This study compared unsupervised and supervised machine learning, and deep learning (U-Net) classifiers on Unmanned Aerial Vehicle (UAV) multispectral imagery to identify nitrogen status in potato crops under nitrogen (N) fertilization treatments, irrigation (I), and their interaction (N × I). The U-Net model outperformed all other methods, achieving accuracies for crop nitrogen status of 65–99% in N, 84–100% in I, and 41–82% in N × I treatments, with variation due to different input data. Supervised machine learning also performed well, with Support Vector Machine achieving 53–87, 66–86, and 32–66% respectively, and Random Forest 61–96, 70–81, and 33–65%. Unsupervised K-means yielded the lowest accuracies (47–58, 9–65, and 8–34%), demonstrating necessity of substantial supervision to delineate crop nitrogen and water status. These findings were confirmed by repeated analyses of UAV imagery acquired later in the growing season with consistent results. Comparable classification performance was observed for crop water status and leaf area index at both time points. Despite being demonstrated in a single-field, single-crop framework, the results provide proof of concept for applying deep learning classifiers to detect subtle nitrogen and water stress under field conditions in precision agriculture. Future research could test diverse agroecosystems and growing seasons, alternative deep learning algorithms, and sensor data fusion to improve classification accuracies.

The high spatio-temporal resolution of UAV sensors requires robust analytical tools to classify subtle agroecosystem variations. This study compared unsupervised, supervised machine learning (ML), and deep learning (U‑Net) classifiers to identify nitrogen (N) and water (I) status, and their interaction (N×I) in potato crops using UAV multispectral imagery. The U‑Net model outperformed all other methods, achieving accuracies of 85% (N), 93% (I), and 70% (N×I). Supervised ML classifiers also performed well and Support Vector Machine achieved 71, 62, and 40% respectively, whereas Random Forest achieved 67, 61, and 40%. The unsupervised K‑means classifier yielded the lowest accuracies (41, 36, and 23%), demonstrating the necessity of substantial supervision to delineate crop N and water properties. These results were confirmed by repeated analysis on UAV imagery acquired later in the season. Deep learning classifiers should be adopted more widely in precision agriculture, as they offer new potential for optimizing N and irrigation co-management under field conditions with subtle spatial variation that is otherwise difficult to capture. Future research should test alternative deep learning algorithms and sensor data fusion to further improve classification accuracies.

Peatlands are significant carbon sinks, heat waves and drought threatens their function as long-term carbon sinks. Detecting stress in Sphagnum-dominated peatlands before damage occurs is critical for conservation and carbon accounting. Here, we present a trial framework combining high-resolution hyperspectral remote sensing with machine learning to identify minimal spectral band sets and develop peatland-specific indices for early stress detection.This study was at the Rzecin peatland in Poland (52°45'N, 16°18'E), a poor fen. Hyperspectral measurements (350–1000nm) were acquired across 13 plots over 22 measurement campaigns, using a Piccolo Doppio dual-field-of-view spectrometer. environmental monitoring included water table depth (WTD) from plots nine using TD divers and meteorological variables (air temperature, vapour pressure deficit, precipitation) recorded at half-hourly intervals from 2020–2024.We defined heat stress events using a compound threshold approach requiring exceedance of the 90th percentile for both daily maximum air temperature (Tair > 29.1°C) and vapour pressure deficit (VPD > 2.90 kPa) for a minimum of three consecutive days. Drought stress was characterised by plot-specific 10th percentile WTD thresholds (site median: −25.4 cm) sustained for at least five consecutive days. Bootstrap resampling (n = 1,000) quantified threshold uncertainty, yielding 95% confidence intervals of 28.6–29.6°C for temperature and 2.71–2.99 kPa for VPD thresholds.To address the hyperspectral multicollinearity, we applied correlation-based filtering (ρ > 0.98), reducing the original 921 spectral bands to 9 representative wavelengths while preserving spectral diversity. Recursive Feature Elimination with Random Forest, validated through leave-one-plot-out cross-validation to ensure spatial independence, identified an optimal subset of eight features: Water Index (WI), Photochemical Reflectance Index (PRI), Peatland Stress-Water Index (PSWI), Normalised Difference Red-Edge Index (NDRE), Peatland Drought Index (PDI), Normalised Difference Vegetation Index (NDVI), reflectance at 800 nm, and the MERIS Terrestrial Chlorophyll Index (MTCI).We are trying to build a peatland-specific spectral indices. The Peatland Drought Index (PDI), calculated as (R705 − R750)/(R705 + R750), exploits the red-edge region's sensitivity to both chlorophyll content and leaf water status. The Peatland Stress-Water Index (PSWI), formulated as (R860 − R550)/(R750 − R670), combines NIR water sensitivity with red-edge slope normalisation. Permutation tests (n = 1,000) demonstrated that PDI significantly outperformed NDVI in detecting VPD-related stress (Δρ = 0.054, p = 0.009), supporting the development of ecosystem-specific rather than generic vegetation indices.Random Forest and XGBoost classifiers achieved strong discrimination between stressed and non-stressed conditions, with areas under the receiver operating characteristic curve (AUC) of 0.836 and 0.851, respectively. The water-related indices (WI, PSWI, PDI) among top-ranked features underscores the primacy of hydrological stress in peatland ecosystems. Sensitivity analysis across varying threshold percentiles (85–95%) and duration requirements (2–7 days) revealed that stress classification varied up to 10-fold, emphasising the critical importance of transparent methodological reporting in peatland remote sensing studies.Our findings demonstrate that reliable stress detection in our peatlands can be achieved with eight spectral features, enabling potential deployment on multispectral sensor platforms. This framework could offer a transferable approach for early-warning systems in peatland conservation, supporting climate adaptation strategies for these critical ecosystem. Acknowledgement: Acknowledgement: Funded by NCN (2020/39/O/ST10/00775), NAWA (BPN/PRE/2022/1/00102), DDSA (2025‑5687), and PANGEOS (CA22136‑80fe26e2).