The integrated use of polarimetric Synthetic Aperture Radar (SAR) and optical image data for land cover mapping using an object-based approach

Image classification has long been used in earth observation and is driven by the need for accurate maps to develop conceptual and predictive models of Earth system processes. Synthetic aperture radar (SAR) imagery is used ever more frequently in land cover classification due to its complementary nature with optical data. There is therefore a growing need for reliable, accurate methods for using SAR and optical data together in land use and land cover classifications. However, combining data sets inevitably increases data dimensionality and these large, complex data sets are difficult to handle. It is therefore important to assess the benefits and limitations of using multi-temporal, dual-sensor data for applications such as land cover classification. This thesis undertakes this assessment through four main experiments based on combined RADARSAT-2 and SPOT-5 imagery of the southern part of Reunion Island. In Experiment 1, the use of feature selection for dimensionality reduction was considered. The rankings of important features for both single-sensor and dual-sensor data were assessed for four dates spanning a 6-month period, which coincided with both the wet and dry season. The mean textural features produced from the optical bands were consistently ranked highly across all dates. In the two later dates (29 May and 9 August 2014), the SAR features were more prevalent, showing that SAR and optical data have complementary natures. SAR data can be used to separate classes when optical imagery is insufficient. Experiment 2 compared the accuracy of six supervised and machine learning classification algorithms to determine which performed best with this complex data set. The Random Forest classification algorithm produced the highest accuracies and was therefore used in Experiments 3 and 4. Experiment 3 assessed the benefits of using combined SAR-optical imagery over single-sensor imagery for land cover classifications on four separate dates. The fused imagery produced consistently higher overall accuracies. The 29 May 2014 fused data produced the best accuracy of 69.8%. The fused classifications had more consistent results over the four dates than the single-sensor imagery, which suffered lower accuracies, especially for imagery acquired later in the season. In Experiment 4, the use of multi-temporal, dual-sensor data for classification was evaluated. Feature selection was used to reduce the data set from 638 potential training features to 50, which produced the best accuracy of 74.1% in comparison to 71.9% using all of the features. This result validated the use of multi-temporal data over single-date data for land cover classifications. It also validated the use of feature selection to successfully inform data reduction without compromising the accuracy of the final product. Multi-temporal and dual-sensor data shows potential for mapping land cover in a tropical, mountainous region that would otherwise be challenging to map using single-sensor data. However, accuracies generally remained lower than would allow for transferability and replication of the current methodology. Classification algorithm optimisation, supervised segmentation and improved training data should be considered to improve these results