Geochemical data from drill core samples is used in mineral exploration to enable geologists to understand subsurface geology. However, it is challenging to analyse these data sets rigorously and consistently. Manual data interpretation relies on the knowledge and experience of logging geologists. When multiple geologists are involved, it is possible to introduce inconsistencies in logged lithology type and scale of logging (‘splitters’ versus ‘lumpers’). Machine learning (ML) is a powerful tool for analysing high dimensional data (ie data with many variables) and large data sets (ie data with many samples). However, ML methods applied to drill hole samples do not incorporate spatial information and this can result in high misclassification rates and small-scale ‘noisy’ results. We demonstrate the use of multiscale spatial analysis (wavelet tessellation) as a pre-ML step to analyse spatial information. This step helps reduce misclassification and filter out unwanted small_x0002_scale variation in the data. In an experiment, we used wavelet tessellation and supervised ML to predict litho-geochemical rock types from geochemistry data. The data set is from the Valhalla uranium deposit near Mount Isa, northern Queensland. First, geochemical knowledge and statistical analysis was used to select appropriate geochemical elements for litho-geochemical classification. Second, the elements were processed using multiscale spatial analysis. Third, we trained a Random Forest model with the data set and found that the model performs worse in predicting rare rock samples than common ones. We applied over-sampling and down-sampling methods in aim to overcome this issue. As a conclusion, the combination of multiscale spatial analysis (to define rock unit boundaries) and Random Forests (to classify rock units) is a good method for rapid and accurate analysis of drill hole geochemistry.