Next Article in Journal
Rice Yield Prediction Based on Simulation Zone Partitioning and Dual-Variable Hierarchical Assimilation
Previous Article in Journal
Granitoid Mapping with Convolutional Neural Network from ASTER and Landsat 8 OLI Data: A Case Study in the Western Junggar Orogen
Previous Article in Special Issue
Analysis of Road Surface Texture for Asphalt Pavement Adhesion Assessment Using 3D Laser Technology
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Remote Sensing
Volume 17
Issue 3
10.3390/rs17030385
remotesensing-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

UAV-Based Phytoforensics: Hyperspectral Image Analysis to Remotely Detect Explosives Using Maize (Zea mays)

by
Paul V. Manley II
1,*,
Stephen M. Via
2 and
Joel G. Burken
1
1
Department of Civil, Architectural, and Environmental Engineering, Missouri University of Science and Technology, Rolla, MO 65409, USA
2
Department of Biology, Norfolk State University, Norfolk, VA 23504, USA
*
Author to whom correspondence should be addressed.
Remote Sens. 2025, 17(3), 385; https://doi.org/10.3390/rs17030385
Submission received: 14 December 2024 / Revised: 14 January 2025 / Accepted: 21 January 2025 / Published: 23 January 2025
(This article belongs to the Special Issue Hyperspectral Imaging and LiDAR Scanning Technology Development and Applications)
Browse Figures
Versions Notes

Abstract

:
Remnant explosive devices are a deadly nuisance to both military personnel and civilians. Traditional mine detection and clearing is dangerous, time-consuming, and expensive. And routine production and testing of explosives can create groundwater contamination issues. Remote detection methods could be rapidly deployed in vegetated areas containing explosives as they are known to cause stress in vegetation that is detectable with hyperspectral sensors. Hyperspectral imagery was employed in a mesocosm study comparing stress from a natural source (drought) to that of plants exposed to two different concentrations of Royal Demolition Explosive (RDX; 250 mg kg−1, 500 mg kg−1). Classification was accomplished with the machine learning algorithms Support Vector Machine (SVM), Random Forest (RF), and Least Discriminant Analysis (LDA). Leaf-level plant data assisted in validating plant stress induced by the presence of explosives and was detectable. Vegetation indices (VIs) have historically been used for dimension reduction due to computational limitations; however, we measured improvements in model precision, recall, and accuracy when using the complete range of available wavelengths. In fact, almost all models applied to spectral data outperformed their index counterparts. While challenges exist in scaling research efforts from the greenhouse to the field (i.e., weather, solar lighting conditions, altitude when imaging from a UAV, runoff containment, etc.), this experiment is promising for subsequent research efforts at greater scale and complexity aimed at detecting emerging contaminants.

1. Introduction

Explosives and other energetic compounds are often found in soil and groundwater in and around Department of Defense (DoD) facilities. Royal Demolition Explosive (hexahydro-1,3,5-trinitro-1,3,5-triazine; RDX) is one of the most common explosive compounds used for military applications and has been detected widely across the United States from its production and use [1]. Millions of hectares of land are contaminated with explosives in the United States from explosives testing at military ranges and production facilities [2,3]. Setting off individual ordnance leaves behind undetonated residue, often in the milligram range, but contamination issues are compounded by multiple detonations in the same areas over years to decades and may affect groundwater sources [4,5]. The detection and monitoring of explosives contamination is a necessity on and around military bases and DoD facilities to limit potential transport and human exposure as 2,4,6-trinitrotoluene (TNT) is a known carcinogen and RDX may be carcinogenic [6,7]. Explosives contamination also exists in more acutely dangerous forms, such as unstable remnant devices from wars and conflicts that are still capable of detonating [8]. Given the dispersed and heterogeneous distribution of fugitive energetic compounds (i.e., pollutants), plants are ideal candidates to serve as detectors of subsurface contaminants because they remain in place for extended periods of time over large areas and serve as passive samplers of their surrounding environment [9].
Remote sensing technologies have enabled the detection of subsurface contaminants (e.g., hazardous waste, explosives, and hydrocarbons) from afar using plants, and vegetation offers a unique view into the subsurface through intimate contact with the soil, groundwater, and soil–vapor phases [10,11,12,13]. Using plants as subsurface samplers of RDX contamination, as with other pollutants, is possible because of the octanol/water partitioning coefficient, log Kow, a value inherent to RDX. The log Kow value is an effective predictor of whether a compound can cross root membranes and be mobile inside a plant. Kow values between 0.5 and 3 are ideal for plant uptake. Consequently, the log Kow of RDX is within that range at 0.87 [14,15,16,17]. RDX enters plants by crossing root membranes via bulk water transport where it is then either stored in roots or translocated to leaves and can accumulate at elevated concentrations [18,19,20]. The analytical chemical testing of plant tissues has demonstrated that plants can be used as sentinels of subsurface pollutants, including explosives, and this is referred to as ‘phytoforensics’. These methods have proven vital to identify species that are able to uptake certain compounds with varying efficiencies of uptake and accumulation [21,22]. Phosphor imaging can be used to detect the uptake and localization of radiolabeled explosives in plants; and high-performance liquid chromatography (HPLC), gas chromatography, and LC with tandem mass spectrometry methods have been developed to quantify concentrations of explosives in plant tissues [23,24,25]. Despite their effectiveness, these methods are not ideal due to their high financial cost and time commitment to retrieve samples from field locations. A method in which large tracts of land can be covered in an efficient manner and quickly screened is imperative.
Decades of research into plant responses to various stressors, both natural and anthropogenic, with spectroradiometry and hyperspectral imagery (HSI) has paved the way for sensors to become smaller, lighter, and cheaper, allowing them to be carried by unmanned aerial vehicles (UAVs). However, water absorption bands for vegetation occur in the short-wave infrared (SWIR), beyond the spectral range of many lightweight hyperspectral sensors, including the sensor used in this study [26,27]. As chlorophyll fluorescence tends to increase with short-term drought conditions and chlorophyll pigments degrade under extended drought periods, indices related to red edge position and chlorophyll content can be used as proxies for water-content-related indices [27]. Using plants as a natural window into the subsurface via remote sensing of plant stress responses is a unique approach to locate pollutants. Once remote sensing can screen for stress in native plant populations, chemical screening of the plants and underlying groundwater can be targeted efficiently for the accurate delineation of contaminated soil and groundwater.
RDX is a phytotoxic nitramine that expresses its effects on plants with chlorosis, leaf curling at the edges, necrosis, and even plant death, not completely dissimilar to plants undergoing drought stress [28,29,30,31,32,33,34]. It is also capable of causing physiological and morphological changes in plants, though not always in consistent ways [31,35,36,37,38]. When mixed in a 60/40 ratio with TNT, known as Composition B (Comp B), synergistic effects have been observed [36,39,40]. The co-contamination of RDX with TNT and other explosives is likely due to the use in munitions and composite explosives. Of the two, RDX is more mobile, recalcitrant, and able to efficiently mobilize within plants. Its detection via a stress response in plant leaves with remote sensing techniques may be effective at delineating contaminated soil and groundwater. To develop these potential methods, an improved understanding of detecting RDX-induced plant stress is required. Very little work in terms of hyperspectral analysis of plants exposed to RDX in a controlled environment has taken place, and we were unable to find any evidence of field-scale experiments in the literature.
Significant efforts have gone into streamlining hyperspectral reflectance spectra from hundreds of wavelengths to a series of vegetation indices (VIs) that rely on a few wavelengths to assess data more rapidly and are often tailored to specific plant species and/or causes of stress. One downfall of this approach is that it requires selecting specific wavelengths and disregarding possibly valuable wavelengths in the analysis [41]. Machine learning (ML) algorithms do not have this limitation; they can learn patterns from the data themselves for anomaly detection, making them adaptable to detect changes caused by new or unknown contaminants across different species and conditions. In this age of big data coinciding with major improvements in computing power in commercially available computers, the added steps to detect and remove highly correlated wavelengths or indices while resulting in no increase in accuracy reaffirm the need to look at “the whole picture”, especially when detecting plant stress from new contaminants in heterogenous ecosystems. Also, our scale of data collection causes concerns in data analysis; real-world environmental data are often highly variable, or ‘noisier’, than controlled-environment data. Explosives-related indices do not exist; this is why we chose to compare ML classification to traditional VI analysis: to assess all the data available instead of trying to rely on VIs developed for different purposes.
The primary objective of this study was to differentiate explosives-induced stress from natural stress (drought conditions) and clean soil (control) using UAV-acquired HSI. The existing, yet limited, body of work was all conducted in greenhouse settings with artificial lighting. It was hypothesized that RDX exposure would affect plant morphology and physiological functioning in a more severe and/or different manner than drought conditions, and that these differences would be discernable via both ground truth data and the ML classification of HSI. The process followed is outlined in Figure 1.

