Hybrid Edge/Cloud Data Architectures for BVLOS Drone Scalability in Agriculture
As commercial adoption of drones began to take off in the mid-2010s, many thought agriculture would be the fastest-growing market segment.
This impression was bolstered by service providers who saw an opportunity to strap a GoPro to a drone, fly fields for farmers, and reinvent global agriculture in the process. But the reality has been more complex, and such predictions have largely failed to materialize.
Though drones have become more prevalent in agriculture, further integration requires challenging technical solutions — making wide-scale adoption a bigger leap and investment than many had anticipated.
While technology is rapidly advancing and the costs of systems are decreasing, one major hurdle remains: Line-of-sight requirements for drone operations.
Overcoming the hurdles placed by line-of-sight requirements is both a regulatory and technological challenge.
In many places in the world, laws stipulate that there must be one pilot per aircraft, and the aircraft must be flown within the pilot’s line of sight. Because agricultural usage typically means surveying large swaths of land, this presents a real restraint on industry growth and adoption.
On the horizon, we see new regulations to enable the growth of beyond visual line-of-sight (BVLOS) drones — which will trigger significantly faster adoption.
But the technology hurdle remains. While a 12-hour flight can gather significantly more information than today’s 30-minute missions, the question becomes how to aggregate and process such large volumes of data.
Breaking Down BVLOS Technology Challenges
As regulations regarding pilots and line-of-sight requirements are removed, the cost of drone programs will fall as well. We anticipate the operational costs will decrease by at least a factor of 10.
Lack of adequate data management tools will remain a major hurdle to widespread adoption of drones in agriculture. To understand why, let’s review a typical agricultural mission today.
A typical flight will collect data at around 100 meters altitude, with a flight time of roughly 30 minutes that covers 20 hectares. The pilot will collect images of about 2 cm resolution, require full-resolution orthomosaicking, and will use an image-stitching software that requires 70-80% overlap.
This typical flight, using a 4-band multispectral system, will produce about 40 GB of raw data.
Going beyond the line of sight, you can collect data with a flight time of about 10 hours, covering 400 hectares. Using the same 4-band multispectral sensor, the flight will produce about 800 GB of raw data.
Manually managing 800 GB of raw data every time you fly your drone is not a scalable solution, nor is it especially manageable on a single-flight basis. So, we need a new solution to manage the data volume these new systems will produce.
As it stands today, we have three general approaches to data management:
- Raw Power: This approach means managing data and creating orthomosaics with the same approach we already take, except on a larger scale with more powerful software and computing architectures.
- Quick Tiling: This approach recognizes that the computing power needed to build orthomosaics is a major hurdle. It also recognizes that collecting images creates a lot of data redundancy. As a solution, this system will look at GPS, IMU, and other navigation data to place images on the ground to create a map, skipping the entire orthomosaic process altogether.
- Sparse Sampling: This approach takes quick tiling a step further. Sparse sampling takes a sampling of images from around the field and adds them to a map based on the metadata. From here, it interpolates the data that’s in the spaces in between.
Each method has its pros and cons, as shown in the graphic above. However, none of these approaches provide a complete system.
Understanding Growers’ Information Needs
When asking ourselves how we solve the problem, we start by looking at what growers need at each level of the organization. For an enterprise-level agricultural management company, there are typically three key users, each with different needs.
- Field Manager: This person is responsible for the day-to-day operations, and is looking to detect or identify problems, track crop development, and allocate resources as necessary. To do that, they need a high density of data collected and processed very rapidly. In most cases, the field manager will also need full-resolution imagery to review the specific conditions identified by the analytics
- Regional Manager: This person doesn’t need to see every single pixel of every single leaf, like the field manager does — but they will still need to see some of the spatial variability of the field. In terms of data, a common method might be to report the statistics on 10 x 10 m grids. This higher-level view of the data reduces the volume from 2 GB per hectare to 0.001 MB per hectare.
- Enterprise-level Management: To understand this person’s needs, you zoom out even further. This person isn’t as concerned with spatial variability within the fields, and instead is looking at field-level statistics across many operational regions.
