Identification of Dynamic Cover Types in Wetlands by using Multitemporal Cross-polarized Sentinel-1 Images

Abstract

Monitoring of long-term land-use and land-cover change patterns may be biased by seasonal changes of different surface properties (e.g. hydrology, phenology, etc.) which become even more prominent in highly dynamic ecosystems such as wetlands (Crews-Meyer, 2008; McClearly, Crews-Meyer and Young 2008; Dronova et al. 2011). These surface dynamics produce transitional states and fine-scale mixtures of classes that may hinder classifications and long-term change detection. Dronova et al. (2015) proposed the term “Dynamic Cover Types" (DCT) to refer to such areas of regimes of periodic or seasonal change. Examples of DCT in the context of wetlands would be seasonally inundated forests, temporal water bodies and waterways, or harvests of reeds and crops such as rice.We assess the spatio-temporal extent of DCT in two study sites; The Camargue, a large coastal wetland in Southern France, and the Lagoon of Fuente de Piedra, a small wetland in Southern Spain. For that we use a multitemporal change detection procedure for polarimetric SAR imagery based on the Complex Wishart distribution developed recently by Conradsen et al (2015), (to be published) and an innovative open source software implementation which makes use of Ipython Notebooks and Docker containers (http://mortcanty.github.io/SARDocker/). The procedure carries out a series of change detection processing routines for the whole time series with a desired significance level. It uses multilook, geocoded and terrain corrected intensity images in C2 matrix. These were generated in the Sentinel Application Platform (SNAP) using 12 Sentinel-1 images (Interferometric Wide, Single Look Complex and cross-polarized) with a monthly resolution.The methodology proposed here for change detection is relatively easy to use and utilizes only open source and free data. It enables an operational monitoring service of short-term change detection. No calibration or validation needed, only interpretation of changes using local knowledge. This has important implications for operational standardized monitoring service such as the ones developed in the –Satellite-based Wetland Observation Service (SWOS) Horizon 2020 project.Besides its easiness to use, this methodology has other important advantages: First, the fine spatial and temporal resolutions of Sentinel-1 SAR data allow us to detect short-time changes for a complete water year regardless of the cloud cover. Second, change detection methods based on classification are affected by classification errors, whose probability of occurrence increases in dynamic and transitional landscapes (Powell et al. 2003). Our approach does not rely on classification and thus is free from such errors. Third, DCT are complex landscapes that often give rise to unique species assemblages (Parrot & Meyer 2012; Watson et. al 2014), and knowing their spatio-temporal extent will assist in biodiversity management. Fourth, annual stable features can be identified and used for training areas, which may facilitate the classification process and improve accuracies. And fifth, estimating the spatio-temporal extent of DCT might shed some light on the wide array of options in classification methodologies available and their different results (Object vs. Pixel based, Support Vector Machines, Random Forest Classifiers, and other algorithms).