Influence of training data variation and biogeochemistry covariates on spatial modeling of acid sulfate soil
2024
Roslund, Tobias
Acid sulfate soils have been described as one of the nastiest soils in the world. As marine sediments rich in organics deposit in an anaerobic environment, hypersulfidic materials can potentially form. As areas where such sediments have been deposited are uplifted and subsequently drained, oxidation can occur and form highly acidic sulfuric acid (H2SO4). Leaching of soils containing H2SO4 can have adverse consequences on ecological systems, infrastructure and potentially human health. It is thus of utmost importance to know where potentially acidifying soils are located. One method that has been used for mapping these types of soils in northern parts of Sweden is machine learning. Improving these models is thus desirable. Using a random forest model and data from Malardalen in Sweden, training data has subsequently been reduced and accuracy evaluated. Data has also been clustered, and a fictional sampling strategy has been tested. Further trying to improve the model, the addition of stream plant biogeochem-ical data as a covariate has been tried. Using ArcGIS, concentrations of different metal ratios have been averaged over drainage areas around Malardalen and added as a covariate. Finally, the best performing metal ratios were applied to the same data variations used in the beginning of the thesis and evaluated for accuracy. The random forest model used seemed to be mostly dependent on amount of training data over other factors. The addition of stream plant biogeochemical data as a covariate seems to potentially increase the accuracy of the model. The addition did however not seem to be able to compensate for changes in training data quantity or quality.
显示更多 [+] 显示较少 [-]