2. Materials and Methods

2.1. Explosives

Four treatment groups were established for this experiment: control (0 mg kg−1), drought (0 mg kg−1), 250 mg kg−1 RDX (referred to as RDX250), and 500 mg kg−1 RDX (referred to as RDX500). A comparison of the effects of RDX to TNT was initially planned; however, plants of both hybrids exposed to TNT were not viable for the study due to toxic impacts as low germination rates were experienced with the exposure. Other studies that utilized maize and TNT reported successful germination from contaminated soil, though the concentrations used were lower than those in this study [20,42].
Each group was contained to its own mesocosm (plastic pools) 1.22 m across, 0.2 m deep, that were filled with 85 kg of sand and locally sourced topsoil (50/50 mixture). Holes were drilled in the bottom of the mesocosms to allow for water drainage, and each mesocosm was placed in a secondary pool to contain any contaminated runoff. Drought plants were watered half as much as other treatment groups to maintain a water deficit state. RDX mesocosms were dosed accordingly with RDX dissolved in acetone, thoroughly mixed into the entire depth of the mesocosm, covered to prevent photodegradation by sunlight, and allowed to dry before the plants were transplanted. Control and drought mesocosms were also dosed with equivalent amounts of acetone to control for any possible effects.

2.2. Plants

The mesocosms were divided in half (oriented east and west) and 50 seeds each of Zea mays L. (maize; hybrids P0688 and P2088; Pioneer®, Corteva Agriscience, Johnston, IA, USA) were planted after dosing and acetone evaporation. The P2088 hybrid was on the west side and the P0688 hybrid was on the east side. A 4.57 m × 2.13 m × 2.13 m portable greenhouse (Outsunny, Aosom LLC, Wilsonville, OR, USA) was constructed and placed over the mesocosms to control the watering regime and to protect plants from storm events and too much sunlight (Figure 2). Due to the shallow nature of the pools, germinated seeds were given three weeks to grow and reach a growth stage between V3 and V5.

2.3. Hyperspectral Imaging

Hyperspectral images were captured with a Headwall Nano-Hyperspec hyperspectral imager (Headwall Photonics, Bolton, MA, USA), which measures light reflectance of 274 bands from 400 to 1000 nm with a 2 nm spectral resolution, which was mounted to a DJI M600 Pro drone (Figure 3; SZ DJI Technology, Co., Ltd., Shenzhen, China). The imagery of all the plants was collected from a single pass at 30 m altitude and a ground speed of 8 m/s with a 12 mm lens, yielding a spatial resolution of approximately 2 cm. Raw HSI was imported into Headwall SpectralView software (version 5.5.1) for pre-processing that consisted of radiometric correction by applying a gain and offset and orthorectification. A XSens MTi-G710 GNSS/IMU (XSens, Los Angeles, CA, USA) was mounted to the Nano and integrated to measure real-time GPS location and sensor orientation (roll, pitch, and yaw) to create an orthorectified hyperspectral image, a process that takes a raw image and applies it to the geoid giving it latitude and longitude as opposed to local x and y coordinates. A 0.254 m × 0.254 m Labsphere (Labsphere, Inc., North Sutton, NH, USA) Spectralon® white reflectance panel, designed to evenly reflect 99% of light, was placed adjacent the mesocosms to be captured in the imagery for atmospheric corrections during post-processing. This was performed by taking a subset of the white reference and dividing the entire scene by its average reflectance [43,44,45].
All groups were imaged at the same time, then broken up into individual subsets by treatment to be processed separately. Regions of Interest (ROIs) were manually created in ENVI (Exelis Visual Information Solutions, NV5 Geospatial, Boulder, CO, USA) for labeling each type of material or category in the scene (vegetation, soil, shadow, or pool). Support Vector Machine (SVM) classification then used the ROIs to classify a training raster for developing parameters that were applied to the remaining rasters for classifying each material type. All but the vegetation class were masked out and the resulting vegetation-only images were diced on a 50 × 50 tile grid in ENVI.

2.4. Field Measurements

Quantum yield (Qy) represents the light-adapted quantum efficiency of photosystem II (Φ PSII) and was measured at the leaf level with a FluorPen FP 100 fluorometer (Photon Systems Instruments, Drásov, Czech Republic) three times per plant on a sample of plants for each treatment (n = 9). An atLEAF chlorophyll meter (FT Green LLC, Wilmington, DE, USA) estimated the chlorophyll content (n = 9). Plant heights were measured for the same plants chosen for fluorescence and chlorophyll measurements. Plant tissue and soil samples were collected for future analysis by liquid chromatography/tandem mass spectrometry (LC-MS/MS) to determine chemical concentrations in the plant biomass.

2.5. Statistics

The Holm–Bonferroni method was used to perform a pairwise comparison between groups in R (R version 4.2.3, Foundation for Statistical Computing, Vienna, Austria). Field measurements and their closest corresponding VIs were plotted together; physical height relates to biomass indices, indices focusing on chlorophyll content are a proxy for leaf chlorophyll measurements, and stress detected remotely can indicate qualitative responses in Qy.
Custom Interactive Data Language (IDL; Exelis Visual Information Solutions, Boulder, CO, USA) scripts were written to prepare datasets for statistical analysis. Each subset was segmented on a 50 × 50 tile grid to increase the image sampling size per treatment. The resulting dice sizes in ENVI are dictated by the number of tiles and image size. Dices containing no data values were omitted from the analysis, as were any with errant index values. Of the remaining dices containing data, the average reflectance spectra of each dice were extracted and outputted to a .csv file for later analysis. Thirty-four VIs (Table 1) were also calculated from every pixel, and the mean values of the raster were exported to a .csv file.
Due to variations in leaf surface area in each image, dicing each subset resulted in slightly different numbers of pixels representing the leaves of each hybrid and treatment. T-tests of each wavelength, calculated in Excel (Microsoft Corporation, Redmond, WA, USA), were plotted along spectral curves to highlight where differences existed in the VIS/NIR.

2.6. Machine Learning

We chose to compare two algorithms widely used in HSI classification, Support Vector Machine (SVM) and Random Forest (RF), along with a lesser used option, Linear Discriminant Analysis (LDA), by applying each to spectral data and VIs. Each has a unique approach to the task of classifying imagery with a potential beneficial application to remotely sensed data for contaminant exposure. SVM is ideal when needing to separate two or more classes because it finds the hyperplane that maximizes the margin between the closest points of the classes, which are referred to as support vectors, and is effective with non-linear datasets such as vegetation spectra [73,74]. RF is also ideal for big data as it works well with thousands of variables, has no expectations on the distribution of the data, and requires little preprocessing to break up the data into decision trees [75,76]. LDA assumes that the data being analyzed are normally distributed and that none of the independent variables exhibit multicollinearity, which does not allow for comparing indices with multiple similar versions (e.g., ARI1 and ARI2) or indices that were developed independently but behave similarly (e.g., LAI and NDVI) [77,78]. Per a review by Sheykhmousa et al. (2020), the mean classification accuracies reported for SVM were higher than those for RF when applied to hyperspectral data, though the performance of each is driven by the number of features and classes [79].
The extracted spectral values and calculated VIs were classified using SVM, RF, and LDA in R with the caret package (version 6.0-94). The resulting dices (Table 2) from segmenting were split 80/20, where 80% of the data was used to train and generate models that were then tested on the remaining 20% for each ML method. The model parameters for each can be found in Table 3.
To compare the spectra to indices derived from spectra, we ran the same ML algorithms separately on the spectra and indices alike. The entire available spectral range (400–1000 nm) was used in each algorithm and all 34 indices were used for SVM and RF. However, LDA required the removal of 11 due to the multicollinear nature of some of the indices. We initially attempted to use a Principal Components Analysis (PCA) for the dimension reduction of the spectra before applying the ML models and saw minimal gains or losses. At our scale, the cost of compute time was not a factor; therefore, we used all the data available. Considering the size and scale of explosives contamination globally, variable selection should be considered for larger datasets. Downsampling was used in each to account for imbalanced datasets. RF utilized cross-validation (CV) because repeated CV resulted in overfitting, while repeated CV was used for SVM and LDA.

