Szczegóły publikacji
Opis bibliograficzny
Thorough statistical comparison of machine learning regression models and their ensembles for sub-pixel imperviousness and imperviousness change mapping / Wojciech DRZEWIECKI // Geodesy and Cartography ; ISSN 2080-6736. — Tytuł poprz.: Geodezja i Kartografia ; ISSN: 0016-7134. — 2017 — vol. 66 no. 2, s. 171–209. — Bibliogr. s. 207–209, Abstr. — Publikacja dostępna online od: 2017-12-29
Autor
Słowa kluczowe
Dane bibliometryczne
| ID BaDAP | 113228 |
|---|---|
| Data dodania do BaDAP | 2018-04-19 |
| Tekst źródłowy | URL |
| DOI | 10.1515/geocart-2017-0012 |
| Rok publikacji | 2017 |
| Typ publikacji | artykuł w czasopiśmie |
| Otwarty dostęp |  |
| Creative Commons | |
| Czasopismo/seria | Geodesy and Cartography (Warsaw) |
Abstract
We evaluated the performance of nine machine learning regression algorithms and their ensembles for sub-pixel estimation of impervious areas coverages from Landsat imagery. The accuracy of imperviousness mapping in individual time points was assessed based on RMSE, MAE and R-2. These measures were also used for the assessment of imperviousness change intensity estimations. The applicability for detection of relevant changes in impervious areas coverages at sub-pixel level was evaluated using overall accuracy, F-measure and ROC Area Under Curve. The results proved that Cubist algorithm may be advised for Landsat-based mapping of imperviousness for single dates. Stochastic gradient boosting of regression trees (GBM) may be also considered for this purpose. However, Random Forest algorithm is endorsed for both imperviousness change detection and mapping of its intensity. In all applications the heterogeneous model ensembles performed at least as well as the best individual models or better. They may be recommended for improving the quality of sub-pixel imperviousness and imperviousness change mapping. The study revealed also limitations of the investigated methodology for detection of subtle changes of imperviousness inside the pixel. None of the tested approaches was able to reliably classify changed and non-changed pixels if the relevant change threshold was set as one or three percent. Also for five percent change threshold most of algorithms did not ensure that the accuracy of change map is higher than the accuracy of random classifier. For the threshold of relevant change set as ten percent all approaches performed satisfactory.