China UAV-Based Hyperspectral Object-Oriented Crop Classification

1. Introduction
Precise mapping of crop distribution is paramount for ensuring food security and enabling sustainable agricultural management. Traditional ground survey methods are prohibitively time-consuming, labor-intensive, and costly, failing to meet the demands of modern precision agriculture. The advent of hyperspectral remote sensing, offering “image-spectrum integration” and high spectral resolution, presents significant potential for crop discrimination. However, satellite and airborne hyperspectral systems often suffer from limitations like coarse spatial resolution, infrequent revisit times, and high operational costs, hindering their applicability for fine-scale agricultural monitoring in regions like China.

The emergence of lightweight imaging spectrometers integrated with Unmanned Aerial Vehicles (UAVs) has revolutionized hyperspectral data acquisition. China UAV platforms offer unparalleled advantages: exceptional operational flexibility, high efficiency for field data collection, significantly lower costs compared to traditional platforms, and crucially, the ability to capture hyperspectral imagery with very high spatial resolution. This capability is essential for discerning subtle spectral differences between crops and managing the fragmented agricultural landscapes prevalent in China UAV operational contexts. While pixel-based classification methods are commonly applied, they struggle with challenges like mixed pixels and spectral confusion. Object-Based Image Analysis (OBIA) provides a robust alternative by segmenting imagery into homogeneous objects based on spectral, spatial, and textural characteristics, effectively mitigating “salt-and-pepper” noise and aligning better with real-world agricultural parcels.

This study focuses on leveraging China UAV-acquired hyperspectral imagery to investigate the feasibility and applicability of an OBIA approach combined with advanced machine learning (ML) algorithms for fine-scale crop classification within a typical Chinese agricultural setting. We specifically address the gap in research concerning OBIA application to China UAV hyperspectral data for detailed crop mapping at the field level.

2. Materials and Methods

2.1. Study Area and Dataset
The research utilized the publicly available WHU-Hi-LongKou dataset, acquired over Longkou Town, Hubei Province, China. Data collection employed a Headwall Nano-Hyperspec sensor mounted on a China UAV platform. Key acquisition parameters were:

Flight Altitude: 500 m
Weather Conditions: Clear sky, ~36°C, ~65% relative humidity
Spatial Resolution: ~0.463 m
Spatial Dimensions: 550 x 400 pixels
Spectral Range: 400 – 1000 nm
Number of Bands: 270

The study area features six primary crop types: Corn (Zea mays), Cotton (Gossypium hirsutum), Sesame (Sesamum indicum), Rice (Oryza sativa), Broadleaf Soybean (Glycine max), and Narrowleaf Soybean (Glycine max).

2.2. Methodology Overview
The methodological workflow comprised four main stages:

Hyperspectral Data Dimensionality Reduction: Mitigating the curse of dimensionality inherent in hyperspectral data.
Multiscale Image Segmentation: Generating meaningful image objects.
Optimal Feature Selection: Identifying the most discriminative features for classification.
Object-Based Classification & Validation: Applying ML classifiers and evaluating performance.

2.2.1. Hyperspectral Data Dimensionality Reduction
The high dimensionality (270 bands) introduces computational complexity and noise susceptibility. Minimum Noise Fraction (MNF) transform was applied using ENVI software to segregate noise and reduce data dimensionality while retaining essential information. The MNF transform involves two principal steps:

Noise Covariance Estimation: Estimating the noise covariance matrix (Σ_N) from the data.
Noise Whitening and PCA: Transforming the original data (X) to whiten the noise, followed by Principal Component Analysis (PCA) on the noise-whitened data.
Let X be the original hyperspectral data matrix (n pixels x p bands). The MNF transformation derives eigenvectors and eigenvalues from the noise-adjusted covariance matrix. The number of significant MNF components was determined by identifying the inflection point (“elbow”) in the plot of eigenvalues versus component number. Components preceding the inflection point contain most of the spatially coherent information, while subsequent components are dominated by noise. Based on this analysis (eigenvalue plot), the first 10 MNF components were selected for subsequent processing. This achieved significant dimensionality reduction (270 -> 10 bands) while preserving critical spectral variance.

