Have you tried counting trees from orthomosaic images? Chances are, at some point or the other, you must— whether it’s for an agricultural survey or even a plot asset evaluation. But, you’d know that it’s an incredibly mundane task that also suffers from human error and so reduced accuracy. Our GIS and A.I. Engineer Sreya Madhavan and her team felt the exact same way, but she decided to do something about it. So, she trained an AI engine on the Picterra Platform to go through orthomosaics and identify palm trees with an accuracy of 97%. But first:
Palm trees are one of the oldest cultivated trees globally— specific to the UAE. It is an important part of the local heritage, contributes to the fight against desertification, and is a crucial economic asset. Across MENA, palm trees cover more than 13,000 sqkm. Knowing where and exactly how many are in a parcel of land benefits crop estimation, irrigation policies, farm management, and forest inventory.
Traditionally, tree counting was an incredibly labour intensive process. Usually, workers would drive and walk through fields manually counting the trees. Apart from being physically demanding, this also contributes to reduced accuracy. Satellite imagery could be an option, but usually, the resolution isn’t sufficient.
With drones in the picture, capturing large parcels of land was made easy. Drones could cover considerable areas in extremely short periods of time, but the counting was still left to humans. While drones drastically cut down the time compared to the traditional methods, the latter half of the process was still time-intensive.
We wanted to automate the process of counting trees from othromosaics which would also significantly improve the turn-around time. But this isn’t as easy as purchasing an off-the-shelf solution; she had to train an AI engine into being able to look through thousands of top-down orthomosaics and effectively identify palm trees.
The AI engine essentially learns through a process of trial and error, so begins the long process of negation and confirmation. We would first manually identify palm trees from data sets and then run them through the system. The team would then verify the result. She would repeat these two steps until the desired detection accuracy was achieved.
Getting the engine to 80% accuracy is the easy part; going from 80 to 95 (the target accuracy level) would be an uphill battle. So over the course of two months, We ran tests on multiple scenarios, different types of palm trees, and varying resolutions to find the optimal parameters, testing the system to see how far she could push it without losing accuracy.
To help the AI on its journey to 95% accuracy, The team would identify 769 unique scenarios of palm trees and manually verify each dataset processed by the database. After verifying the 5,785th tree, the AI was field-ready with an impressive 95% rate of accuracy.
Now we have a working AI engine that can effectively count palm trees from orthomosaics. Let’s compare how the average human stacks against AI. From our testing, the average accuracy rate of a human (manual counting) was 85%, the AI trained by Sreya and her team achieved a final accuracy level of 97%. But maybe the more impressive stat is time. We found that the AI was 11900% faster than a human. I know that number sounds bonkers, but that’s what Precentage calculator returned. From the tests we conducted, our humans (manual counting) averaged 120 hours to count the trees in a parcel of 1 sqkm large land— while the AI managed to do it in just under an hour. An easier way to look at it is that the AI provides a time-save of 99.16%.
I know it sounds astounding, I’ll let the study on it convince you:
Palm tree counting and monitoring also have many applications in forest inventory, crop estimation, irrigation policies, and farm management. Counting and quantifying palm trees provides an inventory of the trees that may help better plan the irrigation process. The exact number of trees is essential for predicting the date of production. It enables farmers and decision-makers to conduct real-time monitoring, improve productivity, and participate in ensuring sustainable production and food security. Providing an accurate evaluation of palm tree plantations in a large region can bring meaningful impacts in both economic and ecological aspects.
Despite their importance, information on the number of palm trees and their distribution across different scenes is difficult to obtain and, therefore, limited. Counting the number of trees in large farms has been a challenging problem for agriculture authorities due to the massive number of trees and the inefficiency and excessive cost of old-style manual counting approaches. The enormous spatial scale and various geological features across regions have made it a grand challenge with limited solutions based on manual human monitoring efforts. The problem becomes even more laborious and tedious when we also need to identify the GPS location of trees for governance purposes and regularly monitor their condition over time. The inefficiency of traditional methods leads to inconsistent data collection about the number of trees. Identifying and counting palm trees over large areas is a very laborious task. To manually demarcate each and every tree is a time-consuming process. It takes days to locate and count palm trees for large plots.
