It should be noted that I deal primarily with geo-spatial image analysis, so there is a not insignificant amount of bias with regards to what data I'm interested in. I like using the USDA NAIP API for imagery, since I can call in imagery using GDAL directly into python or R. I rely heavily on freely available public utility data sets (Parcel level utility data). Beyond that and other than as a starting point, you're training data is always going to be something you've invested in heavily. Good training data is 100% the game. No modeling exercise is going to go well on poor quality training data. Currently as a personal project, I'm trying to develop a platform for developing and training data for geospatial modeling. If you're interested, hit me up on a PM and I can explain it in more detail (after work).