3. Results and Discussion

The plants grown in soil dosed with RDX were morphologically discernible from the controls. There were some variations in responses between the two hybrids used, possibly due to how each genotype was modified based on their marketed characteristics; the product listings highlighted the wide leaf area and greenness of P2088 and emphasized the root growth and development of P0688, which allowed it to be drought tolerant. In fact, in terms of ground data, no drought group significantly differed from the controls. In this study, plants in the RDX groups for hybrid P2088 were significantly taller than their control and drought counterparts and more similar in height to that of all the treatments of hybrid P0688 (Figure 4). The presence of RDX seemed to cause P2088 leaves to be very narrow and rigid and lack the characteristic leaf droop in corn plants. This in turn made the plants appear taller since the leaves used for the measurements as tassels had not yet appeared. Despite the significant height differences of the RDX P2088 groups and the taller, though insignificantly, RDX P0688 plants, the Normalized Difference Vegetation Index (NDVI), an index related to plant biomass and correlated to plant height, was reduced. In the case of P0688, the NDVI of RDX plants was considerably lower relative to the control and drought groups, while all test groups were reduced compared to the control for the P2088 hybrid.
RDX is known to accumulate in leaf tissues at the edges and can cause reductions in chlorophyll concentrations. The exact mechanism is unclear and the extent to which chlorophyll is reduced is species-specific [23,29,38]. For example, the coastal shrub Baccharis halimifolia only showed a significant reduction in total chlorophyll levels at 1500 mg kg−1 and not at lower concentrations [38], while rice (Oriza sativa L.) showed declines in chlorophyll concentrations starting at 500 mg kg−1 RDX [38,80]. In this study, RDX visibly influenced leaf coloration, size, and orientation, and caused necrosis in leaf tips. These visual symptoms were detected at both the leaf level and with remotely sensed data.
Reductions in chlorophyll content were observed at the leaf level and remotely in all explosives groups across hybrids compared to those in the control and drought groups (Figure 5). Chlorophyll reduction is notable because chlorophyll and chloroplasts act as defense mechanisms against xenobiotics and against any resulting stress induced from their presence. Chloroplasts are thought to offer electrons to RDX to photolyze it, and then chlorophylls are capable of absorbing free radicals, or reactive oxygen species (ROS), which can cause physical damage to photocenters [81,82,83,84]. ROS occur naturally in plants during metabolic processes and are produced under stress conditions [85,86,87]. Less chlorophyll implies a potential hindrance in a plant’s natural ability to mitigate stress, which can induce photoinhibition if allowed to progress, thus worsening the cycle of producing structure-damaging ROS and reducing chlorophyll concentrations until eventual leaf senescence [88].
RDX has been shown to limit light-adapted reactions in some species but not others. The quantum yield (Qy) of a coastal shrub, Myrica cerifera, showed a negative linear response to increasing concentrations of RDX, while the Qy of another shrub, Baccharis halimifolia, was not significantly affected at similar concentrations [36,38]. In this experiment, all Qy values of explosives groups were reduced compared to those of the control and drought plants, though the Qy was only significantly impacted negatively for the 250 mg kg−1 RDX group of the P2088 hybrid, possibly due to the deactivation of Φ PSII as a response to RDX uptake (Figure 6). VREI1, a stress index, detected blue shifts of the red edge in both hybrids, indicating negative effects from RDX exposure.
Hyperspectral reflectance is capable of qualitatively identifying chlorophyll deficits faster than leaf-level measurements and in a non-destructive manner. Generally, vegetative stress is observed in hyperspectral imagery (HSI) with increased reflectance in red wavelengths (620–700 nm), a shift in the red edge to shorter wavelengths (blue shift), and/or decreased reflectance in near-infrared (NIR) wavelengths (750–1000 nm) [89]. The amplified signal in red wavelengths occurs due to reductions in chlorophyll concentrations as less pigments are available for light absorption [44,90,91]. In this study, chlorophyll concentrations of hybrids P0688 and P2088 were shown to have been negatively impacted by the presence of explosives. However, the responses of red wavelengths in P0688 reflectance and the red edge of both hybrids were not observed in the HSI data as hypothesized.
Plant morphology and leaf orientation also play a role in signal intensity as the signal becomes lessened with an increased incident or leaf angle, factors that can of course be mitigated in a controlled laboratory setting but present interesting challenges in field applications [92,93]. P2088 reflectance of both RDX groups had a muted signal across all wavelengths, likely due to the vertical orientation of the leaves, an apparent morphological response to RDX uptake (Figure 7). Plants grown in soils dosed with explosives were 57% taller on average, likely due to the more rigid, vertical fashion in which they grew. This morphological response accounted for the flattened spectral response across the assessed wavelength range. Drought plants of the same hybrid displayed a spectral signature more in line with vegetative stress and demonstrated the leaf droop characteristic of maize plants.
For P0688, hyperspectral reflectance in both the 250 and 500 mg kg−1 RDX groups was significantly discernable in the control and drought plants as initially hypothesized (Figure 7). The leaves of P0688 RDX plants visibly did not grow in the same vertical nature as the P2088 counterparts. This hybrid did exhibit the expected increased reflectance in red wavelengths, along with reduced NIR reflectance. These quantifiable changes show promise for future research aimed at expanding the current limited body of work around remote sensing of explosives or other subsurface pollutants in field settings.
We chose to include machine learning in the analysis to better understand if models could predict the presence of explosives from hyperspectral imagery data. SVM and LDA outperformed RF in terms of precision, recall, and accuracy for spectrum training data and when applied to test data. RF performed significantly better for the index training data but was on par with SVM and LDA when applied to the test data. For SVM on the spectral data, the average recall for P0688 and P2088 was 75% and 77%, respectively (Figure 8). The average precision was 73% for P0688 and 77% for P2088. The overall accuracy was 73% for P0688 and 78% for P2088. The RF on spectra demonstrated an average recall of 66% for P0688 and 65% for P2088. The precision was 63% and 66% for P0688 and P2088, respectively. The overall accuracy reached 63% and 67% for P0688 and P2088. LDA when applied to spectra exhibited an average recall of 73% for P0688 and 71% for P2088, with an average precision of 70% and 72%. The accuracy was slightly better, with 71% and 72% for P0688 and P2088, respectively.
The average SVM recall for P0688 and P2088 index data was lower than the hyperspectral data, at 66% and 70%, respectively. The precision was similarly lower at 62% for P0688 and 70% for P2088, with an overall accuracy of 63% and 70%. The RF average recall was 64% for P0688 and 70% for P2088, while the precision was 61% and 72%. The accuracy was 61% for P0688 and 69% for P2088. The index data demonstrated an average recall of 66% for P0688 and 70% for P2088, with precision at 62% and 70%, respectively. The overall accuracy was 62% for P0688 and 70% for P2088.
In general, the models performed better on the hyperspectral data compared to the VIs, particularly in terms of precision and recall. All models performed similarly on index data with precision, recall, and accuracy hovering between 62 and 70%. Across all classifiers, the hyperspectral data for P2088 consistently outperformed P0688, with SVM and LDA proving to be the most robust.

4. Conclusions

This is the first such application of UAV-based hyperspectral image acquisition for remote phytoforensics to locate explosives. The results demonstrate that HSI was able to differentiate between drought-induced stress and chemically imposed stress of RDX at a mesocosm scale, despite the many complex and interacting factors introduced by ambient environmental settings and the varying responses to explosives uptake by plants of different functional types seen in the literature. Leaf-level chlorophyll measurements were the strongest ground truth indicator of RDX exposure. Qy showed reductions across all RDX treatments, but only significantly for hybrid P2088 exposed to 250 mg kg−1 RDX. Plant height was not a good indicator as P2088 plants grew in an unexpectedly rigid and vertical manner. For both hybrids, SVM performed the best overall because it effectively handles high-dimensional hyperspectral data by finding a robust decision boundary, resulting in strong precision and recall across both training and test sets. It was able to generalize well to new data, particularly when working with the original spectra. RF performed well on training data but struggled with generalization, particularly on the test data for indices. While it had high recall, its precision was lower, suggesting overfitting to the training set, which hindered its ability to accurately classify unseen data. LDA, on the other hand, showed moderate performance, with good recall but lower precision, especially when using indices. This was likely due to LDA’s assumptions about normality and equal covariance among classes, which did not hold true for the hyperspectral data, making it less effective at capturing the complexities of the problem and leading to lower performance on the test data.
Currently, no RDX- or explosives-specific vegetation reflectance index exists. Therefore, we compared results from the entire available spectral range at our disposal to existing vegetation indices and found no diminishing returns from utilizing all 274 bands in our analysis. We believe this is foundational research for future large-scale efforts focused on the delineation of explosives contamination. We offer unique findings that strongly suggest varying levels of explosives contamination in soil can be detected from aerial hyperspectral images of plants.

