Implications of changing spatial dynamics of irrigated pasture, California's third largest agricultural water use

Science of the Total Environment 605–606 (2017) 445–453

Contents lists available at ScienceDirect

Science of the Total Environment journal homepage: www.elsevier.com/locate/scitotenv

Implications of changing spatial dynamics of irrigated pasture, California's third largest agricultural water use Matthew Shapero a,⁎, Iryna Dronova b, Luke Macaulay a a b

Department of Environmental Science, Planning, and Policy, University of California at Berkeley, Berkeley, CA 94720, United States Department of Landscape Architecture and Environmental Planning, University of California at Berkeley, Berkeley, CA 94720, United States

H I G H L I G H T S

G R A P H I C A L

A B S T R A C T

• Irrigated pasture (IP) is being converted to other land uses across California. • The accuracy and precision of current land cover metrics that classify IP is poor. • A new methodology is developed that improves the process of classifying IP. • High-resolution imagery and objectbased image analysis improves classification. • Continued loss of IP will likely have broad social and environmental consequences.

a r t i c l e

i n f o

Article history: Received 25 April 2017 Received in revised form 8 June 2017 Accepted 8 June 2017 Available online xxxx Editor: D. Barcelo Keywords: Remote sensing Object-based image analysis (OBIA) Land-use/land-cover (LULC) NLCD Ranching Social-ecological services

a b s t r a c t Irrigated agriculture is practiced on 680 million acres worldwide. Irrigated grazing land is likely a significant portion of that area but estimating an accurate figure has remained problematic. Due to its significant contribution to agricultural water use worldwide, we develop a methodology to remotely sense irrigated pasture using a California case study. Irrigated pasture is the third largest agricultural water use in California, yet its economic returns are low. As pressures mount for the agricultural sector to be more water efficient and for water to be directed towards its most economically valuable uses, there will likely be a reduction in irrigated pasture acreage. A first step in understanding the importance of irrigated pasture in California is establishing a methodology to quantify baseline information about its area, location, and current rate of loss. This study used a novel object-based image analysis and supervised classification on publicly-available, high resolution, remote sensing National Agriculture Imaging Program (NAIP) imagery to develop a highly accurate map of irrigated pasture in a rural county in California's Sierra foothills. Irrigated pasture was found to have decreased by 19% during the ten-year period, 2005–2014, from 4,273 to 3,470 acres. The implications of this loss include potential impacts to wetland-dependent species, groundwater recharge, game species, traditional ranching culture, livestock production, and land conservation. Overall accuracy in classification across years was consistently over 89%. Comparing these results against available measurements of irrigated pasture provided by state and federal agencies reveals that this method significantly improves upon existing metrics and methods of data collection and points to critical needs for new targeted research and monitoring efforts. Broadly, the analysis presented here provides an improved methodology for mapping irrigated pasture that can be extended to provide accurate and spatially-explicit data for other counties in California and other arid and semi-arid regions worldwide. © 2016 Elsevier B.V. All rights reserved.

⁎ Corresponding author at: University of California, 130 Mulford Hall #3114, Berkeley, CA 94720, United States. E-mail address: [email protected] (M. Shapero).

http://dx.doi.org/10.1016/j.scitotenv.2017.06.065 0048-9697/© 2016 Elsevier B.V. All rights reserved.

446

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

1. Introduction Irrigated agriculture is practiced on 680 million acres worldwide. While this accounts for only 20% of cultivated land, these areas provide 40% of the total food produced globally (FAO, 2016). Although studies have attempted to model global consumptive water use of irrigated grazing land (Rost et al., 2008; Postel, 1998), estimating an accurate figure has remained problematic, in part because of the large uncertainties in the global data sets distinguishing croplands and pasturelands in mosaic landscapes (Hannerz and Lotsch, 2008). Due to its likely significant contribution to agricultural water use worldwide, we develop a methodology to remotely sense irrigated pasture using a California case study. California's recent drought (2013–2017) has brought renewed attention to the state's complex and often convoluted system of surface water distribution (Grantham and Viers, 2014). In particular, the irrigation practices of the agricultural sector—which consumes nearly 80% of the state's annual, non-environmental surface water flow (Mount et al., 2014; DWR, 2010)—have been strongly criticized as inefficient (Famiglietti, 2014; AghaKouchak et al., 2015). As agriculture braces for continued cutbacks in surface-water allocation, rising water prices, and requirements to transition to water-efficient irrigation systems or crops (Howitt et al., 2015), the long-term economic and land-use impacts of these changes remain uncertain. Irrigated pasture—or, pastureland artificially irrigated during California's dry summers—is used to graze livestock and is a significant source of forage for the state's cattle industry: without access to irrigated pasture or other expensive supplemental feeds in the dry season, livestock lose body condition due to the low nutritional content of non-irrigated vegetation available on California's rangelands. Irrigated pasture, however, is one of the state's most water intensive and least profitable crops. The most recent figures from the California Department of Water Resources show that irrigated pasture ranked third among crops statewide in amount of water applied (DWR, 2010), but the California County Agricultural Commissioners' Report for 2014– 2015 ranks its agricultural gross value only 52nd out of 70 commodity crops (CDFA, 2016). If considered in purely economic terms, fallowing irrigated pasture or converting it to another crop represents the most rational response to rising water costs and restrictions (Sunding et al., 1997). Yet irrigated pasture is a critical land resource (Huntsinger et al., 2017; Richmond et al., 2010; Earl, 1950); losing substantial acres of it across the state would have broad agricultural, environmental, and land-use consequences that economic analyses to date have not fully considered. A crucial first step in understanding irrigated pasture and its potential reductions in California is measuring its statewide spatial extent, distribution, and current rate of loss. In recent decades, remote sensing has emerged as an indispensable tool in generating land-use/land-cover (LULC) maps (Burkhard et al., 2012; Verburg et al., 2009; Friedl et al., 2002). Analyzing land cover changes across time has allowed environmental managers to measure ecosystem and natural resource dynamics with greater precision (Guida-Johnson and Zuleta, 2013; Sohl et al., 2012; Foley, 2005). The power of analysis, however, has been limited by trade-offs between the spatial and temporal resolution of the available imagery (Ellis et al., 2006). Furthermore, the varied patterning and diverse settings of irrigated pasture across the California landscape offers real challenges in remote identification. As such, this analysis seeks to identify the best ways to address and reconcile these challenges. Currently, a number of county, state, and federal agencies are attempting to measure the acreage of irrigated pasture in California. The California Department of Food and Agriculture (CDFA), (2015) keeps non-spatial records based upon producer self-reporting to county Agricultural Commissioners' offices. The California Department of Water Resources (DWR) provides crop maps for counties across the state. Both the National Agricultural Statistics Service (NASS) and the

