PRecision crop protECtion: deep learnIng and data fuSION

Current farming practices require a uniform application of pesticides in order to protect crop plants from pest and disease. These treatments are typically repeated at regular time intervals. However, it is well known that several pests and diseases exhibit an uneven spatial distribution, with typical patch structures evolving around localized areas, especially at early stages of development.

The detection of early stages of plants infection is crucial: the declination of Precision Agricolture to crop protection operations aims to fulfil this critical task. Precision crop protection encompasses technological, biological, agronomical, agrochemical and modelling knowledge, leading thus to a data fusion approach, which allows to develop an integrated system which

  1. Registers data of the crop in field conditions.
  2. Detects the early signals of pest infection symptoms.
  3. Depicts the most likely scenario of infection spreading.
  4. Foresee the most likely spatial patterns of infection spreading by epidemiology modelling.
  5. Treats initial foci and surrounding areas prone to spreading via the application of protection products.

Data registration is pursued by employing Unmanned Aerials Vehicles (UAVs), equipped with remote sensing tools (multispectral, hyperspectral or thermal sensing systems), allowing thus to integrate visual information (RGB) with deeper insights on crops and fields. Beside the faster and faster technology advances on UAVs, precision application of protection products is almost available.

The detection is early signals of pest infections is achieved by semantic segmentation, employing Deep Learning techniques, namely Neural Networks (NN). In the last decade, Deep Learning has seen a huge development thanks to the Big Data revolution and to great improvements in High Performance Computing (e.g., GPUs). A NN is compound by nodes (neurons) organized in layers: the higher the number of the layers, the deeper the network is. The first layer is the input layer, it receives the data to be classified, while the last layer is the output layer and provides the final classification. A neuron is actually a mathematical function which receives a weighted input from the neurons of the previous layer. It propagates the outcome to the neurons of the successive layer. This outcome is modulated by the activation function (e.g. ReLU, Sigmoid).

Convolutional Neural Networks (CNN) have the potential of dramatically improving the classification capability of very complex patterns (as pest/disease on plants in field conditions), by inserting convolutional filters in the network structure: indeed, this approach has led to promising results in precision crop protection, among other scientific areas.


%d bloggers like this: