Spectral Input Research Articles

A spectrum contextual self-attention deep learning network for hyperspectral inversion of soil metals

The retrieval of heavy metal concentrations in naturally contaminated arable soils through hyperspectral reflectance has become increasingly significant in recent years. Presently, both conventional and deep learning-oriented metal inversion methodologies mandate the development of individualized models for each metallic element, while refraining from leveraging distant dependencies to attain the inherent heterogeneity unique to each metal species. This intricacy renders the attainment of precise inversion outcomes for particular heavy metal concentrations a formidable challenge. To tackle this challenge, the current study introduces the Spectrum Contextual Self-Attention Deep Learning Network (SCSANet), which is created to encompass long-range spectral context dependencies by employing a self-attention network. The model also includes efficient and precise spectral input techniques, as well as simultaneous output of multiple metals. Experiments are carried out in the study area to assess the precision of the proposed model in identifying lead (Pb), copper (Cu), cadmium (Cd), and mercury (Hg) concentrations. The findings reveal that metal inversion does not have a significant impact on pre-processed spectra, and enhancing the input technique of neighbourhood spectra can boost the accuracy of the inversion. The SCSANet model proposed herein achieves the highest inversion accuracy for metals with a similar magnitude of content, and outperforms the compared method in terms of inversion accuracy.

A Traceable Spectral Radiation Model of Radiation Thermometry

Despite great technical capabilities, the theory of non-contact temperature measurement is usually not fully applicable to the use of measuring instruments in practice. While black body calibrations and black body radiation thermometry (BBRT) are in practice well established and easy to accomplish, this calibration protocol is never fully applicable to measurements of real objects under real conditions. Currently, the best approximation to real-world radiation thermometry is grey body radiation thermometry (GBRT), which is supported by most measuring instruments to date. Nevertheless, the metrological requirements necessitate traceability; therefore, real body radiation thermometry (RBRT) method is required for temperature measurements of real bodies. This article documents the current state of temperature calculation algorithms for radiation thermometers and the creation of a traceable model for radiation thermometry of real bodies that uses an inverse model of the system of measurement to compensate for the loss of data caused by spectral integration, which occurs when thermal radiation is absorbed on the active surface of the sensor. To solve this problem, a hybrid model is proposed in which the spectral input parameters are converted to scalar inputs of a traditional scalar inverse model for GBRT. The method for calculating effective parameters, which corresponds to a system of measurement, is proposed and verified with the theoretical simulation model of non-contact thermometry. The sum of effective instrumental parameters is presented for different temperatures to show that the rule of GBRT, according to which the sum of instrumental emissivity and instrumental reflectivity is equal to 1, does not apply to RBRT. Using the derived models of radiation thermometry, the uncertainty of radiation thermometry due to the uncertainty of spectral emissivity was analysed by simulated worst-case measurements through temperature ranges of various radiation thermometers. This newly developed model for RBRT with known uncertainty of measurement enables traceable measurements using radiation thermometry under any conditions.

Prediction of essential oil content in spearmint (Mentha spicata) via near-infrared hyperspectral imaging and chemometrics

Spearmint (Mentha spicata L.) is grown for its essential oil (EO), which find use in food, beverage, fragrance and other industries. The current study explores the ability of near infrared hyperspectral imaging (HSI) (935 to 1720 nm) to predict, in a rapid, nondestructive manner, the essential oil content of dried spearmint (0.2 to 2.6% EO). Spectral values of spearmint samples varied considerably with spatial coordinates, and so the use of averaging the spectral values of a surface scan was warranted. Data preprocessing was done with Multiplicative Scatter Correction (MSC) or Standard Normal Variate (SNV). Selection of spectral input variables was done with Least Absolute Shrinkage and Selection Operator (LASSO), Principal Component Analysis (PCA) or Partial Least Squares (PLS). Regression was executed with linear regression (LASSO, PLS regression, PCA regression), Support Vector Machine (SVM) regression, and Multilayer Perceptron (MLP). The best prediction of EO concentration was achieved with the combination of MSC or SNV preprocessing, PLS dimension reduction, and MLP regression (1 hidden layer with 6 nodes), achieving a good prediction with a ratio of performance to deviation (RPD) of 2.84 ± 0.07, an R2 of prediction of 0.863 ± 0.008, and a RMSE of prediction of 0.219 ± 0.005% EO. These results show that NIR-HSI is a viable method for rapid, nondestructive analysis of EO concentration. Future work should explore the use of NIR in the visible spectrum, the use of HSI for determining EO in other plant materials and the potential of HSI to determine individual compounds in these solid plant/food matrices.

Low-frequency acoustic rotorcraft classification

Rotorcraft emit unique acoustic signatures tied primarily to their blade pass frequencies. In this study, we examine the ability of artificial neural network machine learning methods to successfully classify several distinct types of rotorcraft based on their low-frequency acoustic signatures. Time series, power spectral density, and Welch power spectral density input types are examined, corresponding to convolutional neural networks for the first input type, and multi-layer perceptrons for the latter two input types. A K-fold cross-validation is used to initially assess the generalization performance of the various models, along with confusion matrices for a deeper dive into the model performance. Ablation studies are discussed for insights into model behavior. These insights are used to showcase further enhanced neural network models with increased robustness and accuracy.

Estimation of Chlorophyll Content in Soybean Crop at Different Growth Stages Based on Optimal Spectral Index

Chlorophyll is an important component of crop photosynthesis as it is necessary for the material exchange between crops and the atmosphere. The amount of chlorophyll present reflects the growth and health status of crops. Spectral technology is a feasible method for obtaining crop chlorophyll content. The first-order differential spectral index contains sufficient spectral information related to the chlorophyll content and has a high chlorophyll prediction ability. Therefore, in this study, the hyperspectral index data and chlorophyll content of soybean canopy leaves at different growth stages were obtained. The first-order differential transformation of soybean canopy hyperspectral reflectance data was performed, and five indices, highly correlated with soybean chlorophyll content at each growth stage, were selected as the optimal spectral index input. Four groups of model input variables were divided according to the following four growth stages: four-node (V4), full-bloom (R2), full-fruit (R4), and seed-filling stage (R6). Three machine learning methods, support vector machine (SVM), random forest (RF), and back propagation neural network (BPNN) were used to establish an inversion model of chlorophyll content at different soybean growth stages. The model was then verified. The results showed that the correlation coefficient between the optimal spectral index and chlorophyll content of soybean was above 0.5, the R2 period correlation coefficient was above 0.7, and the R4 period correlation coefficient was above 0.8. The optimal estimation model of soybean and chlorophyll content is established through the combination of the first-order differential spectral index and RF during the R4 period. The optimal estimation model validation set determination coefficient (R2) was 0.854, the root mean square error (RMSE) was 2.627, and the mean relative error (MRE) was 4.669, demonstrating high model accuracy. The results of this study can provide a theoretical basis for monitoring the growth and health of soybean crops at different growth stages.

Circadian rhythm entrainment of the jewel wasp, Nasonia vitripennis, by antagonistic interactions of multiple spectral inputs.

Circadian light entrainment in some insects is regulated by blue-light-sensitive cryptochrome (CRY) protein that is expressed in the clock neurons, but this is not the case in hymenopterans. The hymenopteran clock does contain CRY, but it appears to be light-insensitive. Therefore, we investigated the role of retinal photoreceptors in the photic entrainment of the jewel wasp Nasonia vitripennis. Application of monochromatic light stimuli at different light intensities caused phase shifts in the wasp's circadian activity from which an action spectrum with three distinct peaks was derived. Electrophysiological recordings from the compound eyes and ocelli revealed the presence of three photoreceptor classes, with peak sensitivities at 340 nm (ultraviolet), 450 nm (blue) and 530 nm (green). An additional photoreceptor class in the ocelli with sensitivity maximum at 560-580 nm (red) was found. Whereas a simple sum of photoreceptor spectral sensitivities could not explain the action spectrum of the circadian phase shifts, modelling of the action spectrum indicates antagonistic interactions between pairs of spectral photoreceptors, residing in the compound eyes and the ocelli. Our findings imply that the photic entrainment mechanism in N. vitripennis encompasses the neural pathways for measuring the absolute luminance as well as the circuits mediating colour opponency.

