Effective preprocessing of hyperspectral imagery is a critical and time-intensive step to ensure high data quality, analytical accuracy, and suitability for advanced analyses. In this study, EO-1 Hyperion imagery obtained from the USGS EarthExplorer portal was used. The sensor provides 242 contiguous spectral bands covering the 400–2,500 nm range, with a spectral resolution of approximately 10 nm and a spatial resolution of 30 m (NASA, 2010). EO-1 Hyperion imagery underwent comprehensive preprocessing, including spectral, spatial, radiometric, atmospheric, terrain, and geometric corrections, to ensure accurate reflectance values and spatial consistency. The initial preprocessing step involved identifying and removing non-informative or corrupted spectral bands. Specifically, 75 bands were excluded due to issues such as zero-value data (e.g., bands 1–7, 58–76, and 225–242) and atmospheric absorption artifacts, particularly those caused by water vapor interference (e.g., bands 120–131, 166–182, and 223–224). This resulted in a refined dataset consisting of 167 effective spectral bands, following the preprocessing strategies adopted by Priedītis et al. (2015) and Sun et al. (2019). Subsequently, vertical striping noise commonly present in Hyperion data due to calibration inconsistencies and the imaging mechanism of push-broom sensors was addressed through local Destriping techniques. These systematic vertical striping artifacts, or bad columns, were corrected using Destriping methods incorporating statistical equalization and filtering, wherein the affected pixel values were replaced with neighborhood-based averages to improve image uniformity. Following this, radiometric calibration was conducted to convert digital number (DN) values into radiance units, facilitating the quantification of reflected or emitted energy. The calibrated data were then subjected to atmospheric correction using the Fast Line-of-sight Atmospheric Analysis of Hypercubes (FLAASH) module within the ENVI software environment (Harris Geospatial Solutions, 2015). FLAASH corrects for atmospheric absorption and scattering effects, generating surface reflectance data with smoother spectral profiles compared to alternative methods such as QUAC and ACORN. In this study, the retrieval of water vapor and aerosol information was disabled to streamline the processing workflow. Following atmospheric correction, spatial subsetting was performed to isolate the Hauz Khas region as the area of interest. The resulting spectrally and spatially refined reflectance dataset provides a reliable foundation for further classification and interpretation tasks. To assess the impact of input dimensionality on classification accuracy, four distinct datasets were generated from EO-1 Hyperion hyperspectral imagery.

Four datasets were prepared to evaluate the influence of spectral dimensionality and feature selection on classification performance. The first dataset, termed the Full-Band dataset, comprised 167 spectral bands retained after removing noisy, overlapping and atmospheric absorption bands from the original 242 bands of the EO-1 Hyperion sensor. The second dataset, designated as the Spectrally Separable Subset, was derived using a hybrid feature-selection strategy based on spectral contrast and feature importance. The Average Pairwise Absolute Difference (APAD) metric was employed to quantify inter-species spectral separability and identify the most discriminative hyperspectral bands from the EO-1 Hyperion imagery. For each spectral band, mean reflectance values of the 21 species were used to compute all pairwise absolute differences, which were then averaged across all 210 class pairs to yield a single APAD score per band (Figure 3). Bands were ranked in descending order of APAD values, and the top 30 bands were selected as the most discriminative features. This non-parametric method does not rely on intra-class variance or covariance estimation and is therefore well suited to high-dimensional hyperspectral data with limited homogeneous pixels. Under such conditions, APAD provides a pragmatic, robust separability metric for coarse-resolution hyperspectral data.

In contrast, variance-dependent measures such as the Bhattacharyya or Jeffries–Matusita distance require reliable intra-class statistics, which are often unavailable for natural forest datasets with small sample sizes and overlapping spectra (Zhang et al., 2007). The APAD approach was implemented in Python using a custom script that computed class-wise mean spectra, all pairwise absolute differences, and averaged them to derive APAD scores per band.

