Weed Detection in Crop Fields using Deep Learning Techniques

Abstract

The increasing demand on food leads to applying agriculture methods to increase the overall crop productivity. Most of the farmers decide to use chemical herbicides to get rid off the weeds that grow among the crop in order to avoid crop losses. Most of these chemical products cause an increase on the pollution on soil, on air and on water, damaging the flora, fauna and of course the humans. On the other hand, the concept of organic agriculture is recently being demanded by a lot of people that are conscious of the intrinsic dangers that cause the chemical herbicides used in agriculture. The drawbacks of the organic agriculture is that it takes time and effort, translated into additional costs for the farmer. That’s why farmers with large extension of crop fields are starting to try different tools in order to automate these tasks. Most of these tools are based on robotic systems or autonomous vehicles like Unmanned Aerial Vehicles (UAV) or Unmanned Ground Vehicles (UGV) that are able to perceive the current state of the field like crop health, total crop area, etc., and in some cases, these vehicles are also able to act over the field by performing actions like weeding or the process of getting rid of the weeds. Other methods that can be inside or not of the organic agriculture, are the precision agriculture techniques, mostly based on not only to cover the supervision of an area of crop, but the crop individually. The methods of mapping the position of each crop and to retrieve the health or the volume of a single crop, even to detect and remove any weed detected, are inside the precision agriculture concept. In precision agriculture pipeline, first of all it is necessary to detect what is desired to detect. If this is achieved, then the mapping and other geometric techniques can be performed properly. But the first part based on the detection stage is not as straightforward as it may seem. Classical machine learning techniques based on computer vision like localizing objects from the image, are based on image binarization by threshold and then on counting the blobs of the binary image. This method can only be applied if the physical conditions that are taken into account during the parameter tuning, are the same than the conditions taking place in each of the future situations. Concerning a real-world application, it has to be assumed that there will be changing physical conditions. That’s why it is necessary to use the deep learning techniques in order to handle or generalize the predictions or classifications when the conditions change or when unknown situations are present when the robotic system is supervising the crop. This project is focused on the weed localization or targeting, and it is one of many others that are taking place in Earth Rover Europe, S.L. The main project of the company is to develop a rover or a mobile robot able to supervise and map the crop individually and also to remove weed plants from the crop fields.