Farm management information systems as tools for revealing management zones inside the fields

INTRODUCTION and OBJECTIVES: There is a huge need to increase the productivity in agriculture to feed the world’s growing population. However, this increase needs to be achieved in a sustainable way, without jeopardising the ecosystem and environment. Innovations in AgTech are accelerating this process and providing adequate solutions for optimisation of on-field decision-making, but they are often isolated and inaccessible to the farmers. The objective of our work was to design a comprehensive farm management system that takes scientific achievements and enables farmers to use them in their daily operations. MATERIAL and METHOD: In order to digitally transform the Serbian agriculture, we designed AgroSense farm management information system. It was launched in 2017 and has since gathered more than 20,000 users, whose total area equals one fourth of all farmland in Serbia. The platform has a number of modules for weather forecast, historical weather records, digital field books, satellite image processing etc., while the newest addition is the drone image processing module. This module allows 3rd party drone services to scan the fields and upload the data to the platform, after which, the images are processed and analysed. The analysis is directed towards zone management delineation, which is the first step in application of precision agriculture technologies. Zones are detected within the field as areas with homogeneous soil and elevation properties. This is done by applying k-means, an unsupervised machine learning model for clusterisation of data, i.e. pixels in this case. This algorithm minimises the intra-class variance (variance of pixels within the zone) and maximises the inter-class variance (variance between pixels from different classes. This zone delineation can be done on a pixel-level if the objective of zone delineation is e.g. choosing the right locations for soil sampling, or on the level of the tractor swath if the goal is e.g. the variable-rate application of fertiliser. The number of zones and the swath width are variable parameters, left to the user to choose, according to the size of the field, type of the equipment and other factors. RESULTS and CONCLUSIONS: The resulting platform was deployed in 2021 and tested on a number of users. It yielded excellent results and served for optimising the route and sampling location of unmanned ground vehicles (UGVs), characterisation of fields and variable application of fertiliser. Future work includes development of other algorithms for more complex image recognition tasks, such as row detection, leaf area assessment and disease/weed mapping.