An audit and outlier detection system for a neural network prediction model in the agricultural sector

Abstract

This internship report is based on the work accomplished in Asimetrix S.A.S, a Colombian company focused on the implementation of IoT solutions in the agricultural sector. The main goal of this internship is to provide an outlier detection system that is able to improve the accuracy of a image-based weight prediction system. As part of the discovering phase, Surveys were performed to identify client needs and establish metrics of interest from different points of view. Different roads were explored in order to improve the metrics defined by identifying outliers. Firstly, by removing sub-images that were cropped close to the edges of the main image, followed by implementing a second convolutional neural network and finally, by implementing a statistical analysis into the outcome of the prediction system. This final solution was the one that produced the best results, halving the mean prediction error and improving significantly the performance on every other key parameter. The solution was implemented into a production environment via a cloud service provider (AWS) integrating information from external databases.