Evaluation of Multilevel Thresholding in Differentiating Various Small-Scale Crops Based on UAV Multispectral Imagery

Abstract

Differentiation of various crops in small-scale crops is important for food security and economic development in many rural communities. Despite being the oldest and simplest classification technique, thresholding continues to gain popularity for classifying complex images. This study aimed to evaluate the effectiveness of a multilevel thresholding technique in differentiating various crop types in small-scale farms. Three (3) types of crops were identified in the study area, and these were cabbage, maize, and sugar bean. Analytical Spectral Devices (ASD) spectral reflectance data were used to detect subtle differences in the spectral reflectance of crops. Analysis of ASD reflectance data revealed reflectance disparities among the surveyed crops in the Green, red, near-infrared (NIR), and shortwave infrared (SWIR) wavelengths. The ASD reflectance data in the Green, red, and NIR were then used to define thresholds for different crop types. The multilevel thresholding technique was used to classify the surveyed crops on the unmanned aerial vehicle (UAV) imagery, using the defined thresholds as input. Three (3) other machine learning classification techniques were also used to offer a baseline for evaluating the performance of the MLT approach, and these were the multilayer perceptron (MLP) neural network, radial basis function neural network (RBFNN), and the Kohonen’s self-organizing maps (SOM). An analysis of crop cover patterns revealed variations in crop area cover as predicted by the MLT and selected machine learning techniques. The classification results of the surveyed crops revealed the area covered by cabbage crops to be 7.46%, 6.01%, 10.33%, 7.05%, 9.48%, and 7.04% as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM, respectively. The area covered by maize crops as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM were noted to be 13.62%, 26.41%, 12.12%, 11.03%, 12.19% and 15.11%, respectively. Sugar bean was noted to occupy 57.51%, 43.72%, 26.77%, 27.44%, 24.15%, and 16.33% as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM, respectively. Accuracy assessment results generally showed poor crop pattern prediction with all tested classifiers in categorizing the surveyed crops, with the kappa index of agreement (KIA) values of 0.372, 0.307, 0.488, 0.531, 0.616, and 0.659 for the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and Kohonen’s SOM, respectively. Despite recommendations by recent studies, we noted that the MLT was noted to be unsuitable for classifying complex features such as spectrally overlapping crops.

1. Introduction

The spatial configuration of different types of crops in small-scale farms is a pre-requisite to establishing an efficient approach for sustainable crop production [1,2]. Remote sensing technology offers the prospect of spatially characterizing different crops [3,4]. From a remote sensing perspective, crops are characterized based on their spectral reflectance behavior across several channels of the electromagnetic spectrum. From an optical remote sensing perspective, changes in pixel values across different image bands become the basis for distinguishing a particular type of crop from others in that imagery [5]. Therefore, satellite sensors such as the Moderate Resolution Imaging Spectroradiometer (MODIS) [6], Landsat series [6,7], Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) [8], Satellite Pour l’Observation de la Terre (SPOT) [9], and Sentinel-2 [10] have become the basis for identifying various types of crops due to their ease of access. However, the accurate identification of various types of crops in small-scale farms, which are characteristically less than 10 m wide, using these sensors remains a challenge due to their spatial resolutions, which are lower than individual small-scale farm plots [11].

The advent of unmanned aerial vehicle (UAV) technology has marked an enticing dawn for characterizing small-scale crops due to the ultra-high spatial resolution associated with UAV sensors [12,13,14,15]. Different sensors, such as RGB, multispectral, hyperspectral, or thermal cameras, can be mounted on UAV platforms, with a view matching the specifications of crop variables under investigation. Furthermore, UAV data are less susceptible to atmospheric errors [16]; they recorded the reflected radiation when situated below the atmosphere. However, these ultra-high spatial resolution images are often subjected to noise due to increased discernable features [17]. Though many UAV systems have a higher radiometric resolution (i.e., 32-bits), their coarse spectral resolution makes it difficult to detect crop variations outside their resolution. Moreover, the intricacy of data in UAV imagery upsurges due to the enhancement of sensor radiometric resolution [18]. This further complicates the differentiation of crops, as variations in leaf positions may create an overlapping reflectance intensity across different crops. This spectral entropy is even more challenging when a sensor with a few spectral bands is used. For this reason, the attainment of the accurate classification of crops subjected to these conditions becomes a challenge. Although classification algorithms, such as the maximum likelihood [19,20], minimum distance [21], support vector machine [22], K-Nearest Neighbor [23], and Random Forest [24], were noted to be effective in characterizing various crops based on UAV imagery, these approaches operate under the notion of a crisp clustering, where each pixel spectral property belongs to a single crop type [25]. Moreover, these classifiers, which often require training data to learn patterns, generally classify crops based on maximum accuracy, which often results in the biased categorization of similar pixels towards the majority class and misclassification of the minority class.

Several studies also employed the spectral unmixing approach as a way of improving crop classification accuracy [26,27,28,29,30]. In this approach, the probability of the endmember to belong to any pixel is constructed and subsequently fused with the spectral linear model and classifier to compute crop cover [31]. This is performed under the notion that the heterogeneous land features in the imagery become more apparent as the spatial resolution of the image improves [32]. Although this approach was devised to overcome the shortfalls brought of the empirical method for characterizing crops over [33], it works under the supposition that the spectral reflectance of a particular pixel is a result of the sum of the spectral reflectance of the numerous discernible features within that pixel [34]. Overall, this approach only aims to resolve classification problems that are related to different land cover types occupying the same pixel and not linearly inseparable spectral features. The spectral similarity and difference metric (SSDM) technique was also recommended to minimize land cover misclassification problems [35]. This approach is often used to characterize land features based on the interclass spectral separability or intraclass spectral variability [36]. Moreover, the advent of advanced pattern recognition techniques, such as deep learning, has seen improvement in crop characterization. In comparison with the conventional classification techniques mentioned, deep learning techniques have been shown to be capable of learning linearly inseparable data [37]. Deep learning techniques such as the multilayer perceptron (MLP) [38], radial basis function (RBF) [39], and the Kohonen’s self-organizing map (SOM) [40] have evidently been applied in pattern recognition.