National Land Cover Database (NLCD) keep spatially-explicit records of agricultural land types statewide. NASS also conducts the Census of Agriculture every five years, which gathers county-based estimates through self-reporting (USDA National Agricultural Statistics Service, 2014). All of these sources, however, have different but critical limitations. The CDFA and Census of Agriculture data are not spatially explicit, and they depend upon producer participation statewide as well as producers' accurate knowledge of the number of acres they irrigate. The DWR data are spatially-explicit but requires time-consuming, in-person field work to draw crop maps. Finally, the NASS Cropland Data Layer (CDL) and NLCD spatial datasets have two limitations as they relate to irrigated pasture: 1) both classifications use imagery at course resolution, which limits their accuracy and sensitivity, and 2) neither dataset includes a classification category that would explicitly capture the kind of irrigated pasture that exists in California. The analysis offered here improves the process of identifying irrigated pasture when generating LULC data in California by using supervised classification on National Agriculture Imagery Program (NAIP) imagery for multiple years. The NAIP imagery is of much finer resolution (1 m) than the data used in other classification analyses. As with other landuse studies that use hyperspatial NAIP imagery in heterogeneous landscapes (Moskal et al., 2011, Halabisky, 2011), object-based image analysis (OBIA), rather than a pixel-based approach, was employed to dramatically improve classification accuracy. This study employed OBIA to measure irrigated pasture in one California County to refine the method and to test its accuracy in a setting where its successes and uncertainties were easier to interpret. Nevada County, with its varied terrain, mix of high-density urban areas and low-density exurban sprawl, and range of vegetation cover types was selected as a test case because it offers a broad assortment of remote sensing challenges that apply to irrigated pasture classification in California. More broadly, however, this analysis establishes an improved and repeatable methodology for mapping irrigated pasture that lays the groundwork to generate reliable and accurate data at larger scales in California and ultimately internationally. 2. Materials and methods 2.1. Description of study area Nevada County is a small, rural county located in northeastern California (Fig. 1). Its geographic area is 623,360 acres (974 mile2) and its total population was 98,877 in 2010 (U.S. Census Bureau, 2017). There is a steep elevational gradient as the county spans lower-elevation foothills in the west to the crest of the Sierra Nevada in the east. Accordingly, it includes a diversity of vegetation types, ranging from Valley Grasslands to Oak Woodlands to the Montane and Subalpine Vegetation of the Sierra Nevada Range (Barbour et al., 2007). The climate is Mediterranean, characterized by cool, wet winters and prolonged hot, dry summers. Over 95% of annual precipitation falls between the months of October and May (CIMIS, 2017). Intensive production agriculture is limited due to the county's poor foothill soils and terrain. As a result, the agriculture that has existed since the mid-nineteenth-century California Gold Rush has predominantly been timber and extensive livestock production in range- and woodland settings. In 2015, its leading agricultural commodities were cattle–both fed heifers/steers and cull cows; rangeland and irrigated pasture leases; and truck-farm vegetables (CDFA, 2016). Much of California's irrigation water is managed by quasi-governmental irrigation districts that collect, manage, and distribute water for agricultural and urban use (CA-LAO, 2002). Nevada Irrigation District (NID) is the largest in the county, founded in 1921 to deliver untreated surface water to the region's farmers and ranchers. Much of the complex system of earthen ditches and flumes that NID inherited was built in the nineteenth century to facilitate hydraulic gold mining. Early irrigation efforts within the district focused on gravity-propelled

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

447

Fig. 1. Map of Nevada County.

flood irrigation of alfalfa, hay, and pasture fields (Hyatt, 1931). Nevada County, however, experienced a profound demographic and economic shift in the second half of the twentieth century as retirees and telecommuting urban professionals increasingly replaced resource-extraction-based workers (Duane, 1999). The current landscape within NID's 287,000-acre (448 mile2; 323 mile2 of which is within Nevada County) service area is a complex patchwork of small towns, low-density exurban residences, and hobby and production-oriented agriculture. Today, in addition to providing irrigation waters to ranching operations throughout the county, NID also provides untreated water to exurban residential parcels, housing associations, and golf courses (Huntsinger et al., 2017).

2.2. Data sources This study utilized data from four sources (USDA/NRCS Geospatial Gateway, 2014; USDA NASS, 2014; Homer et al., 2015; Nevada County, California, 2015). NAIP data comes with four raster bands representing the red, green, blue, and near-infrared electromagnetic regions. It is offered either as a digital orthophoto quarter quad tile (DOQQs) or as compressed county mosaics (CCM). Because this analysis was conducted at the county scale, Nevada County's CCM was the most appropriate data choice, and it came from USDA pre-processed so that each tile in the mosaic was individually rectified into the UTM coordinate system, NAD 83. Subsequent files downloaded from other sources were then re-projected into NAD 1983, UTM Zone 10. The county NAIP file was clipped to the boundary of the local water district—the entity that provides raw, agricultural water for irrigation—in order to limit the extent of the study area to improve processing time during classification and to avoid analyzing US Forest Service land (503 mile2, or, 52% of Nevada County). Although there may be some limited pasture irrigated from springs or wells in other parts of the county, clipping to the district boundary should not significantly affect the results of the study, although—if anything—the estimates of irrigated pasture presented here should be considered conservative or slightly lower than actual values.