We can effectively use high-resolution aerial images for automatic palm tree detection. Together with automated processing tools, these images represent a powerful alternative instead of in-situ or manual image counting that requires tedious, time-consuming and costly work. Using an AI solution, locating and counting palm trees can be done very efficiently. This is faster, more accurate, and can easily be done. Farmers and agronomists can also detect Palms of different types and textures easily. This technique opens a new window into asset mapping of large areas where identifying and valuing objects is challenging for humans. In this study, we look into the versatility and reliability of a palm detector built using an AI engine.
Palms identified using object detection:
A process that would once take days can now be accomplished in a matter of minutes. This helps to save both time and money in the form of labor.
We can look at a case study of a 5sqkm farm filled with palm trees and other trees. 35000 palm trees have been detected within 4 hours to an accuracy of 97.5%.
On large farms with over 1sqkm, quantifying palm trees by counting them individually is a tedious task. Drones enable us with aerial data that can be captured in just hours. Human counting on this data is also very time-consuming and laborious. Using AI, assets like palm trees can be counted within minutes and have more than 95% accuracy.
Human errors are prevalent in any manually done work. Comparing the quality of results between both manually done and AI done outputs, we have noticed that many trees have been missed in the former. Even though there are wrong detections in the AI output, overall accuracy is much better with the AI detector.
Precision, recall and F1 scores for palm trees tested are:
|No. of Palm Trees||Counting Precision||Counting Recall||F1 Score|
Overall accuracy is 97% which has been tested on various types of palm trees and found in different environments all over the country. The precision of the model (which tells about how many wrong objects the model is detecting) is 0.97. The recall of the model (which means how many objects have been missed) is 0.97.
Different kinds of palms
Different areas have different types of palms. The object-based detector is so reliable that it will detect any palm trees from anywhere in the world. To test that, we have collected UAV imagery from all over the globe and checked for detector reliability.
The detector has been trained on various types of palms found exclusively in the UAE (red marker). In addition, testing has been done on palms found in Brazil, Ghana, Uganda, Germany, Lebanon, Kuwait, Sri Lanka, Indonesia, Papua, and Australia (blue markers).
Few of the different types of palm tested are shown here:
Accuracy numbers for palm trees identified over different countries:
|Area||Actual # of Palms||Total Detections||Missed Detections||Wrong Detections||Accuracy|
Detection results are commendable with an average accuracy of 95.4%.
UAV images with different resolutions
Datasets of different resolutions ranging from 1cm to 30cm have been tested.
Accuracy numbers for palm trees identified over different resolutions:
|Area||Resolution (cm)||Actual # of Palms||Total Detections||Missed Detections||Wrong Detections||Accuracy|
Palm trees are considered to be a symbolic agricultural heritage in the United Arab Emirates (UAE). Date palms constitute 98% of fruit trees in the UAE, which is one of the world’s top ten producers of dates. This is due to great efforts carried out in planting management and applying best practices in ensuring the health status and maintaining the production rate which indeed requires frequent mapping and monitoring. The traditional way of mapping palm trees was implemented manually, resulting in a lack of accuracy, consuming more time, and requiring human interactions. Remote sensing, including satellites and Unmanned Aerial Vehicles (UAVs), has provided potential solutions in this regard in terms of large areas coverage, spatial and spectral information such data contained.
Date palm trees are important economic crops in many countries, and mapping them in a plantation area is crucial for predicting the yield of date fruits, determining insurance and financial aids, etc. Usually, counting palm trees is carried out manually on site. Such an operation is very tedious and time-consuming. Therefore, the utilization of automatic methods represents an effective alternative for public and private agricultural institutions. Given a UAV image acquired over a palm farm, each palm tree is detected using an AI engine.
Detection of palm trees using an AI detector is a very effective tool for quantifying palm trees. This is a much better alternative compared to traditional methods in terms of speed, accuracy, reliability, and efficiency.
Counting and locating palm trees over large areas has been done quickly and accurately with an accuracy of 97%. Also, the model has been tested across different types of palm trees around the globe which returned an accuracy of 95.4%. Furthermore, datasets of different resolutions ranging from 1cm to 30cm have also been tested with 95.7% accuracy.