14.09.2022

Camera systems in agriculture

Recent hardware and software developments have enabled robust image processing applications. Smartphones can distinguish people based on facial images or classify and interpret complex scenes. Modern image processing (computer vision) is also a prerequisite for autonomous driving. Intensive research is also being conducted on agricultural applications, and some are already in practical use.

Josephinum Research

The field of crop protection, in particular, offers diverse application possibilities for camera systems and image recognition algorithms. The first applications are already emerging in this area. Camera systems that can be used at close range in the field (proximal sensing) offer the advantage of higher resolutions compared to satellite imagery (remote sensing), thus enabling more targeted application of active ingredients or precise control of actuators, such as hoeing elements. Cameras can be used for image acquisition and subsequent control directly on machines like hoeing equipment or sprayers, or for monitoring and map generation on drones (UAVs). Unlike smartphone applications, shorter processing times are essential when using these systems on moving machinery.

Applications for Camera Systems

Systems for identifying plants for the application of pesticides are already available on the market. These systems control a spray nozzle based on the color reflection or near-infrared reflection of the object in front of the sensor. This allows soil to be distinguished from plants, thus reducing the need for herbicides. This simplest form of plant recognition does not require complex computational analysis, but it also cannot be used to differentiate between crops and weeds. Figure 1 (left) shows the result of plant recognition from a camera image; the areas identified as plants are marked in red.

One level higher in terms of technical complexity is row recognition. Cameras for precise row guidance are already widely used in hoeing equipment. These systems detect rows in the image based on green areas or, in the case of 3D cameras, height differences. However, many of the systems available on the market have limited applicability, especially in dense weed stands.

Josephinum Research

$blurhash$

Row detection: detected rows marked in red

Identifying crops and differentiating between plant species is fundamental to successful weeding in rows. Several plant recognition systems are already available for practical use and are deployed on appropriate weeding equipment. The system must identify the area where the plant is located. This area is either marked by a rectangle, as shown in Figure 2, or each pixel is individually marked using image segmentation. When used on a weeding machine, the growing point or stem base must also be identified to enable precise control of the weeding elements. In the simplest case, only crops and weeds are distinguished. More complex systems can also identify and classify different crops and weeds. Similar or slightly higher image resolutions are required for plant recognition as for row detection.

Josephinum Research

Plant species identification (corn plant identification)

Methods for detecting diseases from images are currently still in the research phase. In addition to the visible color spectrum, near-infrared images are often used in this application to further improve detection. Several smartphone applications are already available, but these are limited to spot monitoring. Comprehensive field imaging requires significant storage and computing resources due to the required resolutions. Similarly, methods for detecting pests in nature are still in the research stage. As with disease detection, high image resolution is necessary for identifying insect pests (e.g., slugs, beetles, etc.) and the damage they cause. Initial applications in this area allow for the imaging of yellow sticky traps, from which the insects are then counted and classified. However, several challenges must be overcome before large-scale pest monitoring can be implemented in practice.

Conclusion

Cameras with modern image processing algorithms offer diverse application possibilities in agriculture. Camera systems are already being used successfully, particularly in crop protection, as demonstrated by camera-guided hoeing machines. Similarly, the trend towards autonomous driving will lead to the installation of cameras on tractors for navigation and obstacle detection. This will enable these cameras to be used for a wider range of agronomic parameters.