2.1. TRIP architecture

TRIP design is required to satisfy the following criteria.

1. The locations of each river basin are correct.
2. The river basin size and the drainage area to each gauging station point (GSP) are reasonably accurate.
3. The length from each grid to GSPs or to the river mouth reflects their real length.

The first requirement is obvious, because TRIP will be used to overlay various climatic data. Since most of the continents are covered by major river basins, their locations and boundaries should be realistic. The only exceptions where this becomes difficult are the regions next to arid areas, where major runoff does not occur.

The second requirement, although obvious, is nevertheless important. Using the predetermined drainage area to the GSP, discharge is converted into runoff, which has a dimension of water depth per unit time and can be easily compared with precipitation and evapotranspiration. Therefore, the error in the drainage area directly affects the accuracy of conversion from discharge to runoff. Fortunately, the areas of major river basins have been published, and most GSPs are attributed with their drainage areas. Comparison of the areas obtained from TRIP with these values provides validation necessary for evaluation of TRIP design. This comparison also serves as a verification for the first requirement.

The last requirement is related to the application of TRIP with runoff routing models. The histogram of the length from each grid point to the GSP may have a major influence on the hydrograph formation of a large river basin. Consequently, it is desirable that the length of the river is reasonably accurate.

In constructing the water flow channel network for TRIP, each grid point was idealized to have only one outflow direction with reference to its eight neighboring grid points: N, NE, E, SE, S, SW, W, and NW. It may be classified as “D8” and its disadvantages are discussed in Costa-Cabral and Burges (Costa-Cabral and Burges, 1994). They show that statistical or partial treatment of river channel network, in which each grid may have two or more outflow grids, may be more realistic in some cases, particularly for regions with complicated boundaries. For the global domain, however, the simple format amply satisfies the requirements stated in the previous section; that is, it represents the area size adequately and relates to the lateral flow directions realistically. The numbering of direction is set 1 to 8 clockwise from north. The 0 value is used for a sea grid, and 9 is used for a “river mouth” grid defined as a grid that has no outflow to its neighboring land grids.

2.2 River flow directions in TRIP

There are two known methods of obtaining river flow directions: 1) digitizing water flow directions that can be delineated manually from published maps or 2) extracting the water path directions by a numerical analysis procedure from digital elevation models (DEMs) objectively (automatically). Even though laborious, the first method guarantees accuracy, while the second method is also bound to produce the valid results in a majority of the cases, provided that the quality of the DEM used is accurate over all the target region and the automatic algorithm is able to sort out the pathflows accurately. River channel information on Earth is already available as vector data. Dirmeyer (Dirmeyer, 1995) converted the vector information into raster data. Nevertheless, it is still helpful to extract river channel information from DEMs, because the density and quality of the existing vector data of river channels are not sufficient, as was also pointed out by Dirmeyer (Dirmeyer, 1995).

There have been earlier attempts in making a river channel network from DEMs (e.g., Tarboton et al., 1991), but most of them were applications for high-resolution situations in which DEMs were used on 1–100-m horizontal resolution. On such a small scale, automated techniques, for defining river channels, can extract detailed water flow paths on small river basins. Our goal is to create TRIP for the use in global models that have a much coarser resolution and requires resolution to capture the large-scale picture. Because of the scale difference, the following issues must be addressed.

1. Valley floors are not well represented by DEMs, because a 1° mesh corresponds to approximately 100-km horizontal resolution.
2. The treatment of “pits” or “hollows” in the continents, where the altitude of the grid box is lower than all the neighboring grid boxes, is a major problem to be dealt with. Generally these pits and hollows are artificially filled up in the DEM in order to prevent the dead ends of streams produced by automatic procedures. However, there are inland river basins on the continental scale and the dead ends of streams for such rivers are correct.
3. TRIP design should identify lakes. Some of them are inland lakes, and some of them have outlets for streams that eventually lead to the ocean. On smaller scales, generally lakes were not treated by automatic techniques, but for the larger scales lakes should be included.
4. Most of the previous studies extracted river channels for a specific river basin from DEM, but the number of river basins in TRIP is not prescribed.

