Kawakubo, FS; Morato, RG; Luchiari, A (2013). Use of fraction imagery, segmentation and masking techniques to classify land-use and land-cover types in the Brazilian Amazon. INTERNATIONAL JOURNAL OF REMOTE SENSING, 34(15), 5452-5467.
This work presents a procedure for classifying land-use and land-cover (LULC) types in the Brazilian Amazon. Fraction imagery representing proportions of green vegetation, soil, and shade was estimated using all six reflective bands of the Landsat-5 Thematic Mapper (TM1 to TM5 and TM7) through the linear spectral mixing model (LSMM). The fraction information registered at pixel level was then related to different types of land classes following three principal procedures: (1) selecting an image or image group as input for segmentation; (2) application of sequences of masking techniques to address the segmentation of preselected areas in order to obtain better image partitioning; and (3) application of an unsupervised classifier by region, named Isoseg, to group the segmented regions. Isoseg is a clustering algorithm that calculates the centre of each class using the covariance matrix and the average vector of the regions. An assessment of the classification was performed visually and by error matrix, relating reference data points to classification results. The results showed that fraction images were effective in highlighting the different types of LULC. Several tests were conducted to evaluate the efficacy of the masking technique in the process for extracting information. The results showed that the use of masks significantly improves the segmentation results. However, in the Isoseg classification process, the masking technique was not able to avoid omission and commission errors between classes of similar structures. On comparing the results obtained in this work with a Maximum Likelihood classification, it was found that adopting the procedures described resulted in increases of 10% in average and global accuracy, and 18% in average reliability. Furthermore, a reduction was observed in the variability of errors created in the classification.