Logo do repositório

Percorrer por autor "Amedjar, Mariam"

A mostrar 1 - 1 de 1
  • Random forest effectiveness for Bragança region mapping: comparing indices, number of the decision trees, and generalization
    Publication . Amedjar, Mariam; Castro, João Paulo; Castro, Marina; Naimi, Mustapha; Sebari, Imane
    Remote sensing is a domain that tends to use satellite images for classification and Land Use/Cover (LULC) mapping. For this purpose, classification algorithms are used, which are numerous and diverse, and it is necessary to establish decision criteria when choosing the algorithm. Ultimately, the main decision criterion will be the accuracy obtained in classification because the accuracy of classification may differ from one algorithm to another, even within the same algorithm, according to its variables. But there are other equally important criteria: it depends on the nature of the task, the quantity and types of data available, the type of response expected, the time and computational resources available, the depth of our knowledge about the algorithms. The methodology of each part of the work was described and the criteria for comparison were established. In this research, with the same training data, the same validation data, the same application context (7 classes), and the same image data (Sentinel-2), we tested 15 iterations with the Random Forest classification algorithm, with different tree number decision values, and 3 iterations with vegetation and soil indexes, for the production of the LULC map of the Bragança region (northeast Portugal). Finally, we evaluate the accuracy of the classification, before and after the post-classification tasks (generalization, fragmentation and removal of isolated pixels). The results obtained show that a classification with an nb-trees = 1000, including vegetation and soil indices, and after post-classification tasks, provided excellent precision results (Coefficient Kappa = 0.93, Overall accuracy = 96%, and marginal errors of omission & commission below 4%).
    2020Dissertação de mestrado Acesso aberto Ver mais