5. Future Direction

To better connect the leaf angle and the reflectance impacts, a 3D modeling approach with light detection and ranging (LiDAR) could be beneficial to include when remotely attempting to determine plant stress where the stressor may also cause concurrent morphological changes such as those seen in this study.
In the future, combining remotely sensed data from various sources would be beneficial to better understanding and discerning the variety of potential effects on plants from contaminant-related stress. Enlisting hyperspectral sensors that detect reflectance from a wider range—i.e., 400–2500 nm—would provide insights around how water uptake and use is affected by contaminant exposure and how this is related to measured stress responses. The addition of LiDAR concurrent with HSI would also aid in understanding stress effects on canopy structure, such as what was seen in this study on leaf angle and size. Thermal imagery could also be of importance to detect literal “hot spots” as leaves with lower water content exhibit higher temperatures. This research area is still dependent on ground truthing but when a misstep can literally cause the loss of life or limb, ground truthing may not be possible. Therefore, being able to dependably rely on accurate remote sensing data is crucial to demining efforts in vegetated regions.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/rs17030385/s1, Supplementary Information: Breakdown of precision and recall for each ML algorithm utilized by hybrid for both test and training datasets. Also reported are the model accuracies by hybrid.

Author Contributions

Conceptualization, P.V.M.II and J.G.B.; methodology, P.V.M.II, S.M.V. and J.G.B.; software, P.V.M.II; validation, J.G.B.; formal analysis, P.V.M.II; investigation, P.V.M.II; resources, J.G.B.; data curation, P.V.M.II; writing—original draft preparation, P.V.M.II; writing—review and editing, P.V.M.II, S.M.V. and J.G.B.; visualization, P.V.M.II; supervision, J.G.B.; project administration, P.V.M.II and J.G.B.; funding acquisition, J.G.B. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported in part by the National Science Foundation (IIA-1355406 and IIA-1430427).

Data Availability Statement

The raw data supporting the conclusions of this article will be made available by the authors on request.

