Machine learning based a time series predictive analysis of biomass and carbon for Urban forests: A Case study in Jodhpur, Rajasthan, India

Urban forests are diverse ecosystems that offer tremendous services to the urban environment and the biotic and abiotic components associated with them. The awareness about the carbon storage by urban forest is a prerequisite for understanding their importance for environmental benefits. So a reliable approach for estimation of Urban Forest Biomass and Carbon (UfBC) stocks is important.

The present study assessed the predictive accuracy of three machine learning (ML) algorithm i.e. Linear regression method represented by Partial Least Square Regression (PLSR), Support Vector Machine (SVM) a Non-Linear method and an ensemble algorithm eXtreme Gradient Boosting (XGboost), for quantitative prediction and mapping patterns of Urban UfBC for a part of Jodhpur city, Rajasthan, India between 2013 and 2019, after integrating field detected observations and Landsat 8 data. A combination of 7 Spectral bands and 19 Vegetation indices, total of 26 variables has been used for ML models. Out of three ML models with PLSR and SVM models overestimation and underestimation noted for prediction and mapping patterns of UfBC but XGboost performed well to overcome these problems with R² =0.85 ,MAE =12.62t/ha and RMSE= 15.62t/ha.While with SVM,R² =0.75 ,MAE =10t/ha and RMSE =20.3t/ha have been noted ,least values noted against PLSR as R² =0.70 ,MAE =14.82t/ha and RMSE =22.174t/ha.An increase in area for UfBC is noted between 2013 and 2019. 2-120t/ha Biomass and 0.95-57tC Carbon range noted for 2013 while for 2019 Biomass ranges from 3-138t/ha and carbon1.42-65.5tC. In the current research Urban forest typically grouped into three main classes scattered, Linear and Plantation .Major increase in the area during study period noted for class plantation followed by scattered class. The study showed promising results in the mapping of UfBC based on the ensemble ML approach.