Despite being the ancient image classification technique [41] and the recent developments in pattern recognition through the application of machine learning and deep learning [42,43], recent studies are still emphasizing the significance of thresholding (rule-based classifier) for segmenting complex features [44,45,46,47]. Multilevel thresholding (MLT), in particular, was proposed to partition images into multiple classes [44]. Despite its simplicity, Kumar et al. [48] noted that the multilevel thresholding technique can segment a complex crop image. The advantage of using this approach was based on its ability to use dynamic rules and that it does not rely on human supervision and training data [49]. This technique achieves the partitioning of images by employing either the Otsu class variance method [49] or the Kapur maximum entropy method [50]. Whereas the Otsu thresholding approach uses the maximum variance of classes to partition the image [51], the Kapur approach enhances the maximum entropy of the image histogram to evaluate the uniformity among various classes and determine the best threshold values [52]. When using other classification approaches, Akgün et al. [53] noted that all pixels in the image tend to be allocated to any of the existing classes, unless the thresholding technique is used or an additional class denoting non-crop features is added, in cases of the classification of crop types. Despite its simplicity, flexibility, and global search capabilities [54], studies also revealed that the MLT classifier can achieve the accurate classification of complex features [55,56,57]. This underscores the necessity to evaluate the efficacy of this approach in spectrally differentiating various types of crops. Moreover, we have not found studies that evaluated the efficacy of the multilevel thresholding classifier in differentiating various types of crops. While hyperspectral remote sensing offers a great deal of spectral information pertaining to crop types, the need to spatially partition the surveyed crops in the study area prompted the integration of UAV multispectral imagery and non-imaging field spectrometric data.

2. Materials and Methods

2.1. Study Area

The small-scale farms in which the current study was conducted are found in the Mutale River catchment, situated in the Vhembe District of the Limpopo Province of South Africa. This area is dominated by small-scale farmers who practice substantial farming. The study area is located at the 22°47′41.95″ S, 30°29′08.21″ E and 22°47′58.61″ S, 30°29′19.83″ E grid reference location, at an elevation which ranges between 620 m and 660 m above sea level. The presence of various small-scale crop types that are cultivated under the irrigation system prompted the selection of the study area. In this area, farmers tend to cultivate a variety of crops irrespective of the season. Some of the crops cultivated in the study area include maize, cabbage, sweet potatoes, and sugar bean [11]. Figure 1 presents the location of the study site in South Africa.

2.2. Data Acquisition

The following types of data were collected and utilized to attain the main purpose of the current study.

2.2.1. UAV Multispectral Imagery

The UAV multispectral imagery used in this study was acquired using the MicaSense Red Edge MX imagining sensor, embedded on the DJI Matrice 600 Pro supplied by the SZ DJI Technology Co., Ltd., Shenzhen, China. The focal length of the Red Edge sensor used was 5.5 mm. With the DJI Matrice 600 Pro operating at 120 m altitude, the MicaSense Red Edge-MX imaging sensor captured multispectral imagery of a flat cropland at a spatial resolution of 8.7 cm. This UAV imaging system enabled rapid acquisition of crop data, offering aerial access to inaccessible areas while covering a larger area. The side lap and end lap of 75% was considered during landscape imaging. The maximum distance covered was approximately 15 hectares. Clear sky, wind speed of at most 10 m/s, and midday (i.e., between 12h00 and 14h00 pm) were considered viable environmental conditions for imagery acquisition. Prior to image acquisition, MicaSense Red Edge imaging sensor was calibrated using the Calibrated Reflectance Panel (CRP) placed on flat ground in an open area from about 1 m in height.

Table 1. The UAV spectral band details used to select spectral wavelengths from spectrometric data.

Band Name	Center (µm)	Wavelength Range (µm)
Blue	0.475	0.443–0.507
Green	0.560	0.533–0.587
Red	0.668	0.654–0.682
Red Edge	0.717	0.705–0.729
Near-IR	0.842	0.785–0.899

provides the spectral wavelength center and range for the UAV multispectral sensor used in this study.

2.2.2. Field Spectrometry Data

Analytical Spectral Devices (ASD) FieldSpec 4 Hi-Res Spectroradiometer was used to obtain hyperspectral reflectance data crops under investigation in the study area. This non-imaging system acquired spectral reflectance data for crops using the visible, NIR, and SWIR spectral wavelengths. Prior to collecting data, the ASD Fieldspec4 Spectroradiometer warmed up for 30 min. Moreover, calibration of the ASD FieldSpec4 device was performed after each crop sampling to ensure data quality. The distance of 50 cm was maintained between the light source and the spectral gun during spectral reflectance data collection. A total of five (5) spectral reflectance samples were collected from each sample leaf for each crop type. The centimeter level precision Ashtech^®ProMark2™ Global Positioning System (GPS) device was also used to record the absolute locations where the spectral reflectance samples were collected.

2.3. UAV Imagery Preparation

Prior to image analysis, we ortho-mosaicked the UAV image tiles to cover the study area. This was guided by six (6) ground control points (GCPs) established prior to flying the UAV. We used attached visible white papers in the survey poles and recorded their locations using the GPS. Geometric correction was carried out to assign spatial reference properties to UAV imagery. In this case, the UAV imagery was spatially referenced to the Universal Transverse Mercator (UTM) Zone 36S, based on the World Geodetic System of the year 1984 (WGS84) spheroid. This was carried out using the ArcGIS Drone2Map program supplied by the ESRI South Africa, Midrand, South Africa. The UAV multispectral bands were also radiometrically corrected using the “Radiance” tool in the TerrSet 18.31 Geospatial Modeling and Monitoring program supplied by the Clark Labs, Clark University, Worcester, United States of America. In this study, the $L_{min}/L_{max}$ option was used to convert $Dn$ to radiance using Equation (1):

$$\mathrm{R}\mathrm{a}\mathrm{d}\mathrm{i}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}={\mathrm{L}}_{\mathrm{m}\mathrm{i}\mathrm{n}}+\left[{\displaystyle \frac{\mathrm{D}\mathrm{n}}{\max \mathrm{D}\mathrm{n}}}\ast \left({\mathrm{L}}_{\mathrm{m}\mathrm{a}\mathrm{x}}-{\mathrm{L}}_{\mathrm{m}\mathrm{i}\mathrm{n}}\right)\right]$$

(1)

where $L_{min}$ and $L_{max}$ denote minimum and maximum $Dn$ values, respectively, specified in, i.e., milliWatts per square centimeter per steradian per micron ($\mathrm{m}\mathrm{W}\mathrm{c}{\mathrm{m}}^{-2}\mathrm{s}{\mathrm{r}}^{-1}\mathrm{u}{\mathrm{m}}^{-1})$.

