Using Growing-Season Time Series Coherence for Improved Peatland Mapping: Comparing the Contributions of Sentinel-1 and RADARSAT-2 Coherence in Full and Partial Time Series

Koreen Millard,Amir Behnamian,Patrick Kirby,Sacha Nandlall,Sarah Banks,Fabrizio Pacini

doi:10.3390/rs12152465

Abstract

Differences in topographic structure, vegetation structure, and surface wetness exist between peatland classes, making active remote sensing techniques such as SAR and LiDAR promising for peatland mapping. As the timing of green-up, senescence, and hydrologic conditions vary differently in peatland classes, and in comparison with upland classes, full growing-season time series SAR imagery was expected to produce higher accuracy classification results than using only a few select SAR images. Both interferometric coherence, amplitude and difference in amplitude time series datasets were assessed, as it was hypothesized that these may be able to capture subtle changes in phenology and hydrology, which in turn differentiate classes throughout a growing season. Groups of variables were compared for their effectiveness in Random Forest classification for both Sentinel-1 and Radarsat-2. The Shapley value was used to determine the contribution of each group of variables in thirty scenarios, and Mean Decrease in Accuracy was compared to evaluate its ability to rank variables by relative importance. Despite being dual-pol, the results of classifications using Sentinel-1 coherence (12-day repeat) were significantly better than using fully polarimetric RADARSAT-2 coherence (24-day repeat), likely owing to the difference in baseline and specific acquisition dates of the data in this study. Overall, full growing season Sentinel-1 coherence time series produced higher accuracy results than fully polarimetric quad pol RADARSAT-2 coherence amplitude, difference in amplitude and polarimetric decomposition time series. Using a full growing season of time-series imagery in classification resulted in higher accuracy than using a few dates over a growing season. Using mean decrease in accuracy to rank and reduce variables resulted in a weaker classification than if the entire time series is used.

Highlights

Like all wetlands, peatlands have great economic, societal, and environmental value, including forming a habitat for various unique species and species at risk [1], playing a role in the hydrologic cycle [2] and in sequestering carbon [3]
The main objectives of this study are (1a) to compare a Radarsat-2 (RS2) time series to Sentinel-1 (S1) time series to determine if the shorter temporal baseline of Sentinel-1 leads to improved classification results and (1b) to assess the impact of datasets that were acquired in different seasons to determine if imagery from one particular season contributed more to the classification than another, (2a) determine if coherence, amplitude, or difference in amplitude between image pairs produced better classification results for both S1 and RS2; and (2b) for coherence, amplitude, and difference of amplitude groups of variables, assess different polarization combinations and different RS2 Fine Quad incident angles
The results indicate that growing season short-baseline time-series coherence can produce high accuracy classifications, including using coherence data alone for classification, and contributes the most to all classifications where these variables were included

Summary

Introduction

Peatlands have great economic, societal, and environmental value, including forming a habitat for various unique species and species at risk [1], playing a role in the hydrologic cycle [2] and in sequestering carbon [3]. There are often many similar physical characteristics between different peatland classes (e.g., open bog and poor fen are both dominated by Sphagnum mosses), and there are often many highly variable features within a single class of peatland (e.g., rich fens vary widely in their vegetation composition, owing to local water chemistry). Differences between and within the classes of peatland relate to differences in ecosystem services [5], mapping the extent of different peatland classes is widely studied [6,7] Owing to their waterlogged surfaces and often dense vegetation, field data collection in peatlands can be physically demanding, time-consuming, and expensive. While LiDAR is able to capture these important characteristics at high spatial resolutions, it is expensive to acquire, and not widely collected and not often collected repeatedly over time

Objectives

Results

Discussion

Conclusion