Rule-based land use/land cover classification in ...

Environ Monit Assess (2015) 187:449 DOI 10.1007/s10661-015-4667-3

Rule-based land use/land cover classification in coastal areas using seasonal remote sensing imagery: a case study from Lianyungang City, China Xiaoyan Yang & Longgao Chen & Yingkui Li & Wenjia Xi & Longqian Chen

Received: 1 February 2015 / Accepted: 3 June 2015 # Springer International Publishing Switzerland 2015

Abstract Land use/land cover (LULC) inventory provides an important dataset in regional planning and environmental assessment. To efficiently obtain the LULC inventory, we compared the LULC classifications based on single satellite imagery with a rule-based classification based on multi-seasonal imagery in Lianyungang City, a coastal city in China, using CBERS-02 (the 2nd China-Brazil Environmental Resource Satellites) images. The overall accuracies of the classification based on single imagery are 78.9, 82.8, and 82.0 % in winter, X. Yang : L. Chen (*) Jiangsu Key Laboratory of Resources and Environmental Information Engineering, School of Environment Science and Spatial Informatics, China University of Mining and Technology, Xuzhou 221116, China e-mail: [email protected]

early summer, and autumn, respectively. The rule-based classification improves the accuracy to 87.9 % (kappa 0.85), suggesting that combining multi-seasonal images can considerably improve the classification accuracy over any single image-based classification. This method could also be used to analyze seasonal changes of LULC types, especially for those associated with tidal changes in coastal areas. The distribution and inventory of LULC types with an overall accuracy of 87.9 % and a spatial resolution of 19.5 m can assist regional planning and environmental assessment efficiently in Lianyungang City. This rule-based classification provides a guidance to improve accuracy for coastal areas with distinct LULC temporal spectral features.

X. Yang e-mail: [email protected]

Keywords Rule-based land use/land cover (LULC) classification . Multi-seasonal imagery . Seasonal land use/cover change . Remote sensing (RS) . Coastal area

X. Yang : L. Chen Institute of Land Resources, Jiangsu Normal University, Xuzhou 221116, China

Introduction

L. Chen e-mail: [email protected] Y. Li Department of Geography, University of Tennessee, Knoxville, TN 37996, USA e-mail: [email protected] W. Xi Department of Estate Management, Faculty of Built Environment, University of Malaya, Kuala Lumpur, Wilayah Persekutuan 50603, Malaysia e-mail: [email protected]

Decision making on resource and environment management has been increasingly relying on geo-spatial information, such as land use/land cover (LULC) inventory, digital elevation model (DEM), and other environmental information (Mekasha et al. 2014; Pauleit et al. 2005; Amitrano et al. 2014). In particular, LULC inventory provides a fundamental dataset to assess land resource abundance and environmental impact and to assist managers and policy makers in regional development (Gong et al. 2011). Since higher resolution data may require

449

Page 2 of 15

higher data acquisition cost and more computational power in data processing, selecting an appropriate resolution data suited for a specific decision making is of critical importance in saving money and time. Land resource and environment management in a mesoscale area, such as a county in China with an area of thousands square kilometers, is usually based on thematic maps with scales ranging from 1:50,000 to 1:200,000 and spatial resolutions from 50 to 200 m. In recent years, the LULC inventory derived using remote sensing (RS) imagery has been widely used in regional planning and environmental assessment to assess environmental impact, define zoning function, and design resource restriction for development (Bourgeron et al. 1999; Wu 2001; Zeng 2004; MelendezPastor et al. 2010; Tüshaus et al. 2014; Song et al. 2014). The advantage of RS-based classification is its capability to provide accurate and objective LULC inventory in a relatively cheap and timely manner (Muttitanon and Tripathi 2005; Mas 2004; Kamh et al. 2012). With the continuous advances in sensor and data transmission technology, RS imagery can provide multi-temporal data with suitable spatial resolutions for land resource and environment management. For example, Landsat 7 has provided images with a panchromatic band of 15 m spatial resolution (band 8) and other bands in the spectrum of blue, green, red, near-infrared (NIR), and midinfrared (MIR) of 30 m spatial resolution (bands 1–5, 7). China-Brazil Environmental Resource Satellites (CBERS), launched on October 14, 1999, have provided images with a spatial resolution of up to 19.5 m. The LULC inventory classified from these satellite images could meet the resolution requirements for land resource and environment management in a mesoscale region. LULC types have different impacts on the environment. For example, cropland helps absorb CO2, regulate the precipitation and temperature, and promote the biodiversity compared with construction land, but it also may aggravate the contamination of water and soil due to the improper use of pesticide and chemical fertilizer. Forest reduces sediment and nutrient loads, debris flows, and storm-water runoff (Matteo et al. 2006; Zongjun et al. 2011; Köplin et al. 2012). Grassland may affect soil moisture (Hu et al. 2009), reduce soil erosion, and improve the quality of water systems (Qiu et al. 2011). Wetland may reduce flood damage and provide habitat for species (Gong et al. 2009). Construction land may change the hydrological characteristics and regional climate. Some construction land may also produce pollution substances, such as waste

