O Que É Depressão Geográfica

What depressões geográficas actually are

When you run hydrological flow analysis on a digital elevation model, most cells have a clear downslope direction. Some cells don't. Those are depressões geográficas — also called sinks or pits — and they break flow routing algorithms dead. A depression is any area where water would pool indefinitely because every surrounding cell is higher. In real terrain, this happens in legitimate basins like the Caspian Depression. But in DEM data, most depressions are noise: missing elevation points, interpolation artifacts, or grid resolution limits creating false bowls in what should be a continuous slope.

O que é depressão geográfica e por que ela estraga suas análises hidrológicas

Depressions aren't just a minor inconvenience. If your flow accumulation model hits one, the algorithm either stops routing or routes everything into that one cell, which cascades into garbage stream networks downstream. You'll see zero-flow cells, fragmented watersheds, and outlets that point uphill. It looks wrong immediately if you know what to look for. I hit this once on a LiDAR-derived DEM at 1-meter resolution for a flood modeling project. The terrain was mostly agricultural with gentle slopes. One watershed had a perfectly clean outlet on every other test tile, but this one catchment showed complete flow stagnation in its midsection. After a day of debugging, I found it was a single row of cells — maybe six meters wide across the valley floor — where the laser returns were all filtered out as ground returns and never filled in. The interpolation between adjacent valid cells dropped about two meters, creating an artificial channel that no flow direction algorithm could route out of.

The fix wasn't fancy. I used a filled depression routine with a tolerance threshold of 0.5 meters, then ran a breach algorithm specifically at the identified outlet before re-running flow direction. The breach cost about three minutes. The original fill approach would have taken twenty and potentially altered legitimate flat areas.

How to detect and handle them in practice

The standard workflow in ArcGIS, QGIS, or GRASS is straightforward if you've seen the failure modes before: First, run a sink detection tool. Most GIS platforms have one built in. In QGIS it's under the SAGA toolbox — Grid Geometry > Fill Sinks. In ArcGIS Pro it's the Hydrology > Fill tool. This produces a modified DEM where every depression is raised to its spill point elevation.

Then run flow direction, flow accumulation, and stream network extraction on the filled DEM. Verify the output looks topologically correct by comparing stream order against known features if you have any ground truth data. Here's the thing beginners miss: filling all depressions blindly is often wrong. If your terrain has legitimate endorheic basins — wetlands, salt flats, true internal drainage areas — the fill tool will obliterate them and route water out where it never actually goes. The breach method is more selective. Instead of raising every low cell, it finds the lowest spill point along each depression's perimeter and carves a single-cell notch through the rim. Water can now escape without altering the interior shape of the basin. In WHC/WHAN or WhiteboxTools, this is the FillByBreach algorithm. It's slower than full filling but preserves real geomorphology.

I recommend using FillByBreach when your DEM comes from a modern source like USGS 3DEP LiDAR or EU DEM v2.1. With older SRTM or ASTER GDEM data, the sheer density of artificial depressions from sensor errors sometimes makes brute-force filling the only practical option, even if it sacrifices some realism.

Known limitations you should expect

Depression handling breaks in several common scenarios. Flat terrain with near-zero gradient is the biggest one. When slopes approach the vertical datum precision of your DEM, flow direction becomes ambiguous or random. A 1-meter DEM over a floodplain with a 0.001 slope might produce flow directions that flip between iterations depending on tie-breaking rules in the algorithm. Filling doesn't help here — you need to apply a slope-based damping filter or use a D8-limited approach with a minimum slope threshold. Coastal and riverine inundation zones are another failure mode. The fill algorithm treats a breached levee or channel exactly the same as an elevation artifact. If you're modeling storm surge or river flooding, you absolutely must mask out engineered features like dikes, levees, and culverts before running any depression fill. Otherwise you'll route ocean water inland through a fill-created breach in a levee that exists in reality but not in the bare-earth DEM.

Resolution matters more than people admit. A 30-meter DEM will have far fewer artificial depressions than a 1-meter DEM, but the fill procedure itself introduces elevation errors that scale inversely with resolution. At 1 meter, a fill can alter ground height by half a meter or more in flat areas. At 90 meters (SRTM), the errors are smaller in absolute terms but proportionally larger relative to the actual terrain variation in flat landscapes like the Amazon floodplain. If you're working with very large domains — entire river basins at 30 meters or finer — the fill step becomes a computational bottleneck. The algorithm is O(n) but the memory footprint for the stack-based filling routines can exceed 16 GB RAM on a 50,000 by 50,000 grid. I've run into this and switched to using WhiteboxTools from the command line with tiling enabled, which processes the raster in chunks and writes intermediate results to disk. It cut my runtime from six hours on a single threaded ArcGIS session to about forty minutes on the same machine.

The bottom line: depressões geográficas are a normal part of working with gridded elevation data, and handling them properly requires understanding whether each depression is real or an artifact. Blind application of fill tools produces passable results for rough regional analysis but fails under scrutiny when you need hydrologically accurate stream networks or flood inundation maps.