Mariane Souza Reis

ALOS/PALSAR Data Evaluation for Land Use and Land Cover Mapping in the Amazon Region

In tropical biomes such as the Amazon, the cloud cover is frequent. The use of synthetic aperture radar (SAR) sensor systems is important to monitor and study these biomes because they can acquire data under cloud coverage. In this paper, an advanced land observing system phased array L-band synthetic aperture radar fine-beam dual (ALOS/PALSAR-FBD) image was evaluated for land use and land cover (LULC) classification of an Amazon test site. The features extracted from this image were also evaluated. To perform this task, a method for feature selection, considering the desired classes, was proposed. In order to better understand the applicability of this type of data in Brazilian Government projects (such as DETER-B and TerraClass), the results obtained with SAR images were compared with those from LANDSAT5/TM. The results show that the PALSAR-FBD image and the features selected are not suitable for the discrimination of densely forested classes. They presented, however, a good discrimination among the group of forested and agropastoral classes, as well as among nondensely forested classes (i.e., pastures, bare soil, and new regeneration). Therefore, these data present good applicability for mapping and monitoring of both deforestation and LULC and they can be used in the above mentioned projects. The classification of selected features, with eight classes of interest, achieved an increase of about 133% and 69% in the Kappa index (0.32) and an Overall Accuracy (0.54) regarding the PALSAR-FBD classification (0.136 and 0.322, respectively). This result shows the applicability of the proposed method. It is also expected that the features selected in this paper will improve the classification of similar study sites.

Proposal of a weighted index for segmentation evaluation

Although segmentation is an important process in image classification, selecting among different segmentors and their parameters is a difficult task. This work proposes a reference free index that returns the quality of segmentation, considering the classes that the user intends to classify. Considering a gaussian distribution, this index was tested to evaluate segmentations of an optical simulated image, a LANDSAT5/TM image in a Brazilian Amazon area and its derived fraction image. Index results presented higher values for segmentations more similar to the reference image, and also good agreement with overall accuracy values when classifying the images.

Image segmentation algorithms comparison

This work aims to compare two segmentation tools that are based on the same multiresolution segmentation algorithm. The main significance of this investigation is to assess the feasibility of the use of a free tool (InterIMAGE) instead of a commercial one (eCognition) but still achieving equivalent results. Samples extracted from an optical image (LANDSAT5/TM) were used to fill the segments of a phantom image, creating 100 repetition of 3 simulated image. The obtained simulated images were then segmented using both tools with several parameter configurations and the results were evaluated through a supervised segmentation quality measure. Results obtained by eCognition are better in all tested cases.

Evaluation of SAR-SDNLM filter for change detection classification

This study evaluates the usage of Stochastic Distances Nonlocal Means (SDNLM) speckle filter in an image from Brazilian Amazon. The objective is to evaluate whether the noise reduction improves land cover and change classification. Results obtained from filtered images were compared with those obtained from unfiltered images and images filtered using Gamma Map. Results shows that, when using region based Bhathacharyya Minimum Distance Classifier, land cover and change classification using both speckle filters has accuracy values statistically equal. Analyzing the filtered images themselves, SDNLM obtained better results in terms of visual quality and edges preservation.

The use of land cover change likelihood for improving land cover classification

The likelihood of transitions between pairs of land cover and land use classes in a given time interval and environmental context can be used to impose classification restrictions on an image or to evaluate results. This study presents a methodology for using the likelihood of transitions between classes to improve land cover classification, given a base map (a supposedly accurate map for the same area in another date) and a set of previously classified images. These improved land cover classified images were named conditioned classified images. We aimed to classify one Synthetic Aperture Radar image and an optical one, both from June 2010, using two land cover legends in different level of detail for a region in the Brazilian Amazon. We used both a classified image from 2008 (also in two legends levels) and the data from the Programme for the Estimation of Deforestation in Brazilian Amazon (PRODES) from 2008 as base maps, and presented the likelihood of transitions between the considered classes. The proposed methodology resulted in conditioned classified images with higher Overall Accuracy than the one that does not consider the base maps and the likelihood of transitions. The conditioned classified images presented unlabeled areas due to classification errors in the input data. It is important to highlight that these areas are probably misclassified in maps obtained without using likelihood transition and base maps, since they are impossible to occur in the field.

An application of multiple space nearest neighbor classifier in land cover classification

This work presents a case study in land cover classification using ms-NN, an extension of k-NN classification algorithm. The case study focuses on an area in the Brazilian Amazon region, with data obtained from LANDSAT5 satellite Thematic Mapper (TM) sensor and Advanced Land Observing System satellite (ALOS) Phase Array L-Band Synthetic Aperture Radar (PALSAR), using Fine Beam Dual. The results obtained with ms-NN are compared with k-NN and Support Vector Machine algorithms, considering the use of a single training set, a Monte Carlo procedure for testing and an extensive number of parameterizations for the classification methods. Considering only the best results for each classifier, ms-NN obtained better results than the other methods.