UAV-based multispectral and thermal cameras to predict soil water content – A machine learning approach

•We determine soil water content by UAV multispectral and thermal infrared cameras.•Four machine learning regression methods were investigated for estimating SWC.•Multispectral cameras ensured better input than thermal cameras.•Relationship between SWC and thermal data was exponential.•Single pixel extraction performed with better results instead of buffered areas. Soil water content (SWC) estimation is a crucial issue of agricultural production, and its mapping is an important task. We aimed to study the efficacy of UAV-based thermal (TH) and multispectral (MS) cameras in SWC mapping. Soil samples were collected and the SWC content was determined in a laboratory as reference data and four machine learning regression algorithms (Random Forest [RF], Elastic Net [ENR], General Linear Model [GLM], Robust Linear Model [RLM]) were tested for the prediction efficacy, combined with three pixel value extraction methods (single pixel, mean of 20 and 30 cm radius buffer). We found that MS cameras ensured better input data than TH cameras: R2s were 0.97 vs 0.71, mean-normalized root mean square errors (nRMSE) were 10 vs 25 %, respectively. Best models were obtained by the RF (0.97 R2) and ENR (0.88 R2) in case of MS camera. Relationship between SWC and thermal data was exponential, which was incorrectly handled by the GLM (>40 % nRMSE; furthermore, RLM and ENR was not working with only one variable), thus, TH data was acceptable only with the RF (24.4 % nRMSE). Single pixel extraction provided the best input for the estimations, mean of buffered areas did not perform better in the models. Maps provided appropriate SWC estimations according to the nRMSEs, with high spatial resolution. In spite of potential inaccuracies, visualizing the spatial heterogeneities can be a great help to farmers to increase the efficacy of planning irrigation in precision agriculture.