Acknowledgments

Many thanks to colleagues Catherine Johnson for her support in the field work and facilities, Pengfei Ma for his assistance in validating our statistical methods, and Brian Glass for his assistance with R. The authors also thank the editors and reviewers for their diligent work in maintaining the quality of this journal.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Rylott, E.L.; Jackson, R.G.; Edwards, J.; Womack, G.L.; Seth-Smith, H.M.B.; Rathbone, D.A.; Strand, S.E.; Bruce, N.C. An Explosive-Degrading Cytochrome P450 Activity and Its Targeted Application for the Phytoremediation of RDX. Nat. Biotechnol. 2006, 24, 216–219. [Google Scholar] [CrossRef]
  2. Rylott, E.L.; Lorenz, A.; Bruce, N.C. Biodegradation and Biotransformation of Explosives. Curr. Opin. Biotechnol. 2011, 22, 434–440. [Google Scholar] [CrossRef] [PubMed]
  3. Winfield, L.E.; Rodger, J.H.; D’Surney, S.J. The Responses of Selected Terrestrial Plants to Short (<12 Days) and Long Term (2, 4 and 6 Weeks) Hexahydro-1,3,5-Trinitro-1,3,5-Triazine (RDX) Exposure. Part I: Growth and Developmental Effects. Ecotoxicology 2004, 13, 335–347. [Google Scholar] [CrossRef]
  4. Schoenmuth, B.; Mueller, J.O.; Scharnhorst, T.; Schenke, D.; Büttner, C.; Pestemer, W. Elevated Root Retention of Hexahydro-1,3,5-Trinitro-1,3,5-Triazine (RDX) in Coniferous Trees. Environ. Sci. Pollut. Res. 2014, 21, 3733–3743. [Google Scholar] [CrossRef] [PubMed]
  5. Walsh, M.R. Explosives Residues Resulting from the Detonation of Common Military Munitions: 2002–2006. 2007. Available online: https://www.semanticscholar.org/paper/Explosives-Residues-Resulting-from-the-Detonation-Walsh/44ae86ab79c99a2585247f8b9e2b84cba1498f28 (accessed on 20 January 2025).
  6. Rylott, E.L.; Budarina, M.V.; Barker, A.; Lorenz, A.; Strand, S.E.; Bruce, N.C. Engineering Plants for the Phytoremediation of RDX in the Presence of the Co-Contaminating Explosive TNT. New Phytol. 2011, 192, 405–413. [Google Scholar] [CrossRef]
  7. US Environmental Protection Agency. Technical Fact Sheet—Hexahydro-1,3,5-Trinitro1,3,5-Triazine (RDX); Office of Solid Waste And Emergency Response, EPA 505-F-14-004 2017; US Environmental Protection Agency: Washington, DC, USA, 2017; pp. 1–9. [Google Scholar]
  8. Habib, M.K. Controlled Biological and Biomimetic Systems for Landmine Detection. Biosens. Bioelectron. 2007, 23, 1–18. [Google Scholar] [CrossRef]
  9. Burken, J.G.; Vroblesky, D.A.; Balouet, J.C. Phytoforensics, Dendrochemistry, and Phytoscreening: New Green Tools for Delineating Contaminants from Past and Present. Environ. Sci. Technol. 2011, 45, 6218–6226. [Google Scholar] [CrossRef]
  10. Slonecker, T.; Fisher, G.B.; Aiello, D.P.; Haack, B. Visible and Infrared Remote Imaging of Hazardous Waste: A Review. Remote Sens. 2010, 2, 2474–2508. [Google Scholar] [CrossRef]
  11. Zinnert, J.C.; Via, S.M.; Young, D.R. Distinguishing Natural from Anthropogenic Stress in Plants: Physiology, Fluorescence and Hyperspectral Reflectance. Plant Soil 2013, 366, 133–141. [Google Scholar] [CrossRef]
  12. Lassalle, G.; Credoz, A.; Hédacq, R.; Fabre, S.; Dubucq, D.; Elger, A. Assessing Soil Contamination Due to Oil and Gas Production Using Vegetation Hyperspectral Reflectance. Environ. Sci. Technol. 2018, 52, 1756–1764. [Google Scholar] [CrossRef]
  13. Emengini, E.J.; Blackburn, G.A.; Theobald, J.C. Discrimination of Plant Stress Caused by Oil Pollution and Waterlogging Using Hyperspectral and Thermal Remote Sensing. J. Appl. Remote Sens. 2013, 7, 073476. [Google Scholar] [CrossRef]
  14. Burken, J.G.; Schnoor, J.L. Predictive Relationships for Uptake of Organic Contaminants by Hybrid Poplar Trees. Environ. Sci. Technol. 1998, 32, 3379–3385. [Google Scholar] [CrossRef]
  15. Panz, K.; Miksch, K. Phytoremediation of Explosives (TNT, RDX, HMX) by Wild-Type and Transgenic Plants. J. Environ. Manag. 2012, 113, 85–92. [Google Scholar] [CrossRef] [PubMed]
  16. Via, S.M.; Zinnert, J.C. Impacts of Explosive Compounds on Vegetation: A Need for Community Scale Investigations. Environ. Pollut. 2016, 208, 495–505. [Google Scholar] [CrossRef]
  17. Bagheri, M.; He, X.; Oustriere, N.; Liu, W.; Shi, H.; Limmer, M.A.; Burken, J.G. Investigating Plant Uptake of Organic Contaminants through Transpiration Stream Concentration Factor and Neural Network Models. Sci. Total Environ. 2021, 751, 141418. [Google Scholar] [CrossRef]
  18. Ghosh, M.; Singh, S.P. A Review on Phytoremediation of Heavy Metals and Utilization of It’s by Products. Asian J. Energy Environ. 2005, 6, 214–231. [Google Scholar]
  19. Singh, S.N.; Mishra, S. Phytoremediation of TNT and RDX. In Environmental Science and Engineering (Subseries: Environmental Science); Singh, S.N., Ed.; Environmental Science and Engineering; Springer International Publishing: Berlin/Heidelberg, Germany, 2014; Volume 112, pp. 371–392. ISBN 978-3-319-01082-3. [Google Scholar]
  20. Vila, M.; Lorber-Pascal, S.; Laurent, F. Fate of RDX and TNT in Agronomic Plants. Environ. Pollut. 2007, 148, 148–154. [Google Scholar] [CrossRef]
  21. Burken, J.G.; Shanks, J.V.; Thompson, P.L. Phytoremediation and Plant Metabolism of Explosives and Nitroaromatic Compounds. In Biodegradation of Nitroaromatic Compounds and Explosives; CRC Press: Boca Raton, FL, USA, 2000; pp. 239–275. ISBN 9781420032673. [Google Scholar]
  22. Karnjanapiboonwong, A.; Mu, R.; Yuan, Y.; Shi, H.; Ma, Y.; Burken, J.G. Plant Tissue Analysis for Explosive Compounds in Phytoremediation and Phytoforensics. J. Environ. Sci. Health Tox Hazard. Subst. Environ. Eng. 2012, 47, 2219–2229. [Google Scholar] [CrossRef]
  23. Brentner, L.B.; Mukherji, S.T.; Walsh, S.A.; Schnoor, J.L. Localization of Hexahydro-1,3,5-Trinitro-1,3,5-Triazine (RDX) and 2,4,6-Trinitrotoluene (TNT) in Poplar and Switchgrass Plants Using Phosphor Imager Autoradiography. Environ. Pollut. 2010, 158, 470–475. [Google Scholar] [CrossRef]
  24. Chen, D.; Liu, Z.L.; Banwart, W. Concentration-Dependent RDX Uptake and Remediation by Crop Plants. Environ. Sci. Pollut. Res. 2011, 18, 908–917. [Google Scholar] [CrossRef]
  25. Campbell, S.; Ogoshi, R.; Uehara, G.; Li, Q.X. Trace Analysis of Explosives in Soil: Pressurized Fluid Extraction and Gas and Liquid Chromatography-Mass Spectrometry. J. Chromatogr. Sci. 2003, 41, 284–288. [Google Scholar] [CrossRef] [PubMed]
  26. Sims, D.A.; Gamon, J.A. Estimation of Vegetation Water Content and Photosynthetic Tissue Area from Spectral Reflectance: A Comparison of Indices Based on Liquid Water and Chlorophyll Absorption Features. Remote Sens. Environ. 2003, 84, 526–537. [Google Scholar] [CrossRef]
  27. Kim, D.M.; Zhang, H.; Zhou, H.; Du, T.; Wu, Q.; Mockler, T.C.; Berezin, M.Y. Highly Sensitive Image-Derived Indices of Water-Stressed Plants Using Hyperspectral Imaging in SWIR and Histogram Analysis. Sci. Rep. 2015, 5, 15919. [Google Scholar] [CrossRef] [PubMed]
  28. Best, E.P.H.; Smith, T.; Hagen, F.L.; Dawson, J.O.; Torrey, A.J. Capacities of Candidate Herbaceous Plants for Phytoremediation of Soil-Based TNT and RDX on Ranges. 2007. Available online: https://apps.dtic.mil/sti/tr/pdf/ADA489006.pdf (accessed on 9 June 2024).
  29. Vila, M.; Mehier, S.; Lorber-Pascal, S.; Laurent, F. Phytotoxicity to and Uptake of RDX by Rice. Environ. Pollut. 2007, 145, 813–817. [Google Scholar] [CrossRef] [PubMed]
  30. Douglas, T.A.; Walsh, M.E.; Weiss, C.A.; McGrath, C.J.; Trainor, T.P. Desorption and Transformation of Nitroaromatic (TNT) and Nitramine (RDX and HMX) Explosive Residues on Detonated Pure Mineral Phases. Water Air Soil Pollut. 2012, 223, 2189–2200. [Google Scholar] [CrossRef]
  31. Manley, P.V., II. Plant Functional Trait and Hyperspectral Reflectance Responses to Comp B Exposure: Efficacy of Plants as Landmine Detectors. Master’s Thesis, Virginia Commonwealth University, Richmond, VA, USA, 2016. [Google Scholar]
