This section details the steps in processing satellite imagery of Singapore into a land cover map for natural capital assessment. For the purposes of this study, only the terrestrial, freshwater environments of Singapore will be mapped in detail, while the marine environment that is underwater will be classed as just ‘Marine’. The exception is the mangrove forests, which cross the coastal boundary but generally have at least a part of the vegetation above the water at all times [17
]. The classification uses a hybrid approach; first conducting a supervised machine learning to classify broad land cover types, before adding more detailed sub-classes using secondary sources of spatial data.
Singapore is a city-state located in the tropics with its central point at latitude 1.21° N and longitude 103.49° E. The land area of Singapore measures 724.20 km2
] that consists of one main island and many smaller islands within its territorial boundary. Singapore’s location near the equator means that satellite images are frequently obstructed by cloud cover due to convection of water vapour in the atmosphere [19
]. Hence, it is difficult to find a single cloud-free satellite image of the entire island. Therefore, multiple satellite images acquired over a period of time are needed to provide island-wide coverage. Although the country is highly urbanized, there are many pockets of green spaces in between buildings and tree-lined road networks. Major green spaces in the country are at the Western and Central portions of the main island, and Pulau Ubin and Pulau Tekong in the northeast. As with high-density urban areas around the world, Singapore’s urban form is highly heterogeneous, incorporating industrial commercial and residential land use zones, with a mix of high- and low-rise buildings [20
3.3. Image Classification
A summary of the object based image classification detailed in this section is presented in Figure 3
. To classify high resolution imagery, an object-based approach was used instead of a pixel-based classification technique, to avoid ‘salt and pepper’ effects on the resultant map [23
]. To segment images into objects, a means-shift approach was applied using the segment means-shift function [24
] in ArcGIS. In the parameters of the function, spectral and spatial details were both given maximum importance (set as 20.0 and 1.0, respectively) to discriminate as best as the function can between features in the landscape, for example, trees and grass [25
]. The spectral detail setting was used in means-shift segmentation to discriminate objects based on spectral signatures [25
]. The spatial detail setting was used to discriminate objects based on the shape of the features to produce sharper segments, like buildings and roads within impervious surfaces [25
]. A minimum mapping unit of 300 pixels (approximately 5 m2
) was set to save on processing time and storage space. Since the function can only read three bands in a composite raster to be segmented, the multispectral bands of near infrared (760–900 nm), red (630–690 nm), and green (510–600 nm) were used to focus on discriminating vegetation features. The segmentation produced 25,882,810 objects in the study area that were visually checked to ensure that they enveloped a meaningful object (e.g., building outlines, jetties, and grass patches).
Next, the data from the multispectral satellite image bands were added as attributes to each object to be classified. The zonal statistics function was used to calculate the median of each of the multispectral band pixel values from all pixels within the objects created. Eight multispectral bands were available for the WorldView images [26
], while four were available for the QuickBird images [27
]. The objects were exported to R statistical software 3.5.3 [28
Five broad-level land and water cover types were initially classified using a supervised method using a random forest algorithm 4.6 [29
]. The classes were impervious surfaces; pervious bare surfaces; trees; grass; water. Shadows cast by tall buildings and trees were also classified as a separate class. A separate random forest classification was conducted for each image individually. To train the random forest classifier, at least 150 objects were visually selected for each class by hand on ArcGIS Desktop. Hence each image would have at least 900 training points. The random forests were built with 500 trees with two variables tried at each split [29
]. The out-of-box (OOB) estimates of error rate were all less than 10% for each image classified. The 17 classified images were then mosaicked together starting with the earliest image (Image 901) in ascending date order (Table 3
) with the most recent imagery replacing areas of overlap.
The precision of land cover was further refined into land use classes with data inputs from other sources. Shadows were first dealt with using the zonal statistics to estimate which of the five preceding land cover classes lie in it. The remaining patches of shadows that were not fixed were re-classified manually based on cross referencing to Google Earth Images (that are also high resolution) and local knowledge of the area. Impervious surfaces were refined into buildings with building footprint information downloaded from OpenStreetMap [30
]. Areas of vegetation were also divided between managed and unmanaged vegetation by manually digitizing 2014 SPOT5 satellite image of Singapore based on ground-truthing and a previous lower-resolution vegetation map of Singapore [12
]. Vegetated areas that intersected with swamp and marsh classes from the aforementioned map of Singapore [12
] were reclassified as these classes accordingly. Inland freshwater ecosystems were manually reclassified into water courses (rivers, canals, drains) and water bodies (lakes, reservoirs, swimming pools) based on knowledge of the freshwater network [31
Finally, the map was error-checked and manually corrected with on screen digitization and rectification of errors in classification. This was done systematically with a regular grid laid out across the study area with a size of 1990 m by 1200 m. The map classification within every one of the 649 grids was manually checked for classification errors at a map scale of 1:5500. The erroneous raster pixels were edited using the raster painting [32
] tool in ArcGIS to selectively convert misclassified raster pixels to the correct ones.