Environ Monit Assess (2015) 187:449

water, gas, and other materials, influencing the environmental quality. Thus, it is essential to obtain appropriate LULC information such as arable land, construction land, forest, and water based on the environmental impacts or functions of individual LULC types. RS-based LULC classification was usually based on a single image, while classifications using different seasonal images may result in different LULC classification accuracies (Gong and Wu 2014). Recent studies demonstrated that combining multi-seasonal images could improve the classification accuracy (Ippoliti‐Ramilo et al. 2003; Yuan et al. 2005), especially for vegetation-related classes (Rodríguez-Galiano et al. 2012). For example, Ippoliti‐Ramilo et al. (2003) presented a classification using multi-temporal Landsat TM imagery from the pre-planting (dry) season to evaluate cropland in the planting (rainy) season in Brazil. Yan et al. (2005) applied time-series MODIS NDVI dataset within 1 year to extract double and single cropping systems in China. O’Hara et al. (2003) examined the complex assemblages of vegetation and their seasonal variability based on Landsat imagery acquired from leaf-off and leaf-on conditions in 1991 and 2000. Lu et al. (2011) used multiple resolution images (MSS, TM/ ETM) to extract various LULC types, including forest and orchard, double-cropping land, single-cropping land, no-vegetable land, and water, in the Shandong Peninsula, a coastal region in China, based on temporal characteristics of vegetation, water, barren land, and urban areas. Coastal areas play an important role in the social and economic development around the world (Jack Ruitenbeek 1994; Liu et al. 2008; Syvitski et al. 2005; Small and Nicholls 2003). Almost half of the world’s population (about 3 billion people) live within 200 km from a coastline, and the population density of coastal areas is twice of the world’s average (Creel 2003; United Nations System-Wide Earthwatch 2003). According to the report from the National Bureau of Statistics of China (2011 China statistical yearbook, China Statistics Press), about 43.0 % of the population live in coastal areas in China, accounting for 56.3 % of the gross regional product (GRP) of the whole country. Rapid population growth and socioeconomic development lead to drastic LULC change. For instance, the total urban built-up land in the Lianyungang bay area increased 42.4 % in 2000–2006 with an urban growth rate of 12.8 km2/a (Li et al. 2010). While the ecosystem services of coastal areas have been threatened by the


stressors from urbanization and industrialization, such as sea-level rise, erosion, and anthropogenic pollutants like occasional oil spill, it is of great significance to monitor and assess the environmental quality when certain development strategy is implemented (Wu et al. 2012). However, most of studies have focused on a specific LULC class or classes. Few studies have been conducted to achieve a comprehensive classification for a coastal region. In particular, the classification related to industrial or tidal regime was seldom employed in coastal areas. Different LULC classes in a coastal area may have different temporal characteristics. Vegetationrelated classes are strongly influenced by leaf-on/off seasons, whereas tide-related classes are affected by variations in tide levels. Some classes may be highly related to policy regulations. For example, the agriculture system in parts of eastern China is regulated to plant three crops in 2 years or two crops annually. Under this regulation, the cropland may be represented as barren or cultivated land in different periods and misclassification of these two cropland types may lead to inaccurate or even wrong results in land planning or environmental assessment. Thus, a comprehensive LULC classification in coastal areas needs to combine multi-temporal images that reflect all these perspectives. Some studies combined multi-temporal images from different years in the LULC classification. However, in rapid developing areas, the selection of these multi-temporal images from a short period (e.g., in 1 year) is better for obtaining the LULC information. Rule-based methods have been introduced to classify LULC types using various criterion, such as the characteristics of terrain, soil, or hydrography (Sader et al. 1995); the multi-spectral imagery (Bardossy and Samaniego 2002); the ecological content of single and multi-date imagery and its derived products (Lucas et al. 2007); the feature areas; the relative border length to (adjacency) and number of neighbors (Bauer and Steinnocher 2001); or the signatures of LULC classes, including shape, spectral, and texture features (Zhang and Zhu 2011; Xu 2013; Niu and Ban 2013). However, the seasonal characteristics are seldom taken into consideration in the rule-based classification. Most rulebased classifications focused on certain specific LULC types, such as forest, wetland, specific urban LULC type, or vegetation types, rather than the comprehensive LULC types in a certain region. In addition, the typical LULC types in a coastal area may have specific changing features different from inland regions, leading to the