  32. Manley, P.V., II; Sagan, V.; Fritschi, F.B.; Burken, J.G. Remote Sensing of Explosives-Induced Stress in Plants: Hyperspectral Imaging Analysis for Remote Detection of Unexploded Threats. Remote Sens. 2019, 11, 1827. [Google Scholar] [CrossRef]
  33. Hasanpour, K.; Aalami, A.; Hosseini, R.; Naeimi, S. Effect of Drought Stress on Some Morphological and Physiological Traits in Aegilops Tauschii Genotypes. Iran. J. Genet. Plant Breed. 2021, 12, 890–896. [Google Scholar] [CrossRef]
  34. Zhang, F.; Zhou, G. Estimation of Vegetation Water Content Using Hyperspectral Vegetation Indices: A Comparison of Crop Water Indicators in Response to Water Stress Treatments for Summer Maize. BMC Ecol. 2019, 19, 18. [Google Scholar] [CrossRef]
  35. Via, S.M.; Zinnert, J.C.; Young, D.R. Differential Effects of Two Explosive Compounds on Seed Germination and Seedling Morphology of a Woody Shrub, Morella Cerifera. Ecotoxicology 2015, 24, 194–201. [Google Scholar] [CrossRef]
  36. Via, S.M.; Zinnert, J.C.; Butler, A.D.; Young, D.R. Comparative Physiological Responses of Morella Cerifera to RDX, TNT, and Composition B Contaminated Soils. Environ. Exp. Bot. 2014, 99, 67–74. [Google Scholar] [CrossRef]
  37. Ali, A.; Zinnert, J.C.; Muthukumar, B.; Peng, Y.; Chung, S.M.; Stewart, C.N. Physiological and Transcriptional Responses of Baccharis Halimifolia to the Explosive “Composition B” (RDX/TNT) in Amended Soil. Environ. Sci. Pollut. Res. 2014, 21, 8261–8270. [Google Scholar] [CrossRef] [PubMed]
  38. Zinnert, J.C. Plants as Phytosensors: Physiological Responses of a Woody Plant in Response to RDX Exposure and Potential for Remote Detection. Int. J. Plant Sci. 2012, 173, 1005–1014. [Google Scholar] [CrossRef]
  39. Panz, K.; Miksch, K.; Sójka, T. Synergetic Toxic Effect of an Explosive Material Mixture in Soil. Bull. Environ. Contam. Toxicol. 2013, 91, 555–559. [Google Scholar] [CrossRef]
  40. Via, S.M.; Zinnert, J.C.; Young, D.R. Multiple Metrics Quantify and Differentiate Responses of Vegetation to Composition B. Int. J. Phytoremediat. 2017, 19, 56–64. [Google Scholar] [CrossRef]
  41. Mertens, S.; Verbraeken, L.; Sprenger, H.; Demuynck, K.; Maleux, K.; Cannoot, B.; De Block, J.; Maere, S.; Nelissen, H.; Bonaventure, G.; et al. Proximal Hyperspectral Imaging Detects Diurnal and Drought-Induced Changes in Maize Physiology. Front. Plant Sci. 2021, 12, 640914. [Google Scholar] [CrossRef]
  42. Price, R.A.; Pennington, J.C.; Larson, S.L.; Neumann, D.; Hayes, C.A. Uptake of RDX and TNT by Agronomic Plants. Soil Sediment Contam. Int. J. 2002, 11, 307–326. [Google Scholar] [CrossRef]
  43. Carter, G.A.; Knapp, A.K. Leaf Optical Properties in Higher Plants: Linking Spectral Characteristics to Stress and Chlorophyll Concentration. Am. J. Bot. 2001, 88, 677–684. [Google Scholar] [CrossRef]
  44. Peguero-Pina, J.J.; Morales, F.; Flexas, J.; Gil-Pelegrín, E.; Moya, I. Photochemistry, Remotely Sensed Physiological Reflectance Index and de-Epoxidation State of the Xanthophyll Cycle in Quercus Coccifera under Intense Drought. Oecologia 2008, 156, 1–11. [Google Scholar] [CrossRef]
  45. Zinnert, J.C.; Nelson, J.D.; Hoffman, A.M. Effects of Salinity on Physiological Responses and the Photochemical Reflectance Index in Two Co-Occurring Coastal Shrubs. Plant Soil 2012, 354, 45–55. [Google Scholar] [CrossRef]
  46. Gitelson, A.A.; Merzlyak, M.N.; Chivkunova, O.B. Optical Properties and Nondestructive Estimation of Anthocyanin Content in Plant Leaves. Photochem. Photobiol. 2001, 74, 38. [Google Scholar] [CrossRef]
  47. Gitelson, A.A.; Zur, Y.; Chivkunova, O.B.; Merzlyak, M.N. Assessing Carotenoid Content in Plant Leaves with Reflectance Spectroscopy. Photochem. Photobiol. 2002, 75, 272. [Google Scholar] [CrossRef]
  48. Tucker, C.J. Red and Photographic Infrared Linear Combinations for Monitoring Vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef]
  49. Huete, A.R.; Didan, K.; Miura, T.; Rodriguez, E.P.; Gao, X.; Ferreira, L.G. Overview of the Radiometric and Biophysical Performance of the MODIS Vegetation Indices. Remote Sens. Environ. 2002, 83, 195–213. [Google Scholar] [CrossRef]
  50. Gitelson, A.A.; Gritz, Y.; Merzlyak, M.N. Relationships between Leaf Chlorophyll Content and Spectral Reflectance and Algorithms for Non-Destructive Chlorophyll Assessment in Higher Plant Leaves. J. Plant Physiol. 2003, 160, 271–282. [Google Scholar] [CrossRef] [PubMed]
  51. Sripada, R.P. Measured and Aerial CIR Photographs Were Obtained for Each Site at V7 or VT Prior to N Application. Significant Grain Yield Responses to N PL and N V7, and N VT Were Observed. Ph.D. Dissertation, North Carolina State University, Raleigh, NC, USA, 2005. [Google Scholar]
  52. Louhaichi, M.; Borman, M.M.; Johnson, D.E. Spatially Located Platform and Aerial Photography for Documentation of Grazing Impacts on Wheat. Geocarto Int. 2001, 16, 65–70. [Google Scholar] [CrossRef]
  53. Gitelson, A.A.; Merzlyak, M.N. Remote Estimation of Chlorophyll Content in Higher Plant Leaves. Int. J. Remote Sens. 1997, 18, 2691–2697. [Google Scholar] [CrossRef]
  54. Sripada, R.P.; Heiniger, R.W.; White, J.G.; Meijer, A.D. Aerial Color Infrared Photography for Determining Early In-Season Nitrogen Requirements in Corn. Agron. J. 2006, 98, 968–977. [Google Scholar] [CrossRef]
  55. Crippen, R.E. Calculating the Vegetation Index Faster. Remote Sens. Environ. 1990, 34, 71–73. [Google Scholar] [CrossRef]
  56. Boegh, E.; Soegaard, H.; Broge, N.H.; Schelde, K.; Thomsen, A.; Hasager, C.B.; Jensen, N.O. Airborne Multispectral Data for Quantifying Leaf Area Index, Nitrogen Concentration, and Photosynthetic Efficiency in Agriculture. Remote Sens. Environ. 2002, 81, 179–193. [Google Scholar] [CrossRef]
  57. Daughtry, C.S.T.; Walthall, C.L.; Kim, M.S.; de Colstoun, E.B.; McMurtrey, J.E. Estimating Corn Leaf Chlorophyll Concentration from Leaf and Canopy Reflectance. Remote Sens. Environ. 2000, 74, 229–239. [Google Scholar] [CrossRef]
  58. Haboudane, D.; Miller, J.R.; Pattey, E.; Zarco-Tejada, P.J.; Strachan, I.B. Hyperspectral Vegetation Indices and Novel Algorithms for Predicting Green LAI of Crop Canopies: Modeling and Validation in the Context of Precision Agriculture. Remote Sens. Environ. 2004, 90, 337–352. [Google Scholar] [CrossRef]
  59. Datt, B. A New Reflectance Index for Remote Sensing of Chlorophyll Content in Higher Plants: Tests Using Eucalyptus Leaves. J. Plant Physiol. 1999, 154, 30–36. [Google Scholar] [CrossRef]
  60. Sims, D.A.; Gamon, J.A. Relationships between Leaf Pigment Content and Spectral Reflectance Across a Wide Range of Species, Leaf Structures and Developmental Stages. Remote Sens. Environ. 2002, 81, 337. [Google Scholar] [CrossRef]
  61. Rouse, R.W.H.; Haas, J.A.W.; Schell, J.A.; Deering, D.W. Monitoring Vegetation Systems in the Great Plains with ERTS (Earth Resources Technology Satellites). In Proceedings of the Goddard Space Flight Center Third ERTS Symposium, Washington, DC, USA, 10–14 December 1973; NASA: Washington, DC, USA, 1974; Volume 1, pp. 309–317. [Google Scholar]
  62. Rondeaux, G.; Steven, M.; Baret, F. Optimization of Soil-Adjusted Vegetation Indices. Remote Sens. Environ. 1996, 55, 95–107. [Google Scholar] [CrossRef]
  63. Peñuelas, J.; Filella, I.; Gamon, J.A. Assessment of Photosynthetic Radiation-use Efficiency with Spectral Reflectance. New Phytol. 1995, 131, 291–296. [Google Scholar] [CrossRef]
  64. Gamon, J.A.; Serrano, L.; Surfus, J.S. The Photochemical Reflectance Index: An Optical Indicator of Photosynthetic Radiation Use Efficiency across Species, Functional Types, and Nutrient Levels. Oecologia 1997, 112, 492–501. [Google Scholar] [CrossRef]
  65. Merzlyak, M.N.; Gitelson, A.A.; Chivkunova, O.B.; Rakitin, V.Y. Non-Destructive Optical Detection of Leaf Senescence and Fruit Ripening. Physiol. Plant 1999, 106, 135–141. [Google Scholar] [CrossRef]
  66. Roujean, J.L.; Breon, F.M. Estimating PAR Absorbed by Vegetation from Bidirectional Reflectance Measurements. Remote Sens. Environ. 1995, 51, 375–384. [Google Scholar] [CrossRef]