Vector-Valued Intensity Measures Based on Spectral Shape to Predict Seismic Fragility Surfaces in Reinforced Concrete Buildings

Although some studies have been conducted to compute fragility surfaces of buildings using vector-valued seismic intensity measures (IMs), in all the cases, the first component of the vector usually is the spectral acceleration at first mode of vibration of the structure Sa(T1). In this study, fragility surfaces of three reinforced concrete buildings (RCB) subjected to narrow-band ground motions obtained from soft soil of Mexico City are computed considering vector-valued IMs based not only on Sa(T1), but also the velocity V(T1), pseudo-velocity Sv(T1), and normalized input energy by the mass EI/m(T1) as the first component. As second component of the vector-valued IMs, the Peak Ground Acceleration (PGA), Peak Ground Velocity (PGV), effective duration (tD), earthquake damage potential (ID) and four Np spectral shape-based parameters obtained through spectral acceleration (NpSa), velocity (NpV), pseudo-velocity (NpSv), and input energy (NpEI) have been analyzed. In order to obtain fragility surfaces, Multinomial Logistic Regression (MLR) was applied. It is observed that those vector-valued IMs based on the spectral shape proxies were more efficient to predict the probability of failure of RCB. For this reason, it is important to consider spectral shape vector-valued IMs in order to reduce uncertainty in the structural response of buildings under earthquakes. Thus, the use of two parameters instead of a single intensity measure improves the efficiency. Moreover, the fragility surfaces can be used for the seismic risk evaluation of buildings.

A joint method of spatial–spectral features and BP neural network for hyperspectral image classification

A hyperspectral image (HSI) has also highly correlated and redundant data, in addition to abundant spatial and spectral information. Only the spectral characteristics were usually utilized to perform the HSI classification in most previous studies, which leads to unsatisfactory accuracy and precision. The combination of Principal Component Analysis (PCA), Local Binary Pattern (LBP) and Back Propagation Neural Network (BPNN) (PCA-LBP-BPNN) was used to propose a novel classification method. More specifically, PCA was first used to reduce the dimensionality of HSIs for obtaining independent spectral bands sensitive to the classified objects. LBP was then adopted to extract the spatial texture features. Finally, the feature vectors were formed by fusing spatial–spectral features and input into BPNN for HSI classification. Three key parameters including the number of principal components (p) and the neurons in the hidden layer (l) as well as the learning rate (r) were optimally selected to improve the classification accuracy. Three publicly available hyperspectral datasets including Pavia University (PU), Salinas (Sa) and Botswana (Bo) were selected to validate the performance by comparing the kNN (k-Nearest Neighbor), SVM (Support Vector Machine) and Contextual Deep Convolutional Neural Network (CDCNN). The overall accuracy of PCA-LBP-BPNN reached 93.67%, 98.09% and 92.97%, respectively, for the three datasets. The method had a satisfying performance than kNN, SVM and CDCNN for the PU and Sa, but it had lower accuracies than kNN and CDCNN for the Bo due to extremely similar spectral responses. PCA-LBP-BPNN generally has satisfactory practicability and robustness in adapting to different hyperspectral datasets.

Hyperspectral Meets Optical Flow: Spectral Flow Extraction for Hyperspectral Image Classification.

Hyperspectral image (HSI) classification has always been recognised as a difficult task. It is therefore a research hotspot in remote sensing image processing and analysis, and a number of studies have been conducted to better extract spectral and spatial features. This study aimed to track the variation of the spectrum in hyperspectral images from a sequential data perspective to obtain more distinguishable features. Based on the characteristics of optical flow, this study introduces an optical flow technique for the extraction of spectral flow that denotes the spectral variation and implements a dense optical flow extraction method based on deep matching. Lastly, the extracted spectral flow are combined with the original spectral features and input into a commonly used support vector machine (SVM) classifier to complete the classification. Extensive classification experiments on three benchmark HSI test sets show that the classification accuracy obtained by the spectral flow extracted in this study (SpectralFlow) is higher than traditional spatial feature extraction methods, texture feature extraction methods, and the latest deep-learning-based methods. Furthermore, the proposed method can produce finer classification thematic maps, thereby demonstrating strong practical application potential.

