Growing stock volume from multi-temporal landsat imagery through google earth engine

Growing stock volume (GSV) is one of the most important variables for.forest management and is traditionally- estimated from ground measurements. These measurements are expensive and therefore sparse and hard to maintain in time on a regular basis. Remote sensing data combined with national forest inventories constitute a helpful tool to estimate and map forest attributes. However, most studies on GSV estimation from remote sensing data focus on small forest areas with a single or only a few species. The current study aims to map GSV in peninsular Spain, a rather large and very heterogeneous area. Around 50 000 wooded land plots from the Third Spanish National Forest Inventory (NFI3) were used as reference data, whereas more than 8 000 Landsat-5 TM and Landsat-7 ETM + scenes covering both the study period (1997-2007) and area were chosen as a compromise between availability and suitable temporal and spatial resolution to estimate GSV. Google Earth Engine (GEE) was used to handle the huge amount of remotely sensed data. A total of 805 predictors were calculated from Landsat spectral reflectances. Guided regularized random forests algorithm (RF) was used to deal with the arising multicolinearity and identify the most important predictors by comparing with NFI3 plot-level GSV data. As a result, to model the relationship between spectral information and GSV, the original 805 predictors were reduced to only 29 (highlighting texture metrics, vegetation indices and band ratios involving short wave infrared reflectance) while maintaining the accuracy level (R-2 approximate to 0.4 and RMSE approximate to 60 m(3) ha(-1)). A 30-m spatial resolution wall-to-wall GSV map over Peninsular Spain was obtained from a standard RF with the 29 selected predictors through GEE. Its accuracy was evaluated against NFI3 province-level GSV data, resulting in R-2 approximate to 0.91 and RMSE approximate to 15 m(3) ha(-1).