Detection and classification of nitrogen-fertilized fields using sentinel-2 imagery: an AI-based approach with comparative reflectance analysis of fertilizers types

This study addresses the detection and classification of nitrogen-fertilized areas via Sentinel-2 imagery analysis, a topic of increasing relevance due to the widespread use of nitrogen fertilizers and their considerable environmental, climatic, and human health impacts. The use of these fertilizers is strictly regulated at both EU and national levels, with guidelines specifying application timing and quantities. Currently, law enforcement relies on citizen reports and costly on-site inspections, which are highly inefficient. To address this issue, the present research aims to establish guidelines for developing a near real-time detection system, enabling timely warnings for local authorities. Two approaches were tested: the LightGBM machine learning model and the UNet deep learning architecture. Both models were trained to recognize and label nitrogen-fertilized surfaces and supplied datasets that included fertilized and non-fertilized areas pixels. Additionally, a second LightGBM model learned exclusively from fertilized pixels to enhance classification accuracy. The expected outcome was a mask that temporally categorized fertilizer applications into three predefined classes. Recognizing that regulations also limit the amount of nitrogen that can be applied, the analysis was extended to explore whether different nitrogen fertilizers could be distinguished by their reflectance spectra. This distinction is crucial as nitrogen content varies significantly by fertilizer type. Data from certified applications of nitrogen fertilizers, including bovine, swine, ovine and poultry manure, poultry litter, and both liquid and solid digestate, was collected and their reflectance spectra were analyzed using Sentinel-2 bands 4, 8, and 11, along with NDVI, NDMI, and NDWI indices. Although this research focuses on Emilia Romagna region (Italy), it lays the basis for a method potentially applicable in any other area.