Page 3 of 15 449

necessity of exploring the knowledge of temporal characteristics and building appropriate rules for the rulebased classification to obtain better results. Most rule-based methods are pixel-based supported by commonly used software packages, such as ENVI, ERDAS, and ArcGIS. In recent years, eCognition also implemented object-oriented rule-based classification using fuzzy-rules and contextual and shape information (Bauer and Steinnocher 2001; Lucas et al. 2007). This object-based method could improve the efficiency and accuracy of the LULC classification. However, it requires additional cost for eCognition package to implement this object-oriented rule-based classification, which is not always available for investigators or institutes. In this paper, we apply a rule-based LULC classification approach based on multi-seasonal RS images to classify the LULC inventory for a coastal area in Lianyungang City, China. Detailed objectives are as follows: (1) to establish a rule-based LULC classification based on the seasonal characteristics and regulations suited for the target coastal area and (2) to examine temporal characteristics of LULC types in coastal areas and obtain the LULC information supporting environmental assessment.

Method The spectral characteristic of a LULC class (N) at a certain time is reflected as a set of pixel values in a single RS image with different bands and can be represented as a vector N(C)=N(c1,c2, …, cn), where C is the spectral vector of various bands of the RS image, and the component ci represents the bands i. When the image is taken at a certain time t, the value of C can be represented as a vector Vt =(v1t,v2t, …, vnt), where vit is the value of ci(that is the band i) at time t. Different classes have their own spectral characteristic vectors. A LULC class may have varied characteristic vectors in different time slices, and different classes may show similar characteristic vector at a certain stage due to their spectral similarity. Thus, misclassification may occur when LULC classes are determined simply based on a single vector at one time slice. For example, villages and rural road in the countryside may be classified as vegetation because of the leaf cover when using an image acquired in summer; cropland may be classified as bare land when it is fallow or cleaned after harvest; and

449


Page 4 of 15

tideland may be classified as water in a high tide-level image. Different classes have different temporal characteristics, such as the seasonal change of vegetation, the tidal change of tideland, the changes in salt field due to industrial regulation, and the changes in cropland due to regional agriculture regulation. Thus, the class N can be determined using a combination of multiple vectors in different times. According to the discriminant vector of LULC class i with t temporal state Vi 0, the component of classifier V 0(V10,V20,…,Vk0), if the spectral vector Vj ⊂Vi 0, class N(C, Vj) is determined as LULC class i. Vj is the spectral vector of class j with t temporal slices. Where Vi 0 is described as matrix 0

v011

0 B B 0 V i0 ¼ V i10 ; V i20 ; …V it0 ¼ B v21 @… v0n1

v012

… v01t

1

v022 … v0n2

C … v02t C C; … …A … v0nt

v12 v22 … vn2

… … … …

ð1Þ

and Vj as matrix 0

v11 0 B v21 V j ¼ V j1 ; V j2 ; …V jt ¼ B @… vn1

1 v1t v2t C C: …A vnt

ð2Þ

Note that higher accuracy to distinguish different classes may be achieved by employing more temporal images. But it does not indicate that the more temporal images employed the better the LULC classified. First, the use of more temporal images indicates more image procession tasks. Second, minor difference between images in a short time interval may lead to an inefficient use of discriminant vector in the classification. Thus, an optimal number of temporal images and their corresponding time slices are needed based on the image features and the characteristic of individual LULC classes before conducting the multi-temporal classification. After determining the appropriate time slices and the number of images, it is essential to examine the temporal characteristics of LULC classes to establish the temporal classification rules that can be employed into the LULC classification. Figure 1 illustrates the flowchart of the major steps of our proposed rule-based multi-temporal classification. First, the LULC classification system is defined based on the LULC characteristics, the purpose of the LULC classification, and the spectral features of RS imagery. Then, the temporal classification rules are established based on the temporal characteristics of different LULC

