Carbon stock inversion study of a carbon peaking pilot urban combining machine learning and Landsat images

Abstract

Global warming is a significant challenge, and carbon stocks in terrestrial ecosystems are crucial for reducing the greenhouse effect and increasing sinks. A study was conducted in Zhengzhou City from 2000 to 2020 using Landsat image spectral reflectance to analyze changes in carbon stock. Environmental variables such as surface moisture, salinity, vegetation index, brightness, and soil texture were constructed. Multiple linear regression (MLR), support vector machine regression (SVR), random forest regression (RFR), and long short-term memory (LSTM) models were used to invert the carbon stock. The results showed that the NDCS index, constructed using Landsat's blue band and NIR band, was the best inversion variable for carbon stock, with the clay index (CI) playing a primary role. The LSTM algorithm had the best fitting effect on carbon stock, with an R2 of 0.84 and RMSE of 3.56. The carbon stock in Zhengzhou City decreased by 13.93% between 2000 and 2020, possibly due to the large-scale reduction of arable land. Future land-use planning should focus on protecting arable land, optimizing land-use patterns, and enhancing the ecosystem's carbon sequestration capacity.

Keywords

• Spectral reflectance
• Auxiliary environment variable
• Regression algorithm
• Carbon stock
• Correlation analysis