Enhancing the classification metrics of spectroscopy spectrums using neural network based low dimensional space.

Spectroscopy is a methodology for gaining knowledge of particles, especially biomolecules, by quantifying the interactions between matter and light. By examining the level of light absorbed, reflected or released by a specimen, its constituents, properties, and volume can be determined. Spectra obtained through spectroscopy procedures are quick, harmless and contactless; hence nowadays preferred in chemometrics. Due to the high dimensional nature of the spectra, it is challenging to build a robust classifier with good performance metrics. Many linear and nonlinear dimensionality reduction-based classification models have been previously implemented to overcome this issue. However, they lack in capturing the subtle details of the spectra into the low dimension space or cannot efficiently handle the nonlinearity present in the spectral data. We propose a graph-based neural network embedding approach to extract appropriate features into latent space and circumvent the spectrums' nonlinearity problem. Our approach performs dimensionality reduction into two phases: constructing a nearest neighbor graph and producing almost linear embedding using a fully connected neural network. Further, the low dimensional embedding is subjected to classification using the Random Forest algorithm. In this paper, we have implemented and compared our technique with four nonlinear dimensionality techniques widely used for spectral data analysis. In this study, we have considered five different spectral datasets belonging to specific applications. The various classification performance metrics of all the techniques are evaluated. The proposed approach is able to perform competitively well on six different low-dimensional spaces for each dataset with an accuracy score above 95% and Matthew's correlation coefficient value close to 1. The trustworthiness score of almost 1 show that the presented dimensionality reduction approach preserves the closest neighbor structure of high dimensional spectral inputs into latent space.

Deep learning near-infrared quality prediction based on multi-level dynamic feature

As a fast, efficient, nondestructive, and pollution-free technology, near-infrared spectroscopy (NIRS) has become a practical method for online quality prediction in the process industry. The inherent dynamics of the industrial processes make the NIR spectral data no longer accord with the assumption of independent and identically distributed, which leads to the limitations of the static quality prediction models. In addition, the existing NIRS-based quality prediction methods lack consideration of the possible nonlinear relationship between NIR spectra and quality variables of complex organic samples and the over sensitivity brought by molecular-level detection. This paper proposes a multi-level dynamic feature-based deep learning NIR quality prediction method to solve the above problems. Firstly, short-term feature extraction is mainly based on 2D convolution neural networks (CNNs) to extract short-term dynamic features and eliminate the multi-collinearity and information redundancy with a 2D NIR dynamic spectral matrix input. Thereinto, considering the levels of short-term dynamic features, a combination of dilated CNN and the dense connection is proposed to construct the multi-level features and ensure the continuisty and integrity of the time dimension. Secondly, based on the extracted short-term dynamic features, the long-term dynamic spectral prediction constructed with gate recurrent unit (GRU) is proposed to capture long-term dependence between NIR spectra and predict the NIR spectrum at the next time step. Here, considering the inability of the original GRU to take account of the dynamic feature at all time steps and the possible time lag between industrial laboratory sampling and NIRS detection, temporal attention is utilized to redistribute and fuse the dynamic features of the GRU at all time steps. Finally, a fully connected layer is used to realize the regression from the NIR spectrum of the next moment to the corresponding quality variables. The effectiveness and accuracy of the method are verified by a case of the reaction and distillation process of 2,6 xylenol and corresponding NIR spectral data.

Identifying Nematode Damage on Soybean through Remote Sensing and Machine Learning Techniques

