Evapotranspiration and soil moisture estimation using different remote sensing platforms

• Ano de defesa: 2022
• Autor(a) principal: Safre, Anderson Luiz dos Santos [UNESP]
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Estadual Paulista (Unesp)
• Palavras-chave em Português: Remote sensing; Evapotranspiration; Soil moisture; Eddy covariance; Machine learning; Sensoriamento remoto; Evapotranspiração; Umidade do solo
• Link de acesso: http://hdl.handle.net/11449/235301

Evapotranspiration (ET) and soil moisture (SM) are important parameters for agricultural water management. Sensors mounted on remote sensing platforms such as satellites and Unamed Aerial Vehicles (UAVs) can provide reliable information on crop reflectance at different spatial, temporal and radiometric resolution. Machine learning nonlinear approach have shown the potential of estimating SM using optical images. The Simple Algorithm for Evapotranspiration Retrieval (SAFER) uses data from remote sensing and meteorological stations for the energy balance estimations and then actual evapotranspiration. We evaluated the performance of SAFER to estimate ET. First, we evaluated the results of SAFER using Landsat 8 imagery (30 m pixel size for optical bands and 100 m for thermal within a 18 days revisit) from 2013-2017 and standard coefficients. Then the regression coefficients were calibrated using data from Eddy Covariance (EC) towers and the results from field and remote sensing were compared. The next step was to assess SAFER performance and calibrate the algorithm with Sentinel-2 (10 m pixel size and 5 days revisit) imagery. SAFER results using the thermal band and using only optical bands were compared with six EC flux stations, located at two different sites. We applied Support Vector Regression (SVR), Random Forest (RF) and Artificial Neural Network (ANN) machine learning algorithms to estimate the soil moisture from UAV high-resolution images (2.4 cm/pix). Three bands (G, R, NIR) and NDVI was used as input. After the model calibration SAFER showed good agreement with EC data using Landsat-8 and Sentinel-2. In the Landsat-8 dataset the RMSE was 0.70 mm d-1 using the data from 5 years. The lowest RMSE (0.53 mm d-1) was in 2015 and the highest RMSE (0.89 mm d-1) in 2013. Seasonal ET was estimated and compared with the EC flux towers, showing an R2 that ranged between 0.29 to 0.97. Regarding the SAFER using Sentinel-2 images results, the model RMSE was between 0.62 to 0.84 mm d-1. The model tends to underestimate ET values when there is less water available in the root zone. The seasonal ET estimated using Sentinel-2 images showed a R2 of 0.64, when compared to that from EC measurements. Results show that all three machine learning algorithms had a great performance on the estimation of soil moisture with RMSE < 1%. SVR was the best model with a RMSE of 0.45 % and R2 = 0.71. We conclude that UAVS data and machine learning can be a great tool for soil moisture spatial variability modeling in heterogeneous terrains.