Image Classification for Precision Agriculture: A Coffee Study Case

Abstract

Precision agriculture has emerged as a transformative approachto optimizing crop production and resource management in theagricultural sector. It leverages advanced technologies like remotesensing, geographic information systems, and decision supporttools. In recent years, machine learning has become an integralcomponent of precision agriculture, providing the ability to analyzelarge volumes of complex data and generate valuable insights formore informed decision-making. This study benchmarked differentsupervised algorithms in an image classification dataset of coffeeand non-coffee areas. Algorithms were comprehensively evaluatedand analyzed, with traditional machine learning models trainedin three versions of the original dataset, and compared with twoDeep Learning algorithms. The experimental results were promising,showing that SVMs and RF algorithms can provide accuratepredictions for most images with average accuracy values above0.8. The RF algorithm, in specific, statistically outperformed all theother traditional algorithms. Deep Learning baselines were slightlymore accurate but had a higher computation cost in the trainingstep. VGG16 was the best-evaluated model, followed by the simpleCNN and RF, with no statistical difference between the last two.Results suggest that there is still scope for further improvement inusing the traditional algorithms.