Identifying nematode damage in large soybean areas is not always achievable in a practical way. Multispectral reflectance sensors have not been thoroughly evaluated to detect nematode damage in soybeans (Glycine max L.). The main research aims of this study were to: (i) determine the bivariate relationship between individual spectral bands and vegetation indices (VIs) relative to soybean conditions (symptomatic versus asymptomatic), and (ii) to select the best model for identifying plant conditions using three algorithms (logistic regression—LR, random forest—RF, conditional inference tree—CIT) and three options for data input using bands, vegetation indices (VIs), and bands plus VIs. The trial was conducted in Brazil on three on-farm soybean fields presenting different species of nematode infestation. Multispectral imagery was obtained using a drone-mounted MicaSense RedEdge® sensor. At each sampling, georeferenced point nematode infestation and spectral measurements of soybean plants were retrieved for the classification of symptomatic and asymptomatic areas, according to the threshold level adopted. Bivariate analysis of variance (ANOVA), LR, RF, and CIT were used to select the multispectral bands/VIs that discriminated among symptomatic and asymptomatic plants, assessing the best model via their respective parameters for accuracy, sensitivity, and specificity. The greatest classification accuracy (>0.70) was achieved when using the CIT algorithm with the spectral bands only, with green (560 ± 20 nm) and near-infrared (840 ± 40 nm) included as the main spectral input variables in the model. These results demonstrate the potential of combining remotely sensed data and machine learning to distinguish nematode-symptomatic and asymptomatic soybean plants.

Engaging effort improves efficiency during word recognition in cochlear implant users

Cochlear Implant (CI) users must cope with the degraded spectral input received through their device. This reduced quality leads to changes in how words are recognized: lexical access is delayed, leading to differences in competition between lexical candidates. CI users also report increased effort and fatigue during language processing, but effort is not clearly related to improved accuracy. We investigated the role that listening effort plays in the dynamics of word recognition. We related lexical competition and listening effort in 77 post-lingually deaf CI users with Visual World Paradigm and pupillometry tasks. Subjects also completed measures of peripheral auditory function (spectral ripple discrimination and temporal modulation detection). We used the difference in pupil size during the recognition of spoken and written words to isolate effort for spoken word recognition. Difference in pupil size predicted efficiency in the spoken word recognition task, over and above age and peripheral fidelity. That is, CI users who engage more effort were faster to recognize words and showed lexical competition similar to normal hearing listeners. This interacted with spectral resolution: better discrimination offsets the need to engage effort. This suggests that listening effort can be engaged to improve efficiency during word recognition.

Explanation of Deep Learning–Based Radioisotope Identifier for Plastic Scintillation Detector

Radioisotope identification (RIID) is a representative application of deep learning for radiation measurements. Deep learning-based RIID models have been implemented in various types of radiation detectors; however, very few of these models have been interpreted using explainable artificial intelligence (XAI) methods. This paper presents an explanation of a deep learning–based RIID model for a plastic scintillation detector. The RIID task is defined as a multilabel binary classification problem, and the dataset is generated using a random sampling procedure. The identification performance is verified using experimental data. The experimental results demonstrate that the performance of the RIID models increased with the increase in the total counts of the dataset. Additionally, XAI methods are implemented, and their explanatory performance is verified for the spectral input. The domain knowledge of RIID for the plastic scintillation detector is that patterns near the Compton edge can be used as evidence for the existence of radioisotopes. Among the implemented XAI methods, integrated gradient and layerwise relevance propagation exhibited concurrence with the domain knowledge, with the Shapley value explanation method presenting the most reliable results.

Strong energy component is more important than spectral selectivity in modeling responses of midbrain auditory neurons to wide-band environmental sounds

Modeling central auditory neurons in response to complex sounds not only helps understanding neural processing of speech signals but can also provide insights for biomimetics in neuro-engineering. While modeling responses of midbrain auditory neurons to synthetic tones is rather good, modeling those to environmental sounds is less satisfactory. Environmental sounds typically contain a wide range of frequency components, often with strong and transient energy. These stimulus features have not been examined in the conventional approach of auditory modeling centered on spectral selectivity. To this end, we firstly compared responses to an environmental sound of auditory midbrain neurons across 3 subpopulations of neurons with frequency selectivity in the low, middle and high ranges; secondly, we manipulated the sound energy, both in power and in spectrum, and compared across these subpopulations how their modeled responses were affected. The environmental sound was recorded when a rat was drinking from a feeding bottle (called the ‘drinking sound’). The sound spectrum was divided into 20 non-overlapping frequency bands (from 0 to 20 kHz, at 1 kHz width) and presented to an artificial neural model built on a committee machine with parallel spectral inputs to simulate the known tonotopic organization of the auditory system. The model was trained to predict empirical response probability profiles of neurons to the repeated sounds. Results showed that model performance depended more on the strong energy components than on the spectral selectivity. Findings were interpreted to reflect general sensitivity to rapidly changing sound intensities at the auditory midbrain and in the cortex.