2.2.2. Multiscale Image Segmentation
Image segmentation was performed using eCognition Developer software. Multiresolution segmentation [4] was employed, which iteratively merges pixels or smaller objects based on heterogeneity minimization, governed by a scale parameter (SP), color/shape (w_c / w_s), and compactness/smoothness (w_{cmp} / w_{smt}) weights. The heterogeneity measure (f) for merging is:
f = w_c * Δh_color + w_s * Δh_shape
where Δh_color is the change in spectral homogeneity and Δh_shape is the change in shape homogeneity. The shape homogeneity itself is defined as:
Δh_shape = w_{cmp} * Δh_{cmp} + w_{smt} * Δh_{smt}
Layer weights for the 10 MNF bands were uniformly set to 1. The shape factor (w_s) and compactness factor (w_{cmp}) were optimized through systematic trials, varying both parameters in 0.1 increments from 0.1 to 0.9. Visual assessment of segmentation quality, focusing on accurate delineation of field boundaries and object homogeneity, led to the selection of w_s = 1.0 and w_{cmp} = 0.5.

To objectively determine the optimal Scale Parameter (SP), the Estimation of Scale Parameter 2 (ESP2) tool [5] was utilized within eCognition. ESP2 calculates the Local Variance (LV) and its Rate of Change (ROC-LV) across a range of SP values:
ROC-LV(SP_i) = |LV(SP_{i+1}) - LV(SP_i)| / LV(SP_i)
Peaks in the ROC-LV curve indicate optimal scales where the heterogeneity between potential objects increases significantly. Analysis of the ROC-LV curve (Figure 3a in source) revealed several peaks. Empirical testing confirmed SP = 7 as the optimal scale parameter, yielding well-delineated field objects corresponding to individual crop parcels (Figure 3b in source).

2.2.3. Optimal Feature Selection
A comprehensive set of 46 potential features was extracted for each image object segmented at SP=7:

Spectral Features (20): Mean and Standard Deviation for each of the 10 MNF bands.
Shape Features (18): Area, Border Length, Length, Length/Width, Width, Number of Pixels, Volume, Asymmetry, Border Index, Compactness, Density, Elliptic Fit, Main Direction, Shape Index, Roundness, Rectangular Fit, Max. Ellipse Radius, Min. Ellipse Radius.
*Texture Features (8 – GLCM based):* Mean, Standard Deviation, Homogeneity, Contrast, Dissimilarity, Entropy, Angular Second Moment (ASM), Correlation.

Directly using all features risks overfitting and computational inefficiency. eCognition’s Feature Space Optimization (FSO) module was employed for feature selection. FSO evaluates features based on their correlation with the target class and redundancy with other features. Features exhibiting high correlation with the class label but low correlation with each other are preferred. This process identified the 13 most discriminatory features:

Spectral: Mean (Band 2, Band 3, Band 5), Std. Dev. (Band 2, Band 5)
Shape: Area, Length, Width, Asymmetry, Roundness, Elliptic Fit, Border Index, Max. Ellipse Radius.

