Estimativa da condutividade elétrica por meio de dados hiperespectrais em solos afetados por sais

Abstract

Remote sensing data interpretation is based primarily on the spectral reflectance analysis of materials for wavelength ranging from visible to short wave infrared (400 to 2500nm). For this, one can use reflectance spectroscopy which is a technique capable of measuring, at different wavelengths, the electromagnetic energy reflected from the surface of materials and represent it in the form of a graph called spectral reflectance curve. The analytical power of this technique derives from the spectral information being correlated directly with the chemical composition and physical characteristics of the substances that makes the target. However, the large volume of information contained in a spectral signature increases the difficulty of analyzing it, especially if the dataset is made of images. Thus, computational models are expected to be a viable means of analyzing these spectral curves. The refore, the objective of this thesis is to evaluate the performance of different computational models, such as least squares (LS), multilayer perceptron (MLP) and extreme learning machine (ELM) artificial neural networks, trained on laboratory data to estimate the electrical conductivity of salty soils, and to apply them to a hyperspectral image of the field . This thesis was organized in three parts: first, the ability of computer models to estimate the electrical conductivity of saturation extract (ECse) based on electrical conductivity data from a 1:1 dilution (EC 1:1) is assessed; second, computing strategy for best estimating the electrical conductivity of soil samples using their spectral readings under laboratory conditions are evaluated; and finally, the performance of the best found model applied to an airborne SpecTIR sensor hyperspectral image collected at the Irrigated District of the Morada Nova was evaluated. To evaluate the proposed algorithms, soil samples were collected in the Morada Nova Irrigation District with a history of salinity. These samples were used for model calibration and validation. Spectral data were obtained using the spectroradiometer FieldSpec® 3Hi-Res, from 350 to 2500nm. In an attempt to improve the performance of the models, data transformation was applied using either principal component analysis or derivative analysis. The results show the best performance was produced by the linear model fitted by least squares algorithm applied to the raw data (no transformation), and the spectral bands selected to estimate the electrical conductivity were 395, 1642 and 1717 nm. To estimate the soil's electrical conductivity from SpecTIR's image sensor data, the model calibrated in the laboratory has proved to be feasible, generating a value o f 1.46 for RPD, and 0.80 for the Pearson correlation coefficient. Therefore, one can conclude that the calibrated models using samples in the laboratory are satisfactory for estimating EC based on hyperspectral images.