Linda Pieri1, Dario Mengoli2, Fiorenzo Salvatorelli1, Marco Vignudelli 1, Lorenzo Marconi 3, Francesca Ventura1
1 DipSA – Dipartimento di Scienze Agrarie, Università di Bologna, viale Fanin 44, 40127 Bologna, Italia 2 Drover srl- Via del Greto 2/2, 40132 Bologna, Italia
3 Dipartimento di Ingegneria dell' Energia Elettrica e dell'Informazione «Guglielmo Marconi», Università di Bologna, Viale Risorgimento 2, Bologna, italia
Abstract
The research presented concerns the development and evaluation of innovative technologies for precision farming, aimed at reducing the current use of herbicides. Detailed maps of infestation are useful to localize the herbicides application, and the knowledge of the type of species allows to further reduce the dose of herbicide, using only the specific product. For this purpose an algorithm for processing images, based on convolutional neural networks, has been implemented to recognize monocots species, dicots species and bare soil.
Even if further refinements are in process, the first elaborations led to a recognition accuracy of over 80%. These first results are positive and suggest that the model can already provide information useful for reducing the herbicide dose and apply the specific herbicide. The map of infestation together with the recognition of the type of species allow a site-specific and plant-specific treatment, with advantages for operators, agricultural products and the environment.
Keywords
Weed control, image recognition, neural networks
Parole chiave
Diserbo, riconoscimento immagini, reti neurali
Introduction
Environmental sustainability in agriculture has become a prerequisite. The European Union, through the CAP, has clearly shown that the future agricultural world needs to provide the least environmental impact, reducing the application of pesticides and adjusting doses based on the degree of infestation (Williams, 2011).
It is well known that spraying herbicides homogeneously over the entire field would be inefficient, uneconomical and environmentally damaging. This is because weeds are not distributed uniformly: areas covered by high infestation are followed by others without infestation.
Previous studies have shown that differential treatment, based on the different level of infestation, can result in a saving of the applied herbicides doses up to 80-90% (Gerhards & Oebel, 2006). In particular for extensive crops, site-specific application would lead to significant benefits: an economic advantage for the farmer, as treating only the actual weeds implies a reduction in costs, since herbicides cost is currently about 40% of the total costs of chemicals, and an environmental benefit, since there is a lower environmental impact than conventional application (Peña, 2013).
In addition to the identification of the infestation level, through the creation of weed maps, the recognition of weeds species would be of considerable advantage. This would allow a local application of the herbicide, with active ingredient dependent on the weed. This would produce further decreasing of the applied dose and increasing the effectiveness of treatment.
The aim of this work was to test an algorithm of image processing, able to distinguish monocots species from dicots and bare soil. This study is part of a wider project to
evaluate the use of recent innovations in the field of automation, robotics, remote sensing and electronics for precision agriculture.
Materials and Methods
Specifically, the project concerns the image recognition of corn and its main weeds.
The initial phase of the project consisted in a database creation of images, which are divided in four categories: 1) corn, 2) dicots weeds, 3) monocots weeds, and 4) bare soil.
In particular we collected images of the following dicots:
- Amaranthus retroflexus - Chenopodyum album - Solanum nigrum - Abutilon theophrasti and monocots: - Panicum dichotomiflorum - Setaria glauca - Digitaria sanguinalis - Sorghum halepense
The collection of images has been realized thanks to plants cultivated in greenhouse and grown in open fields.
When the database of pictures was large enough, a model for recognition and classification of images was implemented. The selected model was based on convolutional neural networks, using the framework of "machine-learning" Torch7 (Collobert, 2011).
In order to be suitable to the framework, the images were reduced to the size of 256 x 256 pixel.
A total of 516 images were used for the supervised training of the neural network, while 173 images were dedicated to the recognition test.
In brief, the process for the development and training of the neural network consists of 5 steps:
1. Pre-processing of the images (e.g. convert images to YUV format to better separate luminance and chrominance components of the images (see Fig.1)
2. Construction of the model for the neural network 3. Definition of the “loss” (or “cost”) function
4. Training of the neural network using the training dataset 5. Testing of the result network using the test dataset.
Results and Discussion
Results led to a recognition accuracy of the two categories of weeds (monocots vs dicots) with percentages greater than 80% (86/88 for monocots and 61/73 for dicots). The results obtained from the comparison between the weeds with bare soil and with corn are not presented here, because not sufficiently significative, due to the low number of images of the categories "corn" and "bare soil".
Anyway, further refinements are still underway on the construction of the images database and configuration of the neural network for classification that will confidently lead to an improved performance. Specifically, the expansion of the images set of corn will allow better discrimination between corn and weeds.
However, these initial results are very positive and the possibility to use this new technology to reduce the amount of herbicide sprayed on the ground is realistic. The application of the herbicides specific for weed category (monocot or dicot) only in area with high level of infestation would produce substantial benefits on the environment and safety of agricultural products.
Further field trials will be conducted in the coming months to quantify efficiencies (yields and quality of the product), effects on the field in the following years and economic convenience of the proposed method.
Furthermore, for this first stage of the project, the images were taken manually, but, in the future, this activity may be carried out by automated facilities, such as terrestrial robots and UAV, or camera attached to the tractor. In particular, they may be equipped with the necessary instrumentation to autonomously recognize plants and spray the herbicide in the field only on the weeds.
Conclusions
Finally, these trials are testing innovative methodologies, able to revolutionize the production system of the agricultural sector by working with new technologies. The overall objective is to promote a new agricultural management that ensures a good production by achieving the objectives of reducing environmental impact, ensure healthier products, and allowing the operator to work in conditions of greater security, since the machines can also work autonomously.
Fig.1 - Esempio di immagini pre-processate in formato YUV
Fig.1 - Example of pre-processed images in YUV format References
Collobert, R., Kavukcuoglu, K., & Farabet, C., 2011. Torch7: A matlab-like environment for machine learning. In BigLearn, NIPS Workshop (No. EPFL-CONF-192376). Peña J.M., Torres-Sanchez J., de Castro A.I., Kelly M., Lopez-Granados F., 2013. Weed mapping in early-season maize fields using object-based analysis of unmanned aerial vehicle (UAV) images. PLOS one, 8,10, 1-11.
Gerhards R. and Oebel, H., 2006. Practical experiences with a system for site-specific weed control in arable crops using real-time image analysis and GPS-controlled patch spraying. Weed Research, 46: 185–193.
Williams, 2011. New EU pesticide legislation – the view of a manufacturer. Asp J Appl Biol, 106: 269-274.