Estimating Net Primary Productivity of Semi-Arid Crimean Pine Stands Using Biogeochemical Modelling, Remote Sensing, and Machine Learning

Abstract

The aim of the paper was to predict net primary productivity (NPP) in pure Pinus nigra J.F. Arnold (Crimean pine) stands by consecutively implementing remote sensing, biogeochemical modelling, and machine learning tech-niques. In this context, NPP was estimated using Carnegie-Ames-Stanford Approach (CASA). Following, NPP was re-modelled with spectral characteristics of the P.nigra using multi-temporal remotely sensed images (Landsat 8 OLI and Sentinel-2), land use, soils and meteorological information in a total of 180 temporary sample plots. The model results were validated using litterfall samples from 30 stations for each forest stand, including needle, branch, cone, bark, male flower, and others. The highest relationship was between NPP and male flowers (r =--0.75). In addition, reflectance (R), vegetation indices (VI) and texture (TEX) values (calculated according to filter and degree) for each sample plot were calculated from each sensor. Multiple linear regression (MLR) was applied to define the best subset to model the NPP values with R, VI and TEX values using MLR, support vector machines (SVM) and deep learning (DL) methods. The best prediction accuracy was obtained in TEX data in the SVM method and Sentinel-2 sensor combination. NPP testing determination co-efficiency (R2) values were 0.95. The performance of the male flower litterfall in the validation control was promising for the modelling of NPP in Crimean pine. The TEX properties of the satellite images were well reflected by using different filters, degrees, and functions, resulting in achieving a high success.