- Tytuł:
-
Od Bertina i Hotellinga do Zadeha i Kohonena, czyli o zastosowaniu sztucznych sieci neuronowych w kartografii tematycznej
From Bertin and Hotelling to Zadeh and Kohonen, or about applications of neutral networks in thematic cartography - Autorzy:
-
Kępińska, M.
Olszewski, R. - Powiązania:
- https://bibliotekanauki.pl/articles/204234.pdf
- Data publikacji:
- 2002
- Wydawca:
- Polskie Towarzystwo Geograficzne
- Tematy:
-
kartografia
sieć neuronowa Kohonena
kartografia tematyczna - Opis:
-
W artykule omówiono wybrane współczesne metody klasyfikowania danych oraz pokazano możliwość ich wykorzystania w kartografii. Szczególną uwagę zwrócono na możliwość zastosowania sieci neuronowych Kohonena jako narzędzia nienadzorowanej klasyfikacji danych przestrzennych.
The article discusses selected contemporary methods of multi-feature data and shows their possible applications in cartography. Graphic information processing described by J. Bertin and principal components analysis created by H. Hotelling, which enables the transfer of results from n-dimensional space to three-, two-, and even one-dimensional space, are examples of non-standard classification in cartography. An examples of spatial data classification using L.A. Zadeh's theory of fuzzy sets is presented. In this classification particuler objects belong to different classes, with various levels of subordination. The article draws special attention to possibility of using neural networks (NN) as a tool for unsupervised classification of spatial data. NN using systems are widely applied in branches of knowledge, which research prediction and classification. From the point of view of source data classification, it is interesting to use NN prepared by unsupervised learning. A so called Kohonen's network is an example of such structure. During the learning process this network does not receive feedback on the correctness of particular answers. Not knowing the expected output information, the network selflearns to recognize data structure. The outer surface of the network creates a, so called, Kohonen topological map, which projects the relations of similarity between the features of analyzed objects into one- or two-dimensional space. The article presents two examples of practical applications of Kohonen's network in classification of multi-feature spatial data. Presented multi-feature data classification methods, despite high differentiation of algorithms, show similar approach to the discussed problem. Self-learning of Kohonen's network, like permutation method, consists in revealing the structure of source data. Application of neural networks, similarly to the method of principal components, allows to reduce the dimension of the space of attributes. In neural networks, as in the classification method basing on theory of fuzzy sets, the final interpretation should be preceded by an estimation of the level of activation of particular neurons. Application of one-dimensional out surface of Kohonen's net-work makes it possible to directly present the classification results on a thematic map, which is optimal from a cartographic point of view. - Źródło:
-
Polski Przegląd Kartograficzny; 2002, T. 34, nr 2, 2; 103-114
0324-8321 - Pojawia się w:
- Polski Przegląd Kartograficzny
- Dostawca treści:
- Biblioteka Nauki
Artykuł