Development of a deep learning method for soil moisture estimation at high spatial and temporal resolution using satellite data

Английский. Soil moisture (SM) is an essential climate variable that controls fundamental hydrological and climatic processes. Soil moisture products derived from microwave remote sensing often provide measurements at low spatial resolution and incomplete temporal records. This thesis presents a novel method for estimation soil moisture at both high spatial and temporal resolution by using a deep learning recurrent neural network model. The model relies primarily on Sentinel-1 synthetic aperture radar data but includes additional ancillary data, such as Sentinel-2 vegetation indices, land cover, and weather variables. The model is calibrated and validated on four SM probe networks within continental Europe with data from the International Soil Moisture Network (ISMN) and the Integrated Carbon Observation System (ICOS). The model has been compared with existing SM products and has shown comparable or better results with a mean absolute error of 9.33% and a correlation of r=0.49 with observed measurements. It performs best over agricultural land covers in temperate regions, where satellite observations are most frequent, and poorer over vegetated land surfaces like forests due to the attenuation of microwave signals. The temporal predictions show high accuracy and precision, while the spatial predictions retain a high accuracy but with lower precision. The predictions show satisfactory results overall but warrant further research to test the feasibility of this architecture over larger areas and different climate types.

Английский. This thesis develops a novel method for predicting soil moisture from remote sensing data using deep learning. Remote sensing of soil moisture creates better opportunity for applying the methods at a large spatial scale and deep learning creates the predictions by learning complex relationships from the data. Soil moisture is a hydrological and climatic variable used to express the degree of saturation of the soil. The saturation of the soil affects various hydrological, climatic, and ecosystem functions and hence is an important variable in modelling, forecasting, and analysis within these fields. Soil moisture is traditionally measured using in situ probes, but this method only provides point-based measurements with poor spatial representation. Remote sensing is an increasingly popular alternative that can provide large-scale spatial estimations of soil moisture continuously. However, most remotely sensed solutions do not offer estimations below 1 km spatial resolutions, which is required for applications that focus on land covers with high heterogeneity, such as precision agriculture. Additionally, most of these solutions also do not provide complete time-series, which can be important for many applications. This thesis presents a new method of estimating soil moisture from remotely sensed products, such as Sentinel-1 microwave radar data and Sentinel-2 derived vegetation indices, in addition to climatic and topographic variables. This method relies on a recurrent neural network deep learning algorithm, that specializes in sequence data, where antecedent conditions are predictive of current and future conditions. The model is built as a semi-automatic pipeline that processes the various data types, trains the model, and creates an estimation. The entire pipeline and model are written in Python and uses the Tensorflow and Keras high-level libraries for the assembly of the model. Several variations of the dataset are tested to determine which variables are more beneficial to the model and if reducing the number of parameters in the model can increase the accuracy. The results from the model show that static variables, which pertain variables that do not change over time, such as elevation and soil texture, are less useful to the model and result in lower accuracy when included. This is likely due to low number of stations used in the study, which are not numerous enough to represent a meaningful statistical distribution. The dynamic variables, which do change over time, such as soil moisture and temperature, are the most important in the model, as they contain more information due to their temporal nature. Additionally, the results show that some land covers have worse accuracy than others, hereunder forests due to the dense vegetation and peat bogs because of the high saturation of the soil. The results from the model are compared with existing soil moisture products, where it proves to be a competitive solution in accuracy, while also retaining a higher spatial and temporal resolution. This thesis concludes that deep learning is a promising but under explored method for estimating soil moisture from remote sensing.