This paper addresses the problem of passive source depth discrimination in ocean acoustics using a horizontal line array (HLA). The scope is restricted to low-frequency sources (frequency f < 500 Hz), broadband signals (bandwidth B of a few Hz), deep-water environment (water depth D > 1000 m), and distant sources (range r greater than several km) at the endfire position. In this context, the environment acts as a dispersive waveguide, and one should not use classical source localization methods based on plane waves or any other simplistic wave model. Instead, a method based on the modal behavior driving the propagation is proposed. It notably uses the concept of the waveguide invariant, a scalar that summarizes the waveguide dispersion. In deep water, the waveguide invariant largely depends on source depth, and thus is an interesting input for source depth discrimination. An algorithm is proposed to compute energy ratio in groups of modal interferences. The input data for the algorithm is a range-frequency intensity, as measured on a HLA. The modal interference groups are defined based on their respective waveguide invariant values which in turns depend on source depth. This idea is formalized to propose a source depth discrimination method, which is performed as a binary classification problem. As long as the sound speed profile features a surface thermocline, the algorithm does not require detailed knowledge about the environment and it allows the classification of sources under two hypotheses, above or under a user-chosen threshold depth.