2.4. Assessing Variability in the UAV Spectral Radiance Properties of Crops

A total of three (3) point-based GIS vector layers were created. Subsequently, a total of 200 points were randomly digitized in the vicinity of each type of crop, guided by the RGB image and field knowledge. The digitized points were overlaid on each UAV spectral band and the pixel values on which the points were overlaid were extracted using the “Extract multi-values to points” tool embedded in the Spatial Analysis Tools of the ArcMap 10.8.2 Software package. The point attribute table containing the pixel values for each crop type was exported to Microsoft Excel format for crop spectral radiance and spectral vegetation index values analysis. Variability in response of crops to radiation of different spectra was achieved by using Equation (2), adopted from Levene [58]:

$$W={\displaystyle \frac{\left[\left(N-k\right){\sum }_{i=1}^k{N}_i{\left({\overline{Z}}_i-\overline{Z}\right)}^2{n}_i\right]}{\left[\left(k-1\right){\sum }_{i=1}^k{\sum }_{j=1}^{n_i}{\left(Z_{ij}-{\overline{Z}}_i \right)}^2\right]}}$$

(2)

where $N$ represents the population size; $Z_{ij}=\left|Y_{ij}-{Ȳ}_i\right|$ when $Y_{ij}$ and ${Ȳ}_i$ denote the $Y$ value and the mean for the observation of the $i-th$ subgroup; ${\overline{Z}}_i$ represents the mean of the group $Z_{ij}$; and $\overline{Z}$ represents the overall mean of the $Z_{ij}$.

Consequently, the computed p-value was evaluated to explain the variations in the interpolated UAV spectral radiance properties across crop types.

2.5. Spectral Profiling of UAV Spectral Radiance Properties of Crops

The spectral radiance properties of crops were plotted with a view to identify the UAV spectral band that was suitable for differentiating crops. In this case, the mean radiance values computed using Levene’s k-comparison of equal variance statistical technique was used. The linear graph was plotted to show the spectral radiance behavior of crops.

2.6. Crop Spectral Reflectance Thresholding Selection

In our study, we proposed the thresholding optimization approach, which relied on ground spectral reflectance data between adjacent crop types. Their threshold optimization approach considered descriptive statistics, i.e., minimum and maximum reflectance of each crop type. We proposed Equation (3) to optimize upper and lower thresholds for crop types, such that

$$T_o=\left({\displaystyle \frac{\left(T_{i,u}+T_{i,l}\right)}{2}}\right)$$

(3)

where $T_o$ denotes the optimized spectral band threshold value between crop types; $X_{i_{min}}$ denotes the lower threshold value of the upper class; and $X_{i_{max}}$ denotes the upper threshold value of the lower class.

Crop reflectance data for each wavelength is first arranged in ascending order ($x_1, x_2, x_3,\dots , x_n$). In our study, we adopted the interquartile range threshold selection approach proposed by Tukey et al. [59]:

$$T_{i,l}=Q_{i, l}-c\times IQ{R}_i$$

(4)

$$T_{i,u}=Q_{i,u}+c\times IQ{R}_i$$

(5)

where $IQ{R}_i$ denotes the interquartile distance, computed using Equation (6), and $c$ is the control parameter decided by the user. Bardet and Dimby [60] recommended the use of a 1.5 factor in order to ensure that the probability of a Gaussian random variable being classified as an outlier is approximately 0.07:

$$IQ{R}_i=Q_{i,u}-Q_{i,l}$$

(6)

where $Q_{i,u}$ denotes the upper quartile value of each crop reflectance data, and $Q_{i,l}$ denotes the lower quartile value for each crop reflectance data. $Q_{i,u}$ and $Q_{i,l}$were computed using Equations (7) and (8):

$$Q_{i,u}=x_i\left(p_{i,u}\right)$$

(7)

$$Q_{i,l}=x_i\left(p_{i,l}\right)$$

(8)

$P_{i,u}$ denotes the position of the upper quartile, computed using Equation (9), and $p_{i,l}$ denotes the position of the lower quartile, computed using Equation (10):

$$p_{i,u}={\displaystyle \frac{3\times (n_i+1)}{4}}$$

(9)

$$p_{i,l}={\displaystyle \frac{n_i+1}{4}}$$

(10)

where $n$ denotes the number of samples.

2.7. Multilevel Thresholding of Crop Types

The optimized minimum and maximum thresholds were used as the input data into the multilevel thresholding in differentiating crop types. The multilevel thresholding was then applied on both the corresponding UAV spectral bands, Equation (11), adopted from Jiang et al. [61]:

$$I^T\left(x,y\right)=\left\{\begin{array}{c}{C}_1, if T_o\le I\left(x,y\right) \le T_{o_1}, \\ C_2, if T_{o_1}\le I\left(x,y\right)\le T_{o_2}, \\ C_3 if T_{o_2}\le I\left(x,y\right)\le T_{o_3}, \\ C_4 if T_{o_3}\le I\left(x,y\right)\le T_{o_4} \\ 0, otherwise \end{array}\right.$$

(11)

where $I^T$ denotes the pixel in the $x, y$ location; $C$denotes crop class; and $T$ denotes the minimum and maximum threshold values for each type of crop, respectively.

2.8. Classification of Crops Using Machine Learning Algorithms

In our study, we also selected three machine learning algorithms, viz., MLP, RBFNN, and SOM, to differentiate crop types, with a view that evaluates the performance of the MLT technique against the selected machine learning algorithms. Prior to machine learning model training, training sites were created for the algorithms to analyze in order to learn patterns and make predictions.

Table 2. Sequences followed by MLP, RBFNN, and SOM in differentiating crop types.

