Remote estimation of nitrogen is more accurate at the start of the growing season when compared with end of the growing season in miombo woodlands

Abstract

The growing need to include foliar biochemical measurements in global change models has triggered the landscape level application of broadband satellites in nitrogen (N) estimation. In addition, the determination of seasonal variations in the accuracy of estimating N from remote sensing platforms is critical for understanding the reliability of N data used as an input into these models. However, seasonal differences in the accuracies of landscape level remote sensing models in dry deciduous landscapes such as the savanna woodlands remain poorly understood. Our study was carried out in the dry miombo woodland, one of the most expansive deciduous woodland in Southern Africa. A bootstrapped random forest model in the R environment was used to estimate foliar N from sentinel-2 broadband satellite platform. The relationship between foliar N and spectral reflectance was determined at two key phenological stages; the start and end of the growing season. Our results showed that the start of the growing season model (RMSE = 0.42) estimated N concentration better than the end of the growing season model (RMSE = 0.49). The difference between the two statistics are significantly (p < 0.05) different. In addition, maps of the standard deviation of predictions also indicate that the start of growing season model is more consistent (lower standard deviation overall) in its predictions, when compared with the end of the growing season model. We conclude that compared to the end of the growing period, the start of the growing season is a better phenological stage for remotely estimating N in miombo woodlands. This finding is significant in providing higher accuracy maps of N concentration for input into global change models.