  67. Gitelson, A.A.; Merzlyak, M.N. Spectral Reflectance Changes Associated with Autumn Senescence of Aesculus hippocastanum L. and Acer platanoides L. Leaves. Spectral Features and Relation to Chlorophyll Estimation. J. Plant Physiol. 1994, 143, 286–292. [Google Scholar] [CrossRef]
  68. Curran, P.J.; Windham, W.R.; Gholz, H.L. Exploring the Relationship between Reflectance Red Edge and Chlorophyll Concentration in Slash Pine Leaves. Tree Physiol. 1995, 15, 203–206. [Google Scholar] [CrossRef]
  69. Huete, A.R. A Soil-Adjusted Vegetation Index (SAVI). Remote Sens. Environ. 1988, 25, 295–309. [Google Scholar] [CrossRef]
  70. Birth, G.S.; McVey, G.R. Measuring the Color of Growing Turf with a Reflectance Spectrophotometer. Agron. J. 1968, 60, 640–643. [Google Scholar] [CrossRef]
  71. Broge, N.H.; Leblanc, E. Comparing Prediction Power and Stability of Broadband and Hyperspectral Vegetation Indices for Estimation of Green Leaf Area Index and Canopy Chlorophyll Density. Remote Sens. Environ. 2000, 76, 156–172. [Google Scholar] [CrossRef]
  72. Vogelmann, J.E.; Rock, B.N.; Moss, D.M. Red Edge Spectral Measurements from Sugar Maple Leaves. Int. J. Remote Sens. 1993, 14, 1563–1575. [Google Scholar] [CrossRef]
  73. Rumpf, T.; Mahlein, A.K.; Steiner, U.; Oerke, E.C.; Dehne, H.-W.; Plümer, L. Early Detection and Classification of Plant Diseases with Support Vector Machines Based on Hyperspectral Reflectance. Comput. Electron. Agric. 2010, 74, 91–99. [Google Scholar] [CrossRef]
  74. Souza, A.; Rojas, M.Z.; Yang, Y.; Lee, L.; Hoagland, L. Classifying Cadmium Contaminated Leafy Vegetables Using Hyperspectral Imaging and Machine Learning. Heliyon 2022, 8, e12256. [Google Scholar] [CrossRef]
  75. Agrawal, A.; Petersen, M.R. Detecting Arsenic Contamination Using Satellite Imagery and Machine Learning. Toxics 2021, 9, 333. [Google Scholar] [CrossRef]
  76. Lv, J.; Liu, X. Predicting Arsenic Concentration in Rice Plants from Hyperspectral Data Using Random Forests. Adv. Intell. Soft Comput. 2011, 128, 601–606. [Google Scholar] [CrossRef]
  77. Wang, Q.; Zhang, J.; Chen, J.; Zhang, Y. An Improved Spectral Reflectance and Derivative Feature Fusion for Hyperspectral Image Classification. In Proceedings of the 2011 IEEE International Geoscience and Remote Sensing Symposium, Vancouver, BC, Canada, 24–29 July 2011; pp. 1696–1699. [Google Scholar] [CrossRef]
  78. Zagajewski, B.; Tømmervik, H.; Bjerke, J.W.; Raczko, E.; Bochenek, Z.; Klos, A.; Jarocińska, A.; Lavender, S.; Ziólkowski, D. Intraspecific Differences in Spectral Reflectance Curves as Indicators of Reduced Vitality in High-Arctic Plants. Remote Sens. 2017, 9, 1289. [Google Scholar] [CrossRef]
  79. Sheykhmousa, M.; Mahdianpari, M.; Ghanbari, H.; Mohammadimanesh, F.; Ghamisi, P.; Homayouni, S. Support Vector Machine Versus Random Forest for Remote Sensing Image Classification: A Meta-Analysis and Systematic Review. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2020, 13, 6308–6325. [Google Scholar] [CrossRef]
  80. Vila, M.; Lorber-Pascal, S.; Laurent, F. Phytotoxicity to and Uptake of TNT by Rice. Environ. Geochem. Health 2008, 30, 199–203. [Google Scholar] [CrossRef] [PubMed]
  81. Endo, Y.; Usuki, R.; Kaneda, T. Antioxidant Effects of Chlorophyll and Pheophytin on the Autoxidation of Oils in the Dark. II. The Mechanism of Antioxidative Action of Chlorophyll. J. Am. Oil Chem. Soc. 1985, 62, 1387–1390. [Google Scholar] [CrossRef]
  82. Hsu, C.-Y.; Chao, P.-Y.; Hu, S.-P.; Yang, C.-M. The Antioxidant and Free Radical Scavenging Activities of Chlorophylls and Pheophytins. Food Nutr. Sci. 2013, 4, 1–8. [Google Scholar] [CrossRef]
  83. Wang, E.; Wink, M. Chlorophyll Enhances Oxidative Stress Tolerance in Caenorhabditis Elegans and Extends Its Lifespan. PeerJ 2016, 4, e1879. [Google Scholar] [CrossRef]
  84. Just, C.L.; Schnoor, J.L. Phytophotolysis of Hexahydro-1, 3, 5-Trinitro-1, 3, 5-Triazine (RDX) in Leaves of Reed Canary Grass. Environ. Sci. Technol. 2004, 38, 290–295. [Google Scholar] [CrossRef]
  85. Sharma, P.; Jha, A.B.; Dubey, R.S.; Pessarakli, M. Reactive Oxygen Species, Oxidative Damage, and Antioxidative Defense Mechanism in Plants under Stressful Conditions. J. Bot. 2012, 2012, 217037. [Google Scholar] [CrossRef]
  86. Weisz, D.A.; Gross, M.L.; Pakrasi, H.B. The Use of Advanced Mass Spectrometry to Dissect the Life-Cycle of Photosystem II. Front. Plant Sci. 2016, 7, 617. [Google Scholar] [CrossRef]
  87. Pospíšil, P. Production of Reactive Oxygen Species by Photosystem II as a Response to Light and Temperature Stress. Front. Plant Sci. 2016, 7, 1950. [Google Scholar] [CrossRef]
  88. Aarti, P.D.; Tanaka, R.; Tanaka, A. Effects of Oxidative Stress on Chlorophyll Biosynthesis in Cucumber (Cucumis Sativus) Cotyledons. Physiol. Plant 2006, 128, 186–197. [Google Scholar] [CrossRef]
  89. Buschmann, C.; Lichtenthaler, H.K. Correlation of Reflectance and Chlorophyll Fluorescence Signatures of Healthy and Damaged Forest Trees. In Proceedings of the IGARSS 1988 Symposium, Edinburgh, UK, 12–16 September 1988; ESA Publications Division: Noordwijk, The Netherlands, 1988. [Google Scholar]
  90. Huete, A.R. Remote Sensing for Environmental Monitoring. In Environmental Monitoring and Characterization; Academic Press: Cambridge, MA, USA, 2004; pp. 183–206. [Google Scholar] [CrossRef]
  91. Carter, G.A. Responses of Leaf Spectral Reflectance to Plant Stress. Am. J. Bot. 1993, 80, 239–243. [Google Scholar] [CrossRef]
  92. Zhang, L.; Jin, J.; Wang, L.; Rehman, T.U.; Gee, M.T. Elimination of Leaf Angle Impacts on Plant Reflectance Spectra Using Fusion of Hyperspectral Images and 3D Point Clouds. Sensors 2023, 23, 44. [Google Scholar] [CrossRef] [PubMed]
  93. Lowe, A.; Harrison, N.; French, A.P. Hyperspectral Image Analysis Techniques for the Detection and Classification of the Early Onset of Plant Disease and Stress. Plant Methods 2017, 13, 80. [Google Scholar] [CrossRef] [PubMed]
Remotesensing 17 00385 g001
Figure 1. Workflow diagram for employing two hybrids of Zea mays L. planted in outdoor mesocosms to detect presence of RDX in soil.
Figure 1. Workflow diagram for employing two hybrids of Zea mays L. planted in outdoor mesocosms to detect presence of RDX in soil.
Remotesensing 17 00385 g001
Remotesensing 17 00385 g002
Figure 2. Aerial photograph of field experimental setup. TNT groups had poor, inadequate germination and could not be used in this study. The portable greenhouse was moved from over the plants during data collection.
Figure 2. Aerial photograph of field experimental setup. TNT groups had poor, inadequate germination and could not be used in this study. The portable greenhouse was moved from over the plants during data collection.
Remotesensing 17 00385 g002
Remotesensing 17 00385 g003
Figure 3. Instrumentation for data collection: (a) Headwall Nano-Hyperspec imager. (b) aerial image of Nano-Hyperspec imager mounted to DJI M600 Pro UAV and Ronin-MX gimbal.
Figure 3. Instrumentation for data collection: (a) Headwall Nano-Hyperspec imager. (b) aerial image of Nano-Hyperspec imager mounted to DJI M600 Pro UAV and Ronin-MX gimbal.
Remotesensing 17 00385 g003
Remotesensing 17 00385 g004
Figure 4. Boxplots of (i) P0688 plant height and (ii) P2088 plant height, and Normalized Difference Vegetation Index (NDVI) values for (iii) P0688 and (iv) P2088. Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Figure 4. Boxplots of (i) P0688 plant height and (ii) P2088 plant height, and Normalized Difference Vegetation Index (NDVI) values for (iii) P0688 and (iv) P2088. Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Remotesensing 17 00385 g004
Remotesensing 17 00385 g005