2.3. Classification Due to the composition of the landscape and the occurrence of multiple green vegetation types (Fig. 2a), mapping pasture was performed using a supervised image classification procedure, which uses training samples of each class to statistically evaluate class membership for the unlabeled spatial units. The pilot analysis first considered pixel-based Maximum Likelihood Classification (MLC) in both geospatial software—ArcGIS (Esri Inc.), version 10.2 (Fig. 2b) and remote sensing software—ENVI (Harris Geospatial Inc.), v. 5.2 (Fig. 2c), both of which produced a speckled, heterogeneous appearance to the mapped cover types and visible inaccuracies in spatial extents and contiguity of the target irrigated pasture units. For this reason we used an alternative objectbased image analysis (OBIA), which first uses segmentation to delineate image regions, or “objects,” as relatively homogeneous primitive spatial units and then classifies them into desired land cover types based on the user-selected approach (e.g., Blaschke, 2010). With high-resolution datasets such as NAIP, OBIA offers an improvement over pixel-based classification approaches (Fig. 2d) because it alleviates local spectral noise and, in addition to spectral information, also accounts for shape, form, and texture of the objects during their classification. We implemented OBIA classification in eCognition v.8.8 (Trimble Inc.) software using the four bands of the original NAIP as the image layer inputs. The first step was running a multiresolution segmentation algorithm that merged pixels one by one based on their spectral similarity to one another to form larger aggregates; the process terminated when a user-defined threshold of homogeneity was exceeded. Parameters to determine “shape” and “compactness” were also used. These settings are adjusted iteratively during the segmentation process until objects in the image are satisfactorily grouped. These new, grouped pixels, or objects, carry not only the spectral and statistical information of the pixels of which they consist, but also information on their texture, shape, and position (Rahman and Saha, 2008). Next, office-based, visual interpretation of the 1-m resolution imagery was used to select individual objects as training samples to define the ruleset for each of the classes. Because the emphasis was to classify pasture, we used an abbreviated list of five major classes: 1) irrigated pasture; 2) dry, non-irrigated rangeland; 3) oak and coniferous trees,

448

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

Fig. 2. Comparison of three classification methods. (a) Original 1 m–resolution NAIP imagery of Nevada County. (b) Maximum Likelihood Classification in ArcGIS; notice confusion between shadows in the oaks and irrigated pasture. (c) Maximum Likelihood Classification in ENVI. (d) Classification in eCognition with object-based image analysis.

or forest; 4) water; and 5) impervious surfaces (buildings, roads, rock, gravel, etc.). One hundred training samples were selected to build membership values for each class, for a total of 500 training objects for each image (2005, 2014). Finally, we ran eCognition's nearest-neighbor classifier, which classified the remaining undefined image objects based upon their degree of membership in feature space to the training sample classes (Baatz et al., 2004). 2.4. Accuracy assessment Determining the quality of a classification result, the process known as accuracy assessment demonstrates how well the classification is able to analyze the image and reveals potential shortcomings in the methodology (Congalton, 1991). Accuracy assessment for this study was performed in ENVI, using the tool “Confusion Matrix Using GroundTruthed ROIs.” Separately, in ArcGIS, an office-based assessment of 1 m–resolution NAIP imagery created new point vector files selected with expert knowledge, which were to function as test samples. Each class (irrigated pasture, dry range, oak, water, impervious) had at least one hundred ground-truthed point samples, for a total of 528 test samples for 2005 image and 502 for 2014. These test samples were compared against the classification results from eCognition in a confusion matrix to determine overall accuracy, accuracy within class, user's accuracy, and producer's accuracy (Congalton, 1991; Congalton and Green, 2008). 2.5. Comparison of classification techniques The National Land Cover Database uses 30 m-Landsat imagery to provide consistent, nationwide digital land cover classification data (Wickham et al., 2014). Although it has been shown that nationwide overall accuracy of previous NLCD datasets has achieved upwards of 79% (Wickham et al., 2013), it was unclear how this accuracy would hold up within the context of a smaller area, and by extension, how useful the NLCD dataset is in measuring and mapping LULC within Nevada County. To achieve a comparison between NLCD 2011 and the classification results from eCognition, the NLCD Anderson Level II class codes were consolidated into their Level I codes to generate classes comparable to the five used in the eCognition analysis (Class 81, Irrigated Pasture; Class 71, Rangeland; Class 41, Forest; Class 11, Water; Class 21, Impervious Surfaces). A new point vector file of over 500 groundtruthed test samples was again created in ArcGIS based on 2012 1 m– resolution NAIP imagery. Finally, in ENVI, a confusion matrix was generated for the 2011 NLCD results.