The top 30 bands selected (13, 26, 25, 24, 23, 32, 31, 30, 27, 46, 47, 45, 41, 92, 93, 95, 94, 96, 113, 112, 124, 123, 118, 122, 114, 120, 108, 137, 149, 105, 134) exhibited the highest inter-class separability. The APAD for band i was computed as A P A D i = 1 C ∑ A = 1 k − 1 ∑   B = A + 1 k     μ A , i − μ B , i where:

C =   2 K = k(k−1)/2 is number of unique species pairs

k = 21 is the number of tree species,

n = 167 is the number of bands

μ_A,i and μ_B,i are the mean reflectances of species A and B in band i.

The third and fourth datasets were produced through dimensionality-reduction techniques applied to the Full-Band dataset. Principal Component Analysis (PCA) extracted the first four principal components capturing the majority of spectral variance (Lee and Plaza, 2012), while the Minimum Noise Fraction (MNF) transform retained the four noise-filtered components with the largest eigenvalues (Green et al., 2009).

To derive species-specific band-ratio indices, inter-band correlation and mean reflectance contrast were analyzed. For each species, band pairs with minimal correlation and maximal reflectance difference were selected to form ratio indices that enhanced separability. These ratios were later integrated into classification experiments as additional discriminative features.

Supervised classification techniques were employed to discriminate individual tree species using the pre-processed EO-1 Hyperion surface reflectance data. A total of thirteen supervised classifiers were evaluated, encompassing both traditional statistical and modern machine-learning approaches to ensure a comprehensive comparative assessment.

The statistical classifiers included Binary Encoding, Minimum Distance, Parallelepiped, Maximum Likelihood, Mahalanobis Distance, Spectral Angle Mapper (SAM), and Spectral Information Divergence (SID). The machine-learning algorithms comprised Support Vector Machine (SVM) (Vapnik, 1995; Melgani and Bruzzone, 2004), Artificial Neural Networks (ANN) (Rumelhart et al., 1986; Zhang et al., 2021), Random Forest (RF) (Breiman, 2001; Ma et al., 2021), K-Nearest Neighbor (KNN) (Cover and Hart, 1967; Li et al., 2022), and Decision Tree (DT) (Quinlan, 1986; Breiman et al., 1984). Among these, SVM and ANN have been extensively reported for hyperspectral image classification owing to their effectiveness in handling high-dimensional feature spaces and complex nonlinear class boundaries (Melgani and Bruzzone, 2004; Ballanti et al., 2016).

All classifiers were implemented within the ENVI 5.6 software environment (Harris Geospatial Solutions, 2015) following standardized supervised learning protocols. Ground-truth Regions of Interest (ROIs), delineated from field-collected GPS samples and cross-validated with very high-resolution Pléiades Neo imagery, were used as training data for supervised classification. The mean spectral reflectance profiles of 21 urban tree species, derived from these field-verified ROIs using EO-1 Hyperion hyperspectral data, are presented in Figure 4. These spectra exhibit distinct absorption and reflectance features across the visible, near-infrared, and shortwave-infrared regions, highlighting species-specific spectral variability critical for accurate classification and separability analysis. The dataset was subsequently partitioned into 70% training and 30% validation subsets to ensure unbiased model calibration and performance evaluation. Each classifier was trained using the designated ROIs and subsequently applied to the full hyperspectral dataset to generate thematic tree-species classification maps. Post-classification majority filtering was performed to reduce speckle noise and enhance spatial coherence, thereby improving the overall interpretability of the classification outputs.

Accuracy assessment was performed using the independent validation set by generating confusion matrices for each of the 13 classifiers. Classification performance was quantified using Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and the Kappa coefficient (κ), ensuring a robust comparison of classifier effectiveness at both the dataset and species levels.

