- Title
- Using an artificial neural network to enhance the spatial resolution of satellite soil moisture products based on soil thermal inertia
- Creator
- Senanayake, I. P.; Yeo, I-Y; Willgoose, G. R.; Hancock, G. R.; Bretreger, D.
- Relation
- 23rd International Congress on Modelling and Simulation (MODSIM2019). Proceedings of 23rd International Congress on Modelling and Simulation (MODSIM2019) (Canberra, ACT 01-06 December, 2019) p. 1049-1055
- Publisher Link
- http://dx.doi.org/10.36334/modsim.2019.k10.senanayake
- Publisher
- Modelling and Simulation Society of Australia and New Zealand
- Resource Type
- conference paper
- Date
- 2019
- Description
- High resolution soil moisture information is vital for a number of environmental applications including hydrologic and climatic modelling. However, the available point-scale field measurements and coarse spatial resolution satellite soil moisture products (~10s of km) are unable to provide the spatial resolution requirements for many of these applications, especially at regional scales. Downscaling the L-band satellite soil moisture products appears to be a viable solution for this problem. Many research teams have tested methods and algorithms to downscale the satellite soil moisture retrievals, yet there is no universally applicable model. Among those methods, thermal data based downscaling models have shown promising results over arid and semi-arid regions. The downscaling approach, based on the soil thermal inertia relationship between the diurnal temperature difference (ΔT) and the daily mean soil moisture (μSM), is one of the thermal data-based downscaling methods tested in the United States and Australia. These studies have used this method by building regressions between ΔT and μSM modulated by the vegetation density. However, there are a number of possible factors affecting the linearity of this regression model. Therefore, this study employed a machine learning model to build a more complex algorithm between ΔT, μSM and vegetation density. The Global Land Data Assimilation System (GLDAS) derived 25 km resolution ΔT values (from 2000 to 2017) and aggregated Moderate Resolution Imaging Spectroradiometer (MODIS) derived Normalized Difference Vegetation Index (NDVI) values were used as inputs, whereas GLDAS derived μSM values were used as targets to train an artificial neural network (ANN). The Levenberg-Marquardt algorithm with 50 hidden neurons was used as the network architecture in building this model. Thereafter, 1 km resolution MODIS derived ΔT and NDVI values of November 2005 were input into the model to estimate soil moisture at high spatial resolution (1 km). The estimated soil moisture values were then used to downscale aggregated NAFE’05 airborne soil moisture retrievals. The downscaled soil moisture products were compared with the 1 km resolution soil moisture retrievals from the National Airborne Field Experiment 2005 (NAFE’05). This study has been conducted over two medium-scale catchments, Krui and Merriwa River, located in the Upper Hunter region of the south-eastern Australia. The comparison between downscaled and airborne soil moisture showed root mean square errors (RMSE) of 0.088, 0.072 and 0.058 cm 3/cm 3 on 7 th, 14 th and 21 st November 2005, respectively. The downscaled soil moisture products were able to capture the detailed spatial patterns of soil moisture over the study area, showing a good match with the airborne retrievals. However, the algorithm performed better under dry catchment conditions compared to wet catchment conditions.
- Subject
- artificial neural network; downscaling; soil moisture; Levenberg-Marquardt algorithm; machine learning
- Identifier
- http://hdl.handle.net/1959.13/1460266
- Identifier
- uon:45917
- Identifier
- ISBN:9780975840092
- Language
- eng
- Reviewed
- Hits: 1082
- Visitors: 1082
- Downloads: 0