each irrigated pasture polygon was calculated. Polygons of less than one acre in size were discarded, both to remove small areas that may have been erroneously misclassified in eCognition (shadows, light-colored oaks, ponds, riparian areas, etc.) and because polygons of less than one acre are likely not irrigated pasture but instead irrigation leaks, lawns, or other greenways. The resulting vector files for 2005 and 2014 were unionized to create one master vector file, with three kinds of polygons for irrigated pasture: 1) present in 2005, present in 2014 (no change); 2) present in 2005, absent in 2014 (loss); 3) absent in 2005, present in 2014 (gain). 3. Results 3.1. Irrigated pasture acreage Results from the OBIA supervised classification measured 4273 acres of irrigated pasture in 2005 and 3470 acres in 2014, a loss of 803 acres (19%). These acreage figures are in comparison to the non-spatial figure of 10,000 acres reported by CDFA in 2014; the spatial figure provided by the California Department of Water Resources of 5624 acres from 2005; the non-spatial figures reported by the NASS Census of Agriculture of 4856 acres in 2007 and 4088 acres in 2012; and the 34 acres generated from the NLCD spatial dataset (Class 81, Pasture/Hay). The spatially-explicit Cropland Data Layer by NASS has no analogous, comparable class for irrigated pasture (Table 1), although we combined “alfalfa,” “other hay,” and “clover/wildflowers” as analogues for irrigated pasture, for a result of 8 acres in 2008 and 27 acres in 2014. 3.2. Accuracy of methodological approach Overall accuracy for eCognition classification for 2005 was 89.39% and for 2014 was 89.42% (Table 2). User accuracy percentages for the irrigated pasture class specifically were even higher. To wit, the algorithm correctly classified 96% of the irrigated pasture ground-truthed samples in 2005 and 100% of the samples in 2014. The greatest source of confusion overall was in the impervious class, which was often confused with “Range” and “Oak” classes, likely due to the spectral similarities between impervious surfaces, senesced grasslands, and light-colored oaks. In comparison, overall accuracy for NLCD 2011 was 47.21%. The greatest source of error in the NLCD 2011 classification was in the irrigated pasture class: not one test sample was classified correctly (77% of irrigated pasture test samples were classified as Range, 17% were classified as Oak, and 6% were classified as Impervious). 4. Discussion

2.6. Change detection of irrigated pasture Classification results from the two years, 2005 and 2014, were exported from eCognition to ArcGIS as raster files. Although change occurred between all class types across the ten-year period, change detection analysis was limited to irrigated pasture in this study. To accomplish this, the raster files were vectorized without smoothing, irrigated pasture polygons were selected and exported, and the area of

As populations continue to rise and place increasing pressure on existing water resources, understanding agriculture's full contribution to global water use will be of paramount importance. Irrigated pasture is a significant water use in California, which is likely true of other arid and semi-arid regions around the world. As such, it is critical both to document its distribution and to catalogue its environmental and economic impacts. The spatially explicit quantitative analysis presented

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

449

Table 1 Results.

Classification with National Agricultural Imaging Program (NAIP) imagery CDFA Nevada County Agricultural Commissioner's Report

California Department of Water Resources NASS Census of Agriculture National Land Cover Database (NLCD) NASS Cropland Data Layer a b c

Acres 2005–2008

Acres 2011–2014

Gain/(loss)

Percent change

4273 (2005)a 9700 (2005) 10,000 (2006 & 2008) 7986 (2007) 5624 (2005) 4856 (2007) 20 (2006) 8 (2008)b

3470 (2014)a 10,000 (2011–2014)

(803)a NA

(19%)a NA

– 4088 (2012) 34 (2011) 27 (2014)c

– (768) 14 19

– (16%) 70% 238%

Results generated in this analysis. Class in CDL was “alfalfa.” In 2008, CDL measured no pixels of “other hay/non-alfalfa” or “clover/wildflowers” in Nevada County. Combined classes in CDL were “alfalfa,” “other hay/non-alfalfa,” and “clover/wildflowers.”

here is a crucial first step towards understanding how these systems are changing and identifies the need for additional interdisciplinary study of the human and environmental impacts of this change. Although the loss of irrigated pasture is not surprising given its low economic value and the water allocation pressures created by California's recent drought, the fallowing or conversion of irrigated pasture will most likely have agricultural, cultural, and ecosystem implications that are not accounted for in traditional analyses. Here we provide an analysis of the spatial trends of this land use, the methodology we employed, the potential for broadening to larger geographic areas, and finally discuss the implications for managers and policymakers. 4.1. Irrigated pasture distribution and trends The present analysis establishes more accurate measures of the total number of acres of irrigated pasture, figures which differ substantially from three of the five currently-available metrics: 1) the statistics published by CDFA and the Nevada County Agricultural Commissioner's Report, 2) the LULC data generated by NLCD, and 3) the NASS Cropland Data Layer. Only statistics from the California Department of Water

Table 2 Confusion matrices. Classification

Ground truth (pixels) Irri. past.

Resources (DWR) and the NASS Census of Agriculture align with the present classification (Table 1). Furthermore, results in Nevada County across time demonstrate that there has been a substantial loss of irrigated pasture (19%) in the ten-year period 2005–2014 (see Fig. 3). This trend is corroborated by the NASS Census of Agriculture data, which reported a 16% loss of irrigated pasture in the period between 2007 and 2012. The DWR and Census of Agriculture figures, however, are both limited but for different reasons. The DWR uses in-person, field-based visits to map cover types in each California county. This approach, while accurate, prevents timely production of LULC data: some counties, for example, have not been mapped since 1999 (18 years ago). In comparison, our remote sensing methodology offers substantial time and budgetary savings and the ability to perform change detection across relevant time scales. The Census of Agriculture is limited in that its statistics are non-spatial. The spatially explicit results presented here are an improvement in that they allow for much broader future analysis: in addition to knowing the amount of irrigated pasture that exists in the county and its rate of decline, this study allows managers and policymakers to know where it exists, where losses are happening, and by extension, to begin to answer why and to what it is converting. Furthermore, this information will be useful for managers to better assess how the loss of irrigated pasture impacts other natural resource concerns, such as changes in the hydrologic cycle or threats to sensitive plant or animal species that may rely on irrigated pasture. 4.2. Strengths and limitations of a remote sensing approach

Range

Water

Oak

Impervious

Total

a. eCognition classification (2005) Unclassified 0 Irri. past. 97 Range 2 Water 0 Oak 2 Impervious 0 Total 101 Overall accuracy

0 3 93 0 0 4 100 (472/528)

0 0 0 94 7 0 101 89.39%

0 5 1 5 99 3 113

0 3 11 0 10 89 113

0 108 107 99 118 96 528

b. eCognition classification (2014) Unclassified 0 Irri. past. 100 Range 0 Water 0 Oak 0 Impervious 0 Total 100 Overall accuracy

0 2 97 0 0 1 100 (448/501)

