Evaluation of remote sensing methods to estimate forage quality

This review provides a comprehensive assessment of the performance of various remote sensing methods for estimating forage quality parameters in grassland ecosystems, with a primary focus on nitrogen (N) and crude protein (CP) content. Emphasis is placed on comparing the suitability of different spectral resolution types (i.e. hyperspectral (HSI), multispectral (MSI), and RGB), and different modelling approaches i.e. classical regression techniques versus modern machine learning methods. HSI data emerge as the most frequently used and versatile data, given their high adaptability to cover a wide range of spectra. However, MSI and RGB data deliver comparably accurate estimates. While RGB imagery is often used to derive structural variables such as canopy height and biomass, studies focusing exclusively on these quantity-related parameters were not included in this review. Regarding modelling, random forest is the leading machine learning method, while partial least square (PLSR) dominates regression-based approaches. Both the average machine learning (R²: 0.77) and regression models (R²: 0.72) were slightly better for N estimation than for CP estimation (machine learning: R²: 0.69; regression: R²: 0.64). Overall, the differences between modelling approaches were not significant (Wilcoxon rank-sum test), reinforcing the importance of context-specific method selection, influenced by dataset characteristics, environmental conditions, and modelling configurations. The review also revealed major challenges in cross-study comparability, including heterogeneous study designs, inconsistent units of measurement, variable sample sizes, and differing validation procedures. These findings highlight the urgent need for a standardized protocol for remote sensing–based forage quality assessment. Through this review, we identified several key methodological features that should form the basis of such a protocol, integrating insights from both smart agriculture and computational modelling. From a smart agriculture perspective, transparent reporting of sensor specifications together with clearly documented modelling workflows, validation strategies, and units of measurement is essential to ensure reproducibility and enable meaningful comparisons across studies. Integrating well-described sensor hardware characteristics with computational modelling approaches is therefore central for practical implementation, where budget constraints, dataset size, and local environmental conditions ultimately guide tool selection. Synthesizing these aspects provide a robust and widely applicable foundation for advancing remote sensing applications in grassland forage quality assessment.