COMPUTER VISION IN AGRICULTURE

The future of agriculture will be powered by AI, machine learning, and like-minded teams. Machine learning may not just be useful for complex, data-heavy, all-knowing tools but the future of farming will be far more visual. Plant scanning is already here, and no longer relegated to in-plant pets who peck at notes on a clipboard (Note: One thing I have learned after working in agricultural tech for over a decade is that sometimes the smartest tools are boring). As more video processing power enters our hands, it is likely computer vision will be more present on the farm, and crops will be monitored remotely and photographed as soon as they are rooted. If they need to be harvested, they’ll be seen in their at-the-time purest form in the end, a true sign of maturity.

However, although computer vision is being applied to farming in the field, I don’t imagine it making its way to near-term tech products. Designing for computer vision has too many caveats and complexities that an agricultural developer and user of farm equipment would have to contend with. You need a powerful SQL framework with large-scale access to massive farm datasets. Of course, a farmer’s PC would need access to those massive datasets as well as the development team’s knowledge of crop fungicide effectiveness, labels on fertilizer particles, and other data that could be manually dug up using Google Earth. And it will still be a challenge to have enough power and network connectivity enough to run detailed scans, so in a low-cost dataset like a commercial farm, computers would be starting at a disadvantage compared to a farm with many more detailed farm records. In the event that there is high demand for computer vision in farm maintenance, additional training, and manufacturing, our research will likely leverage the use of real-world photos and public clouds of training data to further the cause.

The software development team at On Device Pro have partnered with growers and our co-workers at On Device Pro. From taking high-resolution portraits to on-farm capacity tests, the farmers that have already made use of computer vision solutions use them to adapt or perform specific tasks in the field. Sometimes they are planting through the use of panoramic images of plants or inspecting the movement of workers or commercial products in the field. Data is tremendously valuable. One of the simplest ways to understand and compare performance in different plants is by looking at how long each is exposed to sunlight and water. Furthermore, all of the farm’s machine learning products from predictive farm analysis to object detection are based on unstructured data derived from those over-data-loaded observation sessions. So, the optimal state of these platforms is data-centric. With that in mind, our work on vegetation sense is very organic and much more education is needed.

Many of the technology can be used in diverse ways, ranging from small nuggets of information that provide visual feedback about the health of the crops being grown, to large-scale image processing. Our work could provide additional data about our crops or make applications in a number of different fields. However, computer vision is likely to face an uphill battle in the right conditions.

A team of researchers from the California Institute of Technology recently published their results which outlines a potential solution to see into plant leaves to inform the correct treatments.

In many settings, such as on a medical bed or on a machine learning board, a computer system is being used to examine deep slices of a piece of tissue to offer diagnostic insights. It can identify abnormal tissue regions. Sometimes, a dermatologist needs to view the same tissue to interpret facial dermatitis diagnoses. Not only does such a technology mark the single role of a regular microscope in skin diagnosis, but it also has the potential to analyse plant and animal tissue differently.

The team initially created this technology as an engineering project. By developing machine vision to detect tissue wound photos (LCMS) — where perforations are detected by the microscope film instead of the cells — and applying such analysis to plant and animal tissue, the team hoped to gain an unprecedented ability to capture skin diseases early.

To make such a system work, the researchers first applied computer vision techniques to tissue grafts such as bone grafts. The images were used to train a computer system that derived a machine-learned tissue map of the wound around the organ graft, and with these maps, the system could classify and distinguish tissue areas from cells.

Computer vision also performs a more complex kind of “landmarking” of foliage. The researchers fixed a landmark in the leaf using mechanical pruning tools and an optical gel probe. Once the researcher fed the images through the system, it successfully identified and segmented with the marker into different tissues.

The computer had to be trained to segment tissues precisely. The system only focused on a fraction of tissues, and less than half of them had consecutive micrometre scales. Because tissue age is difficult to establish without small areas in a larger body, classifying each of them was challenging. One consequence was that each tissue was only segmented into smaller single cells, which blocked them from forming organs. One solution was to classify tissues as single individual tissues.

 

Combining all these images together was a bit like a video game pilot in which the main character tries to manoeuvre into an aircraft using a control system. As the pilot tries to pull off a complex manoeuvre, they see blocky blocks come in and out of the cockpit, meaning they need to aim in the right direction if the manoeuvre is to be successful. To pull off such a manoeuvre is probably extremely difficult. In this case, the pilot needs to carefully measure the position of the patches within the patches — and this required very precise data.

The researchers adopted a particular strategy for classifying tissue. They introduced photos of the tissue to the system as a fixed axial point on which a boundary can be derived from. In this way, the system could see around the tissue and class it as either any state or an outlier with some corresponding cancerous tissue. Using this example, the robot could decide if it needs to apply a poisoner antibiotic, if antiseptic coatings are needed, or if an antibiotic should be applied.

The researchers tested this system on several real research plants such as basil, soybean, leafy greens, and various plant-based lipoproteins. Without an effective system to classify tissues well enough, the researchers did not observe any tissue change, and yet they did receive results from the research plant just fine. Theoretically, the researchers could augment this system with artificial intelligence systems to adjust blood perfusion and surface treatment and perhaps even better detect a tumour.

Most experimental research on the integration of computer vision (and its offspring, Deep Perception), is led by algorithms and practitioners who need more practice in these fields to adapt to applying chemistry and biology. The lesson here, though, is that science is not about staying thin; it’s about weight; it’s about tackling key challenges and honing development. Now that there’s been successful innovation in the form of this technology, there’s the need to build off of it to create systems with greater applications.