When considering the information needs of key users in an organization, we see that field managers need high-resolution imagery, while the other users only need data from which to draw insights, which frames our approach to solving the data collection problem.
How Do We Meet Growers’ Data Collection Needs?
Designing data architectures to get the right information to the right person within an organization requires a mix of edge and cloud computing.
At SlantRange, we’ve spent nearly 7 years working on how to manage the data flow so the right information is going to the right users across the organization.
As we design a data architecture to give the right information to the right person, we’re building a system where every single drone becomes an edge device. That is, a completely self-contained data collection and processing system for the field manager who needs that information immediately.
Accomplishing this means moving all of your data processing onboard the aircraft, and reducing it to a level of statistics that can be pushed to an edge node — which may be a local database or storage with more computing power to run the higher-level analytics.
A global organization might have tens of thousands of hectares spread across a single country. The regional manager might be responsible for that country, and collects all the data from all those fields and ultimately reports to global headquarters
At the highest levels of the organization, there is a full cloud implementation, complete with long-term archival storage of data (including raw data), aggregated analytics, and a variety of visualization tools. In this platform, users might be developing new models based on data coming in, and then pushing new models back out to those edge devices in the field as they're created.
But, of course, actually doing this will require quite a bit of innovation and new technologies. So, let’s break it down by solutions.
Solution 1: Streamlining The Orthomosaicking Process
If you’re covering a field at 80% image overlap, it will take about 1 minute and 40 seconds to collect data on one hectare — which generates about 2 GB of raw data.
If you can reduce the image overlap to 20%, you can reduce the data collection time to about 25 seconds per hectare — and reduce the raw data volume to 125 MB. This approach means you can cut your data collection time by 4x, and cut your data volume by 16x.
On the surface, this sounds a lot like quick tiling, but it’s actually a hybrid approach. Our systems use an advanced navigation system to more accurately place images on the ground. Spatial correlation features ties these elements together, creating a hybrid approach.
Solution 2: Eliminating Image Stitching Errors
With the standard data collection workflow, you collect your images, stitch them together to create an orthomosaic, run your analysis to extract statistics, and aggregate the results.
In this model, the analysis is based on the accuracy of the orthomosaic. The problem here is that if you look at a high-resolution orthomosaic and zoom in closely, you’ll often notice artifacts in the imagery that aren’t visible from a zoomed out macro view
If you’ve ever wondered why stand counts are often inaccurate using this approach , it’s because your software is analyzing orthomosaics that include these types of errors. These errors appear because most imaging software was developed to create panoramas for photography purposes, not quantitative imaging.
SlantRange has inverted the process. With our workflow, you collect imagery, run the analysis on the individual images, generate the statistics, then aggregate the statistics — removing the orthomosaic from the process entirely.
Solution 3: Selective Image Compression
Traditional compression will reduce the amount of information gathered, with one fatal flaw. This method treats every pixel the same, which degrades the high-resolution imagery — and immediately reduces the value of the data and the ability to support more advanced analytics.
In agriculture, not all pixels have equal value. Pixels containing vegetation are significantly more valuable than pixels containing soil. So, SlantRange has developed a process called asymmetric compression that decides which pixels to compress and which to leave at full resolution.
The result? Data processes faster but doesn’t lose its integrity. The manager in the field gets high-resolution, high-value data collected in a timely fashion — so they can get the information they need to make a time-sensitive decision.
New Solutions, New Possibilities
Even if new regulations and aircraft systems are implemented, today’s data architectures will constrain adoption. New technologies and methods that handle data depending on user needs are available to break through those constraints.
BVLOS tools, when combined with improved measurement capabilities, will trigger large-scale adoption of remote sensing techniques in agriculture.
From our point-of-view, developing new data architectures that can scale with the ability of drone systems comes down to understanding the information needs of each user across the organization — and laying out a data plan that works accordingly.
For each user in the organization, new data architectures and BVLOS tools provide new ways to understand their crops, and do so in a much more scalable way.