MLP		RBFNN		SOM
We applied the MLP to differentiate crop types as follows: Network topology: We trained the MLP for crop type characterization using five (5) UAV spectral bands as input layer nodes and two (2) hidden layers, each of which had five (5) nodes, to improve the learning process. Training parameters: We used both automatic training and dynamic learning rates to train models. The learning rate was set to 0.01, with a 0.5 momentum factor and sigmoid constant of 1. Backpropagation training: We trained the MLP for crop type differentiation using Equation (12), adopted from Almeida [62]:		We applied the RBFNN algorithm to differentiate crop types as follows: Basis functions: We used the set of the basis functions (Equation (13)) proposed by Powell [63]:		We applied the SOM algorithm to differentiate crop types as follows:$\mathrm{Coarse} \mathrm{tuning}: \mathrm{We} \mathrm{started} \mathrm{SOM} \mathrm{by} \mathrm{performing} \mathrm{coarse} \mathrm{tuning}, \mathrm{an} \mathrm{unsupervised} \mathrm{classification}, \mathrm{to} \mathrm{achieve} \mathrm{competitive} \mathrm{learning} \mathrm{and} \mathrm{lateral} \mathrm{interaction} \mathrm{neuron} \mathrm{weights} \mathrm{representing} \mathrm{the} \mathrm{underlying} \mathrm{cluster} \mathrm{and} \mathrm{sub}\text{-}\mathrm{cluster} \mathrm{in} \mathrm{input} \mathrm{neurons}. \mathrm{Supposing} x=\left\{x_1,x_2,\dots ,x_n\right\}$ is the n-dimensional feature of SOM, the neuron in the output layer with minimum distance to the input feature vector (known as the winner) is then determined as follows:
$ne{t}_j={\displaystyle \sum _{i=1}^m}{w}_{ij}{O}_i$	(12)	$\phi \left(‖x-x^n‖\right)$	(13)	$Winner=argmin\left({\displaystyle \sum _{i=1}^n}{\left(x_i^t-w_{ji}^t\right)}^2\right)$	(14)
$\mathrm{where} ne{t}_j$$\mathrm{denotes} \mathrm{the} \mathrm{overall} \mathrm{weight} \mathrm{of} \mathrm{the} j\text{-}th$$\mathrm{neuron}; x_i$$\mathrm{denotes} \mathrm{the} \mathrm{input} \mathrm{value} \mathrm{to} \mathrm{the} i\text{-}th$$\mathrm{node} \mathrm{of} \mathrm{the} \mathrm{input} \mathrm{neuron}; w_{ij}$$, \mathrm{computed} \mathrm{using} \mathrm{Equation} \left(15\right), \mathrm{denotes} \mathrm{the} \mathrm{weight} \mathrm{between} \mathrm{the} i\text{-}th$$\mathrm{node} \mathrm{of} \mathrm{input} \mathrm{layer} \mathrm{and} \mathrm{the} j\text{-}th$$\mathrm{node} \mathrm{of} \mathrm{hidden} \mathrm{layer}; \mathrm{and} O_j$$\mathrm{denotes} \mathrm{the} \mathrm{output} \mathrm{from} \mathrm{neuron} j$, computed using Equation (18):		$\mathrm{where} \phi \left(.\right)$$\mathrm{denotes} \mathrm{nonlinear} \mathrm{function}, \mathrm{typically} \mathrm{considered} \mathrm{to} \mathrm{be} \mathrm{Euclidian}, \mathrm{i}.\mathrm{e}., \mathrm{between} x$$\mathrm{and} x^n$. RBFNN training: The training of RBFNN for classifying crops involved two steps. The number of hidden layers were determined through the deployment of an unsupervised k-means classifier, using Equation (16) proposed by Duda and Hart [64]:		$\mathrm{where} x_i^t$$\mathrm{denotes} \mathrm{the} \mathrm{input} \mathrm{to} \mathrm{neuron} i$$\mathrm{at} \mathrm{iteration} t$$; w_{ji}^t$$\mathrm{denotes} \mathrm{the} \mathrm{weight} \mathrm{between} \mathrm{the} \mathrm{input} \mathrm{neuron} i$$\mathrm{and} \mathrm{the} \mathrm{output} \mathrm{neuron} j$.$\mathrm{Whereas} \mathrm{the} \mathrm{weight} \mathrm{outside} \mathrm{the} \mathrm{winner} \mathrm{was} \mathrm{kept} \mathrm{unaltered}, \mathrm{the} \mathrm{weights} \mathrm{of} \mathrm{the} \mathrm{winner} \mathrm{and} \mathrm{its} \mathrm{neighbors} \mathrm{within} \mathrm{radius} \gamma$$\mathrm{were} \mathrm{altered} \mathrm{in} \mathrm{accordance} \mathrm{with} \mathrm{learning} \mathrm{rate} {\propto }^t$, according to Equations (17) and (19), such that
$∆w_{ij(t+1)}=\eta {\delta }_{ji}{O}_i$	(15)	$E\left(c\right)={\displaystyle \sum _i}{\displaystyle \frac{1}{2}}\underset{k}{\mathit{min}}{\left(x_i-c_k\right)}^2$	(16)	$w_{ji}^{t+1}=w_{ji}^t+{\propto }^t\left(x_i^t-w_{ji}^t\right), \forall d_{w_j}\in {\gamma }^t$	(17)
$O_j=f\left(ne{t}_j\right)$	(18)	Then, centers of the RBFs were aligned with the centers of the clusters from the k-means results, using Equations (20) and (23):		$w_{ji}^{t+1}=w_{ji}^t, \forall d_{w_j} \notin {\gamma }^t$	(19)
$\mathrm{where} ne{t}_j$ was computed in Equation (12). The number of hidden layer nodes used in this study were estimated using Equation (13):		${\phi }^2={\displaystyle \frac{1}{p}}{\displaystyle \sum _{k=1}^p}{\left(c_j-c_k\right)}^2$	(20)	$\mathrm{where} {\propto }^t$$\mathrm{denotes} \mathrm{the} \mathrm{learning} \mathrm{rate} \mathrm{at} \mathrm{the} \mathrm{iteration} \mathrm{and} d_{winner}$ and is obtained by deploying Equation (22), adopted from Kohonen [65]:
$N_h=INT\left(\sqrt{N_i\times N_o}\right)$	(21)	$\mathrm{where} p$ denotes the number of radial basis functions:		${\alpha }^t={\alpha }_{max}{\left({\displaystyle \frac{{\alpha }_{min}}{{\alpha }_{max}}}\right)}^{{\displaystyle \frac{1}{t}}}$	(22)
$\mathrm{where} N_h$$\mathrm{denotes} \mathrm{hidden} \mathrm{layer} \mathrm{nodes}; N_i$$\mathrm{denotes} \mathrm{input} \mathrm{layer} \mathrm{nodes}; \mathrm{and} N_0$ denotes output layer nodes.$\mathrm{Learning} \mathrm{error} \mathrm{evaluation}: \mathrm{To} \mathrm{determine} \mathrm{whether} \mathrm{the} \mathrm{learning} \mathrm{process} \mathrm{was} \mathrm{successfully} \mathrm{executed}, \mathrm{the} \mathrm{error} \mathrm{associated} \mathrm{with} \mathrm{the} \mathrm{learning} \mathrm{of} \mathrm{network} \mathrm{was} \mathrm{evaluated} \mathrm{using} \mathrm{the} \mathrm{coefficient} \mathrm{of} \mathrm{determination} (R^2$), such that		$f_n\left(x\right)={\displaystyle \sum _{i=1}^K}{w}_{ni}{\phi }_i\left(x\right)+w_{n0}$	(23)	Fine tuning: We applied fine tuning to optimize the decision boundaries between crop classes based on the training data. We used the learning vector quantization (LVQ) proposed by Kangas et al. [66]. If x is correctly classified, then
$R^2={\displaystyle \frac{{\sum }_{i=1}^n{\left({\widehat{y}}_i-\overline{y}\right)}^2}{{\sum }_{i=1}^n{\left(y_i-\overline{y}\right)}^2}}$	(24)	$\mathrm{where} x$$\mathrm{denotes} \mathrm{input} \mathrm{neurons}; w_{ni}$$\mathrm{denotes} \mathrm{the} \mathrm{connection} \mathrm{between} \mathrm{the} \mathrm{basis} \mathrm{function} \mathrm{and} \mathrm{output} \mathrm{layer}; \mathrm{and} {\phi }_i$ is the radial basis function, computed using Equation (26):		$w_c^{t+1}=w_c^t+{\delta }^t\left(x_i-w_c^t\right)$	(25)
$\mathrm{where} y$$\mathrm{denotes} \mathrm{the} \mathrm{measured} \mathrm{value}; {\widehat{y}}_i$$\mathrm{denotes} \mathrm{predicted} \mathrm{values}; \mathrm{and} \overline{y}$ denotes the mean value of the measured values.		$\phi \left(x\right)=exp\left(-{\displaystyle \frac{{‖x-c‖}^2}{{2\phi }^2}}\right)$	(26)	If x is incorrectly classified, then
		$\mathrm{where} \mathrm{x} \mathrm{is} \mathrm{the} d\text{-}dimensional$$\mathrm{input} \mathrm{vector} \mathrm{with} \mathrm{variable} x_i$$; \mathrm{and} m_j$$\mathrm{denotes} \mathrm{the} \mathrm{vector} \mathrm{which} \mathrm{determines} \mathrm{the} \mathrm{center} \mathrm{of} \mathrm{the} \mathrm{basis} \mathrm{function} w_j$$\mathrm{and} \mathrm{has} \mathrm{variable} m_{ji}$.		$w_c^{t+1}=w_c^t-{\delta }^t\left(x_i-w_c^t\right)$	(27)
				Otherwise
				$w_c^{t+1}=w_c^t, if i\ne c$	(28)
				$\mathrm{where} w_c$$\mathrm{denotes} \mathrm{the} \mathrm{weight} \mathrm{vector} \mathrm{of} \mathrm{the} \mathrm{winner}, \mathrm{and} {\delta }^t$ denotes a gain term, which decreases as time decreases.

