MAPPING EUCALYPTUS PLANTATIONS AND NATURAL FOREST AREAS IN LANDSAT-TM IMAGES USING DEEP LEARNING

Vol 19, 2019 - 96677
Oral
Favoritar este trabalho
Como citar esse trabalho?
Resumo

Automatic mapping of planted and natural forests using satellite images is a challenging task due to spectral similarity issues. In this work, we assessed the use of Convolutional Neural Networks (CNNs) to discriminate between natural forest areas and eucalyptus plantations in a Landsat-TM scene. First, we produced training and testing datasets with data from the MapBiomas project. Then, CNNs were trained with input patches of different sizes (5 x 5, 7 x 7, 9 x 9 and 11 x 11 pixels) to evaluate the influence of patch dimension in the classification accuracy. For comparison, pixel-wise and patch-classification were performed using the Random Forest (RF) algorithm. The best results were obtained using CNNs with 5 x 5 patches. In this scenario, the F-score was of 97.64% for natural forests and 95.49% for eucalyptus plantations. The classification errors reached 9.06% using RF and did not exceed 3% with CNNs.

Instituições
  • 1 Instituto Militar de Engenharia
  • 2 Pontifícia Universidade Católica do Rio de Janeiro (PUC-Rio)
Eixo Temático
  • Floresta e outros tipos de vegetação
Palavras-chave
Convolutional Neural Networks
Patch-classification
Random Forest
Satellite images
Tropical forests