classes. Finally, the rule-based classification will be executed using seasonal imagery and the accuracy of the classification will be assessed. Definition of the LULC classification system The LULC classification system is generally determined by the characteristics of the study area, the purpose of the LULC classification, and the separability in spectral characteristics. The LULC classes in coastal areas usually include cropland, forest, residential, water body, and tideland. Some coastal areas also have salt field. The LULC classes can be further divided into subclasses based on certain purposes. For example, forest can be divided into high- and low-density forest. The tideland in coastal areas usually shows as water body at high tide level and various LULC classes at low tide level based on the characteristics of the exposed land. Thus, it is difficult to classify tideland using a single image. In our proposed classification, tideland is excluded in the classification of a single image, but included in the final classification based on seasonal images by comparing the classification results of images obtained at high and low tidal levels. Temporal analysis of LULC classes to establish classification rules Since our proposed classification is based on multiple temporal images, the selection of these images needs to consider the spectral discrimination of different LULC types. Therefore, we first analyzed the temporal characteristics of seasonal, regulation, and tidal changes to determine the optimal temporal images used for the classification. Then, the temporal and spectral signatures are examined to determine the classification rules based on the temporal spectral characteristics of different LULC classes. In contrast with inland areas, the temporal changes and corresponding spectral features associated with tidal change and salt production are needed in coastal areas. Rule-based classification The classification of individual image was first conducted based on a regular RS classification method, such as the maximum likelihood classification. Then, the classification results from different images are overlapped in GIS and classification rules are applied to assign the


Page 5 of 15 449

Fig. 1 Flowchart of the processing steps for rule-based classification

final LULC class for each overlapped pixel. Finally, the accuracies of individual and final classifications are assessed to test the feasibility of the proposed method and evaluate the quality of classification.

A case study from the Lianyungang City Study area and data sources Lianyungang is a coastal city located in the temperate monsoon climate zone of the northeast Jiangsu Province (118.4–119.8° E, 34.0–35.1° N), China (Fig. 1a). It includes three districts (Xinpu, Haizhou, and Lianyun) with a population of more than 4 million. Lianyungang is a typical rapid growing coastal area with various LULC classes including both typical coastal LULC classes, such as tidelands and salt field, and regular LULC classes, such as cropland, forest, water, transportation, commercial land, industrial land, and residential land. Our study area includes the whole Huaguoshan town and the majority of Punan town and Houzui town in the Xinpu district (Fig. 2). The whole study area is covered by a single CBERS-02 image. The CBERS-02 image was provided by the China Centre for Resources Satellite Data and Application (CRESDA) with spatial resolution of 19.5 m and five spectral bands from visual to near-infrared wavelength: 0.45~0.52 μm (Band 1), 0.52~0.59 μm (Band 2), 0.63~0.69 μm (Band 3), 0.77~0.89 μm (Band 4), and 0.51~0.73 μm (Band 5)

(Huang et al. 2004). The images used in this study were obtained on Feb. 26 (winter), May 12 (early summer), and Oct. 18 (autumn) in 2005 with low cloud cover. The image of May 12 was acquired at a high tide level, while the image of Oct. 18 was at a low tidal level. The 1:5000 land use map, covering the whole Lianyungang area, was obtained based on detailed land survey in 2005. It provides a reliable reference for the accuracy assessment of LULC classification in this area. Therefore, we used this land use map in 2005 to assess the accuracy of the classification. To ensure the quality of the analysis, we first projected the land use map to the same map projection of the RS images, and then registered the RS images to the land use map based on a set of control points (generally more than seven points). The error of the spatial registration is controlled within half pixel. We also processed the images with radiation enhancement using lookup table stretch method to improve the visual contrast of the RS images (Shalaby and Tateishi 2007). LULC classification system We classified eight LULC classes in our study area: high-density forest (HDF), low-density forest (LDF), low-density residential (LDR), high-density residential (HDR), salt field (SF), water body (WB), cropland (CL), and coastal tidelands (CTL). The HDF consists of dense natural and man-made forest and woodland; the LDF consists of sparse woodland, orchard, and tea plantation; the LDR mainly consists of village, road outside of