0 0 0 84 11 5 100 89.42%

0 4 0 2 93 1 100

0 5 10 0 12 74 101

0 111 107 86 116 81 501

c. NCLD (2011) classification (using 2012 NAIP test samples) Unclassified 0 0 0 0 Class 81 (irri. past.) 0 0 0 0 Class 71 (range) 77 90 9 11 Class 11 (water) 0 0 32 0 Class 41 (oak) 17 11 51 89 Class 21 (impervious) 6 1 5 2 Total 100 102 97 102 Overall accuracy (237/502) 47.21%

0 0 41 0 34 26 101

0 0 228 32 202 40 502

Analysis of irrigated pasture distribution with remote sensing is only as sensitive as the available imagery, and classification using remotely sensed images must account for the inherent tradeoffs in the spatial, temporal, and spectral resolutions of data sources. One purpose of the present study was to understand how to reconcile these tradeoffs when classifying irrigated pasture. Compared to NLCD, for example, whose spatial resolution is 30 m as in Landsat imagery, the finer 1 mresolution of NAIP allowed for significantly more detailed and accurate classification (Fig. 4). Alternatively, in comparison to other missions or popular satellite imagery, temporal and spectral resolution for NAIP is more limited. The program's data collection frequency is 1–4 years and images are only taken once at the height of the summer growing season, not necessarily at the same phenological stage across the state. However, because the objectives of this study were to classify pasture and detect change over a decade, NAIP's limited temporal resolution did not adversely impact this analysis. In fact, the within-year timing of the image capture–at the height of the growing season–took full advantage of plant phenology in California. By summer, annual vegetation across the state is senesced, which accentuates spectral differences with the vibrant green of irrigated pastures. And given the strengths of multiresolution segmentation and OBIA analysis, we were able to overcome NAIP's limited spectral resolution, as the algorithm was able to classify successfully with only the four bands offered (red, green, blue, nearinfrared).

450

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

Fig. 3. Change detection in western Nevada County (2005–2014).

There were three potential sources of error in these analyses. The first one was positional error that could result from different accuracy of geo-registration among various datasets. Although the NAIP compressed county mosaics (CCMs) were pre-processed and orthorectified by USDA to be in a consistent coordinate system before download, pixels between years (2005, 2014) did not align precisely. After classification in eCognition, the raster outcomes were resampled in ArcGIS to achieve alignment, however a certain degree of error may have been introduced into the change detection results. Second, classification error, or the incorrect assignment of objects into the candidate cover types, could be a problem, especially between classes with similar spectral signatures (light-colored oaks, irrigated pasture, lawns, riparian areas, etc.). However, overall accuracy figures of close to 90% and producer accuracy figures in the “Irrigated Pasture” class of 96% (2005) and 100% (2014) demonstrate that the level of classification error in this analysis was acceptable. The class confusion of early results that used a pixelbased approach (Fig. 2) further reveals that the success of this classification process relies heavily on OBIA's capacity to enhance class contrasts while smoothing local noise (e.g., Blaschke, 2010; Dronova et al., 2012). Classification for all classes could potentially be improved in the future by refining the segmentation process to achieve greater match of primitive objects to pasture units, increasing the number of classes, or increasing the number of training samples per class. Finally, some of the uncertainly likely resulted from limited temporal availability of the NAIP imagery of California, as previously discussed, which precluded a comprehensive interpretation of land cover and land use transition in the years between 2005 and 2014. Although NLCD classification was performed on Landsat imagery from 2011, the test samples generated in this analysis to perform classification accuracy assessment used

NAIP imagery from 2012. This gap between classification date and accuracy assessment date may have introduced some degree of error. 4.3. Towards a broad-scale regional framework Most critically, these preliminary results suggest that the methodology established here can be repeated to provide accurate and ongoing measurements of irrigated pasture at larger regional scales. Some key challenges, however, will need to be addressed before expanding this framework to broader regional scopes. First, processing time for OBIA segmentation and supervised classification in our study area was long, and substantially increasing the area of interest to include all potential irrigated pastures statewide may prove computationally prohibitive. Solutions to this include: 1) conducting analyses regionally, and thereby decreasing the file size and pixel count of the imagery; 2) exploring parallel processing resources, which would relieve computation strain; or 3) opting to use imagery with lower resolution but greater temporal frequency and spectral sensitivity to irrigated pasture, e.g. 30 m-resolution Landsat. Although these results suggest that NLCD—which uses Landsat imagery—lacked sufficient detail to classify irrigated pasture, a refined OBIA-assisted methodology that uses Landsat imagery and specifically accounts for the unique shape, texture, and spectral characteristics of irrigated pasture would likely improve accuracy. Furthermore, despite its reduced spatial resolution, Landsat's sixteen-day return interval adds a temporal dimension to classification that may also improve accuracy, if, for example, identification of irrigated pasture can be assisted by using differences in seasonal phenology. A second challenge to extending classification statewide is the varied shape, size, and appearance of irrigated pasture. A singular feature class created with test samples

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

451

Fig. 4. Detail of classification results and change detection. (a) 2005 1 m–resolution NAIP imagery, selected area of Nevada County, California. (b) 2014 1 m–resolution NAIP imagery of same selected area. (c) eCognition OBIA classification result of 2005 NAIP image (green: oaks; red: irrigated pasture; grey: impervious surfaces; blue: water; yellow: range). (d) eCognition OBIA classification result of 2014 NAIP image. (e) NLCD classification of same selected area having used 30 m–resolution 2011 Landsat imagery (for comparison). (f) Change detection of irrigated pasture (green: no change; red: loss; blue: gain) having used eCognition classification results. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

