In this previous article, we’ve seen how satellite imagery can be used in agriculture through different vegetation indices. Nevertheless, vegetation indices seem to be only the tip of the iceberg as AI (Artificial Intelligence) in agriculture is providing growing crops with better productivity and respect for the environment. Let’s discover here how AI applied to satellite imagery is bringing an agricultural revolution !
First let’s understand the basis of AI. An AI model is a mathematical algorithm that is designed to be “trained” to learn automatically based on data input.
To simplify, we will take a simple example using machine learning, a subcategory of AI. Let’s say you want to make a computer able to recognize cats on images. You take thousands of images of cats, and then you show the computer where the cat is on the image. This data is therefore integrated into the algorithm, and it will look at all the zones where cats are localized on the images. The algorithm will subsequently be able to identify common visual patterns on each image (two pointy ears, a tail, two eyes, four legs…). With time and a lot of examples, by looking for these common visual patterns, it will be able to reliably detect cats on images that it has never seen before. The more images are analyzed by the algorithm, the more efficient it is. “Simple” right? But how can AI be applied to agriculture?
The limitations of traditional agronomic models
To answer this question, we also need to understand how traditional agronomic models are created. Conventionally, the creation of an agronomic model is time consuming, expensive and unscalable. This is because most agronomic or hydrologic models are heavily parameterized and require extensive calibration. As a result, model development can be incredibly time consuming and still only be applicable for specific regions or crops. For instance, a soil moisture model to determine water availability on a field, requires information on soil properties including soil composition and porosity. Crop type, rooting depth and weather information is also needed for optimal performance. Overall, the amount of input data required for this model prevents scalability.
.jpg)
The new generation of agronomic models using AI
But, what if you could apply machine learning to soil moisture detection ? What if instead of trying to detect cats on images, you try to look for soil moisture information on satellite radar images?
Over the last decade, the multiplication and improvement of satellites along with the progress in big data analytics methods and Artificial Intelligence introduced new opportunities to get more agricultural insights out of satellite imagery, at an affordable cost.
Today, AI is considered as a real game-changer. Similarly to the cat training algorithm, we collected thousands of satellite images of fields on which we had the exact volumetric value of soil moisture at the moment the image was taken. We collected that information from 10 sensors in olive plantations in the south of Spain, over a duration of one year.
So now we have thousands of data points composed of one satellite imagery, and one soil moisture value that belong together. We feed that into our AI model, we make some adjustments and voila! We trained and we adapted our AI model, to be able to detect only from satellite images, the soil moisture level at 30 cm depth for olive plantations in the entire south-east Spain area! And to ensure the accuracy of that model, we took a few other sensors from the region which the model never saw, and we compared their measurements with our model predictions, and we were able to reach 85% accuracy.
The great thing about AI is that, with the correct data, our models can be adapted to different regions and crops of the world in a few hours' time. The algorithm will capture minute details and patterns in crop physiology and, through thousands of samples, it learns which patterns are relevant to soil moisture and which are not. For example when looking at different soil types, our AI algorithm creates an additional layer of information for these differences in soil and adjusts for the combinations of these patterns to get the right soil moisture for any type of soil type.
So if we want to detect the soil humidity level for plum fields in a region of Chile, we just need to feed a dozen sensors’ data for plum fields in this region to the model, and it’ll adapt to be able to predict it.
Overall, with traditional modeling approaches, it might have taken years to define a model that interprets radar imagery to measure soil moisture in one region and for one crop type. It would even have taken a similar amount of time for a different region and crop. And this is where AI shines: With its ability to save time and cost while bringing precision at scale!
.jpg)
AI is often used as a magical wand that can work for all kinds of situations. But, reality is quite different. Building solutions intended for real world usage relies on several factors like the complexity of the task, the data source at hand and the reliability requirements.
At Spacesense, our key differentiator remains our expertise in mixing cutting-edge AI with satellite imagery applied to agritech solutions! We create the best AI model architectures to encapsulate this understanding of the agricultural process from satellite data. And through our ability to measure the accuracy of our solutions, we also bring transparency to ensure that our solutions are up to the standards of our customers. The result is a solution driven by AI and satellite data that has reliable performance for practical use for everyday decision making.
Our mission is to bring this state of art AI-driven satellite solutions to innovative Ag businesses who can then merge it with their core expertise and experience in Agriculture to solve this generation's biggest challenges in food sustainability and climate change.
Want to benefit from geospatial
AI easily?
Check out our upcoming solutions that you can test
Check it out