Estimation of Soil Moisture Index Using Multi-Temporal Sentinel-1 Images over Poyang Lake Ungauged Zone

Yufang Zhang, Jianya Gong, Kun Sun, Jianmin Yin, Xiaoling Chen
2017 Remote Sensing  
The C-band radar instruments onboard the two-satellite GMES Sentinel-1 constellation provide global measurements with short revisit time (about six days) and medium spatial resolution (5 × 20 m), which are appropriate for watershed scale hydrological applications. This paper aims to explore the potential of Sentinel-1 for estimating surface soil moisture using a multi-temporal approach. To this end, a linear mixed effects (LME) model was developed over Poyang Lake ungauged zone, using time
more » ... s Sentinel 1A and 1B images and soil moisture ground measurements from 15 automatic observation sites. The model assumed a linear relationship that varied with both time and space between soil moisture and backscattering coefficient (SM-σ 0 ). Results showed that three LME models developed with different polarized σ 0 images all meet the European Space Agency (ESA) accuracy requirement for GMES soil moisture product (≤5% in volume), with the vertical transmit and vertical receive (VV) polarized model achieving the best performance. However, the SM-σ 0 relationship was found to depend strongly on space, making it difficult to predict absolute soil moisture for each grid. Therefore, a relative soil moisture index was then proposed to correct for site effect. When compared with those of the linear fixed effects model, the soil moisture indices predicted by the LME model captured the temporal dynamics of measured soil moisture better, with the overall R 2 and cross-validated R 2 being 0.68 and 0.64, respectively. These results indicate that the LME model can be effectively applied to estimate soil moisture from multi-temporal Sentinel-1 images, which is useful for monitoring flood and drought disasters, and for improving stream flow prediction over ungauged zones. stream flow prediction [7] , and water resource management [8]. Located in the south bank of the middle reaches of the Yangtze River, Poyang Lake is the largest freshwater lake of china. The ungauged zone of Poyang Lake, which lacks hydrological observations, generally refers to the large alluvial plain stretching from seven stream flow gauging stations to the lake boundary. This region covers a majority of the Poyang Lake eco-economic zone, which is occupied by intensive rice agriculture, supporting the high-density local population. Moreover, the significant seasonal fluctuations of Poyang Lake water levels and the associated inundation areas create extensive wetland ecosystems in this region [9] . However, both droughts and floods have occurred frequently around the lake in recent decades, which are primarily caused by climate change and human activities [10] . For example, the severe continuous drought during 2006-2007 brought the lake water storage down to less than 1% of the lake capacity [11] , and the extrema flood event in 1998 resulted in direct economic losses of about ¥38 billion over the Poyang Lake watershed [12] , with most of the losses occurring within the ungauged zone. Accordingly, soil moisture estimation over Poyang Lake ungauged zone is of vital importance for flood forecasting and drought monitoring, irrigation scheduling, agriculture management, and wetland protection. Besides, knowledge of soil moisture is beneficial to the Predictions in Ungauged Basins (PUB) researches [13] . Soil moisture, especially surface soil moisture, is highly variable in space and time [14] . In situ measurements can provide accurate estimation of soil moisture, but they are both time consuming and expensive, and only represent a small area (few square decimeters). Nevertheless, a number of strategies can be adopted to upscale the spatially sparse ground-based observations [15, 16] , which are invaluable for calibrating and validating land surface models and satellite-based soil moisture retrievals [17] . Microwave remote sensing techniques have been used to obtain surface soil moisture, commonly referred to as the water content of the uppermost soil layer, at various temporal and spatial scales since the 1970s [18] . The strengths lie in that both microwave scattering and emission are directly related to the water content of soil surface and microwave sensors can acquire imagery in all weather, during day and night. In recent years, various soil moisture products have been derived globally from microwave radiometer and scatterometer systems with high temporal resolution [19] [20] [21] [22] , but they are generally too coarse for watershed or field scale applications. Using downscaling techniques, the spatial resolution of those soil moisture products can be largely improved, while uncertainties can be introduced during this process [23, 24] . Conversely, Synthetic Aperture Radar (SAR) systems can achieve much higher spatial resolutions (<1 km) and allow for soil moisture mapping at fine scales directly. However, apart from soil moisture, SAR measurements are also sensitive to several other surface parameters such as surface roughness, vegetation cover, and topography, making it difficult to separate the soil moisture contribution to the backscatter signal from other factors. Therefore, based on different SAR configurations and available surface parameters, various empirical, semi-empirical, and theoretical models have been proposed to retrieve soil moisture at regional scales over the last few decades [25] [26] [27] . Developed by ESA to ensure continuity of C-band SAR data, the Sentinel-1 constellation features a temporal resolution of six days and spatial resolution of 5 × 20 m over land [28] , which are appropriate for soil moisture estimation over Poyang Lake ungauged zone. Currently, several algorithms have been proposed to retrieve soil moisture operationally from Sentinel-1 data, using change detection techniques [29, 30] , Bayesian approach [31], and Artificial Neural Network (ANN) [32], respectively. Both change detection and Bayesian methods exploit the short revisit time of Sentinel-1 data and assume that the backscatter cross section of soil surfaces changes over short timescales, mainly due to variations in soil moisture, while the average characteristics of surface roughness remain almost unaltered [31] . This assumption might be invalidated by the pronounced heterogeneity of soil surfaces when estimating soil moisture from Sentinel-1 data at full spatial resolution. The ANN algorithm trained with theoretical model can retrieve soil moisture accurately from single SAR acquisition, but a robust and extensive reference dataset for the training of the ANN technique is essential, which is lacking for most areas [32] . mode by default, which provides a wide coverage of 250 km with a spatial resolution of 5 × 20 m in range and azimuth directions, respectively [43] .
doi:10.3390/rs10010012 fatcat:kydlz4226bcgphcqc3jddqunvi