1 Introduction

Since the first golf course was opened in 1984, golf course has been undergoing rapid development in China, which has increased from 348 in 2009 to 521 in 2013 with growth rate of 49.71%. By contrast, sharp cropland decline and serious water shortage have drawn more and more attention from both the government and the public. Rapid expansion of golf course is regarded as a non negligible contributing factor leading to cropland and water resource-related issues for two reasons. On one hand, golf course is both land-consuming and water-consuming. On the other hand, most golf courses locate around developed cities where contradiction between people and environment is encountered persistently. In response to threats constituted by golf course, several regulations forbidding golf course construction have been released by the Government since 2004, and special investigations on golf course were carried out in 2009 and 2011 to control golf courses development.

Golf course is characterized by small overall number and large individual area, which make it appropriate to monitor golf course using remote sensing imagery. Medium-resolution multispectral imagery plays an important role in satellite remote sensing applications, the spatial resolution of which ranges from eight to 30 meters with visible and near-infrared spectral bands. Monitoring golf course using medium-resolution multispectral imagery is both efficient and economical. Besides, golf course is a typical kind of composite objects in contrast with simple objects, complex objects [1]. In several researches, golf course is treated as simple object and its detection problem is simplified to be turfgrass extraction. Dimock argued Landsat TM band 5, 4 and 3 could effectively enhance golf course visually, but automatic golf course extraction were unreliable [2]. Frankvich, Chen and Zhang detected golf course based on spectral and textural feature of turfgrass using Landsat TM and SPOT-5 imagery [3,4,5]. However, the low spectral separability between turfgrass and other well-growing vegetations makes these methods noneffective in discriminating golf course, cropland and grassland. To overcome the disadvantage, more recent researches turn to using or combining spatial feature in golf course detection. For example, Bhagavathy used spatial co-occurrence of turfgrass and rough which was termed as texture motif to recognize golf course in aerial image [6]. A hybrid evolutionary algorithm was used in Harvey’s research to select features for golf course extraction in AVIRIS imagery and results proved regional spatial features made the biggest contribution [7]. Guo used Hyerpclique model to describe semantic feature of composite object and took golf course as an example in composite object detection [8]. In Yang’s research on land-use classification using Bag-Of-Visual-Words, golf course was one of the 21 land-use objects of interest [9]. These methods regarded golf course as composite object and spatial relationship between components was taken into consideration, but the small Field-Of-View of aerial and hyperspectral imagery makes them unpractical on a large spatial scale.

In consideration of limitations of state-of-art methods, a practical bottom-up golf course detection approach is proposed in this paper. In terms of bottom-up manner, the approach is a flow composed of elements extraction and a subsequent area combination. In terms of practicality, the approach uses multispectral medium-resolution remote sensing imagery which can observe large area.

2 Study Area and Dataset

A geometrically and radiometrically corrected SPOT-5 HRG multispectral image of Beijing, China acquired on 17th, May, 2007 is used as experimental dataset. Four sub-images with size of 512 × 512 shown in Fig. 1 are selected to clearly show experimental results.

Fig. 1.
figure 1

Sub-images for golf course detection. Golf courses are delineated by green polygons through visual interpretation. (Color figure online)

Full size image

The number of golf course in Beijing increases at a high speed, and there are about 58 standard golf courses by 2013. Meanwhile, landscape characteristics of golf courses in Beijing vary a lot in terms of spatial composition and configuration. Sufficient number and various characteristics make Beijing an ideal experimental region for golf course detection.

As we can see that sub-image #1 and #4 lie in urban area where there exist park, while sub-image #2 and #3 lie in suburban area where there exist cropland. The experimental sub-images will demonstrate the capacity of proposed method in distinguishing golf course from other similar composite objects including park and cropland.

3 Methodology

3.1 Basic Elements for Golf Detection

Basic elements should be determined in advance to detect a composite object in bottom-up manner. As for golf course detection using remote sensing, the basic elements should be determined according to two principles, i.e., they are significant components of golf course and they can be extracted easily from remote sensing data.

According to land-use characteristics [10], turfgrass, water-body and bunker are determined as the basic elements. The former two elements can be extracted easily because of distinguishing spectral response and overwhelming area percentage, i.e., 67% and 7% respectively. While bunker is a significant component for discriminating golf course with other similar composite object.

3.2 Region-of-Interest Extraction

A landscape mosaic must be given to restrict the outward spatial extent in landscape analysis. In golf course detection, the region-of-interest can be broadly defined as a localized region where turfgrass and water-body co-occur, which can be delineated using NDVI and MNDWI [11] respectively.

Morphological closing is used to merge neighboring turfgrass and water-body as a region-of-interest because it can connect adjacent objects, fill small holes and smooth boundary of region when keeping region area. Constraints of region minimum and maximum area as well as area proportion of turfgrass and water-body should be taken into consideration simultaneously. Specifically, region-of-interest area should range from 20 to 200 acreages, and area ratio of turfgrass to water-body should be larger than 1.

3.3 Bunker Extraction

Bunker usually appears as mixed pixels in medium-resolution images because of its small area and complex shape, and bunker cannot be well extracted by multispectral indexes as turfgrass and waterbody. In this paper, spectral mixture analysis (SMA) based on non-negative least squares is used for bunker extraction, which is considered to be a more suitable model for target detection applications [12].

