Estimating Crop Coefficients Using Linear and Deep Stochastic Configuration Networks Models and UAV-Based Normalized Difference Vegetation Index (NDVI)
Haoyu Niu, Dong Wang, YangQuan Chen
Abstract
Crop coefficient (K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> ) methods have been commonly used for evapotranspiration estimation. Researchers estimate K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> as a function of the vegetation index because of similarities between the K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> curve and the vegetation index curve. A linear regression model is usually developed between the K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> and the normalized difference vegetation index (NDVI) derived from satellite imagery. However, the spatial resolution of satellite imagery is in the range of meters or greater, which is often not enough for crops with clumped canopy structures, such as trees, and vines. In this study, the Unmanned Aerial Vehicles (UAVs) were used to collect high-resolution images in an experimental pomegranate orchard located at the USDA-ARS, San Joaquin Valley Agricultural Sciences Center, Parlier, CA. The NDVI values were derived from UAV images. The K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> values were measured from a weighing lysimeter in the pomegranate field. The relationship between the NDVI and K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> was established by using both a linear regression model and a deep stochastic configuration networks (DeepSCNs) model. Results show that the linear regression model has an R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and RMSE value of 0.975 and 0.05, respectively. The DeepSCNs regression model has an R <sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> and RMSE value of 0.995 and 0.046, respectively. The DeepSCNs model showed improved performance than the linear regression model in predicting K <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">c</sub> from NDVI.