Table 1: Optimal Feature Subset Selected for Classification

Feature Type	Specific Features
Spectral	Mean (MNF Band 2), Mean (MNF Band 3), Mean (MNF Band 5), Std. Dev. (MNF Band 2), Std. Dev. (MNF Band 5)
Shape	Area, Length, Width, Asymmetry, Roundness, Elliptic Fit, Border Index, Max. Ellipse Radius

2.2.4. Object-Based Classification and Accuracy Assessment
Four distinct ML algorithms were implemented within eCognition for object-based classification:

K-Nearest Neighbors (KNN): Assigns a class to an object based on the majority vote of its k closest neighbors in the feature space. Distance is typically Euclidean:
d(x, y) = √Σ_{i=1}^n (x_i - y_i)²
where x and y are feature vectors of two objects, and n is the number of features (13). k=5 was used based on common practice and preliminary tests.
Classification and Regression Trees (CART): Builds a binary decision tree by recursively partitioning the feature space to minimize impurity (e.g., Gini Index G = 1 - Σ(p_i²) at each node, where p_i is the proportion of class i at the node). The tree grows until a stopping criterion is met (e.g., minimum node size).
Support Vector Machine (SVM): Finds the optimal hyperplane in a high-dimensional feature space (mapped via a kernel function, e.g., Radial Basis Function – RBF) that maximally separates classes. The optimization problem is:
Minimize: 1/2 ||w||² + C Σξ_i
Subject to: y_i(w·φ(x_i) + b) ≥ 1 - ξ_i, ξ_i ≥ 0
where w is the weight vector, C is the regularization parameter, ξ_i are slack variables, φ(x_i) is the kernel mapping, and b is the bias. RBF kernel parameters (C, γ) were tuned.
Random Forest (RF): An ensemble method constructing multiple decorrelated decision trees (n_tree=500). Each tree is trained on a bootstrap sample of the data and a random subset of features (mtry=√n_features≈4). Final class prediction is by majority vote. RF handles high dimensionality well and provides internal error estimates.

A stratified random sampling approach was used to generate training and validation sets from manually interpreted reference objects. Classification accuracy was rigorously evaluated using:

Overall Accuracy (OA): Percentage of correctly classified validation objects.
Kappa Coefficient (κ): Measures agreement between classification and reference data, correcting for chance. κ = (p_o - p_e) / (1 - p_e), where p_o is observed agreement (OA) and p_e is expected agreement by chance.
User’s Accuracy (UA): For a class, the percentage of objects classified as that class that truly belong to it (Commission Error = 1 – UA). UA = X_{ii} / X_{i+}
Producer’s Accuracy (PA): For a class, the percentage of true reference objects of that class correctly classified (Omission Error = 1 – PA). PA = X_{ii} / X_{+i}
Here, X_{ii} is the diagonal element (correct) for class i, X_{i+} is the row total for class i (classified as i), X_{+i} is the column total for class i (reference objects of i).

3. Results

3.1. Classification Results
Visual analysis of the classified maps revealed distinct differences among the algorithms:

RF: Produced the most visually coherent map. Crop fields were classified as complete, continuous parcels with sharp boundaries, exhibiting strong agreement with the actual ground distribution of crops.
SVM: Showed good results for Sesame and Rice, but suffered from confusion between Corn/Cotton and Broadleaf Soybean, and between Narrowleaf Soybean and Cotton. Field boundaries were often misclassified as weeds.
CART: Performed reasonably well for Sesame but exhibited significant confusion between Cotton and Rice fields.
KNN: Yielded the poorest visual result, characterized by severe misclassification across all crop types. Examples included Cotton areas misclassified as Broadleaf Soybean or weeds, Corn areas misclassified as Broadleaf Soybean, and Rice areas misclassified as Sesame.

3.2. Quantitative Accuracy Assessment

3.2.1. Overall Accuracy and Kappa
The RF classifier achieved the highest overall accuracy (OA) and Kappa coefficient (κ), significantly outperforming the other three algorithms.

Table 2: Overall Classification Accuracy and Kappa Coefficient

Classification Method	Overall Accuracy (OA) %	Kappa Coefficient (κ)
KNN	87.41	0.8432
CART	91.52	0.8950
SVM	93.64	0.9212
RF	95.84	0.9487

3.2.2. Per-Class User’s and Producer’s Accuracy
Analysis of User’s Accuracy (UA) and Producer’s Accuracy (PA) provides detailed insights into the performance for each crop type.

