Crop Identification with Monte Carlo Simulations and Rotation Models from Sentinel-2 Data

Abstract

Crop rotation is a well-established practice that helps reduce nutrient depletion and pressure from pests and weeds. At the same time, the use of artificial intelligence tools to recognize crops from satellite multispectral imagery is gaining momentum as a first step toward automated agricultural monitoring. However, the recognition process is limited by inherent errors and the scarcity of available data. In this paper, we build upon Monte Carlo simulation methods to investigate whether incorporating crop rotation information—encoded as a Markov chain—can improve identification accuracy. To broaden the simulation across diverse datasets, we also synthesize multispectral pixels for underrepresented crop types. Crop rotation is used not only in post-processing, but also integrated into the classifier, where a Gradient Boosting Machine is adapted to penalize learners that predict the same crop as in the previous year. Our evaluation uses Sentinel satellite imagery of agricultural crops, combined with the DACIA5 database from the Brasov region of Romania. We conclude that incorporating accurate prior information and crop rotation models noticeably improves crop identification performance. Synthesized data further enhances recognition rates and enables broader applicability, beyond the original region.