Cart

Drones v Satellites

What are the benefits of drone mapping when satellite data is freely available?

In this post we will compare, drones v satellites using NDVI/NDRE mapping. It also compares the NDVI/NDRE values to the ground truth data, including nitrogen and dry weight. This is a summary of research I carried out for my university dissertation, using the drone data from Leonardo’s absolute nitrogen project with DroneAG.

Data collection

The drone data was collected on winter wheat fields, during growth stages 31 (late April), 39 and 51 (both in late May). A MicaSense RedEdge multispectral sensor was used on the earlier date, mounted on a Storm Agri-Pro drone system. On the later date a Parrot Sequoia sensor was used, mounted on the same drone system.

Ground samples of the crop from randomly selected points within each field, were also collected on the same days. These were geolocated for later reference.

Satellite data

Landsat-8 is a NASA satellite which captures data in 7 bands at a ground resolution of 30 meters per pixel. It has a revisit time of 16 days. It doesn’t have a red-edge band, which limits its use as a comparison. 

Sentinel-2 is a European Space Agency satellite which captures data in 10 bands at varying resolution, the smallest being 10 meters per pixel. It has a revisit time of 5 days. The red, green, blue and NIR bands have a 10m resolution and the red-edge band has 20m resolution.

An issue with using satellite data is that it is heavily affected by cloud cover, due to them orbiting at high altitudes. A drone is not affected by cloud cover as it is mostly flown under the clouds. If the satellite collection date closest to the ground data collection date is cloudy, the next available date can be from a few days to a few weeks from the drone collection date.

Due to the high amount of cloud cover in the Landsat-8 data for GS39/51, only a comparison for the GS31 is used. The Landsat-8 data is taken from 9 days before the drone data. For Sentinel-2 data, GS31 is 4 days before and GS39/51 is 1 day before.

orthophoto

Data Processing

As I conducted the project while at uni, I couldn’t use Pix4D; which is what we use at DroneAG. I used Agisoft Photoscan (now Metashape). Photoscan is photogrammetry software (similar to Pix4D) which aligns the images taken using the drone together and then generates an orthomosaic. (It basically stitches all of the images together to produce one large image.)

An orthomosaic is created using the GPS coordinates and the drone’s altitude at the location the image is taking during the drone flight. The software generates approximately 40,000 key points and tie points, to tie the images together accurately. Due to the accuracy and the shared use of the processing computers this took a long time to process. Multispectral sensors produce a reflectance value for each band or colour. This must be calibrated before the flight and then calibrated during the processing into an orthomosaic for each band.

The satellite data comes at a high level of processing, so almost no extra processing was required. (Other than cropping the images down to the farm area to prevent the program from crashing due to the large file sizes.)

Indices

What Information do you get from the data?

The data was now ready for further processing using spectral indices to allow the comparisons. A spectral index (e.g. NDVI or NDRE) is calculated using the reflectance values of each band. It is a value usually on a scale from -1 to 1.

NDVI (normalised difference vegetation index) is the most common index used to indicate the density of vegetation. NDRE (normalised difference red edge) is a modification of NDVI using the red edge, instead of the NIR band. It helps to detect plant stress, such as nitrogen deficiency or disease.

How are the indices calculated?

I used a satellite and drone imagery processing software called Erdas Imagine.

NDVI = (NIR – Red) / (NIR + Red)

NDRE = (NIR – Red Edge) / (NIR + Red Edge)

Classification

To produce an image which could be used for a visual comparison and analysis, the indices had to be classified into a colour scale. This was done using ArcMap (a mapping analysis software).

An equal spacing for each field and index was used to allow a direct comparison. A scale of 0 to 0.95 was used for NDVI and <-0.2 to 0.75 was used for NDRE.  

