Convolutional neural network based detection of chestnut burrs in UAV aerial imagery

Uncrewed Aerial Vehicles (UAVs) based remote sensing proved to be a valuable tool to acquire extensive data of crops, which are essential for feeding innovative decision support systems aimed at improving growers’ farm management. In this context, the capability to automatically assess chestnut fields production by UAV has been considered a valuable service by sector stakeholders to properly plan the harvesting tasks and the pricing. In this work, an innovative image processing method to automatically detect pixels representing chestnut burrs in aerial multispectral imagery is presented, which is based on U-Net Convolutional Neural Network (CNN). The method was tested on the imagery of a case study chestnut field located in Dronero (Italy). The aerial imagery was acquired on September 2021 at noon, using an airborne MAIA RGB-NIR camera. The flight altitude was 20 m from the terrain, obtaining a ground sample distance of 10 mm. The Matlab Image Segmenter was used to manually classify pixels (burr, canopy, ground) and obtain datasets for CNN training. Sixteen trees were selected from the map, and, for each tree, the results of the CNN segmentation were compared with those manually obtained. The accuracy in the automatic burrs detection was assessed by defining a confusion matrix, which refers to properly detected or missed burrs, and regions of the image wrongly classified as burrs. The obtained results show the feasibility of the CNN with an accuracy in correct detection of 92.3%. The image processing here presented enables the development of new decision support system specifically conceived for chestnut growers, with a focus on the production and harvesting phases.