Gamma rays are often used to characterise isotopes at a distance due to their high penetrating power. The identification of radioisotopes from analysis of gamma emissions can aid informed decision making when handling radioactive material. Analysis of radioactive contaminated underground structures is typically performed using intrusive methods due to large distances between the source and detector. The presented research assessed the suitability of machine learning methods to analyse poor quality gamma spectra collected at locations exterior to the structure. Wavelet analysis was explored as a method of locating low signal-to-noise ratio photopeaks. Different filtering techniques were then implemented to remove unnecessary information from a spectrum, with a technique based on the number of connected local maxima proving to be the most effective. A genetic algorithm was developed to find a weighted combination of filtering techniques that could be applied to any gamma spectrum. This algorithm showed that wavelet analysis cannot resolve multiple photopeaks in close proximity. Neural networks were then explored; hyperparameter tuning was used to compare different network structures and preprocessing techniques. It was determined that a multi-label convolutional neural network had the highest performance when predicting the presence of isotopes in a gamma spectrum. Training sets were created using a novel data augmentation method that involved the addition of multiple experimentally collected spectra using five isotopes, ¹³⁷Cs, ⁶⁰Co, ¹³³Ba, ²⁴¹Am and ²²Na. The testing sets used were entirely composed of spectra collected experimentally, no simulated or augmented spectra were used. A categorical neural network was shown to display signs of shielding invariance and to be able to predict the presence of up to five isotopes simultaneously with high F1 score, accuracy measurements typically greater than 0.85. A regressional neural network was developed capable of determining the percentage of counts in a spectrum caused by gamma emissions from one of the same five isotopes. The average mean absolute error across all isotopes and all tests was typically limited to 10%. Simulations of a waste pipe were created and validated. A simulated and augmented testing set of 100 spectra was created and both neural networks were able to identify the presence of ¹³⁷Cs and ⁶⁰Co in all 100 spectra. Four experimental spectra were collected at the Winfrith nuclear site, the presence of ¹³⁷Cs and ⁶⁰Co was predicted in three of the four spectra demonstrating that neural networks can be used to identify isotopes in challenging deployment scenarios. The presence of ²⁴¹Am was predicted in the fourth spectrum by both neural networks, and it is likely that the networks misassociated the presence of ²⁴¹Am with background radiation in extremely low signal-to-noise ratio spectra.