Random data sampling for an agricultural crop classification task

Abstract

The accurate classification of crops is an important step in sustainable agriculture when it comes to decision-making and optimizing resource management. The effectiveness of machine learning algorithms for this task depends primarily on the quality and representativeness of the training data. In this paper, we propose a random sampling approach to generate training patch sets under optimal conditions for an agricultural crop classification task using Sentinel-2 multi-spectral data. Conventional sampling approaches do not cover all the variability present within agricultural landscapes, thereby reducing the generalization of machine learning models. Our proposed approach is capable of generating more representative patch sets that consider the spatial heterogeneity of crops, as well as the seasonal trends accompanying vegetation growth. We assess the impact of the proposed approach to the classification performed by using the ResNet-18 model. We show experimental results and draw calculations.