A Meta-Methodology for Improving Land Cover and Land Use Classification with SAR Imagery

Marinalva Dias Soares,Rogério Galante Negri,Gilson Alexandre Ostwald Pedro da Costa,Luciano Vieira Dutra,Pedro M. A. Diaz,Raul Queiroz Feitosa

doi:10.3390/rs12060961

Abstract

Per-point classification is a traditional method for remote sensing data classification, and for radar data in particular. Compared with optical data, the discriminative power of radar data is quite limited, for most applications. A way of trying to overcome these difficulties is to use Region-Based Classification (RBC), also referred to as Geographical Object-Based Image Analysis (GEOBIA). RBC methods first aggregate pixels into homogeneous objects, or regions, using a segmentation procedure. Moreover, segmentation is known to be an ill-conditioned problem because it admits multiple solutions, and a small change in the input image, or segmentation parameters, may lead to significant changes in the image partitioning. In this context, this paper proposes and evaluates novel approaches for SAR data classification, which rely on specialized segmentations, and on the combination of partial maps produced by classification ensembles. Such approaches comprise a meta-methodology, in the sense that they are independent from segmentation and classification algorithms, and optimization procedures. Results are shown that improve the classification accuracy from Kappa = 0.4 (baseline method) to a Kappa = 0.77 with the presented method. Another test site presented an improvement from Kappa = 0.36 to a maximum of 0.66 also with radar data.

Highlights

The new generation of orbital Synthetic Aperture Radar (SAR) platforms can deliver several types of all-weather products
Somewhat comparable to the pixel-based methods, the accuracy improvement observed for the MSCAH approach, shows that the classification benefitted from the Region-Based Classification (RBC) strategy, even without tuning segmentation parameters, as a result of the noise reduction brought by the segmentation process
We propose and investigate region-based classification approaches that can boost the performance of SAR classification, the cases where the only remote sensing data available is non-polarimetric SAR data

Summary

Introduction

The new generation of orbital Synthetic Aperture Radar (SAR) platforms can deliver several types of all-weather products. It can be used to observe and analyze several aspects of the land cover, which is often done in conjunction with information obtained by optical sensors. It is common, that SAR data is the only choice available for specific tasks, such as land cover classification in areas constantly covered by clouds. Real classification problems with radar data are generally limited to four or fewer classes [1,2]

Methods

Results

Conclusion