449


Page 6 of 15

Fig. 2 Location of the study area (a). And false color composite image of CBERS-02 (b) (Band 4, 3 and 1. 2005-02-26)

downtown, and some bare land (similar spectral feature as villages in China); and the HDR consists of downtown, road with impervious surface, and the industrial land out of downtown. Since only a very small fraction with an area of 0.17 hm2 in the study area is covered by grassland, we grouped it into CL based on their similar spectral signatures. According to the function of land, the water body used for salt production should belong to salt field. However, since most salt field have similar ecological functions with other water body in coastal area, we classified them as WB instead of SF. LULC temporal and spectral feature analysis Seasonal classification rules were established based on temporal changing and similarity analysis of LULC classes. Temporal changing characteristics are based on the regional climate, mandate regulations, or industrial producing cycle, including the changes of general LULC classes such as cropland and vegetation due to the annual growth cycle of vegetation, tideland based on the tidal level changes,

and salt field based on industrial regulation. Similarity analysis is based on the examination of spectral features in RS imagery according to the typical seasonal signatures generated from training areas of each LULC class in the study area. Cropland The seasonal change of cropland is derived from the regional agriculture system and the regulation. When the cropland is fallow or clear due to harvest, it shows a similar spectral signature as bare land. The major agriculture system for croplands in our study area is 1 year with two crops. Generally, the wheat is cultivated in autumn after the harvest of rice, corn, soybeans, peanuts, or potato. In early summer, it reaches harvest after the dormancy and growth in winter and spring. Other croplands under the system of 2 years three crops show as barren land in winter due to the fallow. Thus, in winter, early summer, and autumn, some croplands appear as similar spectral signatures as barren lands because of no crop cover.


Vegetation and construction land The seasonal characteristics of forest are associated with the species and regional climate. In our study area, the majority of mountains, villages, some part of city, both sides of roads and rivers are planted with deciduous trees such as populus and salix. During autumn and winter, these areas appear as barren land in satellite imagery because leaf fall off trees. While in late spring and summer, they show the spectral signature of forest. Some areas with evergreen trees, such as Japanese red pine and Pinus thunbergii, show the same spectral signature of forest in all seasons. Other vegetation types such as tea plantations, chestnut groves, and vineyards show similar temporal changes with deciduous trees. Most of them are located in mountains, sloping fields, and croplands. Coastal tideland and salt field Tideland and salt field are typical LULC classes in coastal areas. During the periods of low tidal levels, tidelands are exposed from water, making them appear as different LULC classes based on their surface characteristics. During the periods of high tide levels, tidelands show as water body because they are covered by the sea. The spectral change of salt field is determined by the industrial production cycle. Salt field generally consists of transpiration plots, brines plots, crystallization plots, and salt piles. When the precipitation is low in dry months, some crystallization plots show as the typical spectral feature of salt field because of low water content (Wang et al. 2005; Yan 2004). In other months, it appears as water body. Spectral signatures of LULC classes and classification rules The spectral signature and its temporal variation of each LULC class can be analyzed based on its training areas. Figure 3 illustrates the three seasonal mean values of each LULC class in the study area. Since cropland, tea plantations, and deciduous trees have different impacts on the environment, it is of importance to differentiate these three types. According to the CBERS-02 imagery, cropland, tea plantations, and deciduous trees all have similar spectral characteristics in early summer, but different spectral signatures in winter. In winter,

