Soil visible and near-infrared spectroscopy provides a non-destructive, rapid and low-cost approach to quantify various soil physical and chemical properties based on their reflectance in the spectral range of 400–2500 nm. With an increasing number of large-scale soil spectral libraries established across the world and new space-borne hyperspectral sensors, there is a need to explore methods to extract informative features from reflectance spectra and produce accurate soil spectroscopic models using machine learning. Features generated from regional or large-scale soil spectral data play a key role in the quantitative spectroscopic model for soil properties. The Land Use/Land Cover Area Frame Survey (LUCAS) soil library was used to explore PLS-derived components and fractal features generated from soil spectra in this study. The gradient-boosting method performed well when coupled with extracted features on the estimation of several soil properties. Transfer learning based on convolutional neural networks (CNNs) was proposed to make the model developed from laboratory data transferable for airborne hyperspectral data. The soil clay map was successfully derived using HyMap imagery and the fine-tuned CNN model developed from LUCAS mineral soils, as deep learning has the potential to learn transferable features that generalise from the source domain to target domain. The external environmental factors like the presence of vegetation restrain the application of imaging spectroscopy. The reflectance data can be transformed into a vegetation suppressed domain with a force invariance approach, the performance of which was evaluated in an agricultural area using CASI airborne hyperspectral data. However, the relationship between vegetation and acquired spectra is complicated, and more efforts should put on removing the effects of external factors to make the model transferable from one sensor to another.:Abstract I Kurzfassung III Table of Contents V List of Figures IX List of Tables XIII List of Abbreviations XV 1 Introduction 1 1.1 Motivation 1 1.2 Soil spectra from different platforms 2 1.3 Soil property quantification using spectral data 4 1.4 Feature representation of soil spectra 5 1.5 Objectives 6 1.6 Thesis structure 7 2 Combining Partial Least Squares and the Gradient-Boosting Method for Soil Property Retrieval Using Visible Near-Infrared Shortwave Infrared Spectra 9 2.1 Abstract 10 2.2 Introduction 10 2.3 Materials and methods 13 2.3.1 The LUCAS soil spectral library 13 2.3.2 Partial least squares algorithm 15 2.3.3 Gradient-Boosted Decision Trees 15 2.3.4 Calculation of relative variable importance 16 2.3.5 Assessment 17 2.4 Results 17 2.4.1 Overview of the spectral measurement 17 2.4.2 Results of PLS regression for the estimation of soil properties 19 2.4.3 Results of PLS-GBDT for the estimation of soil properties 21 2.4.4 Relative important variables derived from PLS regression and the gradient-boosting method 24 2.5 Discussion 28 2.5.1 Dimension reduction for high-dimensional soil spectra 28 2.5.2 GBDT for quantitative soil spectroscopic modelling 29 2.6 Conclusions 30 3 Quantitative Retrieval of Organic Soil Properties from Visible Near-Infrared Shortwave Infrared Spectroscopy Using Fractal-Based Feature Extraction 31 3.1 Abstract 32 3.2 Introduction 32 3.3 Materials and Methods 35 3.3.1 The LUCAS topsoil dataset 35 3.3.2 Fractal feature extraction method 37 3.3.3 Gradient-boosting regression model 37 3.3.4 Evaluation 41 3.4 Results 42 3.4.1 Fractal features for soil spectroscopy 42 3.4.2 Effects of different step and window size on extracted fractal features 45 3.4.3 Modelling soil properties with fractal features 47 3.4.3 Comparison with PLS regression 49 3.5 Discussion 51 3.5.1 The importance of fractal dimension for soil spectra 51 3.5.2 Modelling soil properties with fractal features 52 3.6 Conclusions 53 4 Transfer Learning for Soil Spectroscopy Based on Convolutional Neural Networks and Its Application in Soil Clay Content Mapping Using Hyperspectral Imagery 55 4.1 Abstract 55 4.2 Introduction 56 4.3 Materials and Methods 59 4.3.1 Datasets 59 4.3.2 Methods 62 4.3.3 Assessment 67 4.4 Results and Discussion 67 4.4.1 Interpretation of mineral and organic soils from LUCAS dataset 67 4.4.2 1D-CNN and spectral index for LUCAS soil clay content estimation 69 4.4.3 Application of transfer learning for soil clay content mapping using the pre-trained 1D-CNN model 72 4.4.4 Comparison between spectral index and transfer learning 74 4.4.5 Large-scale soil spectral library for digital soil mapping at the local scale using hyperspectral imagery 75 4.5 Conclusions 75 5 A Case Study of Forced Invariance Approach for Soil Salinity Estimation in Vegetation-Covered Terrain Using Airborne Hyperspectral Imagery 77 5.1 Abstract 78 5.2 Introduction 78 5.3 Materials and Methods 81 5.3.1 Study area of Zhangye Oasis 81 5.3.2 Data description 82 5.3.3 Methods 83 5.3.3 Model performance assessment 85 5.4 Results and Discussion 86 5.4.1 The correlation between NDVI and soil salinity 86 5.4.2 Vegetation suppression performance using the Forced Invariance Approach 86 5.4.3 Estimation of soil properties using airborne hyperspectral data 88 5.5 Conclusions 90 6 Conclusions and Outlook 93 Bibliography 97 Acknowledgements 117

