Toward Scalable Nitrogen Monitoring: A Sentinel-2 and AI-Based Feasibility Study

After every harvest, soil productivity decreases unless nutrients are replenished. Nitrogen (N) is essential for crop production, but excessive application has well-documented negative effects on the environment, climate, and human health. Consequently, N-fertilizer use is strictly regulated across the EU, governing both application timing and quantities. Currently, compliance monitoring relies on citizen reports and resource-intensive on-site inspections. This research explores the potential of Sentinel-2 imagery for classifying N-fertilizer applications according to the time elapsed since application and the type of fertilizer used. LightGBM and U-Net models were trained for temporal classification, while statistical methods, including Principal Component Analysis (PCA), were used to investigate differences between fertilizer types. Using spectral values from a single date, the models were unable to reliably classify N-fertilizer applications according to the time elapsed since application. In contrast, statistical analysis provided evidence that observations of different N-fertilizer types form distinct groups. These findings suggest that the evolution of N in the soil, from the perspective of Sentinel-2 spectral signatures, may depend not only on the time elapsed since fertilizer application but also on the fertilizer type and environmental conditions. The results further suggest that different N-fertilizer types may be distinguishable through their spectral signatures.