Soil Moisture Retrieval from Airborne Multispectral and Infrared Images using Convolutional Neural Network

Abstract This paper deals with the modeling of soil moisture retrieval from multispectral and infrared (IR) images using convolutional neural network (CNN). Since it is difficult to measure the soil moisture level of large fields, it is essential to retrieve soil moisture level from remotely sensed data. Quadrotor unmanned aerial vehicle (UAV) is considered as sensing platform in order to acquire data with high spatial resolution at anytime by non-experts. With considerations both on the availability of sensors for the platform and the information needed to overcome the effects of the canopies covering soil, IR and multispectral images are selected to be used for soil moisture retrieval. In order to prevent information loss by the calculation of parameters from measurements and enhance the applicabiliy for online operations, CNN is applied for the construction of soil moisture retrieval model to use the sensor measurement images directly as input data. Training and testing are conducted for the proposed CNN-based soil moisture retrieval model using the data from actual quadrotor flight over an agricultural field.