Edge Computing for Palm Trees
Why intelligent sensors make business sense for date palm farmers?
In many countries around the world, dates is a primary dietary element. In the Middle East and Mediterranean region, there are tens of millions of dates-producing trees with overall annual volume topping USD 50B according to Statista.
Depending on the type and quality of produce, the book value of a mid-age date palm tree is typically around USD 3–4K. This is indeed a non-trivial value. However, it is quite surprising that date palm farmers still rely on sporadic human field checks to monitor health and growth conditions of their trees!
In the age of ubiquitous connectivity and prevalent internet-of-things (IoT) technology, a legitimate question would be: why isn’t each palm tree out there monitored with an IoT device?
The reason: over-the-counter IoT technologies available today do not scale well in the case of palm farms!
Moving the needle to the edge
The conventional wisdom today is to dumb down IoT devices as much as possible and keep intelligence in the back-end, e.g. server in the cloud. Such a paradigm, however, would be unfavorable due to the following reasons:
- It results in too much communications to the back-end thus increasing the capital and operational cost of the communications solution itself.
- Too much communication in return depletes IoT devices batteries very quickly. With tens of thousands of trees per orchard, who wants to replace so many batteries so often?! For a date palm orchard owner, the math simply does not add up.
As such, there is a dire need for moving intelligence to the edge, i.e. empower those IoT sensors to make decisions on their own and only send low-key keep-alives or red-flags to the cloud.
As an example, let’s take the notorious problem of detecting the existence of red palm weevil (RPW) in a palm tree. The RPW is a vicious pest that wickedly sneaks into the trunk of a palm tree and starts chewing on it from inside. If not detected early on, it causes irreversible damage in a matter of weeks.
The chewing activity of the RPW produces an acoustic signal that has been widely characterized in literature. However, detection algorithms proposed and implemented mostly rely on a very limited set of features. With machine learning techniques becoming more efficient, we are able today to develop computationally light inference models. Such models can be entirely carried out at the edge, i.e. at the sensor IoT device.
Fortunately, the amount of computational power available on low-end low-cost microcontrollers is actually sufficient to run somehow sophisticated machine inference models. Our initial investigation shows that simple recurring neural network (RNN) or long short-term memory (LSTM) architectures can do a very good detection job. The resulting inference model can fit well into a LoRa system-on-chip (SoC) like the STM32WL from ST Microelectronics could .
The need to drive cost down
Battery depletion rate is not the only hurdle to scalability. Another major one is the cost per tree. Taking the RPW example again, current solutions in the market today are offered at no less than USD 50–100 per sensor (i.e. tree). This is most likely driven by the invasive nature of the IoT device. It typically entails a highly sensitive acoustic sensor with a coupling spike that needs to be drilled into the tree trunk.
So what is the price point per tree that would make business sense for a RPW detection solution? We carried out a series of interviews and collected data from a few farms to answer that question. Our investigation indicated that the price point per tree needs to be brought down below USD 10.
Obviously, we needed to figure out a design that employs low cost sensors yet does not compromise detection performance. Taking this forward, we quickly realized that the financial ramifications of “missed detections” outweigh the those associated with “false alarms”. Infested trees which go undetected result in direct earnings losses for that year as well as devaluation of the farm. On the other hand, false positives are less painful (as long as there are not too many of them). A false positive does entail flushing some OPEX down the drain. However, it is negligible in value compared to the adverse impact of missed detections.
The design of a smart tree “watch”
Exploiting the interplay between detection and false positive performance, we took a stab at the problem by playing out the game of diversity. For example, if a high-performance acoustic sensor is worth ~USD 50, then there is a good chance its performance can be matched with an array of 3–5 lesser quality sensors, each costing less than USD 2.
In signal processing terms, using an array of low-cost sensors offers a tangible gain in signal quality while keeping the price way below that of the expensive high-performance sensor. Such a tactic allows us to boost detection performance while incurring only a modest rate of false positives. At this point, the concept of a “smart tree watch” was born.
Hitting the sweet
The Smart Tree Watch is still a nascent project that needs to be scaled up and tested more intensively in the field. Nonetheless, we believe the way it is designed strikes the right balance between probability of detection and probability of false alarm.
Advanced-tech solutions such as those based on fiber reflectometry promise to achieve better performance characteristics. However, that definitely comes at the expense of increasing the price per tree by an order of magnitude (we think it is no less than 5X compared to our design). Instead, our design methodology exploits the tolerance we have in false alarm performance to hit a sweet spot as shown in the performance evaluation curves below.
Ushering the era of tree “wearables”
We are very excited at the prospects of this project. Together with many others in this domain, we’re creating the “internet of trees”. We believe low-cost long-endurance IoT devices will have a disruptive effect on the agricultural sector. In some sense, it is quite analogous to effect which wearable technology is having on us.
Acknowledgements: Mark Tester, Babar Khan, Andrew Yip, Ahmad Alrefai, Mohammad Abu-Alasal, and Samer Al-Reqeb.