*3.1. Pre-Processing of Hyperspectral Images*

Typical processing of hyperspectral imagery includes geometric correction, orthorectification, radiometric correction, and atmospheric correction. For satellite- and airplane-based hyperspectral images, the geometric and orthorectification correction are generally performed by data providers, and the radiometric and atmospheric corrections can be done following standard image processing steps available in remote sensing software. For UAV-based images, in contrast, the users need to conduct these processing steps and decide on appropriate processing methods and associated parameters. For instance, a digital elevation model (DEM) and ground control points (GCPs) are usually needed for performing the orthorectification and geometric correction [12]. If the sensor mounted on UAV is pushbroom based, accurate sensor orientation information recorded by an IMU will be needed for these corrections, and the IMU needs to be integrated into the UAV and well-calibrated [12,27]. Software packages commonly used in previous studies for performing these corrections on UAV-based hyperspectral images include ENVI (Exelis Visual Information Solutions, Boulder, CO, USA) and PARGE (ReSe Applications Schläpfer, Wil, Switzerland) [12,26,117].

Radiometric correction is conducted to convert image digital numbers to radiance using calibration coefficients that are provided by the sensor manufacturer [11]. These coefficients may need to be updated over time due to the degradation of spectral materials used to construct the hyperspectral sensors. Regarding atmospheric correction, although the UAVs are flown at low altitudes, the signals acquired are still subjective to the influence of various atmospheric absorptions and scatterings, such as oxygen absorption at 760nm; water absorption near 820, 940, 1140, 1380, and 1880 nm; and carbon dioxide absorption at 2010 and 2060 nm [12,13,26,150]. Therefore, atmospheric correction is critical for obtaining good-quality spectral information. However, Adão et al. [11] suggest that this process might be skipped if the UAVs are operated close to the ground. Therefore, the application of atmospheric correction will depend on specific flight missions and research purposes (e.g., flight altitudes, if atmosphere-influenced spectral bands are needed). Software or methods commonly used in previous studies for performing atmospheric correction on UAV-based hyperspectral images include the MODTRAN model (Spectral Sciences Inc.), ENVI FLAASH (L3Harris Geospatial), PCI Geomatica (PCI Geomatics Corporate), SMARTS model (Solar Consulting Services), and empirical line correction [12,19,27,32,33,116].

Hyperspectral images typically have hundreds of bands, and many of them are highly correlated. Therefore, dimension reduction is also an essential procedure to consider in the pre-processing of hyperspectral imagery. Many previous studies using hyperspectral imagery have discussed the challenges of data redundancy and have used different methods for dimension reduction. For instance, Miglani et al. [151] performed principal component analysis (PCA) on hyperspectral images and indicated that 99% of the information could be explained in the first 10 principal components. Amato et al. [152] discussed a few previous methods of dimension reduction, such as PCA, minimum noise fraction (MNF), and singular value decomposition (SVD), and proposed a dimension reduction algorithm based on discriminant analysis for supervised classification. Teke et al. [38] reviewed several dimension reduction methods and summarized them based on transformation techniques. Thenkabail et al. [153] discussed the problems of high dimensionality and listed a number of spectral bands that are more important for investigating crop features. Sahoo et al. [4] reviewed different methods for dimension reduction, such as PCA, uniform feature design (UMD), wavelet transforms, and artificial neural networks (ANNs), and discussed their features of operation. Wang et al. [154] proposed an auto-encoder-based dimensionality reduction method that is a deep learning-based approach. Of these different methods, the wavelet transform is one of the most widely used ones for

dimension reduction. This technique decomposes a signal into a series of scaled versions of the mother wavelet function and allows the variation of the wavelet based on the frequency information to extract localized features (e.g., local spectral variation) [155,156]. It has also been successfully used for image fusion, feature extraction, and image classification [156–158].

In addition to dimensionality reduction, band sensitivity analysis and band selection have also been widely used in hyperspectral remote sensing to reduce the data size by selecting only the bands that are sensitive to the object of interest. Different algorithms have been proposed in previous studies for band selection, such as a fast volume-gradient-based method that is an unsupervised method and removes the most redundant band successively based on the gradient of volume [159], a column subset selection-based method that maximizes the volume of the selected subset of columns (i.e., bands) and is robust to noisy bands [160], and a manifold ranking-based salient band selection method that puts band vectors in manifold space and selects a band-based ranking that can tackle the problem of inappropriate measurement of the band difference [161]. With the sensitivity analysis, previous studies have identified spectral bands that are sensitive to different crop properties, for instance, ~515, ~550, ~570, ~670, 700–740, ~800, and ~855 nm for investigating chlorophyll content; ~405, ~515, ~570, ~705, and ~720 nm for evaluating nitrogen status; ~970, ~1180, ~1245, ~1450, and ~1950 nm for assessing water content; ~682, ~855, ~910, ~970, ~1075, ~1245, ~1518, ~1725, and ~2260 nm for estimating biomass; and ~550, ~682, ~855, ~1075, ~1180, ~1450, and ~1725 nm for crop classification [36,44,153,162]. Overall, pre-processing is an essential step for improving the quality of hyperspectral images and preparing for further data analysis. After the pre-processing, the analytical methods to be discussed below can be used for analyzing the hyperspectral information and investigating various agricultural features on the ground.