from the whole state may not be able to adequately account for the variety across regions. As discussed above, conducting regional analyses to improve processing time would have the added benefit of allowing for region-specific test samples. 4.4. Implications for management and policy The inaccuracy of existing measurements—with the exception of the Census of Agriculture and the Department of Water Resources—highlights how irrigated pasture is not sufficiently monitored or its loss prioritized by many government agencies. Because water is such a crucial resource for California, this study provides a mechanism to better understand the implications of changes to irrigation on pasture. From a purely economic perspective, irrigating pasture as a land practice may appear an unwise use of water (Sunding et al., 1997) but reducing acreage statewide would likely have significant environmental, agricultural, and policy repercussions that demand further consideration. Previous research has catalogued the important ways that irrigated pasture supports or facilitates wetland-dependent wildlife species (Richmond et al., 2010; Swolgaard et al., 2008; Ivey and Herziger, 2001), groundwater recharge and riparian area promotion (Peck and Lovvorn, 2001; Wiener et al., 2008), game species (Earl, 1950), and wildfire abatement (Huntsinger et al., 2017). However, there has been no systematic effort to quantify the value of these ecosystem services or incorporate them into an economic model of water use in California. To this end, irrigated pasture is an excellent example of a “social-ecological service,” in which the complex and frequently unrecognized services it provides are cogenerated by human activity and ecosystem process (Huntsinger & Oviedo, 2014; Hruska et al., 2015). Additionally, land managers and policy experts have not adequately identified the economic, social, and policy dimensions of lost irrigated pasture, more specifically the anticipated impacts and disruption to ranching livelihoods, livestock production, and land conservation of working landscapes. Ranchers rely on irrigated pasture as a forage resource, which fills a critical gap in the livestock production calendar when non-irrigated rangelands elsewhere in the state are unable to support the operation's stock numbers or the nutritional demands of the animals. As a result, many ranching operations migrate seasonally

between non-irrigated rangelands in the wetter winter months to irrigated pasture or montane meadows during the dry summer months (Huntsinger et al., 2010). The loss of irrigated pasture is likely to erode a rancher's ability to productively and profitably operate in the state, and could mean a major reduction in production capacity for livestock operations. Irrigated pasture loss would also result in elevated grazing pressure on rangelands or increased need to truck cattle out of state, both of which would upset the current balance achieved by ranchers who migrate seasonally between irrigated pasture and rangelands. Irrigated pasture acreage then has an amplification effect: if an acre of irrigated pasture can support two cows for the summer, those same two cows will then graze twenty to forty acres of non-irrigated rangeland in the winter (Drake and Phillips, 2006). If land managers and landuse planners prioritized conserving irrigated pasture, it could have the indirect benefit of ensuring appropriate levels of continued grazing on rangelands across the state. While poorly managed grazing can have detrimental impacts to an ecosystem, well-managed grazing provides important ecosystem benefits to rangelands in California, including the prevention of invasive plant establishment and spread, the promotion of floral and faunal biodiversity, and the reduction of fuels to support fire prevention efforts (Hayes and Holl, 2003; Marty, 2005; Sulak and Huntsinger, 2007). Finally, the advances in remote sensing provide an opportunity for governments and international development agencies to gather precise statistics on irrigated pasture use without the major investment of resources that is required to implement a survey or census of agriculture. While this kind of traditional statistics gathering remains an important part of understanding the agricultural sector in many developed countries, other developing countries may not have the resources to implement sophisticated efforts to gather high-quality data. This remote sensing methodology provides policymakers, administrators, and international development practitioners the ability in certain contexts to leap-frog traditional data gathering methods and to achieve similar if not more reliable results (Van Eekelen et al., 2015). 5. Conclusions The present analysis reveals that irrigated pasture as a land use is on the decline in this region, which likely portends a broader statewide and

452

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453

potentially global trend. Additional interdisciplinary research is needed to determine the implications of this loss to the environment, the economy, and to agriculture. Successful management in the future of this human-natural system will require timely and accurate data on land use practices across large spatial scales, a methodology for which is offered here. The inaccuracies evident in existing products such as NLCD highlights a critical point: namely, that many remotely-sensed products adequately identify landscape features, or land cover, but have inherent limitations identifying human activity, or land use (Irwin and Geoghegan, 2001). Traditional maps and remote sensing analyses alone are often a “snapshot” and are thus are less powerful in informing dynamic, on-the-ground management over time. The effort here to establish a management-informed methodology that captures LULC and detects change of a particular land use across time, in this case irrigated pasture, is a first step towards a new type of map that is able to illuminate the agricultural, economic, and environmental dimensions at work in land and water use politics in California. To that end, the approach here follows Nagabhatla et al. (2015) who argue that geospatial datasets can serve as a foundational and unifying platform for researchers from multiple disciplines in order to address otherwise complex human and natural resource dilemmas. A next step towards understanding the implications of irrigated pasture conversion in California and beyond will require a “transdisciplinary” approach, one that brings together policymakers, ecologists, ranchers, and economists. Funding This research received funding support from the University of California Division of Agricultural and Natural Resources. Additional support was provided by the Department of Environmental Science, Policy, and Management at the University of California, Berkeley, through its Graduate Student in Extension fellowship program. References AghaKouchak, A., Feldman, D., Hoerling, M., Huxman, T., Lund, J., 2015. Water and climate: recognize anthropogenic drought. Nature 524 (7566):409–411. http://dx.doi. org/10.1038/524409a. Baatz, M., Benz, U., Dehghani, S., Heynen, M., Höltje, A., Hofmann, P., ... Willhauck, G., 2004. eCognition User Guide. Definiens Imaging GmbH, Munich, Germany. Barbour, M.G., Keeler-Wolf, T., Schoenherr, A.A., 2007. Terrestrial Vegetation of California. University of California Press. Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS J. Photogramm. Remote Sens. 65 (1):2–16. http://dx.doi.org/10.1016/j.isprsjprs.2009.06.004. Burkhard, B., Kroll, F., Nedkov, S., Müller, F., 2012. Mapping ecosystem service supply, demand and budgets. Ecol. Indic. 21:17–29. http://dx.doi.org/10.1016/j.ecolind.2011.06. 019. California Department of Food and Agriculture, 2015. California County Agricultural Commissioners' Reports, Crop Year 2012–2013. California Department of Food and Agriculture, 2016. California County Agricultural Commissioners' Crop Year 2014–2015. California Department of Water Resources (DWR), 2010. Agricultural Land and Water Use Estimates. Retrieved March 3, 2017, from. http://www.water.ca.gov/landwateruse/ anlwuest.cfm. California Irrigation Management Information System (CIMIS), 2017. California Department of Water Resources. Browns Valley, California. http://www.cimis.water.ca.gov. California Legislative Analyst's Office (CA-LAO), 2002. Water Special Districts: A Look at Governance and Public Participation. Retrieved March 3, 2017, from. http://www. lao.ca.gov/2002/water_districts/special_water_districts.html. Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 37 (1), 35–46. Congalton, R.G., Green, K., 2008. Assessing the Accuracy of Remotely Sensed Data: Principles and Practices. CRC Press. Drake, D.J., Phillips, R.L., 2006. Fundamentals of Beef Management. vol. 3495. UCANR Publications. Dronova, I., Gong, P., Clinton, N., Wang, L., Fu, W., Qi, S., Liu, Y., 2012. Landscape analysis of wetland plant functional types: the effects of image segmentation scale, vegetation classes and classification methods. Remote Sens. Environ. 127, 357–369. Duane, T.P., 1999. Shaping the Sierra: Nature, Culture, and Conflict in the Changing West. Univ of California Press. Earl, J.P., 1950. Production of mallards on irrigated land in the Sacramento Valley, California. J. Wildl. Manag. 14 (3), 332–342. Ellis, E.C., Wang, H., Xiao, H.S., Peng, K., Liu, X.P., Li, S.C., ... Yang, L.Z., 2006. Measuring longterm ecological changes in densely populated landscapes using current and historical