Table 3: Per-Class User’s Accuracy (UA) and Producer’s Accuracy (PA) (%)

Crop Type	Metric	KNN	CART	SVM	RF
Corn	UA	79.20	94.50	100	97.73
	PA	88.28	91.32	87.22	96.99
Cotton	UA	77.57	91.41	95.04	75.28
	PA	79.20	82.80	98.10	93.92
Sesame	UA	63.65	100	100	85.91
	PA	83.83	92.42	83.83	92.42
Rice	UA	100	75.65	100	93.43
	PA	83.63	79.81	97.27	97.40
Narrowleaf Soybean	UA	100	74.61	70.00	89.13
	PA	73.80	93.82	97.00	90.83
Broadleaf Soybean	UA	87.40	94.01	89.81	99.44
	PA	89.30	96.10	92.95	93.30

Key Observations:

RF: Excelled in detecting crops (high PA) for Corn (96.99%), Rice (97.40%), and Sesame (92.42% – tied best). Its UA was also high for Corn (97.73%) and Broadleaf Soybean (99.44%), indicating reliable predictions for these classes. However, RF exhibited the lowest UA for Cotton (75.28%), signifying significant commission error (other crops misclassified as Cotton).
SVM: Demonstrated perfect UA for Corn, Sesame, and Rice (100%), meaning no pixels of other classes were misclassified into these categories. However, its PA for Corn (87.22%) and Sesame (83.83%) was the lowest among all algorithms, indicating substantial omission errors (true Corn/Sesame missed). SVM achieved the highest PA for Cotton (98.10%) and Narrowleaf Soybean (97.00%). Its UA for Narrowleaf Soybean was notably low (70.00%).
CART: Achieved perfect UA for Sesame (100%) and the highest PA for Sesame (92.42%, tied with RF). However, it showed relatively low UA for Rice (75.65%) and Narrowleaf Soybean (74.61%), and low PA for Rice (79.81%).
KNN: Achieved perfect UA for Rice and Narrowleaf Soybean (100%) but suffered from low PA for all classes, especially Narrowleaf Soybean (73.80%) and Sesame (83.83% – tied lowest). Its UA for Sesame was also very low (63.65%).
Corn Classification: RF produced the best balance (UA=97.73%, PA=96.99%). SVM had perfect UA (100%) but the lowest PA (87.22%).

4. Discussion

This study demonstrates the significant potential of China UAV-based hyperspectral remote sensing coupled with OBIA and ML, specifically RF, for achieving fine-scale crop classification in the context of Chinese agriculture. The high spatial resolution (~0.5m) provided by the China UAV platform was critical for resolving small, fragmented fields typical of the region. The OBIA approach effectively generated meaningful field-scale objects, overcoming the limitations of pixel-based methods.

4.1. Algorithm Performance Analysis
The superior performance of the RF algorithm (OA=95.84%, κ=0.9487) aligns with its known characteristics: inherent resistance to overfitting through ensemble learning, robustness to noise and outliers, ability to handle high-dimensional data efficiently, and implicit feature weighting. Its effectiveness for Corn and Broadleaf Soybean (dominant crops) is consistent with its strength in handling classes with larger sample sizes. The relatively lower UA for Cotton (the smallest sample class) suggests RF might be more sensitive to class imbalance when sample sizes are very small, leading to commission errors.

SVM showed excellent ability to avoid commission errors (high UA) for Corn, Sesame, and Rice, and high detection rates (PA) for Cotton and Narrowleaf Soybean. This reflects SVM’s strength in finding optimal separation boundaries, especially for smaller classes, given appropriate parameter tuning (C, γ). However, its lower PA for Corn and Sesame indicates difficulty in fully capturing the spectral variability within these classes using the selected features and kernel.

CART performed well for Sesame but showed limitations with other classes, likely due to its tendency to overfit and create complex trees sensitive to training data specifics. KNN, while simple, was clearly outperformed, suffering from noise sensitivity and the “curse of dimensionality” despite feature reduction.