Problems 1 and 2 are related to the horizontal resolution of the DEM used, while 3 and 4 are related to the target region being global. In studies of small scales, results from automated techniques were validated vis-à-vis published maps or air photos, but such a validation is not viable for TRIP, because our goal is to extract the broad features of water flow that emerges from subgrid-scale water flow structures. We believe any river channel network produced by automated techniques needs to be verified against real-world atlases, and wherever the outcome differs from the actual direction of river flow, we introduce manual correction. As a consequence, the river channel network extracted by an automated technique was viewed as a first guess, and it was modified by manual corrections to address the difference between real versus ideal world.

2.3. Automated extraction of TRIP from a DEM

To produce the first guess of TRIP, river flow directions were determined from a global DEM called ETOPO5 (Edwards, 1986). ETOPO5 has 5′ × 5′ horizontal resolution; thus, it has 12 × 12 cells in each 1° × 1° grid box. In each grid box, mean elevation of 5′ × 5′ cells above 0 m was calculated and used for the extraction of river channel network. The values of medium, maximum, and minimum elevations in each 1° grid box were also prepared.

The delineation of land and sea followed the land–sea mask of the ISLSCP Initiative I CD-ROM (Meeson et al.,1995; Sellers et al.,1995). Large lakes were classified as sea in the land–sea mask, but some of them are parts of drainage systems of rivers. Therefore, such a sea grid box in the land–sea mask was changed into a land grid box, and its outflow direction was attributed during the manual correction step. Thus the lakes in current TRIP are grid boxes that have flow directions but are classified as sea in the land–sea mask of the ISLSCP Initiative I CD-ROM.

As for most D8-type river channel networks, the idea of “hydrological flow modeling method,” as illustrated by Chorowicz (Chorowicz, 1992), was employed. The outflow direction was toward the lowest land point of the eight neighboring grids, provided the point was lower than the originating point. It might appear that the direction toward the steepest slope should be the obvious choice but considering that there is meandering and unevenness within the subgrid scale, it is assumed that somehow water will find a way to reach the lowest level. We shall show that the choice between steepest slope and lowest neighbor was not a clear one because both methods produced a similar number of grid cells requiring subjective correction (see section 3.2). In fact, approximately 70% of the directions determined by these two algorithms were identical. Therefore we opted for the simplest choice, that is, lowest neighbor method. If more than one outflow directions were found, the median, maximum, and minimum values of neighboring 1° × 1° grid boxes were examined in this order. If a grid point was lower than all neighboring grids, it was marked as “hollow” and was left for the subjective evaluation and manual correction. If such a hollow grid box had a sea grid in its immediate neighborhood, it naturally became a river mouth.

Antarctica was processed differently. Lateral flow in this area is glacier flow and not river (water) flow; therefore, the river channel network merely represents internal consistency of TRIP. If ice height was lower in the grid box to the north, northeast, or northwest, outflow direction was set to that direction. In all other cases, outflow direction was set to the north.

The result obtained from the above algorithm showed fairly realistic pathways as compared to the observed pathways. Rough stream structure and the drainage system to the ocean were well represented by the automatic procedure. The hollow grids were found to be located appropriately in areas such as at marshes and lakes or at the confluence of large tributaries.

2.4. Manual correction of TRIP

A manual correction was applied to the TRIP network, obtained from the automated method, by comparison with published geographical information on water flow networks. This is done by overlaying TRIP on the actual river channels using the General Mapping Tools of Wessel and Smith (Wessel and Smith, 1995). Global continents were divided into 27 regions, and the river channel network was carefully examined and corrected. Refer to Figure 10 as an example of overlaid maps. Additional information to verify and correct TRIP was obtained from two atlases (Teikoku-Shoin,1985; Rand McNally,1995).

