Soil surface roughness measuring method based on neural network and decision tree

Abstract: Soil surface roughness is one of the important indices commonly used to describe soil hydrological characteristics and Lambert characteristic. In microwave quantitative remote sensing application, it affects the microwave scattering values and therefore impacts the accuracy of soil moisture retrieved using microwave sensing data. Therefore, measuring soil surface roughness has become one of the research hotspots in the field of microwave remote sensing. Two kinds of techniques are used to calculate soil surface roughness, including contact method, such as the pin meter and profile meter, and non-contact method, such as ultrasonic measurement, laser scanning, three-dimensional photography, infrared measurement and radar measurement method. All these methods need some special device. The development of image processing technology and the popularization of digital camera provide a simple measuring method which only needs a reference whiteboard and a camera. However, the detailed scale information commonly used on the reference whiteboard increases the requirements for data acquisition and data processing. The purpose of this study is to provide a method to obtain the soil surface image with a simplified reference whiteboard and then to measure soil surface roughness in the presence of field environmental noise. Therefore, a simple image acquisition method is introduced and then an image processing method combining the neural network and the decision tree is proposed. The neural network is built to detect image edge points. To reduce the environmental noise effect, the input characteristic parameters of the neural network are selected carefully, which include not only gradient information, but also image direction and neighborhood consistency information. The cutting of the background section on the original image based on image edge detection result improves the computing speed effectively. A decision tree model is introduced to divide image segments into 4 classes including soil, whiteboard, reference square and vegetation, which are not easy to classify correctly using other classification methods. Considering the effects of weeds and light which are inevitable in field environment, the decision criteria of decision tree integrate the texture and color information. The texture information used is the entropy, the correlation and the first-order invariant central moment, while the color information includes gray value and a component value in Lab color space. To assess the effect of the proposed method under the conditions of different illumination and with different line widths of reference square (1 and 2 mm), 6 photos taken at 11:00, 15:00 and 18:00 on September 14, 2014 are used to measure soil surface roughness. Experiments show that the proposed measuring method combining the neural network and the decision tree can calculate soil roughness from the complex field photos efficiently. The error of root mean square height error can be controlled under 5%, and the calculation error of correlation length less than 1%. Considering the photo-taken distance and the illumination condition, the width of 2 mm for reference square on the whiteboard will be more suitable for high-precision soil roughness measuring. The method proposed in this paper is easy to understand and easy to implement. Its accuracy which can meet the requirement of soil surface roughness measuring makes it widely applicable. The suggestion based on experimental results will further improve the measuring accuracy for soil surface roughness.