[Test] DJI Phantom 4 Multispectral vs. Micasense RedEdge MX
Comparative test of DJI’s new multispectral camera drone with industry-standard Micasense RedEdge MX.
Multispectral cameras specifically designed for agricultural use have countless features that distinguish them from a conventional camera. For example, they are able to record different wavelengths of light reflected from leaves, thus we can diagnose the current physiological condition of the culture.
The most frequently used cameras in agriculture are conventional (RGB) cameras because they make it easy to create physical condition assessment of plants. These cameras are the most suitable devices to answer the question of whether there is a plant, how many pieces are in an area, how much damage caused by wildlife. However, it is also possible to achieve much more forward-looking and precise solutions so that we can also examine a culture from a physiological point of view.
The spread of these methods are in early phase than that of analyses based on conventional cameras. However, it is not difficult to see that the future will in all respects belong to devices that can simultaneously collect data on which physical and physiological analyses can be based (RGB and multispectral camera at the same time).
Differentiated application and close monitoring of vegetation allow us to use countless intervention methods that were not previously known to be necessary. Nutrient application, weed control, spraying are based on remote sensing data with high reliability, and these data come from either satellites or drones. The need for accurate and reliable data sources and their professional processing will become paramount. In recent years, the number of multispectral camera solutions available on the market, whether they are single or multi-lens sensors, calibrated or uncalibrated devices, has grown significantly.
AGRON pays close attention to these devices in order to make our self-developed processing and testing system available with as many reliable drones / sensors as possible.
Accordingly, we regularly conduct tests and professional comparisons in order to select instruments from the market that provide reliable and accurate results in conjunction with AGRONmaps. In our most recent test, we placed the DJI Phantom 4 Multispectral drone camera on the autopsy table, which was subjected to depth testing with professional examinations and statistical analysis of remote sensing data.
The basis of comparison
Our company has been using the Micasense RedEdge-MX multispectral camera for more than a year, during which time device underwent detailed tests at the Martonvásár research institute, based on set experiments. The accuracy of the data collected by the camera was verified by more than 5,000 manual measurements and an outstanding correlation was found for chlorophyll content. The device met our pre-formulated expectations. Outstanding distortion-free optics, design, calibratability, the presence of a light sensor and separate cameras applied per band all provided a perfectly reliable solution.
The DJI Phantom 4 Multispectral has a much friendlier price, moreover has promising bands and sensor data, that can bring close to everyone a good value-for-money data collection tool for physiological analyses.
The Phantom 4 drone is a sophisticated and reliable device, so we will not discuss it in our article, we will only examine the features and usability of the multispectral camera, thus we performed various kind of tests. Accordingly, the comparison will be made with the already proven RedEdge-MX multispectral camera.
Before performing more detailed tests, it is worth getting a little familiar with the properties of the sensors. The pixels of both sensors are read at the same time (global shutter), which is one of the basis of distortion-free mapping. The RedEdge-MX is a narrow-angle (47.2 °) camera that produces distortion-free and sharp shots, but you have to fly a relatively large amount to map a field. The Phantom 4 Multispectral camera has a wide viewing angle (62.7 °), which results in slightly more distortion and thus less sharp shots. On the plus side, you have to fly less because of the wide viewing angle.
A multispectral sensor will be a good sensor if it has more optics and the light filters built into the optics are as specific as possible (Figure 1, Table 1). In this respect, the benchmark is the RedEdge-MX, which is actually being developed for research purposes, making it particularly sensitive and accurate. This camera ‘sees’ exactly the wavelength ranges needed to collect accurate physiological data. In contrast, the Phantom 4’s Multispectral camera also scans similar ranges, but is less sensitive, as the light filter built into the optics allows a wider wavelength range into the sensor.
Figure 1. Comparison of wavelength ranges covered by Micasense RedEdge-MX and DJI Phantom 4 Multispectral sensors. (Values in Table 1.)
Table 1. Wavelength ranges covered by Micasense RedEdge-MX (REMX) and DJI Phantom 4 Multispectral sensors
How accurate data can be collected with the Phantom 4 Multispectral camera? Based on the results, this is a good tool, but proper calibration of the recordings plays a key role.
The study area
A full day was set aside for the study, which was based on a Micasense RedEdge MX running on a DJI Matrice 210 V2 platform. The multispectral Phantom 4 was provided by one of our key partners, Tokaj Viridis. Measurements were performed at multiple flight altitudes with both equipment at nearly the same time. We used the same flight plan for each flight.
Figure 2. Flight plan programmed for the experimental area in case of RedEdge-MX (left) and Phantom 4 Multispectral (right).
The test was performed on a 5-hectare winter wheat crop that was heterogeneous enough to be suitable in all respects for the study. After all, the greater the standard deviation of the data, the easier it is to assess the capabilities of a sensor. Significant rooting damage was observed on the field. In addition, the effects of drought stress caused by significant drought in the recent period have been seen on the plants, so the health status of the crop proved to be quite low according to the AGRON’s classification (4, on a scale of 1 to 10).
A total of 207 (15x15m) sampling zones were defined on the map from the field, which were examined on five differentally calibrated recordings. The entire database thus consists of 1035 data points. The data points have 27945 values combined from the raw data and the calculated vegetation indexes.
Figure 3. Table defined in Agron Maps software
Figure 4. Composite RGB (left) and GNDVI (right) orthomosaics of the examined field. The GNDVI image were colored with the standard look up table (LUT) approved by AGRON.
