Comparison of Optical and SAR Data in Tropical Land Cover Classification for REDD
Sirro, Laura1; Häme, Tuomas1; Rauste, Yrjö1; Antropov, Oleg1; Hämäläinen, Jarno2
1VTT Technical Research Centre of Finland, FINLAND; 2Arbonaut Oy, FINLAND
The Measurement, Reporting and Verification (MRV) activity of the REDD (Reducing Emissions from Deforestation and Forest Degradation) process requires reliable and robust methods for the assessment of land cover classes and their changes. Research projects have reported variable accuracies that are based on different approaches in image interpretation and accuracy assessment procedures. Such variability makes it difficult to evaluate the robustness and performance of the approaches. Optical earth observation data are frequently contaminated by clouds which makes SAR imagery an attractive data source. The REDD services would benefit from using several alternative image types by augmenting cloudy data with SAR, for instance.
A comparison study on a study site of approximately 100 km by 100 km in Mexico in state Chiapas was conducted to evaluate the performance of Landsat, RapidEye, Envisat ASAR and ALOS PALSAR data in the classification of IPCC compatible land cover classes.
The data for the accuracy assessment were from a statistical sample of 50 m by 50 m plots that had been located on Kompsat-2 VHR data and evaluated visually by a party that did not conduct the image analysis. The test data were not distributed to the parties that conducted image analysis. Kompsat-2 data were also used for model training but the training data were not included in the accuracy assessment data set.
The baseline image analysis method was the in-house Probability fuzzy classification method. It is an unsupervised approach with some features of supervised methods. Additionally, supervised non-parametric methods such as decision trees and random forest were tested. The training reference data were divided into two sets. Testing the initial model with training data that were not used for its computation helped to improve the model iteratively. The other part of the training data was introduced just in the second stage of modeling. The iterative approach also helped to reduce bias in the final classification.
The results of the testing were not available at the time of writing this abstract but will be computed by the end of February 2013 and those figures will naturally be reported in the actual paper. With the training data the overall accuracies in forest and non-forest classification varied between 80 % of SAR data and over 90 % of RapidEye data. The largest differences between user’s and producer’s accuracies were around 10 percentage units with SAR data that tend overestimate forest area. The separation of all the six IPCC compatible classes did not succeed well with SAR data. With RapidEye data the accuracy was approximately 80 %. Cropland and grassland were largely mixed.
The study was conducted as part of the ReCover project of the Framework Program 7 of the European Union.