TL;DR
This paper presents a method for accurately estimating and mapping nitrogen concentration in apple canopies using UAV-based hyperspectral imagery. By combining spectral indices and red-edge parameters, and employing a backpropagation neural network (BPNN), the study achieves high accuracy in predicting nitrogen levels, offering a valuable tool for precision agriculture.
What Problem Does This Paper Address?
Accurate monitoring of nitrogen (N) levels in crops is crucial for effective agricultural management, directly impacting growth, yield, and quality. Traditional ground-based methods are labor-intensive and limited in scale. While remote sensing offers a faster, non-destructive alternative, satellite-based approaches can suffer from low spatial and temporal resolution, and multispectral sensors lack the spectral detail needed for fine-grained analysis. Hyperspectral sensors, with their high spectral resolution, hold promise, but applying them to complex tree canopies presents challenges, including distinguishing canopy from background (bare soil, shadows) and the need for robust inversion models. This study aims to overcome these limitations by leveraging UAV-mounted hyperspectral imaging for precise apple canopy nitrogen concentration (CNC) estimation.
Methods
The study employed a multi-step approach:
- Canopy Extraction: To isolate the apple canopy's spectral information, a thresholding method was used. This involved combining the Normalized Difference Vegetation Index (NDVI) to differentiate vegetation from bare soil and the Normalized Difference Canopy Shadow Index (NDCSI) to remove shadows. This approach was found to be more refined than using a Random Forest Classifier (RFC) alone.
- Spectral Parameter Selection: A range of spectral parameters were investigated, including:
- Vegetation Indices (VIs): Standard indices like mSR705, mND705, CIred-edge, CIgreen, and DCNI.
- Random Two-Band Spectral Indices: NDSI, RSI, and DSI, formed by combining different spectral bands.
- Red-Edge Parameters: REP, Dr, Drmin, NDDr, RDr, DDr, and SDr, which capture spectral features in the red-edge region, known to be sensitive to plant physiological status.
- Inversion Modeling: Two modeling techniques were used:
- Partial Least Squares Regression (PLSR): A statistical method for analyzing multiple predictor variables.
- Backpropagation Neural Network (BPNN): A powerful machine learning algorithm capable of capturing non-linear relationships. Models were built using single spectral parameters, combinations of random two-band indices, combinations of red-edge parameters, and a comprehensive combination of random two-band indices and red-edge parameters.
Datasets
The study utilized hyperspectral imagery collected by a Cubert UHD 185-Firefly sensor mounted on a DJI Matrice 600 PRO UAV. Field samples of nitrogen concentration were collected from 92 'Red Fuji' apple trees in Qixia City, Shandong Province, China. These samples were divided into a modeling set (69 trees) and a validation set (23 trees) for model development and evaluation.
Key Insights and Ideas
- Thresholding for Canopy Extraction is Key: Effectively separating the apple canopy from bare soil and shadows using NDVI and NDCSI thresholds proved crucial for accurate spectral analysis, outperforming a pure RFC approach in refinement.
- Random Two-Band Indices Shine: These indices, particularly RSI, NDSI, and DSI, showed stronger correlations with CNC than traditional VIs and red-edge parameters, suggesting their utility in capturing nitrogen-related spectral features.
- BPNN Dominates for Complex Relationships: The BPNN model, especially when combining random two-band spectral indices and red-edge parameters, significantly outperformed PLSR, highlighting its ability to model the complex, non-linear relationships between spectral data and nitrogen concentration.
- Combination is Better than Single Parameters: Models built using combinations of spectral parameters consistently achieved higher accuracy than those using single parameters, underscoring the value of integrating diverse spectral information.
Key Results
- The optimal canopy extraction was achieved with NDVI thresholds of 0.65 and NDCSI thresholds of 0.45.
- Random two-band spectral indices exhibited strong correlations with CNC, with RSI (466, 542 nm) showing the highest at 0.72.
- The best inversion model was a BPNN that combined random two-band spectral indices and red-edge parameters. This model achieved a modeling coefficient of determination ($R^2$) of 0.77 and a root mean square error (RMSE) of 0.16.
- On the independent validation dataset, this optimal BPNN model yielded an $R^2$ of 0.75 and a residual predictive deviation (RPD) of 1.92, indicating good predictive ability.
- The generated spatial distribution map of CNC revealed variations across the orchard, with higher concentrations generally observed in the south and at the top of individual trees, likely due to differential light exposure.
Limitations and Future Work
- Canopy Complexity: While the study addressed canopy extraction, the inherent three-dimensional structure of apple trees and varying illumination conditions can still influence spectral data.
- Data Redundancy: Hyperspectral data, while rich, is also redundant, requiring complex processing that can be a challenge for large-scale applications.
- Future Work: The authors suggest that the developed methodology provides a theoretical basis and technical support for large-scale, rapid, and non-destructive monitoring of apple nutritional status. Future work could focus on optimizing processing for larger areas and exploring the transferability of these models to different apple cultivars or regions.
Summary updated