To maintain methodological rigor, several preprocessing steps were undertaken prior to classification, including the removal of noisy and atmospheric absorption bands, destriping correction, and atmospheric calibration using the Fast Line-of-sight Atmospheric Analysis of Spectral Hypercubes (FLAASH) module. Independent field samples were cross-validated with Pléiades imagery to confirm that training and validation ROIs represented pure or dominant tree crowns, minimizing mixed-pixel errors. Parameter settings for each classifier followed established hyperspectral remote sensing standards (Melgani and Bruzzone, 2004; Ballanti et al., 2016), ensuring that results reflected true algorithmic performance rather than dataset inconsistencies or implementation biases.

Accuracy assessment was conducted using the independent validation set by computing confusion matrices and deriving standard accuracy metrics—Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and the Kappa coefficient (κ) for each classifier. This ensured rigorous, quantitative evaluation of classifier performance at the species level. All parameter settings were verified with reference to established hyperspectral classification literature (e.g., Melgani and Bruzzone, 2004; Ballanti et al., 2016). Data preprocessing included the removal of noisy bands, atmospheric correction using FLAASH, and destriping correction. Field and Pléiades Neo data were jointly used to cross-validate class labels, ensuring that the comparative results reflected algorithmic performance rather than preprocessing or dataset inconsistencies.

The performance of all classification algorithms was rigorously evaluated using confusion matrices and standard accuracy metrics, following the accuracy assessment framework proposed by Congalton (1991). Each supervised classification algorithm was applied to the prepared datasets, and its performance was quantified based on Overall Accuracy (OA), Producer’s Accuracy (PA), User’s Accuracy (UA), and the Kappa coefficient (κ). The classifier that achieved the highest overall and class-wise accuracies for the given dataset was identified as the optimal model for subsequent analysis.

To further validate the classifier’s capability to discriminate individual tree species, both UA and OA were analyzed to determine classification reliability and inter-class confusion. Species exhibiting the highest classification confidence and spectral separability were considered optimally extractable under the selected model. To enhance class separability and minimize spectral overlap, a correlation matrix of species-wise spectral responses was generated. Statistical descriptors, including mean, standard deviation, and coefficient of variation, were derived and integrated within a decision-tree framework to construct species-specific band ratio indices. These indices facilitated the extraction of discriminative spectral features, thereby improving inter-species separability and overall classification performance.

Table 1 summarizes the classification performance (Overall Accuracy, OA, and Kappa coefficient) of the thirteen classifiers applied to the four datasets. Following the evaluation framework of Ma et al. (2021), several traditional classifiers, particularly spectral and statistical methods, exhibited relatively low accuracies across all dimensionality-reduced datasets. These results underscore the limited ability of older algorithms to manage the high-dimensional and noise-prone nature of Hyperion hyperspectral data.

In contrast, ensemble-based classifiers—particularly Random Forest (RF) and Decision Tree (DT)—demonstrated consistently high accuracies across all datasets, confirming their robustness and adaptability. Among all classifiers, RF achieved the highest OA of 82.56% and Kappa coefficient of 0.81 using PCA-transformed data, indicating strong agreement between predicted and reference classes. DT also performed competitively, yielding OA values above 73% for all datasets and peaking at 79.49% with the Spectral Separability subset.

Traditional classifiers such as Binary Encoding, Parallelepiped, and Spectral Angle Mapper (SAM) produced considerably lower accuracies (OA < 45%), reaffirming their limited suitability for complex hyperspectral scenes. SVM exhibited moderate performance (OA ≈ 44%), while ANN and Neural Networks showed dataset-dependent variability. For example, Neural Networks achieved 73.33% OA on the Spectral Separability dataset but dropped to 5.13% on the Full Band data, likely due to overfitting and insufficient training samples in high-dimensional feature spaces. The lower accuracies observed for several classifiers are not attributed to preprocessing or implementation errors, as workflows and parameters were verified across all tests. Rather, they reflect inherent algorithmic limitations when handling mixed-pixel and high-dimensional data characteristic of EO-1 Hyperion imagery.