Soil spectroscopy has experienced a tremendous increase in soil property characterisation, and can be used not only in the laboratory but also from the space (imaging spectroscopy). Partial least squares (PLS) regression is one of the most common approaches for the calibration of soil properties using soil spectra. Besides functioning as a calibration method, PLS can also be used as a dimension reduction tool, which has scarcely been studied in soil spectroscopy. PLS components retained from high-dimensional spectral data can further be explored with the gradient-boosted decision tree (GBDT) method. Three soil sample categories were extracted from the Land Use/Land Cover Area Frame Survey (LUCAS) soil library according to the type of land cover (woodland, grassland, and cropland). First, PLS regression and GBDT were separately applied to build the spectroscopic models for soil organic carbon (OC), total nitrogen content (N), and clay for each soil category. Then, PLS-derived components were used as input variables for the GBDT model. The results demonstrate that the combined PLS-GBDT approach has better performance than PLS or GBDT alone. The relative important variables for soil property estimation revealed by the proposed method demonstrated that the PLS method is a useful dimension reduction tool for soil spectra to retain target-related information.

Soil spectroscopy has experienced a tremendous increase in soil property characterisation, and can be used not only in the laboratory but also from the space (imaging spectroscopy). Partial least squares (PLS) regression is one of the most common approaches for the calibration of soil properties using soil spectra. Besides functioning as a calibration method, PLS can also be used as a dimension reduction tool, which has scarcely been studied in soil spectroscopy. PLS components retained from high-dimensional spectral data can further be explored with the gradient-boosted decision tree (GBDT) method. Three soil sample categories were extracted from the Land Use/Land Cover Area Frame Survey (LUCAS) soil library according to the type of land cover (woodland, grassland, and cropland). First, PLS regression and GBDT were separately applied to build the spectroscopic models for soil organic carbon (OC), total nitrogen content (N), and clay for each soil category. Then, PLS-derived components were used as input variables for the GBDT model. The results demonstrate that the combined PLS-GBDT approach has better performance than PLS or GBDT alone. The relative important variables for soil property estimation revealed by the proposed method demonstrated that the PLS method is a useful dimension reduction tool for soil spectra to retain target-related information.

Visible and near-infrared diffuse reflectance spectroscopy has been demonstrated to be a fast and cheap tool for estimating a large number of chemical and physical soil properties, and effective features extracted from spectra are crucial to correlating with these properties. We adopt a novel methodology for feature extraction of soil spectroscopy based on fractal geometry. The spectrum can be divided into multiple segments with different step–window pairs. For each segmented spectral curve, the fractal dimension value was calculated using variation estimators with power indices 0.5, 1.0 and 2.0. Thus, the fractal feature can be generated by multiplying the fractal dimension value with spectral energy. To assess and compare the performance of new generated features, we took advantage of organic soil samples from the large-scale European Land Use/Land Cover Area Frame Survey (LUCAS). Gradient-boosting regression models built using XGBoost library with soil spectral library were developed to estimate N, pH and soil organic carbon (SOC) contents. Features generated by a variogram estimator performed better than two other estimators and the principal component analysis (PCA). The estimation results for SOC were coefficient of determination (R2) = 0.85, root mean square error (RMSE) = 56.7 g/kg, the ratio of percent deviation (RPD) = 2.59; for pH: R2 = 0.82, RMSE = 0.49 g/kg, RPD = 2.31; and for N: R2 = 0.77, RMSE = 3.01 g/kg, RPD = 2.09. Even better results could be achieved when fractal features were combined with PCA components. Fractal features generated by the proposed method can improve estimation accuracies of soil properties and simultaneously maintain the original spectral curve shape.

