An Artificial Intelligence System for Combined Fruit Detection and Georeferencing, Using RTK-Based Perspective Projection in Drone Imagery

TL;DR

This paper introduces an AI system using Faster R-CNN for detecting and georeferencing apples in drone imagery, employing RTK data and a novel hyperparameter tuning approach to improve accuracy in orchard monitoring.

Contribution

It presents a novel preprocessing stage for individual tree cropping and a k-means based anchor box design for Faster R-CNN, enhancing apple detection accuracy in aerial imagery.

Findings

Achieved a calibrated mAP of 0.7627 for apple detection.

Optimal hyperparameters include 225 box proposals, learning rate 0.001, and ResNet 101 backbone.

K-means clustering significantly improved detection performance.

Abstract

This work presents an Artificial Intelligence (AI) system, based on the Faster Region-Based Convolution Neural Network (Faster R-CNN) framework, which detects and counts apples from oblique, aerial drone imagery of giant commercial orchards. To reduce computational cost, a novel precursory stage to the network is designed to preprocess raw imagery into cropped images of individual trees. Unique geospatial identifiers are allocated to these using the perspective projection model. This employs Real-Time Kinematic (RTK) data, Digital Terrain and Surface Models (DTM and DSM), as well as internal and external camera parameters. The bulk of experiments however focus on tuning hyperparameters in the detection network itself. Apples which are on trees and apples which are on the ground are treated as separate classes. A mean Average Precision (mAP) metric, calibrated by the size of the two classes, is devised to mitigate spurious results. Anchor box design is of key interest due to the scale of the apples. As such, a k-means clustering approach, never before seen in literature for Faster R-CNN, resulted in the most significant improvements to calibrated mAP. Other experiments showed that the maximum number of box proposals should be 225; the initial learning rate of 0.001 is best applied to the adaptive RMS Prop optimiser; and ResNet 101 is the ideal base feature extractor when considering mAP and, to a lesser extent, inference time. The amalgamation of the optimal hyperparameters leads to a model with a calibrated mAP of 0.7627.