Some of the problems encountered in the river channel network are as follows. The first problem relates to the diversion of a river channel into two or more major streams. This is a common occurrence at a delta near the river mouth. There are only a few rivers whose deltas are resolved on a 1° × 1° grid; however, when it gets resolved, it is impossible to assign the flow directions because TRIP design allows only one outflow direction. For example, this situation arose in the Amazon, the Nile, and the Lena River pathways. For all these cases, one direction of outflow was assigned as a major pathway referring to the atlases referenced above. Since there are no major gauging stations in these areas, these approximations are not likely to cause a serious problem for simulating/evaluating river flows from runoff at grid points.

Most of the other problems were encountered in arid and dry land regions. There are a number of wadis in dry regions, and small streams from these regions occasionally run into a main stream. It is difficult to decide whether to include such a stream flow in a major river basin, but lateral flow of water in this region would be inactive for all practical purposes and the decision of flow direction would have an inconsequential effect, particularly for GCM-scale models. Some grid boxes with inland lakes were attributed to river mouths, and river mouths are also arbitrarily put into arid areas in order to terminate water flow from mostly dry channels into episodic ponding. This is indeed realistic, but the automatic algorithm was unable to handle it appropriately.

The following features of TRIP were checked in order to correct possible inaccuracies caused by manual operations.

1. No river channels are allowed to crisscross.
2. All river channels flow from one land grid box to another.
3. Every land grid box has one, and only one, river mouth toward its downstream, which eliminates the possibility of grid boxes counterflowing toward each other.

2.5. Postprocessing of TRIP

TRIP divided the entire land into river basins. Each river mouth was given a temporary river basin number (RBN). To assign an RBN to a grid point without an RBN, the downstream grid point was examined, and the upstream grid point received the same RBN as the downstream grid point. This procedure was followed for all grid points and repeated until all the land grid points had an RBN. The chosen values of RBNs were allocated in the descending order of basin sizes. Thus, a larger river basin has a smaller RBN.

In order to demarcate the river basin boundaries, color-coded numbers were attributed to each river basin. Diagonal grid boxes were considered to be in contact and they were given different color codes if they belong to different river basins. In the current TRIP, seven colors were sufficient to distinguish adjoining river basins.

The order of streams was determined by an automated procedure. We define headwaters as the grid points with no inflow from any of their eight neighbors; the streams from the headwaters were assigned the river order of 1. The headwater grid cells were identified first, then each stream was given the grid distance from headwaters. All the grid points with 1 grid distance from the headwater grids were tested for water inflow. If two or more streams of river order 1 flow into the grid box, the outflow stream was attributed the river order of 2; otherwise, it continued to be river order 1. The confluence of two or more streams of n order produces an (n + 1)-order stream. All the grid boxes were examined in this way in order from the shorter grid distance to the headwaters. Because more than three tributaries can make a confluence, the river orders of all inflow streams were stored during these steps, and the river order was increased by +1 when the highest river order and the second highest river order were the same.

A river catalog has been prepared that identifies the river basins (see Table 1). The location of river mouth, its name, and its official river basin size are given in the catalog. Major river basins with areas larger than 100,000 km2 are all included in this catalog. Since we recognized the possibility of estimation errors, we extracted river basins with official areas of more than 50,000 km2 from Korzun (Korzun, 1978, hereafter referred to as KU78) and compared with the basin sizes in TRIP.

A similar catalog, but for selected gauging stations of river discharge, has been prepared as well. The selections and the locations of gauging stations are from the Global Runoff Data Centre catalog (GRDC, 1992). Some of them are from ISLSCP Initiative I datasets. Drainage area numbers were attributed to each grid point by an algorithm similar to the one used for the river basin numbering described above.

The distance between grid points was computed by an algorithm described in appendix A. The length of the stream from the grid point to its river mouth was also computed. The stream length at the grid box of the river mouth was set to zero. The longest stream length in each river basin was taken to be the representative length scale of the river basin. Indeed, that is not the case in some river basins such as the Mississippi. Itasca Lake, in Minnesota, is defined as the origin of the Mississippi River, and the length from Itasca to the river mouth, New Orleans, is approximately 3800 km. However, the Missouri River has the longer stream length, with more than 6200 km for the entire Mississippi River.