Soil respiration estimation in desertified mining areas based on UAV remote sensing and machine learning

Timely and accurate monitoring of soil respiration (Rs) in desertified mining areas is helpful to understand its spatial and temporal distribution and changes, which is crucial for assessing the carbon cycle of ecologically fragile open-pit mining ecosystems. In this study, we acquired multispectral and thermal infrared images of five experimental sites in Hongshaquan mining area including Dump reclamation area, Plantation forest, Tamarisk forest, Southern line, and Hongsha spring by unmanned aerial vehicle (UAV) and collected ground soil respiration data by gas chamber method. The spectral indices were constructed based on the UAV spectral information, and the relevant combinations of independent variables affecting soil respiration were identified by Pearson correlation analysis and Random Forest (RF) importance assessment. Multiple Linear Regression (MLR), Random Forest Regression (RFR), Support Vector Regression (SVR), Back Propagation Neural Network (BPNN), and Particle Swarm Optimization Support Vector Regression (PSO-SVR) were used to construct soil respiration remote sensing inversion model, and through determination coefficient (R 2 ), root mean square error (RMSE) and Akaike’s information criterion (AIC) to evaluate the accuracy. The results indicated that the accuracy of MLR was weaker than machine learning methods, and the highest accuracy model was the PSO-SVR soil respiration inversion model based on the combination of vegetation indices (VIS): Green band chlorophyll vegetation index (CI green ), Red-edge band chlorophyll vegetation index (CI rededge ) and Green wave atmospheric resistivity index (GARI), surface temperature (ST), and salinity index (SI) variables (R 2 = 0.959, RMSE = 0.497, AIC = -0.561). High-resolution UAV multispectral and thermal infrared systems combined with machine learning methods can better estimate soil respiration in desertification mines and provide reference information and basic data for soil respiration and ecosystem carbon cycle monitoring in mining areas.