explains the sequences employed to achieve crop type differentiation using the selected machine learning algorithms.

2.9. Classification Accuracy Assessment

The performance of thresholding, along with selected machine learning classifiers, was evaluated. In this case, the ground truth image, classified based on GPS points collected in the vicinity of each crop type, was compared with spectrally classified crops. This was achieved using the overall kappa index of agreement (KIA), computed using Equation (29) and adopted from Cohen [67]:

$$k={\displaystyle \frac{p_0-P_e}{1-p_e}}$$

(29)

where $p_0$ denotes the proportion of pixels assigned to the correct class, and $p_e$ denotes the proportion of pixels expected to be assigned to the correct class. A KIA value ranges from 0 to 1, with the value closer to 0 indicating poor classification performance, and the value closer to 1 indicating accurate classification results.

3. Results

3.1. Identification of Crops Cultivated in the Study Site

During the field observation of the study area, a total of three (3) types of crops were identified in the study area, and these were cabbage, maize, and sugar bean. Figure 2 shows various crop types that were identified in the study area, as imaged by the UAV system.

3.2. Spectral Characterization of Crops from UAV Imagery

Levene’s k-comparison of equal variance statistics results for analyzing variability in crop patterns imaged using a UAV multispectral sensor are provided in

Table 3. Levene’s k-comparison of equal variance statistics for crop radiance variability obtained from UAV imagery.

	Blue			Green			Red
	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean
N	200	200	200	200	200	200	200	200	200
α	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05
μ	0.34	0.45	0.36	0.42	0.62	0.49	0.25	0.33	0.25
σ	0.11	0.14	0.09	0.12	0.11	0.07	0.09	0.06	0.08
Var.	0.01	0.02	0.1	0.01	0.01	0.01	002	0.02	0.01
Min.	0.23	0.089	0.247	0.209	0.164	0.319	0.166	0.141	0.113
Max.	0.855	0.376	0.413	0.966	0.833	0.811	0.617	0.792	0.327
p-value	<0.001			<0.001			<0.001
	Red edge			NIR
	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean
N	200	200	200	200	200	200
α	0.05	0.05	0.05	0.05	0.05	0.05
μ	0.55	063	0.67	0.72	0.70	0.68
σ	0.11	0.09	0.12	0.06	0.11	0.11
Var.	0.01	0.01	0.01	0.01	0.01	0.01
Min.	0.37	0.15	0.46	0.41	0.341	0.46
Max.	0.94	0.88	0.89	0.93	0.91	0.92
p-value	<0.001			<0.001

.

From Table 3, the UAV spectral radiance properties were noted to vary across the surveyed crops. At N = 200, $\alpha$ = 0.05, Levene’s k-comparison test produces the p-value < 0.001, which was less than 0.05 significance alpha, signifying variability in the ASD spectral reflectance properties across the surveyed crops.

3.3. Hyperspectral Reflectance Patterns of the Identified Crops

In this study, field spectrometric data were used to analyze spectral distinction across the surveyed crop types in the study area. Figure 3 presents the reflectance spectrum of the surveyed crops, i.e., cabbage, maize, and sugar bean.