Page 7 of 15 449

cropland can be separated from deciduous trees in the fourth band and from tea plantations in the fifth band. In autumn, cropland can be separated from tea plantations in the second, third, and fifth bands. The spectral difference between urban and salt field is higher in winter than that in autumn. These results suggest that although some classes may have similar spectral characteristics in a certain seasonal image, they may have different spectral signatures in other seasonal images. The classification rule for each class can be established based on different spectral signatures of different seasonal images. Table 1 lists the major rules used to classify the LULC classes in our study area. For example, if some patches are classified into LDR in winter and LDR or CL in early summer, while CL in autumn, they are determined as CL. If patches are classified into WB, LDR, or HDR in winter, and WB in early summer (high tide level), but LDR or HDR in autumn (low tide level), they are classified as CTL. For temporal LULC changes that are not represented in Table 1, the majority of the classified types from different seasons were chosen as their final LULC classes. For example, if the classification of a certain pixel in two out of three seasonal images is the same, the final class for this pixel will be assigned as this class. For pixels with three different classes from different seasons, the classification of the season with the highest accuracy was determined as the final LULC class. Rule-based classification We first classified each seasonal image (Fig. 4a–c) using the supervised Maximum Likelihood Classifier (MLC) with ERDAS imagine. As one of the most widely used methods, the MLC first establish the Gaussian probability density function for each class using training sites, and then judge each pixel based on the maximum probability of the pixel belonging to each class. As mentioned earlier, tideland is excluded in the classification of each single image. Then, we overlapped the three seasonal classifications in ArcGIS 9.3 and used the rules listed in Table 1 with Con function in raster calculator to assign the final class for pixel (Fig. 6a). Finally, we assessed the accuracies of the seasonal and final classifications using 256 points randomly selected from the land use map in 2005 (Tables 2 and 3). At least 10 points were selected in each LULC type to ensure its accuracy assessment. As mentioned earlier, the 1:5000 land use map in 2005 was used for the accuracy assessment. In

449


Page 8 of 15

Fig. 3 Seasonal signatures of typical LULC types with CBERS-02 imagery

this land use map, land use was mainly classified as paddy field, dry farm, bare land, rural residential, urban residential, road, forest, shrub land, tea plantation, river, pond, salt field, coastal tideland, mine land, and others. We combined the land use classification into several classes to match the LULC types we classified based on RS imagery. The paddy field and dry farm were combined as cropland, bare land, urban residential, and road as high-density residential, rural residential as low-density residential, forest as highdensity forest, and shrubland and tea plantation as lowdensity forest. Since this LULC map was composited based on detailed land survey, we treated it as the Bground truth^ to assess the LULC classification accuracies. Seasonal classification and the LULC change The classification results of each seasonal image and the seasonal changes are shown in Figs. 4 and 5. Most croplands are distributed in the western (Punan town) and mid

(Huaguoshan town) area, high-density residential in the central area, water body mainly in Houzui town, salt field in the northern area (Houzui town), high-density forest and high-density residential mainly in the eastern area (Huaguoshan town), and low-density residential scattered within the whole area. The minority of high-density forest are distributed among cropland and high- or low-density residential in early summer and autumn results. Some high- and low-density residential are distributed in the northern area in winter and autumn results. The changes between high- or low-density residential and high-density forest are found mainly in the outside of villages, both side of rural roads and woodlands, and the changes between the high- or low-density residential and the water body mainly in the northern area. Table 2 lists the accuracy assessment of the classification based on each seasonal image, including the accuracy of each class, the overall classification accuracy (OCA), and the overall Kappa coefficient (OKC). The classification of the winter image shows a relatively lower accuracy (OCA 78.9 %, OKC 0.72), while the

Table 1 LULC classification rules with seasonal imagery No

Classification rules

Final classification type

1

If LDR in winter, LDR or CL in early summer, and CL in autumn

CL

2

CTL

3

If WB in winter, WB in early summer, and LDR or HDR in autumn If LDR in winter, WB in early summer, and LDR or HDR in autumn If HDR in winter, WB in early summer, and LDR or HDR in autumn. If WB or SF in all three seasons, and SF at least in one seasonal classification

4

If CL in winter, HDF or CL in early summer, and LDF in autumn

LDF

5

If three seasonal different LULC

Classification type with highest accuracy

6

If at least in two seasonal classification result shown as same LULC type

Majority of the result

SF

HDF high-density forest, LDF low-density forest, LDR low-density residential, HDR high-density residential, SF salt field, WB water body, CL cropland, CTL coastal tidelands


Page 9 of 15 449

Fig. 4 The LULC classification results (a, b, and c) based on supervised method

early summer and autumn classifications have relatively higher accuracies. Low-density forest has the lowest classification accuracy among all seven classes, even the highest accuracy of three seasonal classifications is only 0.44 in the Kappa coefficient (KC). The highest KC are 0.65 for low-density residential, 0.78 for high-

density residential, and 0.68 for salt field; all of them are