A Reflection on AI Model Selection for Digital Agriculture Image Datasets

Seth Ockerman, John Wu, Christopher Stewart, Zichen Zhang

Keywords: Computer Vision, Digital Agriculture, Machine Learning, Deep Learning

TL;DR: We studied whether widely used image processing datasets should be used as benchmarks for model selection on digital agriculture datasets and proposed an approach to improve model selection for digital agriculture datasets

Abstract:

Cameras, sensors, and autonomous vehicles deployed in agricultural
settings are producing large, complex, and highly multidimensional datasets.
Artificial intelligence techniques can extract insights hidden within
these datasets to automate crop management and develop better farming practices.
In particular, recent studies have shown that neural networks can accurately
characterize crop health conditions within digital agriculture datasets.
However, choosing between neural network architectures is challenging;
One must select from multiple architectures and hyper parameters.
Benchmark datasets, i.e., datasets that represent a class of similar datasets,
are often used to select models for digital agriculture datasets.
However, if benchmark datasets are not faithful representatives for
digital agriculture datasets, their use could lead to poor model selection.
This paper demonstrates the danger of using standard vision benchmarks to
inform model selection for digital agriculture datasets.
We then propose a gradient-boosting prediction approach that would
significantly reduce costs to benchmark digital agriculture datasets directly,
which could improve the fit between model and dataset.