The importance of light sensor and calibration
In plant physiology analyses, we work primarily with numerical information, not visual ones. We collect data from the leaf surface, thus we want to know as much information about the plant as possible and assess its stress status, for example. This value is greatly influenced by the strength of the irradiation coming from the sun to the leaves and the amount reflected. How would we know what ‘comes back’ from a leaf surface and what they absorb or let through if we don’t know the amount of light? So the key is to have a light sensor that captures incident light.
However, this is only one side of the coin. The other is to measure the amount of incident light so that we can normalize the data. In this way we can record the state of the given global brightness and compare it with the valuable results of later measurements. It will not matter, therefore, that after two weeks we recorded the data under different light conditions, because with the help of normalization the data remain comparable.
To illustrate this, we performed a principal component analysis (PCA) that reveals the changes in the data by exploring the data structure in the presence or absence of different calibrations (Figure 5). The article is basically about the Phantom 4 Multispectral, so we illustrate the importance of calibration with the data from this sensor. Data were compared to a calibrated and normalized recording with a RedEdge-MX light sensor and a Micasense Calibration Panel.
Figure 5. Comparison of raw data collected by RedEdge-MX (REMX) and Phantom 4 Multispectral (P4M) using principal component analysis. The recording taken by REMX was also calibrated with a calibration panel and a light sensor. (F – recording calibrated with light sensor; K – recording calibrated with calibration panel; KF – recording calibrated with calibration panel and light sensor; nK – recording uncalibrated)
In Figure 6, the blue dot set is spectacularly separated because that image was not calibrated against the others. Remarkably, the values of the recording calibrated with a light sensor and a calibration panel (green dots) approach the values of RedEdge-MX as opposed to the data of the recording calibrated with a light sensor only (pink dots). Data calibrated with the calibration panel only (greenish-brown dots) fall between these two.
A total of 11 indexes were examined in the statistical analysis. For each data set, we calculated and examined how their values are affected by calibration. We obtained a similar trend as for the raw data. As a result of full calibration, the indices calculated from the Phantom 4 Multispectral data came very close to the index values calculated from the Red-Edge-MX data. However, the lack of calibration resulted in unusable dataset.
Figure 6. Comparison of index values calculated from data collected by RedEdge-MX (REMX) and Phantom 4 Multispectral (P4M) using principal component analysis. The recording taken by REMX was also calibrated with a calibration panel and a light sensor. (F – recording calibrated with light sensor; K – recording calibrated with calibration panel; KF – recording calibrated with calibration panel and light sensor; nK – recording uncalibrated)
Based on the above, we can state as a short summary that proper calibration is essential for performing precise physiological studies. This is particularly important as precision methods are emphasized with the development of agricultural methods. For this reason, plant protection treatments, for example, will be based on camera data. An improperly calibrated recording and a treatment plan elaborated from it can cost a lot later!
The data were also subjected to a correlation analysis that provided an easier-to-interpret result, in which Pearson’s correlation coefficients were determined (Table 2). This study shows the relationship between two data sets, for example in an understandable way, how similar they are and how much one can express one with the other. Values can range from -1 to 1 (-1 strong negative correlation, 0 no correlation, 1 strong positive correlation).
The differently calibrated values of the Phantom 4 Multispectral were compared to the fully calibrated data of RedEdge-MX. The weakest correlations were given by the uncalibrated recording (nK), where zero or weak positive correlations were obtained. Calibration can greatly improve the situation. Recordings separately calibrated with the calibration panel and light sensor show weak, medium, and strong positive correlations along a similar trend. The strongest correlation was measured on the recording calibrated by both methods. This result also illustrates the importance of proper calibration.
Table 2. Correlation analysis of RedEdge-MX (REMX) and Phantom 4 Multispectral (P4M) sensor data and the most common indexes calculated from them. The recording taken by REMX was also calibrated with a calibration panel and a light sensor. Colors vary along a defined color gradient from -1 to 1. (F – light sensor calibrated recording; K – calibration panel calibration recording; KF – calibration panel and light sensor calibration recording; nK – uncalibrated recording; FR – far red; NIR – near infrared)
The indices calculated from the data collected by Phantom 4 Multispectral, when properly calibrated, are very similar to the data from a camera designed for research work. However, it is important to note that due to the different sensitivity of the light filters, there is such a difference that the data of the two cameras are not completely compatible. This is well illustrated by the fact that the correlation of near-infrared data is very high, as practically the two cameras “see” almost the same band (Table 1). However, in the case of the other filters, there are such differences or the course of the reflectance spectrum is so variable that the correlation coefficients also decrease visibly. However, this phenomenon does not dramatically affect vegetation indexes with normalized values.
Overall, Phantom 4 Multispectral is perfect for practical use and a good value for money alternative. However, the parallel use of the two sensors is not recommended in research work. During the analyzes, we could see what capabilities the Phantom 4 Multispectral camera has. The sensor is basically a good construction. AGRON’s professional tests can be fully performed with it.
However, it is recommended that you obtain a calibration panel to achieve proper results. One of the best panel on the market is manufactured by Micasense, which is perfect for any other multispectral sensor. It is needed to determine the calibration values of the panel for each multi-lens multispectral camera that AGRON can do. However, wide-angle optics make mapping more difficult. These types of lenses have a so-called edge distortion. This is typically seen in a thin band at the edges of the shots. As a result, the resulting map may be blurred or smeared. In extreme cases, the software cannot even generate an orthomosaic. For this reason, great attention must be paid to flight planning, as the camera is sensitive to changes in individual flight parameters. With proper design, erroneous recordings can be almost completely eliminated. The optimal programming of the drones can be mastered during the training of the AGRON agricultural monitoring drone pilot program.
Following the test, the DJI Phantom 4 Multispectral has been added to the AGRON’s list of trusted devices, which will be available on AGRON Maps.
AGRON Analytics Ltd.
All rights reserved!