The species-level classification maps generated by the top-performing classifiers—Random Forest (RF) and Decision Tree (DT)—using their respective optimal datasets (PCA, MNF, Spectral Subset, and Full Band), illustrating spatial distribution and inter-species variability across the Hauz Khas urban forest shown in Figure 5. Figure 6 presents the comparative performance of all thirteen classifiers across the four datasets—Full Band, Spectral Separability Subset, PCA, and MNF—based on Overall Accuracy (OA) and Kappa coefficient (κ). These findings confirm the effectiveness of dimensionality reduction and feature optimization in enhancing classification reliability and inter-class separability. Among the compared models, Random Forest consistently outperformed others in both accuracy and robustness, confirming its suitability for operational tree species mapping with hyperspectral data.

Table 2 summarizes the species-level classification performance results using the optimal classifier–dataset combinations, showing Overall Accuracy (OA), Producer’s Accuracy (PA), and User’s Accuracy (UA) for 21 urban tree species ranked by mean accuracy. The results highlight the best-performing classifier–dataset combinations and demonstrate the influence of dimensionality reduction techniques on species discrimination shown in Figure 7.

Ensemble-based classifiers—Random Forest (RF) and Decision Tree (DT)—dominated the overall performance, achieving ≥86% Overall Accuracy (OA) for 16 out of 21 species. Among these, Peepal and Indian Banyan achieved the highest accuracies (95% OA and PA) using PCA-transformed datasets with RF and DT classifiers, respectively. This finding reinforces the strength of ensemble learning methods in modeling non-linear spectral responses inherent to hyperspectral data. Species such as Neem and Eucalyptus also exhibited strong classification performance, achieving >92% OA with MNF and Spectral Subset (SS) datasets, respectively. These outcomes indicate that both PCA and MNF effectively reduce spectral redundancy and noise while retaining essential discriminatory information for species-level classification. Moderate-performing species, including Tree of Heaven, Multiya, and Christ’s Thorn, attained OA values around 87% using Full Band or MNF datasets. The lower accuracies observed for these species can be attributed to their overlapping spectral signatures and mixed-pixel effects due to canopy intermixing within 30 m Hyperion pixels.

Overall, RF consistently produced stable and high accuracies across multiple datasets, attributed to its ensemble averaging and resistance to overfitting (Breiman, 2001). DT, while slightly more sensitive to noise, provided comparable performance and computational efficiency, making it particularly suitable for rapid classification of large hyperspectral datasets. Neural Networks demonstrated satisfactory performance (90% OA) for Northern Cottonwood, suggesting their potential for species with distinct spectral–structural features.

The comparative analysis confirms that PCA and MNF transformations substantially improve classification reliability by enhancing inter-class separability and minimizing within-class variability. Furthermore, the APAD-based Spectral Subset effectively supported species with subtle spectral differences, such as Eucalyptus and Fig, highlighting its utility for targeted band selection. In summary, the integration of dimensionality reduction with ensemble-based classifiers achieved the highest overall classification efficiency. The combination of RF and PCA emerged as the most effective strategy for species-level mapping in heterogeneous urban forest conditions, ensuring both accuracy and computational practicality.

Using the Average Pairwise Absolute Difference (APAD) metric, 210 unique species-pair separability scores were computed across 167 bands. The analysis revealed several species pairs—such as Cassie and Christ’s Thorn and Multiya and Neem—with higher APAD values, indicating enhanced inter-species separability. These pairs were instrumental in guiding band selection and the development of species-specific ratio indices.

Although APAD does not incorporate intra-class variance, it effectively captured meaningful inter-species contrast. Species pairs such as Cassie–Christ’s Thorn, Multiya–Christ’s Thorn, and Neem–Tree of Heaven exhibited correlation coefficients between 0.4 and 0.6, suggesting substantial potential for discrimination via band ratios or supervised classifiers. The least correlated species pairs—Cassie–Christ’s Thorn, Multiya–Christ’s Thorn, Tree of Heaven–Christ’s Thorn, Cassie–Neem, Multiya–Neem, Cassie–Blackboard Tree, and Orchid Tree–Cassie—represent the most promising candidates for targeted separability enhancement. Although no pair achieved orthogonal spectral distinction, these moderate decorrelation levels demonstrate the usefulness of APAD in identifying separable spectral regions.

