Neural network for fruit counting

In 2017, the I2P group is giving its methodological support to the PixYield initiative (2016 Scientific Creativity and Innovation Incentive Action) led by Émile Faye an agroecologist specialising in the design and evaluation of innovative, productive and environment-friendly cropping systems – of the CIRAD’s HortSys RU (Agroecological Functioning and Performances of Hortical Systems). This exploratory action is part of an applied research project conducted by HortSys to develop, in collaboration with ISRA – Institut Sénégalais de Recherche Agronomique – tools for spatialised estimation of mango crop yields in the horticultural region of Les Niayes in Senegal, a subject whose economic stakes are undeniable.

For this exploratory action, the I2P group, called upon for its expertise in automated visual data processing, is moving towards Deep Learning methods that can rapidly process large quantities of images with highly heterogeneous content. In particular, he chose to evaluate the interest and relevance of deep convolutional neural networks (DCNs), described in the scientific literature as suitable for « high-speed » processing of images taken under natural conditions, presenting de facto a strong heterogeneity. This choice is motivated by the constraints induced by the study area selected by HortSys: an area of approximately 500 km² between the cities of Dakar, Noto and Thiès, comprising some thirty orchards selected to take into account the spatial and pedoclimatic heterogeneity of the farms; 300 trees will be monitored monthly for 3 to 5 years by digital land and air acquisitions to estimate the fruit load over the fruiting period, which represents nearly 2000 photos to be processed each month.

The Faster Region-based CNN (F-RCNN) neural network, published in 2016 by Inkyu Sa, is deployed and specialized to identify mangoes in standing tree photographs, regardless of their location in the mango tree and their stage of maturity: this counting is essential for evaluating tree productivity and, ultimately, for evaluating farm performance.

 

Network accuracy

With an average (prediction) accuracy of 62% on heterogeneous data sets, this network is well above the usual Machine Learning methods used in the field (Hung 2013, Payne 2014, Linker 2016, etc.): for example, the KNN classifier (k closest neighbours, Machine Learning) a prediction rate of 40% on the same data sets, scores however much lower than the results published in the literature.

détection des mangues

Expert annotation and network predictions: mangoes are characterized by enclosing boxes (red for annotation and blue for prediction) – 1.2 : false positives = unannotated predictions ; 3 : false negatives = unpredicted annotations. Note that 2 is actually an annotation defect, a fruit not annotated by the expert but detected by the network.

The differences in precision between experiments and published results are currently the subject of a complementary study jointly conducted by the I2P group and the HortSys UR. in particular on the constitution of the training game and the impact of the network configuration on its prediction performance.

Publication

forthcoming

Contact

Ecological aspects , Image aspects

 

L'image au service de l'étude des Plantes et des Paysages