Spectral imaging is a key method for high-throughput phenotypic analysis, which can be used to assess various biological parameters. The Normalized Difference Vegetation Index (NDVI), calculated using specific wavelengths, is commonly used to analyze the health of crops. Developing a compact, low-cost, and easy-to-install spectral imaging system is of great significance in precision agriculture. This article describes a method using a dual-camera system connected to a Raspberry Pi to generate NDVI images, referred to as NDVIpi. By calibrating image reflectance using spectral reference targets and then calculating NDVI, its accuracy is improved compared to systems using a single reference/standard. The NDVIpi imaging system demonstrates robust performance compared to standard spectral methods and is compared with a relatively more expensive commercial camera (Micasense RedEdge), with both cameras showing comparable performance in measuring NDVI. There are differences in NDVI values between NDVIpi and RedEdge, possibly due to the different wavelengths used in NDVI calculations by each camera. The results indicate that the wavelengths used by NDVIpi exhibit higher sensitivity to changes in chlorophyll content than those measured by RedEdge. We propose a Raspberry Pi-based NDVI imaging system that utilizes low-cost, off-the-shelf components and a calibration method that can accurately measure NDVI values with strong robustness. Despite being only a fraction of the cost of commercial systems, our system achieves precise NDVI values. Our results also emphasize the importance of selecting red wavelengths in NDVI calculations, which leads to differences in sensitivity between different camera systems.

Figure 1: Schematic diagram of the connection of Raspberry Pi Compute, PiCamera, and NoIR PiCamera, along with the metal camera bracket. The camera bracket keeps both cameras in the same plane, allowing for good alignment of the captured images.

Figure 2: Raspberry Pi calibration plate, composed of six diffuse reference materials with known relative reflectance.

Figure 3: Calibration process, showing each step from captured images to generated output NDVI images, with dashed arrows indicating optional steps.

Figure 4: Example data. A: Calibrated raw numbers from images directly captured by the camera, using known reflectance of the six calibration materials to find the relationship between numbers and reflectance. Once the relationship is established for the image, the raw numbers of the entire image can be converted to reflectance values. B: Normalization of reflectance data by rescaling the converted reflectance values (from 0 to 100%) to utilize the full 16-bit (0–65535) range. C, D: Typical NIR images, showing visual differences between images before (C) and after (D) calibration and normalization processes.

Figure 5: A: RGB image of legume plants in a greenhouse. B: Corresponding color NDVI image. C: RGB image of legume leaves taken under controlled lighting. D: Generated NDVI image, selecting a color rendering scheme visually corresponding to the RGB image.

,

Figure 6: Comparison of NDVI values calculated using a spectrometer versus those calculated using Raspberry Pi. The red dashed line represents the ideal curve, with wheat (yellow), barley (blue), and legumes (green). In A, n=181, while in B, legume dataset is omitted with n=59. The NDVI values calculated from Raspberry Pi images show a good relationship with those calculated using a spectrometer (R2 > 0.89).

Figure 7: A series of wavelengths calculated NDVI values measured by a spectrometer to determine which wavelengths calculated NDVI values are closest to those obtained in Raspberry Pi images. The percentage (%) difference between the two NDVI values is measured. In Figure A, the visible red wavelength range for NDVI is 600-700nm, with NIR wavelength fixed at 750nm. In Figure B, the NIR wavelength range is 700-800nm, with visible red fixed at 620nm. The results indicate that the NDVI value calculated using 620nm (red)/ 750nm (NIR) matches best with the NDVI obtained in Raspberry Pi images (n = 150).

Figure 8: NDVI images of legume leaves taken in greenhouse under ambient lighting. Figure A is the NDVI image from the NDVIpi system; Figure B is the NDVI image from Micasense RedEdge.

Figure 9: (Black dots) NDVI measured with Raspberry Pi, (Red dots) NDVI measured with Micasense RedEdge, compared with NDVI calculated using 620nm wavelength (red) and 750nm wavelength (NIR).

Figure 10: Relationship between NDVI calculated with NDVIRaspPi and NDVIMicasense measured by the spectrometer.

Video source from Plant Phenotyping Circle

Click here to read the original article