Weed management is an important part of the investments in crop production. Cost of herbicides accounts for approximately 40% of the cost of all the chemicals applied to agricultural land in Europe. In order to increase the profitability of crop production and to reduce the environmental concerns related to chemicals application, it is needed to develop site-specific weed management strategies in which herbicides are only applied in the crop zones were weeds spread. Moreover, these strategies should be implemented in early season, just when post-emergence herbicide treatments are usually applied. Until now, obtaining weed infestation maps in early season has been a great challenge due to the reduced size of the weed and crop seedlings and the spectral similarity between weeds and crop. And this challenge has been especially hard in winter cereals, in which the narrow space between crop lines complicate the weed-crop discrimination.  Today, this challenge can be overcome by the use of ultra-high spatial resolution images taken by Unmanned Aerial Vehicles (UAV) and the possibility of acquiring images in-time just a few days before the herbicide treatment is scheduled.

 

This article describes the complete workflow developed to achieve the weed patch mapping in wheat fields, as paradigm of winter cereal. The workflow can be divided in three main steps: 1) configuration of the UAV and design of the flight route to acquire a set of overlapped images of the wheat field, 2) mosaicking of these images to create a georeferenced ortho-image of the whole crop field and 3) automatic object based image analysis (OBIA) procedure developed for generating weed patch maps and herbicide prescription maps. The UAV used in this research was a Microdrone MD4-1000 equipped with a commercial camera Olympus PEN E-PM1 which takes images in the visible range of the spectra. The vehicle was programmed to overflow automatically the wheat field naturally infested with grass and broad-leaved weeds, and to trigger the camera at the moment required to supply images with a previously fixed overlapping between them. Then, images were mosaicked using a commercial software that needs the overlapping to create an accurate georeferenced ortho-image of the entire crop fields. At last, the mosaicked images were analyzed using a robust and completely automatic OBIA procedure developed by our research group. The OBIA analysis algorithm combines object-based features such as spectral, position, orientation and hierarchical relationships, and is based on the idea that plants growing between row crops are supposed to be weeds; so, it detects the crop rows by the application of a dynamic and auto-adaptive classification process in which crop rows orientation is taken into account, and then classify the vegetation objects outside the rows as weeds. At last, the weed infestation map is divided into a grid of user-adaptable size, and every one of the generated squares is classified in function of its weed density, obtaining a prescription map ready to be used by a spatially variable herbicide application sprayer.

 

The described workflow was applied in two wheat crop parcels of about 1 ha each. The wheat fields were located in southern Spain, and were naturally infested with grass and broad-leaved weeds. Crop and weeds were at two-four leaves unfolded (Principal stage: leaf development; BBCH extended scale) . Imagery to generate the ortho-mosaic was taken 50 days after sowing. Accuracy of the weed detection was evaluated by comparing the weeds detected by the algorithm with the weeds visually identified in 49 georeferenced metallic frames distributed over the experimental fields. In order to study the influence of spatial resolution, the OBIA algorithm was applied to the ortho-images generated flying at three different altitudes: 30, 60, and 100 m. Finally, the herbicide savings were calculated based on more accurate weed maps.