high resolution imagery. Remote Sens. Environ. 100 (4):457–473. http://dx.doi.org/ 10.1016/j.rse.2005.11.002. Famiglietti, J.S., 2014. The global groundwater crisis. Nat. Clim. Chang. 4 (11):945–948. http://dx.doi.org/10.1038/nclimate2425. FAO, 2016. AQUASTAT Main Database - Food and Agriculture Organization of the United Nations (FAO) (Website accessed on [05/04/2017 19:14]). Foley, J.A., 2005. Global consequences of land use. Science 309 (5734):570–574. http://dx. doi.org/10.1126/science.1111772. Friedl, M.A., McIver, D.K., Hodges, J.C.F., Zhang, X.Y., Muchoney, D., Strahler, A.H., Woodcock, C.E., Gopal, S., Schneider, A., Cooper, A., Baccini, A., Gao, F., C. S., 2002. Global land cover mapping from MODIS: algorithms and early results. Remote Sens. Environ. 83. Grantham, T.E., Viers, J.H., 2014. 100 years of California's water rights system: patterns, trends and uncertainty. Environ. Res. Lett. 9 (8):10. http://dx.doi.org/10.1088/17489326/9/8/084012. Guida-Johnson, B., Zuleta, G.A., 2013. Land-use land-cover change and ecosystem loss in the Espinal ecoregion, Argentina. Agric. Ecosyst. Environ. 181, 31–40. Halabisky, M., 2011. Object-based classification ofsemi-arid wetlands. J. Appl. Remote. Sens. 5 (1):53511. http://dx.doi.org/10.1117/1.3563569. Hannerz, F., Lotsch, A., 2008. Assessment of remotely sensed and statistical inventories of African agricultural fields. Int. J. Remote Sens. 29 (13), 3787–3804. Hayes, G.F., Holl, K.D., 2003. Cattle grazing impacts on annual forbs and vegetation composition of mesic grasslands in California. Conserv. Biol. 17 (6), 1694–1702. Homer, C.G., Dewitz, J.A., Yang, L., Jin, S., Danielson, P., Xian, G., Coulston, J., Herold, N.D., Wickham, J.D., Megown, K., 2015. Completion of the 2011 National Land Cover Database for the conterminous United States-representing a decade of land cover change information. Photogramm. Eng. Remote Sens. vol. 81 (5), 345–354. Howitt, R., Medellín-azuara, J., Macewan, D., 2015. Economic Analysis of the 2015 Drought for California Agriculture. Center for Watershed Sciences. University of California, Davis, California (20 pp.) Retrieved from. http://watershed.ucdavis.edu. Hruska, T.V., Huntsinger, L., Oviedo, J.L., 2015. An accidental resource: the social ecological system framework applied to small wetlands in Sierran foothill oak woodlands. In: Standiford, Richard B., Purcell, Kathryn L. (Eds.), Proceedings of the Seventh California Oak Symposium: Managing Oak Woodlands in a Dynamic World (tech. cords. Gen. Tech. Rep. PSW-GTR-251). Huntsinger, L., Forero, L.C., Sulak, A., 2010. Transhumance and pastoralist resilience in the Western United States. Pastor. Res. Policy Pract. 1 (1):9–36. http://dx.doi.org/10. 3362/2041-7136.2010.002. Huntsinger, L., Oviedo, J.L., 2014. Ecosystem services are social-ecological services in a traditional pastoral system: the case of California’s Mediterranean rangelands. Ecol. Soc. 19 (1), 8. Huntsinger, L., Hruska, T.V., Oviedo, J.L., Shapero, M.W.K., Nader, G.A., Ingram, R.S., 2017. Save water or save wildlife? Water use and conservation in the Central Sierran foothills oak woodlands of California, USA. Ecol. Soc. 22 (2). Hyatt, E., 1931. Report on Irrigation Districts in California For the Year 1930, Bulletin No. 21-B. Publications of the Division of Water Resources. Irwin, E.G., Geoghegan, J., 2001. Theory, data, methods: developing spatially explicit economic models of land use change. Agric. Ecosyst. Environ. 85 (1), 7–24. Ivey, G.L., Herziger, C.P., 2001. Distribution of Greater Sandhill Crane Pairs in California, 2000. California Department of Fish and Game. Marty, J.T., 2005. Effects of cattle grazing on diversity in ephemeral wetlands. Conserv. Biol. 19 (5), 1626–1632. Moskal, L.M., Styers, D.M., Halabisky, M., 2011. Monitoring urban tree cover using objectbased image analysis and public domain remotely sensed data. Remote Sens. 3 (10): 2243–2262. http://dx.doi.org/10.3390/rs3102243. Mount, J., Freeman, E., Lund, J., 2014. Water Use in California. Public Policy Institute of California. http://dx.doi.org/10.5066/F7KD1VXV (July). Nagabhatla, N., Padmanabhan, M., Kühle, P., Vishnudas, S., Betz, L., Niemeyer, B., 2015. LCLUC as an entry point for transdisciplinary research - reflections from an agriculture land use change study in South Asia. J. Environ. Manag. 148:42–52. http://dx. doi.org/10.1016/j.jenvman.2014.03.019. Nevada County, California, 2015. GIS Open Data Portal. Retrieved December 1, 2015 from. https://www.mynevadacounty.com/nc/igs/gis/Pages/OpenData.aspx. Peck, D.E., Lovvorn, J.R., 2001. The importance of flood irrigation in water supply to wetlands in the Laramie Basin, Wyoming, USA. Wetlands 21 (3):370–378. http://dx.doi. org/10.1672/0277-5212(2001)021[0370:TIOFII]2.0.CO;2. Postel, S.L., 1998. Water for food production: will there be enough in 2025? Bioscience 48 (8), 629–637. Rahman, M.R., Saha, S.K., 2008. Multi-resolution segmentation for object-based classification and accuracy assessment of land use/land cover classification using remotely sensed data. J. Indian Soc. Remote Sens. 36 (2):189–201. http://dx.doi.org/10.1007/ s12524-008-0020-4. Richmond, O.M.W., Chen, S.K., Risk, B.B., Tecklin, J., Beissinger, S.R., 2010. California black rails depend on irrigation-fed wetlands in the Sierra Nevada foothills. Calif. Agric. 64 (2):85–93. http://dx.doi.org/10.3733/ca.v064n02p85. Rost, S., Gerten, D., Bondeau, A., Lucht, W., Rohwer, J., Schaphoff, S., 2008. Agricultural green and blue water consumption and its influence on the global water system. Water Resour. Res. 44 (9). Sohl, T.L., Sleeter, B.M., Sayler, K.L., Bouchard, M.A., Reker, R.R., Bennett, S.L., ... Zhu, Z., 2012. Spatially explicit land-use and land-cover scenarios for the Great Plains of the United States. Agric. Ecosyst. Environ. 153, 1–15. Sulak, A., Huntsinger, L., 2007. Public land grazing in California: untapped conservation potential for private lands? Working landscapes may be linked to public lands. Rangelands 29 (3), 9–12. Sunding, D., Zilberman, D., Howitt, R., Dinar, A., MacDougall, N., 1997. Modeling the impacts of reducing agricultural water supplies: lessons from California's Bay Delta