From Figure 3a, cabbage showed a higher reflectance in the visible–NIR spectral region when compared with other surveyed crops, with sugar bean showing a high reflectance than other crops in the SWIR spectral region. From Figure 3b, clear reflectance disparities were noted in the four labeled spectral regions (i.e., a–c in Figure 3b). By implication, reflectance data generated from these spectral regions could be ideal for thresholding the surveyed crops.

3.4. Reflectance Behavior of Maize Cultivars in ASD Spectral Regions Corresponding to the UAV Spectral Wavelengths

The hyperspectral reflectance curves for the surveyed crops were analyzed, with specific focus on the wavelengths corresponding to the wavelengths of UAV spectral bands. Figure 4 provides the reflectance patterns of the surveyed crops in the spectral regions corresponding to wavelengths of the UAV spectral bands.

Analysis of spectral reflectance properties of the surveyed crops in spectral wavelengths corresponding to UAV spectral bands showed that these crops can be spectrally differentiated in the Green and NIR spectral wavelengths. In the Blue spectral band, only cabbage could clearly be identified due to its higher reflectance compared to other surveyed crops, while maize had a similar reflectance with sugar bean (Figure 4).

Moreover, Levene’s k-comparison of equal variance was computed for crop reflectance data acquired from the ASD spectral wavelengths corresponding to UAV spectral bands (see

Table 4. Levene’s k-comparison of equal variance results for ASD spectral variance across crops.

	0.443 µm–0.507 µm			0.533 µm–0.587 µm			0.654 µm–0.682 µm
	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean
N	100	100	100	100	100	100	100	100	100
α	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05	0.05
μ	0.38	0.48	0.34	0.39	0.59	0.55	0.22	0.37	0.2
σ	0.12	0.12	0.05	0.09	0.12	0.09	0.06	0.12	0.03
Var.	0.01	0.01	0	0.01	0.01	0.01	0	0.02	0
Min.	0.2	0.21	0.23	0.22	0.36	0.36	0.12	0.16	0.12
Max.	0.85	0.83	0.47	0.98	0.87	0.85	0.58	0.8	0.34
p-value	<0.001			<0.001			<0.001
	0.705 µm–0.729 µm			0.785 µm–0.899 µm
	Cabbage	Maize	Sug. bean	Cabbage	Maize	Sug. bean
N	100	100	100	100	100	100
α	0.05	0.05	0.05	0.05	0.05	0.05
μ	0.68	0.67	0.71	0.68	0.67	0.71
σ	0.09	0.07	0.08	0.09	0.07	0.08
Var.	0.01	0.01	0.01	0.01	0.01	0.01
Min.	0.41	0.45	0.46	0.41	0.45	0.46
Max.	0.97	0.85	0.95	0.97	0.85	0.95
p-value	<0.001			<0.001

).

From Table 3, the ASD spectral reflectance properties were noted to vary across the surveyed crops. At N = 50, $\alpha$ = 0.05, Levene’s k-comparison test produces the p-value < 0.001, which was less than 0.05 significance alpha, signifying variability in the ASD spectral reflectance properties across the surveyed crops. Therefore, the computed mean reflectance values were used to spectrally profile various crops in the study area. The mean spectral reflectance values obtained from the ASD spectrometer were compared with the mean spectral radiance values acquired from UAV imagery to assess the degree of agreement between crop reflectance and radiance properties (Figure 5).

From Figure 5, both the UAV mean radiance and ASD mean reflectance properties of the surveyed crops were noted to follow similar pattens across the specified spectral wavelengths. However, the UAV mean radiance showed a slight deviation from the ASD mean reflectance of the cabbage in the 0.533 µm–0.587 µm and 0.785 µm–0.899 µm spectral wavelengths (Figure 5a). In Figure 5b,c, the discrepancy between the UAV mean radiance and ASD mean reflectance for maize was noted in the 0.785 µm–0.899 µm spectral wavelength.

3.5. Reflectance $Q_1$, $Q_3$, and $IQR$ for the Surveyed Crops

The ${\mathrm{Q}}_1$, ${\mathrm{Q}}_3$, and $\mathrm{I}\mathrm{Q}\mathrm{R}$ were computed for each surveyed crop, based on the spectral wavelengths in which a reflectance discrepancy among the crops was observed.

Table 5. ${\mathrm{Q}}_1$, ${\mathrm{Q}}_3$, and $\mathrm{I}\mathrm{Q}\mathrm{R}$ values for each crop type computed from selected spectral wavelengths.

	0.443 µm–0.507 µm			0.533 µm–0.587 µm			0.785 µm–0.899 µm
Surveyed crops	$Q_{i,l}$	$Q_{i,u}$	$IQ{R}_i$	$Q_{i,l}$	$Q_{i,u}$	$IQ{R}_i$	$Q_{i,l}$	$Q_{i,u}$	$IQ{R}_i$
Cabbage	0.33	0.43	0.1	0.18	0.25	0.07	0.62	0.75	0.13
Maize	0.5	0.68	0.18	0.29	0.44	0.17	0.63	0.72	0.09
Sugar bean	0.49	0.61	0.12	0.18	0.22	0.04	0.66	0.77	0.11

shows the ${\mathrm{Q}}_1$, ${\mathrm{Q}}_3$, and $\mathrm{I}\mathrm{Q}\mathrm{R}$ values computed for each surveyed crop type.

The upper reflectance and the lower.

Based on the computed ${\mathrm{Q}}_1$, ${\mathrm{Q}}_3$, and $\mathrm{I}\mathrm{Q}\mathrm{R}$ values, the upper and lower threshold for each surveyed crop type were computed and are presented in

Table 6. Upper and lower thresholds computed for characterizing crops.

	0.443 µm–0.507 µm		0.533 µm–0.587 µm		0.785 µm–0.899 µm
Surveyed crops	$T_{i,l}$	$T_{i,u}$	$T_{i,l}$	$T_{i,u}$	$T_{i,l}$	$T_{i,u}$
Cabbage	0.18	0.58	0.075	0.36	0.425	0.745
Maize	0.48	0.79	0.335	0.695	0.495	0.855
Sugar bean	0.55	0.95	0.12	0.28	0.645	0.935

.

Upon the computation of $T_{i,l}$, $T_{i,u}$, and $IQ{R}_i$ values, the thresholds were computed to separate the surveyed crop types.

