You don't mention scale or ...

You don't mention scale or if you already have appropriate land use data classified by others, but you do mention deployment.  If you want open source tools to create web services based on land cover change, you might consider the Django-raster package (https://github.com/geodesign/django-raster).  It lets you create tiled map services feeding slippy-maps, or raster polygon area statistics as tables or json.  If you don't need that level of control, you could use a spatial web service like ArcGIS Online (ESRI), or Data Basin (https://databasin.org). 

If you don't actually have the change data classified, but instead have raw pixels, then you need classification tools.  For programmatic classification of multispectral imagery, we use and recommend scikit-learn (http://scikit-learn.sourceforge.net).  Its got a steep learning curve, but once over the hump supports a huge range of state-of-the-art classification techniques (SVM, Random Forests, etc).   You can get excellent land cover results at 10m resolution with Sentinel-2 (free 13-band imagery from ESA). Better resolution than Landsat and 3x more frequent coverage, but same principles apply.  

Keep in mind that scikit-learn and similar classifiers work well only when spectral color at pixel level is enough to separate classes.  If your 'change' is harder, imagery shallower, or in a place with lots of clouds and atmospheric interference, then you need approaches like atmospheric correction and deep learning.  PlanetLabs and similar high-res imagery fit here.  Online platforms for this are only just emerging, but check out Google's Earth Engine (https://earthengine.google.com)  and DigitalGlobe's GBDX (https://platform.digitalglobe.com/gbdx/).    

Lastly, for hydrological classifications, we've had some success that using vegetation and impervious cover indices to generate basic hydro model inputs.  You can get a long way with the NDVI (normalized vegetation difference index), for example.  This avoids the supervised classification step, and just uses "band math" within a single image to show relative greenness.  If you ultimately need impervious cover plus vegetation aggregated per sub catchment, such as for SWMM, then you can save substantial time simply thresholding indices "by eye."  You don't get the confidence intervals of a formal supervised classification, but can get a near-immediate results on very recent imagery.