Mapping spatial variability of foliar nitrogen and carbon in Indian tropical moist deciduous sal (Shorea robusta) forest using machine learning algorithms and Sentinel-2 data

Foliar Nitrogen (N) and Carbon (C) are two vital leaf biochemical components that can indicate forest health. In this study, the spatial distribution of foliar N and C was mapped using Sentinel-2 data in a tropical moist deciduous sal (Shorea robusta) forest of northwest Himalayan foothills of India. Empirical relationships were established between satellite data-derived spectral indices, band reflectance and ground measured foliar N and C using machine learning algorithm (MLA). Performance of MLAs viz. Random Forest (RF), Artificial Neural Network (ANN) and Support Vector Machine (SVM) were assessed. Using all the independent spectral variables, RF performed better than ANN and SVM in explaining the variation in foliar N and C. RF was further used to optimize the independent variables for identifying the best predictor variables. It was found that foliar N is strongly related to shortwave infrared-1 (SWIR-1) band and Normalized Difference Red Edge Index (NDRE). Foliar C was found to be strongly related to SWIR-1 band and spectral vegetation indices: Green Normalized Difference Vegetation Index (GNDVI), Green Red Vegetation Index (GRVI), Green Chlorophyll Index (GCI), and Ratio of Modified Chlorophyll Absorption in Reflectance Index and Optimized Soil Adjusted Vegetation Index (MCARI/OSAVI). Using these best predictor variables, the spatial variability of foliar N and C was mapped using RF algorithm. On validation of the predicted N and C with ground measured foliar N and C, R (2) of 0.85 (RMSE = 0.04%) for N and R (2) of 0.86 (RMSE = 0.26%) for C were observed. It can be concluded that MLAs have great potential to map the spatial variability of foliar N and C in the tropical forests using the broad bands of Sentinel-2 imagery.

Automated summary

This study mapped the spatial distribution of foliar Nitrogen and Carbon in a tropical moist deciduous sal forest using Sentinel-2 data and machine learning algorithms. The study established a strong relationship between Nitrogen and SWIR-1 band and NDRE, and between Carbon and SWIR-1 band along with various spectral vegetation indices. The results showed that Random Forest algorithm performed better than Artificial Neural Network and Support Vector Machine in explaining the variability in foliar Nitrogen and Carbon.