Article Preview
Top1. Introduction
Soil surface is a starting point when investigating ecological, hydrological, thermodynamic and meteorological processes (Felix M Riese & Sina Keller, 2018). Soil temperature, as a critical parameter, has great influence on soil water, nutrients, enzymes, air, microbes and plants (Tian et al., 2018). In addition, soil temperature availability can serve as a relative indicator of a potential rate of fire spread, fire intensity, and fuel consumption (Girardin, Wotton, & Climatology, 2009). And thus precise data about spatial distributions and dynamics of soil temperature is valuable. Soil temperature estimation under almost real-world conditions is demanded. However, traditional situ point measurements are time-wasting and labor-consuming and cannot provide accurate estimation over large areas. Spectral techniques have been proven to be a key technology and provide non-invasive techniques in monitoring physical parameters (Borra, Thanki, & Dey, 2019) such as soil moisture (Chatterjee, Dey, & Sen, 2018), land cover (Baghbaderani, Wang, Stutts, Qu, & Qi, 2019), water stress (Jessica, Stephan, Wenxuan, James, & Andy, 2014) and crop diseases (Xavier et al., 2019) over larger areas. Hyperspectral sensors are mounted nowadays on unmanned aerial vehicles (UAVs) and satellites and can be also installed on hand-held devices, which enables hyperspectral data acquisition becomes available and affordable (Dey, Bhatt, & Ashour, 2018). In this work, hyperspectral data is used to estimate soil temperature.
The modeling of predicting soil temperature with hyperspectral data is a high-dimensional and non-linear problem. It is difficult to establish analytical equations with parameters in practice to demonstrate the relationship among complex and high-dimensional hyperspectral data. Artificial neural networks (ANNs) are notable for strong learning ability of nonlinear mapping, tolerance to errors and low computational cost, which have been applied in the forecasting tasks with various structures, e.g., feed forward neural network (FFNN), radial basis function neural network (RBFNN), and wavelet neural network (WNN). ANNs have so excellent generalization ability and robustness that they excel in many areas: pattern classification, function approximation, intelligent control, fault detection, signal processing and system analysis etc. The superior forecasting performances of ANNs own to the capability of extracting features from the input variables. In other words, the excellent feature representation can ensure forecasting accuracy (Xie, Wang, Liu, & Bai, 2018). However, it is usually expensive and time-consuming to manually extract domain-specified features for the traditional shallow ANN models. In addition, when ANNs are used as predictors, improper initial weights may affect the learning convergence speed and make learning trap at local optima, which are problems of premature convergence (Zhang & Hong, 2019). Therefore, it is imperative to develop a new method for feature learning in the ANN for the soil temperature forecasting. Deep learning has obvious advantages when dealing with a large number of samples and nonlinear data. As one of deep learning architectures, SAE obtains a reduced representation of the inputs, which is much smaller but still preserving the integrity of the original data (Lan et al., 2018). Moreover, SAE is more efficient because its objective function can be solved by fast backward propagation.