Precise Estimation of NDVI with a Simple NIR Sensitive RGB Camera and Machine Learning Methods for Corn Plants

Liangju Wang, Yunhong Duan, Libo Zhang, Tanzeel U. Rehman, Dongdong Ma, Jian JinSensors 2020 · DOI

TL;DR

This paper introduces "Ncam," a low-cost, NIR-sensitive RGB camera system combined with machine learning, capable of accurately estimating NDVI and plant nitrogen content. This approach offers a more accessible alternative to expensive multispectral and hyperspectral cameras for plant phenotyping.

What Problem Does This Paper Address?

The Normalized Difference Vegetation Index (NDVI) is a crucial metric for monitoring plant health and growth. Traditionally, calculating NDVI requires specialized multispectral (MSC) or hyperspectral (HSC) cameras that capture near-infrared (NIR) and red light. However, these cameras are prohibitively expensive, bulky, and demand specialized expertise for operation and data processing. While modified RGB cameras (RGBN) have been explored, their NDVI estimates often fall short of the accuracy achieved by MSC/HSC. This gap limits the widespread adoption of precise plant monitoring technologies, especially for smaller research groups or commercial applications.

Methods

The core innovation is the "Ncam" system, a low-cost ($70-$85) imaging setup built around a Raspberry Pi Noir camera (which is sensitive to NIR light) and a microcontroller. Two versions of Ncam were developed:

  • Ncam lp: Utilizes a long-pass filter to block visible light below 580 nm, allowing it to capture red and NIR reflectance.
  • Ncam tp: Employs a two-band-pass filter to specifically capture narrow bands around 640 nm (red) and 850 nm (NIR).

To overcome the limitations of traditional NDVI calculation with RGBN cameras (where red and NIR signals can be mixed within channels), the researchers leveraged machine learning. They trained models (Multiple Linear Regression - MLR, Gaussian Process - GP, and Support Vector Regression - SVR) using preprocessed RGB values from Ncam and compared their performance against ground truth NDVI data from a high-end hyperspectral camera. This approach allows the models to learn the complex relationship between the camera's sensor responses and the actual NDVI, rather than relying on a simplified formula.

Datasets

The study used data collected from 56 corn plants grown under controlled greenhouse conditions with varying nitrogen and water treatments. The plants were imaged at the V6 stage.

  • Ground Truth Data: Hyperspectral images from an MSV 500 HSC were used to derive the "ground truth" NDVI values (NDVIHSC).
  • Ncam Data: Images from both Ncam lp and Ncam tp systems were captured simultaneously with the HSC.
  • Nitrogen Content Data: Ground truth nitrogen content measurements for corn leaves were obtained using a FlashEA 1112 Nitrogen and Carbon Analyzer.

Key Insights and Ideas

  • Machine Learning Bridges the Gap: The primary insight is that machine learning, specifically regression models, can effectively compensate for the spectral mixing issues inherent in RGBN cameras, enabling accurate NDVI estimation. This moves beyond simple formula-based approaches.
  • Filter Choice Matters: The Ncam tp system, with its two-band-pass filter, demonstrated superior performance in NDVI estimation compared to Ncam lp (with a long-pass filter). This suggests that more precise spectral band selection is critical for accurate NIR and red signal capture.
  • Beyond NDVI: The study highlights that Ncam isn't limited to NDVI. It also shows promise in predicting plant nitrogen content, a critical indicator of plant health and nutrient status.
  • Cost-Effectiveness is Key: The paper strongly emphasizes the affordability and accessibility of the Ncam system, making advanced plant phenotyping tools available to a broader range of researchers and agriculturalists.

Key Results

  • NDVI Estimation: The Ncam tp system, when paired with MLR, achieved a high NDVI estimation performance with an R² of 0.96 and an RMSE of 0.0079, closely matching the hyperspectral camera's results.
  • MLR vs. Typical Formula: Machine learning models (especially MLR) consistently outperformed the traditional NDVI calculation formula using RGB channels, achieving higher R² values (6-7% improvement for Ncam lp and Ncam tp, respectively).
  • Nitrogen Content Prediction: Ncam lp demonstrated high accuracy in predicting nitrogen content (R² = 0.94), comparable to the HSC, while Ncam tp showed lower accuracy (R² = 0.84). This difference is attributed to Ncam lp's filter capturing the red edge region, which is important for nitrogen estimation.
  • System Advantages: Ncam offers significant advantages over HSC, including faster imaging speed (less than 1 second vs. 1 minute), a more compact structure, and much lower cost.

Limitations and Future Work

  • Field Applicability: The current study was conducted in a controlled greenhouse environment. The authors acknowledge that NDVI measurements can be significantly impacted by light changes and weather conditions in the field. Future work should involve field testing to adapt Ncam parameters and machine learning models for real-world conditions.
  • Nitrogen Prediction Scope: For nitrogen content prediction, the authors suggest that more nitrogen level treatments and a wider variety of crop species are needed to fully evaluate Ncam's performance.
  • Comparison with Standard RGB: The paper proposes a future study to compare the performance of a standard RGB camera (without NIR sensitivity) against the NIR-sensitive Ncam for NDVI estimation, especially when combined with advanced algorithms like genetic algorithms.
  • Generalized Formula: The researchers aim to develop a generalized formula for RGBN cameras as an alternative to the typical NDVI formula, based on their machine learning models.

Summary updated

Abstract

The normalized difference vegetation index (NDVI) is widely used in remote sensing to monitor plant growth and chlorophyll levels. Usually, a multispectral camera (MSC) or hyperspectral camera (HSC) is required to obtain the near-infrared (NIR) and red bands for calculating NDVI. However, these cameras are expensive, heavy, difficult to geo-reference, and require professional training in imaging and data processing. On the other hand, the RGBN camera (NIR sensitive RGB camera, simply modified from standard RGB cameras by removing the NIR rejection filter) have also been explored to measure NDVI, but the results did not exactly match the NDVI from the MSC or HSC solutions. This study demonstrates an improved NDVI estimation method with an RGBN camera-based imaging system (Ncam) and machine learning algorithms. The Ncam consisted of an RGBN camera, a filter, and a microcontroller with a total cost of only $70 ~ 85. This new NDVI estimation solution was compared with a high-end hyperspectral camera in an experiment with corn plants under different nitrogen and water treatments. The results showed that the Ncam with two-band-pass filter achieved high performance (R2 = 0.96, RMSE = 0.0079) at estimating NDVI with the machine learning model. Additional tests showed that besides NDVI, this low-cost Ncam was also capable of predicting corn plant nitrogen contents precisely. Thus, Ncam is a potential option for MSC and HSC in plant phenotyping projects.
BibTeX
@inproceedings{wang2020precise,
  author = {Liangju Wang and Yunhong Duan and Libo Zhang and Tanzeel U. Rehman and Dongdong Ma and Jian Jin},
  title = {{Precise Estimation of NDVI with a Simple NIR Sensitive RGB Camera and Machine Learning Methods for Corn Plants}},
  booktitle = {Sensors},
  year = {2020},
  doi = {10.3390/s20113208},
}
Figure