Estimating crop type and yield of small holder fields in Burkina Faso using multi-day Sentinel-2

Remote Sensing affords the opportunity to monitor and evaluate data scarce regions where field collection efforts are costly. A particular challenge is monitoring and evaluation in regions with smallholder agricultural systems (∼1 ha) that are often subsistence focused, vulnerable to food insecurity and data scarce. Using multi-day moderate resolution Sentinel-2 and Random Forest models, this study shows that crop type and rice yields in Burkina Faso can be predicted with greater than ∼80% accuracy in the rainy season. Model optimization using varying spectral and vegetation index inputs can increase crop type and yield prediction accuracy in the dry season where denser cultivation is a challenge for the 10–20 m resolution of Sentinel-2. However, there is a trade-off between opting for very high-resolution imagery (<2 m) or the number of bands offered by Sentinel-2 as the bands that occupy and vegetation indices that utilize the red through NIR ranges were most important across all models. In addition, model type, linear Regression or nonlinear Random Forest, matters little when estimating yield in these landscapes, unless Harmonic regression is utilized for the linear model. This study also showed that a model trained with high quality 2019 dry season crop cut data can predict the subsequent dry season's interannual crop type with overall accuracy as high as 60%, comparable to crop type models trained with 2020 survey data and used to estimate crop type in the concurrent season, as the survey collection. This indicates some utility in leveraging the calibrated Random Forest models to make skillful predictions of interannual crop type and ultimately food availability of nearby communities for years with no training data. Given increasing global food prices and restricted commodity trade, understanding local agricultural productivity using affordable and timely remote sensing-based methods is essential for ensuring appropriate humanitarian interventions.

Modeling and performance analysis of the LUVOIR coronagraph instrument

Future space missions such as the Large UV/Optical/Infrared Surveyor (LUVOIR) and the Habitable Exoplanet Observatory, when equipped with coronagraphs with active wavefront control to suppress starlight, will allow the discovery and characterization of habitable exoplanets. The Extreme Coronagraph for Living Planetary Systems (ECLIPS) is the coronagraph instrument on the LUVOIR Surveyor mission concept, an 8- to 15-m segmented telescope. ECLIPS is split into three channels, namely, UV (200 to 400 nm), optical (400 to 850 nm), and near IR (850 nm to 2 μm), with each channel equipped with two deformable mirrors for wavefront control, a suite of coronagraph masks, a low-order/out-of-band wavefront sensor, and separate science imagers and spectrographs. The apodized pupil Lyot coronagraph and the vector vortex coronagraph are the baselined mask technologies for ECLIPS to enable the required 10 − 10 contrast for observations in the habitable zones of nearby stars for LUVOIR-A (15-m telescope) and LUVOIR-B (8-m telescope), respectively. Their performance depends on active wavefront sensing and control, as well as metrology subsystems to compensate for aberrations induced by segment errors (e.g., piston and tip/tilt), secondary mirror misalignment, and global low-order wavefront errors. Here, we present the latest results of the simulation of these effects for the LUVOIR coronagraph instrument and discuss the achieved contrast for exoplanet detection and characterization after closed-loop wavefront estimation and control algorithms have been applied. Finally, we show simulated observations using high-fidelity spatial and spectral input models of complete planetary systems generated with the Haystacks code framework.

Empirical Correlations between the Spectral Input Energy and Spectral Acceleration

ABSTRACT Correlating different intensity measures (IMs) enables the joint consideration of multiple IMs in various performance-based earthquake engineering applications, such as ground motion selection. The correlations between the spectral input energy and spectral accelerations have not been previously examined. This study develops a correlation model quantifying the uncertainties due to GMPE selection and a finite sample size. The results show a strong correlation between the spectral amplitudes at the same period, and the correlation decreases as the period difference increases. With the proposed model, ground motion characteristics reflected by the input energy can be explicitly considered in performance-based earthquake engineering applications.

