The development of artificial intelligence in recent years, together with the increase in computing power and the immense amount of information available, has led to truly amazing achievements in fields such as language recognition and computer vision. One of the keys to these successes has been the techniques of the so-called deep learning. These techniques are included in the machine learning, which is, in short, a series of models and algorithms that allow computers to “learn” to perform a task based on data, rather than explicitly programming what they should do.
This technology has a huge potential to massively analyze drone images and perform in a first phase the counting of trees or other crops such as lettuce. The objective is to develop analysis systems that allow us to detect diseases, crop estimation, etc.
The development of artificial intelligence in recent years, together with the increase in computing power and the immense amount of information available, has led to truly amazing achievements in fields such as language recognition and computer vision.
One of the keys to these successes has been the techniques of the so-called deep learning. These techniques are included in the machine learning, which is, in short, a series of models and algorithms that allow computers to “learn” to perform a task based on data, rather than explicitly programming what they should do.
With these techniques, things like the automatic transcription of a video or an autonomous car detecting a person crossing a street can be achieved. As you can see, these are tasks that, until recently, had to be performed by a person.
Example of training image of Mapsens® Agro algorithm for tree inventory.
At Mapsens we are implementing this technology in a first phase to perform crop inventories automatically. This is a repetitive and tedious task, difficult to automate using traditional programming techniques, but very useful for the farmer’s management of a farm. Therefore, we are using the data we have been collecting so that our computers learn to do it instead.
For this purpose, we are using algorithms based on a convolutional neural network, the type of model of deep learning which has achieved the best performance in terms of artificial vision. With this technique it is possible to achieve robust object detection without the need to implement any a priori knowledge in the program. In order for the network to learn to detect and locate trees, we trained it with images taken by drones with examples of inventories already performed.
After training the algorithm is able to identify the trees that are part of the crop and place them spatially in images that it has not seen before.
Example of tree detection performed by the network in Mapsens® Agro
It can be seen how the network detects trees of different sizes, with different shades in the image or overlapping. It is this variability of the images that makes this task difficult to perform with other traditional methods.
This functionality is also adaptive, as new data is obtained, its operation can be adjusted to adapt to new crops or improve its performance, for example. The detections made by the algorithm are processed and georeferenced to obtain an inventory that is integrated into the Mapsens GIS environment.
The idea is to establish a fast and efficient service that serves our customers for the automatic counting of feet to estimate the harvest. But our goal goes beyond that, looking for quick solutions for crop estimation, pest detection, fertilizer estimation and a long etcetera.
We keep working!
