Multispectral UAV and satellite images for digital soil modeling with gradient descent boosting and artificial neural network

Sensor technology and machine learning (ML) enable rapid and accurate estimation of soil properties. This study aimed to estimate some soil characteristics with different ML algorithms using unmanned aerial vehicle (UAV) and Sentinel 2A optical satellite images. Four spectral indices and soil data were statistically compared to assess the performance of estimation. The ML algorithms including Multi-Layer Perception Artificial Neural Network (MLP-ANN) and Gradient Descent Boosting Tree (GDBT)ML were employed to improve the estimation. Bayesian optimization was used to optimize the hyperparameters of the GDBT ML algorithm. The relationships between vegetation indices calculated using the UAV and Sentinel 2A (S2A)satellite images were examined. Total of 122images were taken for 1.66 ha land with a spatial resolution of 3.99 cm. The Green Normalized Difference Vegetation Index (GNDVI), Normalized Difference Vegetation Index (NDVI), Modified Soil Adjusted Vegetation Index (MSAVI), and Transformed Soil Adjusted Vegetation Index (TSAVI) from UAV in rangeland and olive orchards were highly correlated with the same vegetation indices calculated using the S2A image. The RMSE values improved between 23.23 and 35.66% for sand, silt and soil organic matter content in MLP-UAV networks, while the MLP-S2A networks provided 9.73 to 19.85% improvement for pH, clay and soil moisture (SM). The RMSE values in UAV-based GBDT ML algorithms were more successful in estimation of pH, sand, silt, CaCO3, and SM than the S2A models and the relative improvement was between 12.16 and 93.66%. The results showed that (i) estimation success is affected by the spectral response of the soil property as well as statistical characteristics of the observation values, (ii) different optimization techniques as well as the estimation algorithm affect the estimation accuracy, (iii) land use types play an important role in the estimation variance, and (iv) the estimation performance of UAV based models is compatible with the S2A.