Determining decision thresholds to separate landmines from clutter is a critical but difficult task in landmine identification procedures because such thresholds depend upon local geology, distribution of landmines and clutter, data acquisition, and cultural and environmental noise at a given landmine cleanup site. We describe a learn-as-you-go approach to build an electromagnetic induction spectroscopy (EMIS) library and decision thresholds for humanitarian demining operations, which include uncertainties regarding the target/sensor position and orientation, target depth, geological and environmental noise, and measurement errors. This approach is then simulated mathematically using electromagnetic (EM) data obtained over the calibration plot at a test site. The resulting EMIS library and thresholds are tested on the EM data colleted in the test squares at the same site. Apparent conductivity converted from the data is used as a detector function, and both spectral shape and amplitude response are used quantitatively in the matching algorithm. Our results show that the detection rates are increased from 92% to 98% and identification rates from 76.2% to 95%, and the false alarm rate is reduced from 13% to 5%, compared with the previous results obtained from the same dataset.