Table 7. Optimized thresholds for the surveyed crop types.

	0.443 µm–0.507 µm	0.533 µm–0.587 µm	0.785 µm–0.899 µm
Cabbage–Maize	0.53	0.348	0.62
Maize–Sugar bean	0.67	0.24	0.75

provides the computed thresholds that classify crops in the study area.

From Table 7, the reflectance threshold between cabbage and maize and between maize and sugar bean, as computed from a 0.443 µm–0.507 µm spectral wavelength, were found to be 0.53 and 0.67, respectively. The reflectance threshold between cabbage and maize and between maize and sugar bean, as computed from a 0.533 µm–0.587 µm spectral wavelength, were found to be 0.24 and 0.35, respectively. Lastly, the reflectance threshold between cabbage and maize and between maize and sugar bean, as computed from a 0.785 µm–0.899 µm spectral wavelength, were found to be 0.62 and 0.75, respectively.

3.6. Optimized Threshold Values Selected from 0.443 µm–0.507 µm and 0.533 µm–0.587 µm Wavelengths

The computed thresholds were applied to differentiate the surveyed crops. Moreover, machine learning techniques were also applied to classify crops. This was performed to ensure the performance evaluation of the MLT against machine learning algorithms. Figure 6 provides the surveyed crops differentiated using MLT and selected machine learning algorithms.

The accuracy assessment results of the classified crop types using thresholding, along with selected machine learning techniques, are presented in

Table 8. Accuracy assessment summary.

Classifier	Overall Accuracy	KIA
MLT on Blue band	0.435	0.372
MLT on Green band	0.333	0.307
MLT on NIR	0.496	0.488
MLP	0.594	0.531
RBFNN	0.662	0.616
SOM	0.643	0.659

.

From Table 8, analysis of the KIA generally showed the inefficiency of all the image partitioning techniques in differentiating the surveyed crops. The MLT approach was generally ineffective in categorizing the surveyed crops. This was apparent from the KIA values of 0.372, 0.307, and 0.488 for the MLT on Blue band, MLT on Green band, and MLT on NIR, respectively. The MLP, RBFNN, and Kohonen’s SOM produced KIA values of 0.531, 0.616, and 0.659, which were lower than the acceptable KIA value of 0.7. This generally confirms the significant proportion of misclassified crops. However, Kohonen’s SOM produced promising results when compared with all the evaluated techniques.

3.7. Area Determination of the Surveyed Crops

The area covered by each surveyed crop type, as predicted by the MLT and selected machine learning techniques, are presented in Figure 7.

From Figure 7, the area covered by cabbage crops, as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM, was noted to be 7.46%, 6.01%, 10.33%, 7.05%, 9.48%, and 7.04%, respectively. The area covered by maize crops, as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM was noted to be 13.62%, 26.41%, 12.12%, 11.03%, 12.19%, and 15.11%, respectively. Sugar bean was noted to occupy 57.51%, 43.72%, 26.77%, 27.44%, 24.15%, and 16.33%, as predicted by the MLT on Blue band, MLT on Green band, MLT on NIR, MLP, RBFNN, and SOM, respectively. Generally, the misclassification of crops was noted in all crop classes predicted by all the deployed classification techniques. In Figure 6a, MLT on the Blue spectral band predicted maize crop, even in areas occupied by bare surface that was classified as maize crop. In Figure 6c, the MLP predicted cabbage in areas covered by sugar bean.

4. Discussion

The ability of multispectral remote sensing technology to characterize crops has not gone unnoticed [68,69]. As such, crop phenotyping offers traits that are critical for distinguishing various crops by analyzing the response of their traits to radiation received in different electromagnetic spectral wavelengths [70]. Ndou et al. [71] stated that UAV imaging systems can provide the successful acquisition of high spatial resolution imagery with detailed information regarding crop characteristics. However, information provided by UAV imaging systems is very complex, because these systems are also capable of detecting even the smallest undesired features [72]. UAV imagery often employs discrete spectral bands, which may lack specific spectral information outside the specified bandwidths [73]. Torres et al. [74] noted that spectral-based classification yields accurate results when utilizing sensors with many spectral bands. Nevertheless, non-imaging hyperspectral sensors offer continuous spectral information in numerous narrow bands (350–2500 nm), which allows for the detailed analysis of different crop properties [75]. They are employed in real-time, in situ leaf or canopy reflectance measurements, thus reducing the effect of background noise [76]. However, Potgieter et al. [77] noted that distinguishing spectrally overlapping crops remains a challenge. Therefore, the differentiation of crop types by thresholding UAV imagery based on ASD reflectance data was tested in this study.

In this study, true color composite (TCC) imagery, coupled with field knowledge, served as the base from which several crops were identified in the study area. Duranona Sosa et al. [78] noted that using TCC imagery in crop identification is relevant, because working with other imagery such as single band or spectral vegetation indices is not easy, due to their inability to facilitate image morphological filtering. Levene’s k-comparison of equal variance facilitated the variability analysis of spectral radiance/reflectance across the surveyed crops. Levene’s results produced p values that were <0.01, signifying variability in spectral radiance/reflectance across the surveyed crops. This technique test is usually applied to overcome challenges associated with the analysis of variance (ANOVA), which is to determine whether k-populations have a common mean $\mu$ [79]. It facilitates the diagnosis of the normality of data [80]. Otherwise, the ANOVA test would be used in case the computed p-values were >0.05. Although Serbin et al. [8] noted that maize and other various crops exhibited similar reflectance signatures in the visible (400–700 nm) and near-infrared (700–1100 nm) spectral wavelengths, the current study found that the maize cultivar can be spectrally distinguished from other crops in the Blue, Green and NIR spectral channels of spectrometric wavelengths. Angulo and Pamboukian [81] noted the efficacy of the NIR channel in distinguishing maize cultivars against other types of crops such as soybean, oats, and soy can be achieved in the NIR region and in the Green region of the spectrometer. Maize, known for its broad, long leaves, has a unique NIR reflectance profile that distinguishes it from other crops [82]. Through the combination of the UAV multispectral imagery data and non-imaging hyperspectral remote sensing techniques, Sudu et al. [83] noted that crop can be effectively and quantitatively analyzed.