Figure 5. Boxplots of leaf-level chlorophyll measured from (i) P0688 and (ii) P2088, and the Green Chlorophyll Index (GCI) extracted from (iii) P0688 and (iv). Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Figure 5. Boxplots of leaf-level chlorophyll measured from (i) P0688 and (ii) P2088, and the Green Chlorophyll Index (GCI) extracted from (iii) P0688 and (iv). Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Remotesensing 17 00385 g005
Remotesensing 17 00385 g006
Figure 6. Boxplots of quantum yield (Qy) of (i) P0688 and (ii) P2088, and the Vogelmann Red Edge Index 1 (VREI1) calculated for (iii) P0688 and (iv) P2088. Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Figure 6. Boxplots of quantum yield (Qy) of (i) P0688 and (ii) P2088, and the Vogelmann Red Edge Index 1 (VREI1) calculated for (iii) P0688 and (iv) P2088. Lowercase letters indicate statistically significant differences between groups, based on a Holm-Bonferroni pairwise comparison. Treatments sharing the same letter are not significantly different from each other, while groups with different letters show a significant difference.
Remotesensing 17 00385 g006
Remotesensing 17 00385 g007
Figure 7. Average reflectance values by treatment for hybrids (i) P0688 and (ii) P2088. Red segments indicate significantly different wavelengths from those of the control. Note: spectra may be entirely red if the full range differed from control.
Figure 7. Average reflectance values by treatment for hybrids (i) P0688 and (ii) P2088. Red segments indicate significantly different wavelengths from those of the control. Note: spectra may be entirely red if the full range differed from control.
Remotesensing 17 00385 g007
Remotesensing 17 00385 g008
Figure 8. Precision vs. recall scatterplots of SVM, RF, and LDA from (i) 80% training and (ii) 20% test hyperspectral data and (iii) 80% training and (iv) 20% test vegetation index data. A table of the values can be found in Supplementary Information.
Figure 8. Precision vs. recall scatterplots of SVM, RF, and LDA from (i) 80% training and (ii) 20% test hyperspectral data and (iii) 80% training and (iv) 20% test vegetation index data. A table of the values can be found in Supplementary Information.
Remotesensing 17 00385 g008
Table 1. Vegetation indices calculated on image dices using ENVI.
Table 1. Vegetation indices calculated on image dices using ENVI.
IndexRelated toFormulaReference
Anthocyanin Reflectance Index 1 (ARI1)Anthocyanin Content(1/R550) − (1/R700)Gitelson et al., 2001 [46]
Anthocyanin Reflectance Index 2 (ARI2)Anthocyanin ContentR800 [(1/R550) − (1/R700)]Gitelson et al., 2001 [46]
Carotenoid Reflectance
Index 1 (CRI1)		Carotenoid Content(1/R510) − (1/R550)Gitelson et al., 2002 [47]
Carotenoid Reflectance
Index 2 (CRI2)		Carotenoid Content(1/R510) − (1/R700)Gitelson et al., 2002 [47]
Difference Vegetation
Index (DVI)		BiomassRNIR − RRedTucker, 1979 [48]
Enhanced Vegetation
Index (EVI)		Biomass2.5 [(RNIR − RRed)/(RNIR + 6 × RRed − 7.5 × RBlue − 1)]Huete et al., 2002 [49]
Green Chlorophyll Index (GCI)Chlorophyll content(RNIR/RGreen) − 1Gitelson et al., 2003 [50]
Green Difference
Vegetation Index (GDVI)		BiomassRNIR − RGreenSripada, 2005 [51]
Green Leaf Index (GLI)Biomass[(RGreen − RRed) + (RGreen − RBlue)]/[(2 × RGreen) + RRed + RBlue]Louhaichi et al., 2001 [52]
Green Normalized Difference Vegetation Index (GNDVI)Biomass(RNIR − RGreen)/(RNIR + RGreen)Gitelson & Merzlyak, 1997 [53]
Green Optimized Soil
Adjusted Vegetation Index (GOSAVI)		Biomass(RNIR − RGreen)/(RNIR + RGreen + 0.16)Sripada, 2005 [51]
Green Ratio Vegetation Index (GRVI)PhotosynthesisRNIR/RGreenSripada, 2006 [54]
Green Soil Adjusted
Vegetation Index (GSAVI)		Biomass1.5(RNIR − RGreen)/(RNIR + RGreen + 0.5)Sripada, 2005 [51]
Infrared Percentage
Vegetation Index (IPVI)		BiomassRNIR/(RNIR + RRed)Crippen, 1990 [55]
Leaf Area Index (LAI)Biomass3.618 × EVI − 0.118Boegh et al., 2002 [56]
Modified Chlorophyll
Absorption in Reflectance Index (MCARI)		Chlorophyll[(R700 − R670) − 0.2(R700 − R550)] × (R700/R670)Daughtry et al., 2000 [57]
MCARI2LAI1.5 [2.5(R800 − R670) − 1.3(R800 − R550)]/SQRT[(2 × R800 + 1)2 − (6 × R800 − 5 × SQRT(R670)) − 0.5]Haboudane et al., 2004 [58]
Modified Red Edge Normalized Difference Vegetation Index (MRENDVI)Biomass(R750 − R705)/(R750 + R705 − 2 × R445)Datt, 1999 [59], Sims & Gamon, 2002 [60]
Modified Red Edge
Simple Ratio (MRESR)		Biomass(R750 − R445)/(R705 − R445)Datt, 1999 [60], Sims & Gamon, 2002 [60]
Modified Triangular
Vegetation Index (MTVI)		Biomass1.2 [1.2(R800 − R550) − 2.5(R670 − R550)]Haboudane et al., 2004 [58]
MTVI2LAI1.5 [1.2(R800 − R550) − 2.5(R670 − R550)]/SQRT[(2 × R800 + 1)2 − (6 * R800 − 5 * SQRT(R670)) − 0.5]Haboudane et al., 2004 [58]
Normalized Difference
Vegetation Index (NDVI)		Biomass(R800 − R670)/(R800 + R670)Rouse et al., 1973 [61]
Optimized Soil Adjusted
Vegetation Index (OSAVI)		Biomass(RNIR − RRed)/(RNIR + RRed + 0.16)Rondeaux et al., 1996 [62]
Photochemical Reflectance
Index (PRI)		 Carotenoids chlorophyll (R531 − R570)/(R531 + R570)Peñuelas et al., 1995 [63],
Gamon et al., 1997 [64]
Plant Senescence
Reflectance Index (PSRI)		Stress(R680 − R500)/R750Merzlyak et al., 1999 [65]
Renormalized Difference
Vegetation Index (RDVI)		Biomass(RNIR − RRed)/SQRT(RNIR + RRed)Roujean & Breon, 1995 [66]
Red Edge Normalized
Difference Vegetation
Index (RENDVI)		Biomass(R750 − R705)/(R750 + R705)Gitelson & Merzlyak, 1994 [67], Sims & Gamon, 2002 [60]
Red Edge Position Index (REPI)StressMax d/dR690–740Curran et al., 1995 [68]
Soil Adjusted Vegetation Index (SAVI)Biomass1.5(RNIR − RRed)/(RNIR + RGreen + 0.5)Huete, 1988 [69]
Simple Ratio (SR)BiomassRNIR/RRedBirth & McVey, 1968 [70]
Transformed Chlorophyll
Absorption Reflectance Index (TCARI)		Chlorophyll content3 [(R700 − R670) − 0.2(R700 − R550)(R700/R670)]Haboudane et al., 2004 [58]
Triangular Vegetation
Index (TVI)		LAI[120(R750 − R550) − 200(R670 − R550)]/2Broge & Leblanc, 2000 [71]
Vogelmann Red Edge
Index 1 (VREI1)		StressR740/R720Vogelmann et al., 1993 [72]
Vogelmann Red Edge
Index 2 (VREI2)		Stress(R734 − R747)/(R715 − R726)Vogelmann et al., 1993 [72]
Table 2. Number of image dices by group used for spectra and reflectance index extraction.
Table 2. Number of image dices by group used for spectra and reflectance index extraction.
nControlDroughtRDX250RDX500
P0688908856349641
P2088534467446409
Table 3. Model parameters in R using the caret package.
Table 3. Model parameters in R using the caret package.
ModelParameters
SVMCross-validation: method = “repeatedcv”, number = 10, repeats = 3, sampling = “down”; Training: method = “svmPoly”, metric = “Accuracy”
RFCross-validation: method = “cv”, number = 2, verboseIter = TRUE, returnResamp = “all”, classProbs = TRUE, summaryFunction = multiClassSummary, sampling = “down”; Training: method = “rf”, tuneLength = 2, metric = “Accuracy”, type = “Classification”
LDACross-validation: method = “repeatedcv”, number = 10, repeats = 3, sampling = “down”; Training: method = “lda”
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Manley, P.V., II; Via, S.M.; Burken, J.G. UAV-Based Phytoforensics: Hyperspectral Image Analysis to Remotely Detect Explosives Using Maize (Zea mays). Remote Sens. 2025, 17, 385. https://doi.org/10.3390/rs17030385

AMA Style

Manley PV II, Via SM, Burken JG. UAV-Based Phytoforensics: Hyperspectral Image Analysis to Remotely Detect Explosives Using Maize (Zea mays). Remote Sensing. 2025; 17(3):385. https://doi.org/10.3390/rs17030385

Chicago/Turabian Style

Manley, Paul V., II, Stephen M. Via, and Joel G. Burken. 2025. "UAV-Based Phytoforensics: Hyperspectral Image Analysis to Remotely Detect Explosives Using Maize (Zea mays)" Remote Sensing 17, no. 3: 385. https://doi.org/10.3390/rs17030385

APA Style

Manley, P. V., II, Via, S. M., & Burken, J. G. (2025). UAV-Based Phytoforensics: Hyperspectral Image Analysis to Remotely Detect Explosives Using Maize (Zea mays). Remote Sensing, 17(3), 385. https://doi.org/10.3390/rs17030385

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop