P-Band radar data classification by neural network for Amazonin land cover assessment / Uso de rede neural artificial não supervisionada na classificação de dados de radar na Banda-P para mapeamento de cobertura da terra em floresta tropical

AUTOR(ES)
DATA DE PUBLICAÇÃO

2003

RESUMO

The applicability of P-band radar data for land cover mapping using the unsupervised artificial neural network Fuzzy-ART (Adaptive Resonance Theory) is evaluated. The study area is located near Tapajós National Forest in the State of Para, Brazil. The radar data was acquired during an airborne mission conducted by AeroSensing RadarSystem GmbH in september 2000. A 2.4 km x 7.4 km image strip was selected for the study. The input parameters for the neural network Fuzzy-ART were optimized by genetic algorithm. It was investigated the speckle reduction efficiencies of Map Gamma filter (5x5 pixels) and the combination of Frost and Median filters (3x3 pixels). The following images were analyzed individually and combined in pairs: backscatter in the polarizations HH, HV, VV, Average cross section (ACS), and the biophysical indices Biomass Index (BMI), Canopy Structure Index (CSI) and Volume Scattering Index (VSI). The combination HH/HV/VV was also evaluated. The clusters discriminated by the neural network were related with the land cover classes of sites previously observed in field work. The eight reference classes are: Bare Soil (BS), Pasture and Agriculture (PA), Upland Forest Regrowth - Pioneer Stages (R1), Upland Forest Regrowth Early Intermediate Stages (R2), Upland Forest Regrowth Late Intermediate Stages (R3), Upland Forest Regrowth Advanced Stages (R4), Primary Upland Forest (PF) and Primary Floodplain Forest (FF). Part of the reference data set was used for cross tabulation to map unsupervised clusters set onto the land cover class set and the other part for estimating the global accuracy and the Kappa coefficient. The discrimination of the eight land cover classes was not satisfactory. Best global accuracy (56%) was obtained with PT. Based on the degree of confusion among reference classes, the combinations of classes and corresponding clusters were reduced to five and four classes. The best results of global accuracy were obtained in the discrimination of four classes (BS/PA; R1/R2; R3/R4/FP and FF). The following global accuracies were obtained for the individually classified images: 84%, 73%, 78%, 83%, 74%, 79%, and 76% for HH, HV, VV, PT, BMI, CSI and VSI, respectively. It was obtained the following global accuracies for the classifications of combined images: 84,9%, 84,5% 83,7%, 81,2%, 79,6%, 76,5%, 74,4%, and 72,8% for CSI/HV, HH/HV/VV, HH/HV, HH/VV, CSI/VV, VSI/HV, BMI/HV and VV/HV, respectively. As a general result of the analyses, the best result (global accuracy of 84,9%) was obtained with the combination of CSI and HV pre-filtered with the Map Gamma filter for the discrimination of the classes BS/PA; R1/R2; R3/R4/FP and FF. It was concluded that the utilization of co-polarized and cross-polarized images contributes for the improvement of the classification result, and that the applicability of P-band radar data for land cover assessment in tropical forest landscape is only reliable for broadly defined land cover classes.

ASSUNTO(S)

sensoriamento remoto genetic algorithms processamento de imagem reconhecimento de padrão imaging radar image classification cobertura da terra image processing p band redes neurais remote sensing classificação de imagens pattern recognition imagem de radar land cover neural networks algoritmos genéticos

ACESSO AO ARTIGO

Documentos Relacionados