A key limitation lies in the coarse 30 m spatial resolution of EO-1 Hyperion, where each pixel may encompass multiple tree crowns, introducing mixed-pixel effects that reduce separability. Ground-truth field data, APAD-based band selection, and band-ratio indices were therefore critical in improving classification reliability. Despite residual uncertainty, this approach successfully enhanced inter-species separability under sensor limitations.

Optimal band-ratio indices were developed to further enhance discrimination between spectrally similar species. Ratios were derived by pairing bands with low inter-correlation and high mean reflectance contrast, as determined from 167 bands of Hyperion imagery. These ratios provide interpretable, physically meaningful features for vegetation index design and decision-rule formulation. The following ratios were found to be most discriminative:

Christ’s Thorn: B23/B92; Ipil-Ipil: B92/B134; Neem: B2/B93; Blackboard Tree: B52/B166; Cassie: B69/B92; Karanj: B54/B166; Fig: B79/B131; Peepal: B53/B132; Medlar: B24/B92; Jaggery Palm: B1/B92; Indian Banyan: B53/B92; Northern Cottonwood: B3/B138; Australian Silky Oak: B51/B163; Orchid Tree: B37/B92; Ashok: B1/B53; Golden Shower Tree: B122/B132; Eucalyptus: B24/B105; Putranjiva: B92/B140; Buddha Coconut: B2/B92; Multiya: B80/B93; Tree of Heaven: B1/B95.

Among these, Band 92 (near-infrared region) frequently appeared in denominator positions, reflecting its strong discriminative ability due to high reflectance sensitivity to vegetation structure and moisture content.

The classification results provide ecological insights relevant to urban forest monitoring and management. Accurate species-level mapping enhances understanding of urban forest composition, resilience, and biodiversity distribution. The ability to distinguish native species such as Neem (Azadirachta indica) and Peepal (Ficus religiosa) from exotic or invasive taxa supports informed decision-making in reforestation and conservation initiatives, thereby contributing to ecological balance and urban sustainability. High-accuracy classification maps also enable improved estimation of urban biomass, canopy health, and carbon sequestration potential—key parameters for green GDP assessment and climate-resilient urban planning. The demonstrated success of ensemble classifiers highlights the potential of hyperspectral remote sensing as a reliable decision-support tool for urban forest management, guiding targeted species planting, stress detection, and biodiversity restoration.

The integration of ensemble classifiers—particularly Random Forest (RF) and Decision Tree (DT)—demonstrated superior accuracy and robustness, confirming their suitability for scalable species-level classification in spectrally complex urban ecosystems. The combination of Average Pairwise Absolute Difference (APAD)-based band selection and species-specific band-ratio indices proved effective in enhancing spectral separability, offering a transferable framework for vegetation discrimination even when using moderate-resolution hyperspectral data such as EO-1 Hyperion.

Despite these promising results, two methodological limitations were identified. First, the single-date Hyperion acquisition constrains analysis to a single phenological phase, limiting the ability to capture seasonal dynamics such as leaf emergence, flowering, or senescence. Second, while APAD effectively quantified inter-species spectral separability, it does not account for intra-class variance, partly due to mixed pixels in the 30 m Hyperion imagery limiting homogeneous samples for variance-based statistics. Although deep learning architectures—such as Convolutional Neural Networks (CNNs), deformable CNNs, and hybrid CNN–Transformer networks (Wang et al., 2021; Chen et al., 2023; Li et al., 2023; Sun et al., 2024)—have shown excellent potential for hyperspectral classification, their effective deployment requires large labeled datasets and substantial computational resources. Given the limited ground-truth samples and mixed-pixel constraints, this study prioritized computationally efficient fast-learning classifiers. Future research could integrate deep learning with high-resolution multispectral or hyperspectral data and LiDAR-derived structural metrics to further enhance species-level discrimination and support scalable, operational urban forest monitoring.