M. Shapero et al. / Science of the Total Environment 605–606 (2017) 445–453 problem. Decentralization and Coordination of Water Resources Management. Swolgaard, C.A., Reeves, K.A., Bell, D.A., 2008. Foraging by Swainson's Hawks in a vineyard-dominated landscape. J. Raptor Res. 42 (3):188–196. http://dx.doi.org/10. 3356/JRR-07-15.1. U.S. Census Bureau, 2017. Quick Facts, Nevada County, California, Population estimates base, July 1, 2015. Retrieved March 3, 2017, from. https://www.census.gov/ quickfacts/table/PST045215/06057. USDA National Agricultural Statistics Service, 2014. Census of Agriculture. Retrieved February 25, 2017 from. https://www.agcensus.usda.gov/Publications/2012/. USDA National Agricultural Statistics Service Cropland Data Layer, 2014. National Cropland Data Layer. Retrieved December 1, 2015 from. https://nassgeodata.gmu.edu/ CropScape/ USDA-NASS, Washington, DC. USDA/NRCS Geospatial Gateway, 2014. National Ag. Imagery Program Mosaic. Retrieved December 1, 2015 from. https://datagateway.nrcs.usda.gov/GDGOrder.aspx. Van Eekelen, M.W., Bastiaanssen, W.G., Jarmain, C., Jackson, B., Ferreira, F., Van der Zaag, P., ... Dost, R.J.J., 2015. A novel approach to estimate direct and indirect water withdrawals from satellite measurements: a case study from the Incomati basin. Agric. Ecosyst. Environ. 200, 126–142.

453

Verburg, P.H., van de Steeg, J., Veldkamp, A., Willemen, L., 2009. From land cover change to land function dynamics: a major challenge to improve land characterization. J. Environ. Manag. 90 (3):1327–1335. http://dx.doi.org/10.1016/j.jenvman.2008.08. 005. Wickham, J.D., Stehman, S.V., Gass, L., Dewitz, J., Fry, J.A., Wade, T.G., 2013. Accuracy assessment of NLCD 2006 land cover and impervious surface. Remote Sens. Environ. 130:294–304. http://dx.doi.org/10.1016/j.rse.2012.12.001. Wickham, J., Homer, C., Vogelmann, J., McKerrow, A., Mueller, R., Herold, N., Coulston, J., 2014. The multi-resolution land characteristics (MRLC) consortium—20 years of development and integration of USA national land cover data. Remote Sens. 6 (8), 7424–7441. Wiener, J.D., Dwire, K.A., Skagen, S.K., Crifasi, R.R., Yates, D., 2008. Riparian ecosystem consequences of water redistribution along the Colorado Front Range. Water Resour. IMPACT 10 (3), 18–21.