The performance of a prompt gamma neutron activation analysis (PGNAA) system for lower-weight landmine detection is investigated in this study. A total of 2880 characteristic gamma-ray spectra of 10 buried samples (five explosives and five nonexplosives), within a weight range of 0.01 to 10 kg and a hidden depth of 2.5 to 15 cm, under 0%, 10%, and 20% soil moisture conditions, were generated using Monte Carlo N-Particle Code 5 (MCNP5). The conventional characteristic peak analysis method was not applicable to lower-weight sample detection. The discrimination accuracy was acceptable only under 0% soil moisture when explosives exceeded 2 kg with the discrimination accuracy exceeding 80%. Four machine learning models, including radial basis function (RBF) neural network, fully connected neural network, XGBoost, and LightGBM, were used to perform whole-spectrum analysis, and better performance was demonstrated. The discrimination accuracy exceeded 90% in most cases, and the RBF neural network was demonstrated to be the best performance (96.6% for explosives and 95.1% for nonexplosives). All four of these models were insensitive to soil moisture. The minimum detectable weight of 0.02 kg for the simulation data provided valuable reference for experimental applications. These results indicate that machine learning was an effective method for lower-weight landmine detection using PGNAA under complicated conditions.