4.2. Implications for China UAV Applications
The success of RF in this China UAV study highlights its suitability for operational crop mapping tasks requiring high accuracy. The achieved fine-scale classification accuracy (~96%) is highly relevant for practical applications within China UAV agricultural monitoring frameworks:

Precision Agriculture: Enabling targeted field management (irrigation, fertilization, pest control) based on precise crop type and location.
Crop Area Estimation: Providing accurate statistics for yield forecasting and policy planning.
Crop Condition Monitoring: Serving as a baseline for detecting stress or disease within specific crop parcels.
Land Use/Land Cover (LULC) Mapping: Contributing to detailed agricultural land use inventories.

4.3. Limitations and Future Work
While demonstrating the viability of the approach, this study has limitations:

Dimensionality Reduction: Only MNF was explored. Comparative analysis of other techniques (e.g., PCA, SPA, band selection algorithms) could identify the optimal method for China UAV hyperspectral crop classification.
Algorithm Parameter Optimization: Tuning was primarily empirical. Systematic hyperparameter optimization (e.g., grid search, Bayesian optimization) using cross-validation for all algorithms (especially SVM and RF) could yield further accuracy gains.
Feature Engineering: While FSO was used, exploring more advanced spectral indices, temporal features (if multi-temporal data available), or deep learning features extracted from the hyperspectral data itself could enhance separability, particularly for spectrally similar crops.
Spatial Generalization: Validation was confined to one study area. Testing the robustness of the model across diverse agricultural landscapes in China is essential.
Class Imbalance: The impact of sample size imbalance (e.g., Cotton) warrants investigation using techniques like SMOTE or cost-sensitive learning within the RF framework.

Future research using China UAV hyperspectral data should focus on integrating multi-temporal imagery to capture phenological differences, combining hyperspectral data with LiDAR or high-resolution RGB for improved structural/textural features, and exploring deep learning models (e.g., 1D/2D/3D CNNs, Transformers) within an OBIA framework or for end-to-end pixel/object classification. Investigating transfer learning to apply models trained in one region to others within China would enhance operational efficiency.

5. Conclusion
This research successfully established a robust workflow for fine-scale crop classification utilizing China UAV-acquired hyperspectral imagery. Key conclusions are:

China UAV platforms provide hyperspectral data with the necessary spatial resolution (~0.5m) for effective fine-scale agricultural monitoring and crop discrimination in fragmented landscapes.
Object-Based Image Analysis (OBIA) is highly effective for processing China UAV hyperspectral data, generating meaningful field-level objects that mitigate pixel-level noise and align with management units.
The MNF transform efficiently reduced data dimensionality (270 bands -> 10 components) while preserving critical information for classification within the China UAV data context.
Systematic feature selection identified 13 optimal spectral and shape features crucial for discriminating between the six crop types using the China UAV data.
Among the four ML algorithms evaluated (KNN, CART, SVM, RF), Random Forest (RF) delivered the highest overall classification accuracy (95.84%) and Kappa coefficient (0.9487). RF produced the most spatially coherent maps with minimal class confusion for most crops, demonstrating its robustness and suitability for this China UAV application.
Detailed per-class accuracy analysis revealed algorithm strengths and weaknesses: RF excelled in detection (PA) for Corn, Rice, and Sesame; SVM achieved high UA (avoiding commission errors) for Corn, Sesame, and Rice and high PA for Cotton and Narrowleaf Soybean; CART performed well for Sesame; KNN was consistently the weakest performer.
The RF-based OBIA approach using China UAV hyperspectral data proves to be a highly feasible and applicable method for fine-scale crop classification in Chinese agricultural settings. This methodology offers significant potential for enhancing crop monitoring precision, supporting precision farming initiatives, and improving agricultural resource management using China UAV technology.