Drones v Satellites- GS31

Drone

You can see how certain colour differences are highlighted in the same areas. This is due to there being patches of ground visible between plants, the map to the right shows this

  • Red areas indicate tramlines and patches of dead or unestablished ground.
  • Green areas indicate plants that are further along in their growth.
  • Yellower areas indicate a mid range value, possibly plants with disease or a nutrient deficiency.

A few patches of vegetation within the field are represented with a darker shade of green than the rest of the field. This is not necessarily outside the expected range, but could also indicate weeds. 

Landsat-8

The satellite imagery does not have the same level of detail as the drone imagery. This is due to the resolution being almost half the resolution for Sentinel-2 and Landsat-8 being almost as low as a quater. This is why tramlines are not shown as clearly.

There are some similarities between the drone NDVI image and the Landsat-8, low areas are similarly recognised. For example, the blue circle is identified with a mid value (yellow), when compared to the same area the drone provides a refined detail of some low (red) and some mid values (yellow).

Sentinel-2

Sentinel-2 is better at identifying similar areas to the drone than Landsat-8 is, due its improved resolution. It has identified the same low values in the NDVI as the drone, shown by the red/yellow areas. This is very clear in the fields Greymere and Big Dunsdale. There are large areas of low values, see the blue circles as a comparison. This shows that there are benefits of Sentinel-2 in helping to identify key areas of disease or poor crop health. 

Drones v Satellites- GS39/51

NDVI v NDRE

Drone

The NDRE index is typically better than NDVI in the later growth stages of crops. This is because the red band in NDVI is mostly absorbed by the leaves at the top of the plant. These leaves are often more healthy and green than the lower leaves, which causes the NDVI value to be high. Whereas NDRE uses the red-edge band which is not absorbed as strongly by the top leaves. This means that the NDRE value is a representation of more of the plant rather than just the top. The two areas circled in blue in the images to the left highlight this. The NDVI map shows areas with high values, where they are lower in the NDRE map.

Sentinel-2

The NDVI maps show an increase in the areas of green for GS39/51, due to the growth of the crop. The low areas are now more easily identifiable due to the higher contrast. Again the drone maps show a higher detail, highlighting more specific areas of no crop or crop with poor health.

For example in Sentinel-2, the top of the field is generalised as being low. Whereas the drone map shows the more specific areas with varying low NDVI values. This can be seen within the blue circles.

Drones v Satellites – A Statistical Comparison

To provide a statistical comparison between the indices from the drone and satellite data, I generated 100 random points per field, using a process in arcMap. I then generated a few different statistics to provide a comparison between the values from the drone and the satellites. The statistics which compare the accuracy the best are the RSME (root mean square error) and the r2 value (or coefficient of determination).

RSME is a value which measures the difference between the index values of the satellites and the values of the drone, this therefore shows how much error there is between two datasets.

An r2 value in simple terms, is how statistically similar values in the two datasets are (using a simple linear regression model). It gives a value between 0 and 1, with 0 being no similarity and 1 being identical, generally a value of above 0.6 is considered as showing similarity between the datasets.

Drone v Landsat-8

The statistics comparing the drone and Landsat-8 NDVI map, shows the differences between them. The r2 value indicates that there is a low similarity between the values.  This shows that there is no significance of using Landsat-8 imagery when comparing to drone imagery at a higher resolution.

NDVI
r20.51
RSME0.14
Max Difference0.20
Min Difference-0.06
Mean Difference0.11
Drone v Sentinel-2

In GS31 the r2 values show that there is a high positive relationship between both the NDVI and NDRE maps, due to the value being above 0.7. However, this decreases in the later growth stage. The other statistical values reflect this, such as the RSME value increasing for the later growth stages.

GS31GS39/51
NDVINDRENDVINDRE
r20.790.710.630.56
RSME0.070.070.130.26
Max Difference0.120.050.270.40
Min Difference-0.1-0.160.020.09
Mean Difference0.030.060.120.25

Drones v Satellites- NDRE v Nitrogen

GS31

I compared the NDRE to the points with a nitrogen value of less than 40Kg/ha. NDRE is good at showing nitrogen deficiency due to the change in the reflectance of the red edge band. This helps to estimate and identify the chlorophyll content which then indicates the nitrogen content. In the drone image to the left you can see that most of the points are identified in a lighter shade of green or by pale yellow. 

