom right field survey photo with a zoom-in inset in Fig. 5(b)] typically have very low fractional cover and were misclassified as fallowland by the ACCA, due to the low proportion of green vegetation. Conversely, some fallowland fields identified in the field surveys had sparse to medium levels of green shrub cover [see the upper right field survey photo in Fig. 5(b)], whereas ACCA defined fallowland as cropland areas exhibiting bare soil-like spectral signatures (lower NDVI). Furthermore, agriculture in California is characterized by homogenous cultivated fields [Fig. 5(a)], and therefore the MODIS-based ACCA worked very well in mapping cultivated area, but the relatively coarse spatial resolution of MODIS data compared with the survey fields may not be sufficient to delineate smaller and/or isolated fallowlands. Yet, ACCA can still provide valuable information for assessing drought impacts in agricultural regions with accuracy standard of $\pm 25\%$ or better for mapping fallowland extent for various months within the growing season.

Fig. 5

Comparison of (a) early season (April to June) and (b) late season (September to October) field surveys with ACCA-derived (a) cropland map of August and (b) fallowland map of October 2012. Example field survey photos for cultivated and fallow fields are presented.

Table 5

Summary of field survey-based accuracy assessments (# of fields) of ACCA-derived cropland and fallowland maps of 2012.

		ACCA-derived cropland (August 2012)
		Crop	Total	Producer’s accuracy
Field survey (early season)	Crop-present	1233	1305	95%
	Total	1278
	User’s accuracy	96%

		ACCA-derived fallowland (October 2012)
		Fallow	Total	Producer’s accuracy
Field survey (late season)	Fallow	160	191	84%
	Total	210
	User’s accuracy	76%