Data Cleaning & CRS Transforms

Data reaches you from scans, downloads, other people's projects and field devices, and each source adds its own mess. Geometries can be invalid — a polygon that crosses itself (a "bowtie"), a ring that does not close, duplicate vertices. Attributes can be inconsistent — the same land use spelled three ways, numbers stored as text, nulls where zeros belong, units mixed. And whole layers can be in the wrong or mismatched coordinate reference system, so they will not overlay.

Cleaning is unglamorous and it is most of the job. The payoff is that everything downstream — the joins of the last lesson, the buffers and overlays of Module 4, the terrain work of Module 5 — assumes clean, valid, consistently-projected data. Skip the bench and every later result inherits the dirt.

An invalid geometry is a shape that violates the rules of what a geometry may be — most commonly a polygon whose boundary crosses itself, or one with a self-touching ring. Invalid shapes make area calculations wrong and make overlay tools (clip, intersect, union) fail or return garbage.

The workflow is simple: detect, then repair. A validity check lists each problem feature and often names the fault ("self-intersection at…"). A fix-geometries operation then rebuilds the shapes into valid, simple rings — splitting a bowtie, closing an open ring, removing duplicate points. Run the check first so you know what you are fixing, run the fix, then re-check to confirm zero errors remain. Do this before any overlay or area work, not after you get a strange result.

Geometry is only half of clean. The attribute table usually needs work too. Standardise category values so "Resi", "residential" and "RESIDENTIAL" become one class — inconsistent spellings fracture every count and join. Fix field types so numeric fields really are numeric (recall last lesson's silent-join failures). Distinguish a genuine null (unknown) from a zero (known to be none), because averaging over nulls versus zeros gives different answers. Trim whitespace, reconcile units into one system, and remove duplicate records.

Use the field calculator and find-and-replace for bulk fixes, and — crucially — keep a copy of the raw data. Cleaning is editing, and editing is destructive; a preserved original is your only way back if a "fix" turns out to be a mistake.

Coordinate work has two operations that beginners fatally confuse. Defining (assigning) a CRS tells the software what system the existing numbers are already in — it relabels, it does not move a single coordinate. You do this only when a layer's CRS is missing or wrong (a lost .prj from Module 0). Reprojecting (transforming) actually recomputes every coordinate from one CRS into another — the numbers change so the features land in the same real place under a new system.

Get these backwards and you cause disasters. Reproject a correctly-defined layer and it is fine; but re-define a layer that was already correct and you have lied about where it is — it will jump across the map. The rule: if a layer sits in the wrong place because its CRS label is missing/incorrect, define. If it sits correctly but you need it in a different system to match others or to measure, reproject.

Most GIS shows layers of different CRSs together using on-the-fly reprojection — a convenience that displays them aligned without changing the files. Do not mistake that for done: many geoprocessing and every honest measurement should run on data actually stored in one appropriate projected CRS.

Tie this back to Module 0's codes. Store and share in geographic EPSG:4326 (WGS 84) — but never measure area in degrees. To measure length or area over a region, reproject into the local UTM zone — EPSG:32643 / 32644 / 32645 for much of India — which is in metres. For pan-India work in one metric frame, ISRO's national EPSG:7755 (WGS 84 / India NSF LCC) is a Lambert Conformal Conic, not a Transverse Mercator despite a widespread myth. Reproject a working copy for the task, keep your master in a stable CRS with its .prj intact, and record the CRS in the metadata.