Soil spectroscopy is a promising technique for soil analysis, and has been successfully utilized in the laboratory. When it comes to space, the presence of vegetation significantly affects the performance of imaging spectroscopy or hyperspectral imaging on the retrieval of topsoil properties. The Forced Invariance Approach has been proven able to effectively suppress the vegetation contribution to the mixed image pixel. It takes advantage of scene statistics and requires no specific a priori knowledge of the referenced spectra. However, the approach is still mainly limited to lithological mapping. In this case study, the objective was to test the performance of the Forced Invariance Approach to improve the estimation accuracy of soil salinity for an agricultural area located in the semi-arid region of Northwest China using airborne hyperspectral data. The ground truth data was obtained from an eco-hydrological wireless sensing network. The relationship between Normalized Difference Vegetation Index (NDVI) and soil salinity is discussed. The results demonstrate that the Forced Invariance Approach is able to improve the retrieval accuracy of soil salinity at a depth of 10 cm, as indicated by a higher value for the coefficient of determination (R2). Consequently, the vegetation suppression method has the potential to improve quantitative estimation of soil properties with multivariate statistical methods.

Visible and near-infrared diffuse reflectance spectroscopy has been demonstrated to be a fast and cheap tool for estimating a large number of chemical and physical soil properties, and effective features extracted from spectra are crucial to correlating with these properties. We adopt a novel methodology for feature extraction of soil spectroscopy based on fractal geometry. The spectrum can be divided into multiple segments with different step–window pairs. For each segmented spectral curve, the fractal dimension value was calculated using variation estimators with power indices 0.5, 1.0 and 2.0. Thus, the fractal feature can be generated by multiplying the fractal dimension value with spectral energy. To assess and compare the performance of new generated features, we took advantage of organic soil samples from the large-scale European Land Use/Land Cover Area Frame Survey (LUCAS). Gradient-boosting regression models built using XGBoost library with soil spectral library were developed to estimate N, pH and soil organic carbon (SOC) contents. Features generated by a variogram estimator performed better than two other estimators and the principal component analysis (PCA). The estimation results for SOC were coefficient of determination (R2) = 0.85, root mean square error (RMSE) = 56.7 g/kg, the ratio of percent deviation (RPD) = 2.59; for pH: R2 = 0.82, RMSE = 0.49 g/kg, RPD = 2.31; and for N: R2 = 0.77, RMSE = 3.01 g/kg, RPD = 2.09. Even better results could be achieved when fractal features were combined with PCA components. Fractal features generated by the proposed method can improve estimation accuracies of soil properties and simultaneously maintain the original spectral curve shape.

Soil spectroscopy is a promising technique for soil analysis, and has been successfully utilized in the laboratory. When it comes to space, the presence of vegetation significantly affects the performance of imaging spectroscopy or hyperspectral imaging on the retrieval of topsoil properties. The Forced Invariance Approach has been proven able to effectively suppress the vegetation contribution to the mixed image pixel. It takes advantage of scene statistics and requires no specific a priori knowledge of the referenced spectra. However, the approach is still mainly limited to lithological mapping. In this case study, the objective was to test the performance of the Forced Invariance Approach to improve the estimation accuracy of soil salinity for an agricultural area located in the semi-arid region of Northwest China using airborne hyperspectral data. The ground truth data was obtained from an eco-hydrological wireless sensing network. The relationship between Normalized Difference Vegetation Index (NDVI) and soil salinity is discussed. The results demonstrate that the Forced Invariance Approach is able to improve the retrieval accuracy of soil salinity at a depth of 10 cm, as indicated by a higher value for the coefficient of determination (R2). Consequently, the vegetation suppression method has the potential to improve quantitative estimation of soil properties with multivariate statistical methods.

Soil spectra are often measured in the laboratory, and there is an increasing number of large-scale soil spectral libraries establishing across the world. However, calibration models developed from soil libraries are difficult to apply to spectral data acquired from the field or space. Transfer learning has the potential to bridge the gap and make the calibration model transferrable from one sensor to another. The objective of this study is to explore the potential of transfer learning for soil spectroscopy and its performance on soil clay content estimation using hyperspectral data. First, a one-dimensional convolutional neural network (1D-CNN) is used on Land Use/Land Cover Area Frame Survey (LUCAS) mineral soils. To evaluate whether the pre-trained 1D-CNN model was transferrable, LUCAS organic soils were used to fine-tune and validate the model. The fine-tuned model achieved a good accuracy (coefficient of determination (R²) = 0.756, root-mean-square error (RMSE)= 7.07 and ratio of percent deviation (RPD) = 2.26) for the estimation of clay content. Spectral index, as suggested as a simple transferrable feature, was also explored on LUCAS data, but did not performed well on the estimation of clay content. Then, the pre-trained 1D-CNN model was further fine-tuned by field samples collect in the study area with spectra extracted from HyMap imagery, achieved an accuracy of R² = 0.601, RMSE = 8.62 and RPD = 1.54. Finally, the soil clay map was generated with the fine-tuned 1D-CNN model and hyperspectral data.