Multi-scale convolution neural network with residual modules for determination of drugs in human hair using surface-enhanced Raman spectroscopy with a gold nanorod film self-assembled by inverted evaporation

Detection of illegal drug users is crucial in controlling drug-related crimes, reducing drug prevalence, and protecting human lives to ensure social stability. In this study, surface-enhanced Raman spectroscopy (SERS) and deep learning networks were employed to determine methamphetamine, ketamine, and morphine in human hair. Drugs were obtained from hair through alkaline hydrolysis and liquid–liquid extraction, and gold nanorods were employed to prepare the self-assembled film as the SERS substrate by inverted evaporation. The film showed good uniformity and excellent sensitivity, with a relative standard deviation of 15.6% and a detection limit of at least 10-10 M in the SERS detection of crystal violet. The spectra of methamphetamine, ketamine, and morphine at 0.05–1.0, 0.1–2.0, and 0.1–2.0 ng/mg were obtained, and the three drugs could be detected. Inception, a multi-scale feature extraction network, was combined with residual modules (Inception–ResNet) to develop the identification models of drugs, and the effect of spectral input form as a vector or matrix was explored. Inception–ResNet with input form of matrix outweighed other methods with 100.00%, 100.00%, and 99.23% accuracies in the training, validation, and prediction sets, respectively. In brief, SERS and Inception–ResNet with the spectra in matrix form provide an efficient and accurate determination of drugs in human hair, enabling the retrospective evaluation of drug use, and the method will be anticipated to detect excitant, poison, and toxic chemicals in human hair.

Estimating dry biomass and plant nitrogen concentration in pre-Alpine grasslands with low-cost UAS-borne multispectral data – a comparison of sensors, algorithms, and predictor sets

Abstract. Grasslands are an important part of pre-Alpine and Alpine landscapes. Despite the economic value and the significant role of grasslands in carbon and nitrogen (N) cycling, spatially explicit information on grassland biomass and quality is rarely available. Remotely sensed data from unmanned aircraft systems (UASs) and satellites might be an option to overcome this gap. Our study aims to investigate the potential of low-cost UAS-based multispectral sensors for estimating above-ground biomass (dry matter, DM) and plant N concentration. In our analysis, we compared two different sensors (Parrot Sequoia, SEQ; MicaSense RedEdge-M, REM), three statistical models (linear model; random forests, RFs; gradient-boosting machines, GBMs), and six predictor sets (i.e. different combinations of raw reflectance, vegetation indices, and canopy height). Canopy height information can be derived from UAS sensors but was not available in our study. Therefore, we tested the added value of this structural information with in situ measured bulk canopy height data. A combined field sampling and flight campaign was conducted in April 2018 at different grassland sites in southern Germany to obtain in situ and the corresponding spectral data. The hyper-parameters of the two machine learning (ML) approaches (RF, GBM) were optimized, and all model setups were run with a 6-fold cross-validation. Linear models were characterized by very low statistical performance measures, thus were not suitable to estimate DM and plant N concentration using UAS data. The non-linear ML algorithms showed an acceptable regression performance for all sensor–predictor set combinations with average (avg; cross-validated, cv) Rcv2 of 0.48, RMSEcv,avg of 53.0 g m2, and rRMSEcv,avg (relative) of 15.9 % for DM and with Rcv,avg2 of 0.40, RMSEcv,avg of 0.48 wt %, and rRMSEcv, avg of 15.2 % for plant N concentration estimation. The optimal combination of sensors, ML algorithms, and predictor sets notably improved the model performance. The best model performance for the estimation of DM (Rcv2=0.67, RMSEcv=41.9 g m2, rRMSEcv=12.6 %) was achieved with an RF model that utilizes all possible predictors and REM sensor data. The best model for plant N concentration was a combination of an RF model with all predictors and SEQ sensor data (Rcv2=0.47, RMSEcv=0.45 wt %, rRMSEcv=14.2 %). DM models with the spectral input of REM performed significantly better than those with SEQ data, while for N concentration models, it was the other way round. The choice of predictors was most influential on model performance, while the effect of the chosen ML algorithm was generally lower. The addition of canopy height to the spectral data in the predictor set significantly improved the DM models. In our study, calibrating the ML algorithm improved the model performance substantially, which shows the importance of this step.