Combing transfer learning with the OPtical TRApezoid Model (OPTRAM) to diagnosis small-scale field soil moisture from hyperspectral data

Abstract

Accurate, timely, and continuous soil moisture information is helpful for crop stress diagnosis and irrigation management decision. OPtical TRApezoid Model (OPTRAM) based on optical satellite data has been proven to be an effective method for assessing soil moisture status. However, the applicability of OPTRAM to small-scale field soil moisture assessment remains to be explored. In this study, we propose a strategy for the genetically parameterized OPTRAM and evaluate its applicability on Unmanned Aerial Vehicle (UAV) high-resolution hyspectral data. The results showed that: (1) When OPTRAM was used to genetically parameterized with PROSAIL generated dataset, 46 characteristic narrowband bands (|R|= 0.52–0.78) were determined in the spectral region of near infrared (NIR) (750–850 nm) and SWIR (1060–1080 and 1450–1500 nm); (2) By fine-tuned soil moisture estimation model using transfer learning strategy, the reliable soil moisture estimation was achieved in three crops (R2=0.57–0.64; RMSE=0.008–0.022 m3m−3);(3) Compared to soil moisture estimation model using a single spectral region (NIR or SWIR), the DSWC model that combine NIR and SWIR was more effective for tracking soil moisture; (4) The scale effect was observed when the fine-tuned soil moisture estimation model was applied on the high-resolution UAV images. The model performance was stable in pixel size of 1–7 cm and began to drop at pixel size of 11 cm. The above results advance the application of OPTRAM on small farmland soil moisture assessment and demonstrate the application potential of OPTRAM on narrow-band hyperspectral data. This study provides a new candidate for the use of hyperspectral data to estimate soil moisture, and scientific support for precision agriculture and irrigation scheduling.