<p>The estimation of the volumes of contaminated soil to be treated is a crucial step in soil remediation. Numerous techniques exist for estimating the distribution of pollutants in soils, such as inverse distance weighting, kriging, Gaussian sequential simulation, and sequential indicator simulation. Unfortunately, these methods require significant computational resources to achieve precise estimations. Moreover, both kriging and Gaussian simulation require the transformation of non-normal distributions, often seen in hydrocarbon contamination, to produce accurate results. In this paper, we propose a generative neural network to generate three-dimensional maps of contaminant distributions without prior training, and to estimate the contaminated volumes. This differentiates this work from other deep learning approaches that necessitate training data. The proposed method relies on a convolutional neural network for image reconstruction and inpainting. Rather than solely depending on the concentration of chemicals determined in the laboratory, we utilize hyperspectral imaging data from soil cores to achieve a more precise depiction of soil contaminants. We assess the proposed method using a synthetic three-dimensional dataset and a real case of hydrocarbon pollution on a polluted site in France. The method demonstrates competitive performance with efficiently managed computation time, achieved through the use of a GPU accelerator. This study offers a new, practical way to improve soil pollution management using fast, data-driven techniques.</p>