The MLT was selected for classifying crops in this study, and its selection was prompted by several recent studies that noted its reliability in image classifying complex features [84,85]. Kumar et al. [48] noted that multilevel thresholding showed the ability to characterize crops and minimize non-crop background. Houssein et al. [86] also noted that the thresholding technique has the advantage of efficiency and versatility. In this study, we evaluated the MLT performance against selected machine learning techniques, viz., MLP, RBFNN, and SOM. Mndela et al. [11] noted that various types of crops in small-scale farms imaged using the UAV multispectral system can be distinguished by mean spectral profiling of sampled crop leaves. The advantage of using the mean-based thresholding is that its structure entirely relies on the normal distribution [87]. However, using spectral reflectance mean values as the threshold for categorizing crop types undermines the crop reflectance variability under varying illumination intensity. In the current study, we used minimum and maximum thresholds to categorize crop types. Pixel values which were outside the defined thresholds were classified as non-crop features.

The comparison of MLT classification results revealed shortcomings of the MLT against selected machine learning algorithms (Table 8). Specifically, the MLT failed to produce the desired classification results. This is despite the statement by Mahmoud et al. [85] that thresholding based on statistical values can provide accurate image segmentation results. However, this challenge was not only with MLT; generally, even selected machine learning approaches did not yield satisfactory classification results. However, Faridatul and Wu [88] noted that optimization of the threshold for a specific land feature class is important when classifying based on spectral vegetation indices. Despite the deployment of the IQR optimization technique on the MLT in the current study, as recommended by Zhao et al. [89], the results achieved were still poor. Despite the existence of numerous innovative and efficient image partitioning approaches in recent studies [90,91,92], a well-adapted classification technique applicable to areas with diverse crops remains elusive, owing to the challenges posed by multiple crop types, complex background information, and heterogeneous and altered leaf morphology due to crop infection.

The study did not account for crop water content, as variations in water content may also cause variations in crop chlorophyll content, which may lead to spectral overlap. Genc et al. [93] noted that reflectance of water-stressed sweet corn decreases in the red reflectance in red and increases in the NIR spectral region. It is worth noting that the surveyed crops in this study were also subjected to stress due to pests and pathogens, and this might have altered their natural spectral reflectance properties and created spectral overlap with one another. During the collection of hyperspectral reflectance data, crop leaves that were fully exposed to solar energy were targeted. However, some crop leaves were subjected to partial occlusion by the top-most leaves of the same plant. As such, partially occluded cabbage leaves had a similar reflectance as maize when observed from the UAV imagery. Zhou et al. [90] noted that aerial imagery comes with a variety of information, such as shapes and hues, which may create variations in the brightness levels of the same crop. Therefore, it remains intriguing to whether actual patterns can be achieved under the occlusion-free scenario. Although mature maize cultivars and other types of crops were considered, there was no clear scrutiny used to determine variations in crop stages. Therefore, future studies must attempt to spectrally differentiate maize cultivars from other crops under different growth stages. Due to similar spectral properties with crops, weeds were also classified as crops. This was also noted by Hlaing and Khaing [91], who noted that weeds tend to be included as part of crops due to the similarity in spectral reflectance properties.

Tufail et al. [92] noted that spectral radiance/reflectance properties of the same crop type can substantially differ, owing to variations in phenology and environmental conditions. These issues remain unresolved, despite the advancement in pattern recognition. It is therefore intriguing to understand how these factors influenced differentiation of crop types in the study area. Moreover, the study did not consider the variation in crop reflectance because of the variation in crop growth stages. As such, future studies must attempt to determine the effect of reflectance variations due to crop growth stage on crop classification using thresholding.

The main issue with thresholding classification is the selection of the parameters that correctly define the thresholds. This also includes accounting for variables that create a complex image, such as noise, heterogeneity in intensity, and texture. These highlighted issues make the practicality of the MLT approach across different seasons and geographical settings a challenging task, because crops often exhibit variability in these parameters, owing to their response to seasonal and environmental settings. Moreover, selection of thresholds must also take into account the influence of polarimetric scattering of different crops to attain optimum thresholds. Aitkenhead and Dyer [94] revealed that the potential issue with regard to generalizability of the MLT lies in its lack of transferability of winner thresholds from one imagery to another. While Sun et al. [95] recommended the inclusion of a synthetic aperture radar (SAR) sensor to improve the differentiation of crops using multispectral imagery due to the ability of SAR to account for polarimetric scattering traits [96], the medium spatial resolution associated with these sensors deem them unviable for application in small-scale farms of less than 2 hectares in size. This problem is not only inherent to MLT but also machine learning algorithms. Therefore, these existing challenges underscore the necessity to further search for other alternative methods for optimizing threshold selection.

5. Conclusions

The current study aimed to spatially partition different types of crops in the small-scale farms based on thresholding UAV imagery based on ASD hyperspectral reflectance data. This study noted that UAV imagery can help in the identification of plots occupied by crops. However, the UAV imaging system proved ineffectiveness in distinguishing different crops. The field spectrometric reflectance analysis of crops also revealed that crops can be spectrally differentiated in the red, NIR, and SWIR spectral wavelengths. The thresholding of UAV spectral bands showed an inability to classify multiple crops. Sugar bean had distinct reflectance from other crops in the SWIR. This spectral wavelength could be ideal for demarcating this crop alone. The results were worse compared to the ones produced by machine learning algorithms. The thresholding of UAV imagery based on optimized minimum and maximum thresholds did not yield accurate results. The study revealed several spectral-related challenges that inhibit the accurate characterization of crops by integrating UAV imagery and non-imaging field spectrometric data. Despite recommendations by recent studies, we noted that the MLT was noted to be unsuitable for classifying complex features such as spectrally overlapping crops. As long as various crops still exhibit spectral overlaps, classification of these crops by thresholding will remain a challenge unless new, robust thresholding optimization techniques are proposed. As such, a framework for facilitating both the selection of the most suitable thresholding method for a particular domain and the application of thresholds to facilitate the classification of spectrally overlapping crops is inevitable. Overall, the results of this study underscore the ongoing contribution of remote sensing technology in addressing agricultural sustainability and food security goals. Perhaps the future study should explore the prospect of calibrating images to thresholds for individual crops to evaluate the separability of the individual. Hyperspectral reflectance analysis of the surveyed crops also revealed reflectance disparities across crops in the shortwave spectral wavelength. The inclusion of this spectral band in the UAV platform or simulation of synthetic shortwave infrared bands based on visible–NIR spectral bands of the UAV is also suggested as a possible future research direction.