When comparing the same points on the Sentinel-2 map, the outcome is similar with most points showing as a light green or yellow. However, as you can see in the image due to the lower resolution of the satellite data, the identification of a single point is less accurate if the point is near the corner of a pixel, as you have to decide which pixel the point represents.  On the other hand, when the point is in the center of a pixel, it provides the benefit of making it easier to identify the point’s value without having to zoom in or use any software to pin point the exact pixel value.

GS39/51

In the later growth stages, the drone NDRE map also identifies some of the areas of low nitrogen (less than 100T/ha). The Sentinel-2 image identifies a few of the same points with a low NDRE value, shown by the blue circles. Again, some points have been identified as higher values than expected, for example, the point on the far right. However, this could be due to the point being near the edge of the pixel. 

drones v satellites, NDRE v Nitrogen drones v satellites NDRE v Nitrogen

Statistics

The graphs highlight that there are some similarities between the nitrogen and NDRE values. Which is shown by the clustering of the points near the line. Sentinel-2 has slightly higher NDRE values than the drone. This is shown by the movement of the points and trend line upwards in the comparison. It also shows a general trend- the higher the nitrogen content, the greater the NDRE value. 

Drones v Satellites- NDVI v Dry Weight

GS31

I have compared the NDVI images to the dry weight ground values, to see if NDVI is accurate at determining the areas of poor biomass. This had some interesting results- some areas with a dry weight value of less than 1 have a low NDVI value but not all do. This is demonstrated by the two blue circles, the circle at the top shows a high NDVI value yet the circle at the bottom shows a low value.

The Sentinel-2 NDVI map produced a similar result to the drone. The same areas have been identified as a similar NDVI value/colour. For these points it is easy to identify what colour/value the points represent as they are in the center of a pixel.

Drone v Sentinel-2

satellites v drone, NDVI v dry weight

GS39/51

For GS39/51, the results are similar for the drone, as shown by the image. However it is hard to distinguish the low values. This is because the crops are at a later growth stage so therefore have a thicker canopy. The satellite imagery shows similar results to the drone imagery. It identifies the same points with a similar high/low value. The two blue circles show this, the left shows a mid range value and the right shows a high value. Again there are some uncertainties when identifying which pixel represents the point. 

Statistics

The graphs show that there is a lower similarity between NDVI and dry weight, than there is between NDRE and nitrogen. Which can be seen by the higher dispersion of points from the line. The accuracy of the drone compared to Sentinel-2 is highlighted by the difference in the clustering/dispersion of the points around the line. The drone has detected higher NDVI values than Sentinel-2.

Drones v Satellites NDVI v dry weight drones v satellites NDVI v dry weight

Drones v Satellites – Key Findings

1

High resolution NDVI/NDRE maps can show the parts of fields which are bare or have poor crop health.
2

Landsat-8 is not suitable for this application in agriculture, as it does not contain a red-edge band and has a poor resolution.
3

NDVI is a good measure of the biomass/dry weight at early growth stages but not at later growth satges.
4

NDRE maps show a measure of the nitrogen content.
5

Satellites are affected by cloud cover, limiting the days available.
6

Sentinel-2s accuracy decreases at later growth stages.

Summary

  • Satellite imagery is beneficial when used alongside drone imagery, it can be used with indices like NDVI or NDRE to identify key areas in the crop cover. These areas can then be plotted using skippy scout for further investigation.
  • There are historical records of the satellite data which can be used to help identify any trends over the years within a field.
  • This shows the benefits of the Phantom 4 Multispectral, which will produce both NDVI and NDRE maps along with other indices.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.