Land Use / Land Cover Classification

Every surface reflects sunlight differently across the spectrum. Water swallows most near-infrared light; healthy vegetation reflects a lot of it; concrete and tin roofs are bright almost everywhere. A satellite sensor records these differences as numbers in several bands (wavelength ranges). Two pixels that look the same shade of grey on screen can have quite different band values - and that is what a classifier reads.

Classification is the process of assigning each pixel to a class based on its band values. The output is a new raster where the numbers no longer mean brightness - they mean category: 1 for built-up, 2 for cropland, 3 for forest, and so on. This is what people mean by a Land Use / Land Cover (LULC) map. Strictly, land cover is what is physically there (trees, water, concrete) while land use is what people do with it (a park, a farm, a market); a satellite sees cover directly and infers use, so honest maps say LULC.

There are two broad routes. In supervised classification, you teach the machine first. You draw small training samples (regions of interest, ROIs) over places you are sure about - this patch is forest, that one is water - and the algorithm learns each class's spectral signature, then labels every other pixel by similarity. A common method is maximum likelihood, which asks 'which class is this pixel most probably in?'

In unsupervised classification, you ask the computer to group pixels into a number of natural clusters by their band values alone (methods like ISODATA or k-means). It hands you, say, 15 spectral clusters with no names; you then inspect them and label each - cluster 4 is water, clusters 7 and 9 are both cropland. Supervised needs field knowledge up front and gives you exactly the classes you want; unsupervised needs no prior labels but leaves the interpretation to the end. Most real projects blend the two.

Your class list should follow your question, not the other way round. A green-cover study might need only three classes (vegetation, water, everything else); a master-plan base layer needs built-up split from cropland from wasteland. India has a ready reference here: the NRSC national LULC mapping programme classifies the whole country on a standardised, hierarchical scheme - broad Level-I classes such as built-up, agricultural land, forest, wasteland, water bodies and wetlands, each nesting finer sub-classes. Borrowing an agreed scheme means your 'forest' means the same thing as everyone else's.

Resolution decides what you can even attempt. Sentinel-2 (free, 10 m in the key visible and near-infrared bands) resolves individual large plots and neighbourhood green space. Landsat (free, ~30 m) is coarser but reaches back decades - priceless for history. India's own Resourcesat LISS-III (~23.5 m) and AWiFS (~56 m) add regional coverage. A 30 m pixel cannot 'see' a single house or a footpath, so match the sensor to the smallest thing you must map.

A classified map always looks convincing - it is crisp and coloured. That is exactly why you must test it. Accuracy assessment compares the map against independent reference points you know the truth for (field visits, high-resolution imagery), and tabulates agreement in an error matrix (also called a confusion matrix). From it you read overall accuracy, plus per-class producer's and user's accuracy, and often the kappa statistic.

The error matrix is honest in a way the map alone is not: it shows which classes get confused (young plantation misread as cropland; bright bare soil misread as built-up) so you can go back, add training samples, and re-run. A LULC map quoted without an accuracy figure is an opinion dressed as a measurement. Report the number, and report how you got it.