Potato SPAD Estimation via UAV Hyperspectral & Machine Learning
author: Callum
2025-03-11
Introduction
Potato is China’s fourth-largest food crop and plays a crucial role in ensuring food security. Chlorophyll (Chl-a) is a key biochemical parameter in crops, essential for light absorption during photosynthesis. Its content reflects potato photosynthetic capacity and nitrogen status. Traditional chlorophyll measurement involves field sampling and chemical analysis, which are labor-intensive, destructive, and time-consuming.
The SPAD-502 portable chlorophyll meter provides a non-destructive method for leaf chlorophyll measurement. However, as a point-based tool requiring multiple measurements across different leaves, it is impractical for large-scale monitoring.
To enable real-time chlorophyll monitoring on a regional scale, remote sensing offers a rapid, non-destructive, and efficient alternative. UAV-based hyperspectral imaging has gained importance in crop biochemical parameter monitoring due to its high spatial resolution, timely data acquisition, and cost-effectiveness. Compared to multispectral and thermal imaging, hyperspectral imaging provides more spectral bands and higher spectral resolution, capturing subtle crop variations.
However, full-spectrum modeling introduces redundant spectral data, affecting model accuracy. Feature selection algorithms, such as Competitive Adaptive Reweighted Sampling (CARS), Uninformative Variables Elimination (UVE), Successive Projections Algorithm (SPA), and Random Frog (RF), improve model efficiency by identifying sensitive spectral bands. Selecting the optimal feature selection method enhances UAV-based hyperspectral chlorophyll estimation models, improving accuracy and stability.
Technical Approach and Key Content
2.1 UAV Hyperspectral Data Acquisition and Preprocessing
A DJI M350 UAV equipped with the ATH1010 hyperspectral camera (developed by Optosky) was used for hyperspectral imaging. The UAV has a maximum payload of 5 kg, a takeoff weight of 15.5 kg, a flight speed of up to 4 m/s, and a flight endurance of approximately 50 minutes. The ATH1010 camera, weighing <700 g, covers the 400–1000 nm spectral range with a spectral resolution of 1.7 nm, capturing over 400 bands.
Data collection was conducted under clear, windless, and cloud-free conditions from 11:00 to 13:00, with the UAV flying at 100 m altitude and 4 m/s speed. Waypoint navigation was manually planned, utilizing hover scanning mode with an 80% forward overlap and 70% side overlap. The camera was set to nadir view with a 38° FOV, achieving a ground resolution of 0.039 m.
After data acquisition, OptoskyHIPS software was used for lens calibration, reflectance calibration, and atmospheric correction, generating calibrated hyperspectral images. The images were then stitched in OptoskyHIPS and further processed in ENVI 5.3 using Subset Data from ROIs to extract target regions. A 0.2 m radius ROI was established around sampling points, and the mean spectral reflectance within each ROI was computed as the potato spectral reflectance.
Figure 1 Administrative Map of Gansu Province
2.2 SPAD Measurement
SPAD measurements were conducted synchronously with UAV hyperspectral imaging during the potato tuber formation (July 29, 2022) and tuber expansion (August 19, 2022) stages. The SPAD-502Plus handheld chlorophyll meter was used to measure SPAD values. For each sampling point, 10 fully expanded healthy potato leaves were selected, and the average SPAD value was recorded. Meanwhile, real-time kinematic (RTK) positioning was used to log sampling point coordinates. A total of 100 SPAD data sets (50 per stage) were collected.
2.3 Hyperspectral Data Processing and Feature Selection
To reduce background interference and enhance spectral-data correlations, the raw spectral reflectance (R) was transformed into its inverse reflectance (1/R) and logarithmic inverse reflectance [log(1/R)], forming the basis for SPAD modeling.
Feature selection is essential for handling high-dimensional hyperspectral data, extracting relevant information, and improving model efficiency. Three algorithms—Competitive Adaptive Reweighted Sampling (CARS), Uninformative Variables Elimination (UVE), and Random Frog (RF)—were used for sensitive spectral band selection.
CARS, integrating Monte Carlo sampling and PLS regression, sequentially evaluates and eliminates spectral wavelengths, making it highly effective for high-dimensional spectral feature selection.
