• er under ombygning - vi er tilbage mandag med nyt udseende. Henover weekenden er alt vores indhold åbent, men man kan ikke logge ind og debattere.

Satellites and AI in agriculture

Illustration: Privatfoto

OPINION PIECE: The content of this article is solely the author's own opinion.

In the past decade, we have seen great development in space launches. In addition to the fact that humanity has managed to send a red Tesla to Mars, we now also have Earth observation satellites, whose images have led to a wide range of new applications. In agriculture among other sectors, and farmers now have the option of monitoring their fields from space. For example, using the European Space Agency’s Sentinel-2 satellites, which provide free images of any area on Earth every five days.

Satellite imagery has helped make agriculture far more data-driven—an industry that has historically been based on human experience and intuition. On a small scale—that is, on the scale of individual fields—we now see satellite images being used for precision agriculture, in which images of fields are, for example, used to allocate nutrients to the areas most in need. On a large scale, satellite images help to provide an understanding of crops over huge areas of land, e.g. which different crop types are planted throughout Denmark or Europe, how well they are doing, and what the expected yield is.

Automated analysis of satellite data

However, the huge amount of satellite data to be analysed at such a scale requires automatic and scalable methods, as manual analysis of such large amounts of data is unfeasible. In this area, deep learning has enabled automatic and accurate analysis of satellite data, and the vast amount of data can even be an advantage when training data-hungry models. However, any satellite image used for supervised learning also requires appropriate labels corresponding to the task at hand. Getting accurate labels for satellite images of fields usually requires sending agricultural experts out to the fields and collecting the relevant data, as it is difficult for humans to analyse satellite images directly.

This is in contrast to other data sources in a more “human format”, such as photographs or text, where labels can be relatively easily annotated by people comfortably sitting in front of a computer. Fortunately, the collection of relevant labels in agriculture is often a process that is already ongoing.

One example is that farmers have to report which crops they have planted to receive agricultural support on an annual basis, and this data is now publicly available. Unfortunately, this data is often scattered around different sources. For example, farmers definitely know the yield of their fields, but these numbers are often hidden away in a notebook or on a USB drive somewhere, and not in a single database in a clean format, which is the desired scenario for deep learning. Accessible labels for satellite images are therefore a goldmine for solving many problems in agriculture, as they both enable the training of accurate deep learning models to solve a specific agricultural problem and also reduce the necessary number of physical control visits for the farmer.

The right training data

The interesting thing about using deep learning to solve problems in agriculture is that with the right training data, these models can be trained by programmers without a deep knowledge of the underlying agronomic processes. As computer scientist Richard Sutton writes in his blog post The Bitter Lesson, it is very tempting to build expert knowledge into the methods we create, but this approach has historically proven to be beaten time and time again by methods that use brute force computing power, such as deep learning.

However, the key part of deep learning is still “the right training data”. While it is usually a bad idea to build knowledge about how we think an agricultural problem should be solved into the model—e.g. by using hand-crafted features that we humans use to understand satellite images—it is often much better to simply feed the model raw data and let it extract relevant features itself.

However, specific knowledge about the domain is still relevant during the actual selection of the data that the model will be fed. For example, the satellite spectral bands outside the visible spectrum have a large correlation with the growth of crops, and it has also proven advantageous to use them in models for e.g. crop classification. In addition, we also know that the weather has a great importance for the development of crops, and weather data has also been shown to be an important input for improving the generalization of deep learning models to adapt to new areas and seasons.

Agriculture, an industry traditionally marked by human experience and intuition, is now facing a digital revolution thanks to the collection of big ag data from space and new, data-driven deep learning methods. This is an important element in optimizing agricultural processes so that we can secure the food supply for a growing population while minimizing the climate footprint.