An automated endmember selection method combining pixel purity index (PPI), categorical map and endmember spectral signature is designed to select pure pixels as endmember of basic elements. In each region-of-interest, endmember is selected by following steps:

  1. 1.

    Region-of-interest PPI is calculated and pixels with high value are extracted as spectral pure pixels.

  2. 2.

    Turfgrass endmember pixels are intersection of pure pixels and turfgrass area, and water-body endmember pixels are intersection of PPI and water-body area.

  3. 3.

    Bunker endmember pixels are intersection of pure pixels and pixels with high digital numbers in red band.

Thus, region-of-interest bunker abundance given by spectral unmixing is used to extract bunker.

3.4 Clustering Based on Landscape Metrics

Region-of-interest derived based on turfgrass and water-body may contain other similar composite objects such as parks (co-occurrence of grass and lake) and cropland (co-occurrence of dry and paddy cropland). Fortunately, spatial composition and spatial pattern quantified by landscape metrics make it possible to distinguish golf course from other similar composite objects. Spatial composition refers to what elements constitute a composite object and their area percentage, while spatial pattern refers to how elements distribute in a composite object. In this paper, seven patch-level and seven landscape-level metrics shown in Table 1 are chosen to analyze landscape characteristics of region-of-interest. The “scale” in Table 1 denotes the level at which landscape metrics are calculated. Class-level metrics refers to metrics characterizing a specific land-cover class of interest, i.e., turfgrass, water-body and bunker, existing in region-of-interest, while landscape-level metrics refers to metrics characterizing region-of-interest on the whole. More detailed description of these landscape metrics can be found in FragStats [13].

Table 1. Metrics for landscape analysis.
Full size table

Following sequential forward searching (SFS) strategy in feature selection, J-M distance based filter is used to select optimal feature subset from the 14 metrics. Golf course is detected by two successive steps. Firstly, region-of-interest where there doesn’t exist any patch of the three basic elements, i.e., turfgrass, water-body and bunker, are eliminated. Secondly, region-of-interest are grouped into two categories, i.e., golf course and other objects, using fuzzy C-means (FCM).

4 Experiment and Discussion

Thresholds for turfgrass and water-body extraction based on NDVI and MNDWI respectively are determined empirically. Experiments show that it’s unnecessary to tune the two thresholds carefully. After golf course region-of-interest is derived by morphological closing of turfgrass and water-body images, SMA is used to unmix region-of-interest and pixels with bunker abundance higher than 0.3 are considered to be bunker. The region-of-interest land-cover thematic maps composed of turfgrass, water-body and bunker are shown in Fig. 2. There are 17, 9, 15 and 15 region-of-interest in sub-image #1, #2, #3 and #4 respectively.

Fig. 2.
figure 2

Categorical thematic maps of region-of-interest. Turfgrass, water-body and bunker are shown by green, blue and white pixels respectively. (Color figure online)

Full size image

J-M distance based filter gives an optimal feature subset, i.e., LSIT, SHAPEAMT, NPW, LPIB, EDB, SHAPEAML and ENNMNL, where subscripts T, W, B and L denote metrics of turfgrass, water-body, bunker and landscape respectively. Each region-of-interest thus is quantified as a 7-dimension vector, based on which golf courses are recognized using FCM. Golf course detection results are shown in Fig. 3.

Fig. 3.
figure 3

Golf course detection results. Red, green, blue and yellow areas denote true positive, false negative, true negative and false positive golf courses respectively. (Color figure online)

Full size image

Visual interpretation of 4 sub-images shows there exist 15 golf courses in 56 region-of-interest, among which there are 13 true positive golf courses, 2 false negative golf courses, 33 true negative golf courses and 8 false positive golf courses. The overall detection rate is 86.67% and the overall false alarm rate is 38.10%.

False negative golf course in sub-image #1 is a golf course where bunker can be hardly recognized even by visual interpretation, and the other false negative golf course in sub-image #3 is a small part of an entire golf course. In both cases, false negative samples show non-typical golf course landscape characteristics. The main reason contributing to false alarm lies in confusion between water-body and other dark objects in urban area, e.g., building shadow, asphalt surface and dense vegetation. Experiments show false alarm rate may decrease sharply when water-body can be extract more accurately in urban area.

5 Conclusion

A practical bottom-up golf course detection approach based on landscape metrics using multispectral remote sensing imagery is proposed in this paper. Experiments carried on SPOT-5 image achieve an acceptable golf course detection result with detection rate of 86.67% and false alarm rate of 38.10%. The bottom-up workflow enables detection of a specific composite object when its spatial composition and spatial pattern have been well understood. In other words, by presenting a case study of golf course, the proposed approach is proven to be promising for composite object detection which includes but is not limited to golf course.

Advantages of the proposed approach lie in two aspects. On one hand, the approach is essentially an unsupervised method because it doesn’t need any training samples. Only three thresholds involved in extraction of turfgrass, water-body and bunker are indispensable. On the other hand, the proposed method is more practical than existing methods since it can work well on multispectral medium resolution imagery instead of aerial or hyperspectral imagery.

In the future, accuracy of basic elements extraction and its effect on golf course detection will be further explored. Meanwhile, the proposed approach will be compared or combined with deep learning models such as convolutional neural network.