UVE, based on PLSR regression coefficients, introduces a variable stability index to filter out uninformative variables, enhancing feature selection accuracy.
RF, a weighted regression-based feature selection method, removes irrelevant spectral bands while improving model prediction performance and robustness.
2.4 Model Construction and Evaluation Metrics
To estimate potato SPAD values, three machine learning algorithms were applied: Partial Least Squares Regression (PLSR), Support Vector Regression (SVR), and Backpropagation Neural Network (BP-NN).
PLSR is widely used for high-dimensional spectral regression, effectively handling collinearity while extracting meaningful spectral features and minimizing noise.
SVR applies kernel functions to transform input data, optimizing the regression function by minimizing errors.
BP-NN, a multi-layer feedforward neural network, captures nonlinear relationships between spectral data and SPAD values.
Model performance was evaluated using the coefficient of determination (R²) and root mean square error (RMSE), where higher R² and lower RMSE indicate better model accuracy and reliability.
Conclusion
Using the 1/R-CARS-SVR model, spatial inversion mapping of potato SPAD values was performed for two growth stages (Figure 6). The color gradient in the map represents different SPAD levels. The SPAD values ranged from 40 to 50, with:
45% of the study area showing SPAD values between 44–46 during the tuber formation stage
43.5% of the study area showing SPAD values between 46–48 during the tuber expansion stage
Overall, SPAD values were higher during the tuber expansion stage than in the tuber formation stage. This inversion mapping approach effectively visualizes potato growth conditions across different areas, providing valuable insights for agronomic decision-making.
Hyperspectral Prediction of Potato SPAD
This study focused on Gansu Province, China, one of the major potato-producing regions, to assess the capability of UAV-based hyperspectral remote sensing in SPAD estimation. Previous studies have shown that mathematical transformations of raw spectra enhance spectral sensitivity to vegetation chlorophyll, and our results confirm this.
Figure 2 Potato Prediction Results
Using full-spectrum modeling across three machine learning algorithms (PLSR, SVR, BP-NN):
1/R transformation improved R² by 0.28, 0.34, and 0.46 and reduced RMSE by 0.21, 0.24, and 0.35 compared to the original reflectance (R).
log(1/R) transformation improved R² by 0.16, 0.48, and 0.45 and reduced RMSE by 0.13, 0.37, and 0.33.
These findings highlight that spectral transformation techniques combined with UAV hyperspectral imaging and machine learning can significantly enhance potato SPAD estimation accuracy, optimizing precision agriculture applications.
Product Introduction | ATH9030 UAV‑Based Hyperspectral Real‑Time Imaging System: Get Answers While in Flight
UAV Hyperspectral Detection of Citrus Huanglongbing
Related Article
Discover how UAV‑based hyperspectral imaging and deep learning can monitor SDI, TOC, and TEP – key membrane fouling indicators for seawater desalination during harmful algal blooms. Based on a recent Water Research study, this article explores how Optosky's ATH9010 enables spatial risk mapping and proactive intake management.
Applications | ATH9010 Enables Water Quality and Membrane Fouling Monitoring for Desalination During Algal Blooms
Choosing between 785 nm and 1064 nm for Raman? This guide explains the physics of fluorescence, compares signal strength and fluorescence suppression, and provides a step‑by‑step decision tree. Real‑world drug detection and pesticide examples show why wavelength matters – and how dual‑wavelength coverage offers the ultimate solution.
785 nm or 1064 nm? Fluorescence Is the Dividing Line
Inspired by the 2026 Science paper from Zhang Jun’s team on a video‑rate on‑chip hyperspectral microsystem, Optosky’s ATH series moves beyond lab‑only tools. We deliver tailored, real‑world hyperspectral systems that integrate compact hardware, edge AI, and application‑specific algorithms—turning raw spectral data into actionable decisions, right where you need them.
From Science to the Field: Custom Hyperspectral Solutions by Optosky
See how 40 Optosky NY3300Pro multispectral units were deployed across a large‑scale demonstration farmland in Northwest China. With 9‑band imaging, solar power, 4G transmission, and automated data analytics, this solution enables precise growth monitoring, early pest detection, and water‑fertiliser optimisation – a true leap toward smart agriculture.
Case Study | Bulk NY3300Pro Installation Transforms Smart Farming