Respectfully, I disagree with above comment; most of parametric classification approaches such as MLC are sensitive to the number and mathematical distribution (e.g. normality) of prototypes. As a results, they demonstrate poor results in an intricate case study. On the other hand, machine learning (non-parametric)  algorithms have emerged as more accurate and efficient choices to the parametric algorithms, when faced with large dimensional and complex data spaces and have been used for large area mapping. Of particular value, the curse of dimensionality are considerably mitigated in machine learning methods. The best methods in this realm are "SVM', "MLP" and "Random Forests". They have their own disadvantage and benefits. I would recommend you read the following paper to gather a general picture: http://dx.doi.org/10.1016/j.isprsjprs.2012.04.001

Dear friend,

The Maximum likelihood image analysis is the best method for land use / land cover classification, but, it is a probability value and the occurrences of paramedic value of multispectral wave length ranging from visual to microwave. You have choose a visual spectrum DN image value of blue, green, and red band for minimum likelihood image analysis with self sample training sets for LULC classification, and choose red and infrared band image value for land cover analysis, using Maximum likelihood image analysis, and also you may use cluster analysis for urban landscape environment or special interest with a particular LULC phenomenon. It is for your kind information that each one technique has its own merits and demerits. However, comparatively, Maximum likelihood classification is used by most of scientists or researcher who are not well acquitted with LULC analysis and they are depend on machine knowledge, and if the researcher is well knowledge in LULC image value, they are using Minimum likelihood image analysis with self sampling training sets for LULC classification. 

I think there are no "best" methods for lu and lc classification. Actually it depends on what the spatial resolution of your remote sensing image is, what the quality and number of your samples for each class, and even where your study area locates. Maybe sometimes using unsupervised classification(samples not required, relatively simple) method  would give you better results, or sometimes you need an object-oriented method(image segmentation and feature extraction, relatively complicated).

It depends upon the resolution of satellite data and terrain features. MLC should give good results. If not, you can improve upon by using spectral signature obtained on the basis of ground truth data. Use supervised classification. You make use of appropriate band data.

It depends upon what kind of area you are classifing, if you have no idea about the area then unsupervised classification is the only option, otherwise if you have significant knowledge and have numerous GCP  of that area then you can go through supervised classification (maximum likelihood) in ERDAS or ENVI software.  https://www.youtube.com/watch?v=I1FwzIC7XlU, https://www.youtube.com/watch?v=v40fLxxMSI4 or https://www.youtube.com/watch?v=n91PnT3ielw&list=PLCEBF53A31190DABE

Land use cannot be directly detected in standard RS imagery; land cover can. As above, the best cover classification method depends on number, type and spatial complexity of cover classes, imagery and ground truth. For a reality check see attached.

Beyond image classification, machine learning methods fed by landscape variables (e.g. elevation) can improve classification accuracy in complex cases by 10-15% (unpublished yet).

Article A comparison of fine resolution census and image-based natio...

I strongly supported the answer given by  Hein Van Gils.