ESTIMATION OF SOIL ORGANIC CARBON STOCK IN AGRICULTURE AREA USING SATELLITE BASE REMOTELY SENSED DATA AND MACHINE LEARNING ALGORITHMS: A CASE STUDY IN THE LOWER PART OF NORTHEASTERN THAILAND

Abstract

The significance of soil organic carbon (SOC) in mitigating climate change is important. It aids in the removal of carbon from the atmosphere. Soil organic carbon (SOC) affects soil dynamics, a crucial component of soil organic matter (SOM) and soil fertility. SOM enhances soil structure, microbial activity, and plant nourishment. The rapid conversion of biological systems to agroecosystems is stressing the world's soils. Agriculture depletes soil organic matter (SOM) by changing its physical, chemical, and biological qualities. Thus, global scientists want to estimate soil carbon storage. Carbon content was inferred using soil profiles and plant area ratios. Global soil carbon storage estimates intrigue scientists. Soil profile data, spatial interpolation, and soil-landscape model-based DSM predict SOC. Early studies evaluated average carbon density and carbon content using soil profiles and soil-vegetation area ratios. Estimating soil carbon is difficult. Because soil carbon varies by soil type, plant type, and terrain, there are many assessment methods with different results. Tropical landscapes have varied vegetation, soil, and topography. This complicates SOC mapping and characterization across land cover categories and management regimes. This study aims to using remote sensing and machine learning methods to assess soil organic carbon stock in agriculture areas in lower Northeastern Thailand in 2020. We also want to analyze the relationship between environmental variables and remote sensing data in the research locations and compare soil organic carbon estimation machine learning techniques.

Soil organic carbon storage estimation and monitoring often involve significant fieldwork, soil sampling, and laboratory analysis. Remotely sensed data and machine learning exceed traditional methods. Remote sensing can estimate regional or global SOC. This helps understand SOC distribution and variations among land cover types and management regimes. In this study, multisource remote sensing datasets, including Sentinel-1 Radar data (VV, VH), Sentinel-2 MSI (Spectral index, band red, and red edge band), SMAP (soil moisture), NASA/CGIAR (topography), NASA/GLDAS-2 (meteorological data), and Land Development Department (LDD) of Thailand (land use map), were used as feature variables in this investigation. Additionally, three machine learning algorithms—RF, SVR, and XGBOOST—must be used to estimate soil organic carbon, not only to improve the accuracy of SOC estimates but also to process large amounts of data rapidly, thereby enhancing the efficiency of SOC estimation. This can be beneficial for estimating SOC over large regions or for monitoring frequently over time. The most essential component of data is the soil sampling dataset; 234 soil samples collected via ground survey and laboratory method from LDD in 2020 with difference land use type and soil series were evaluated using a machine learning model.

Multiple sources of remote sensing data were gathered using the open-source Google Earth Engine Python API software which is a highly effective platform for geospatial analysis and remote surveillance. All remote sensing data were analyzed utilizing cloud-based processing, mean value calculation, soil and vegetation index extraction, etc., and resampling at a resolution of 30 meters. Similarly, soil sampling datasets were utilized to calculate the total soil organic carbon stock (ton C/ha-1) in the study area based on soil series data using the kriging method. XGBOOST and RF models processed the model using the cuML GPU package, which provides speed, scalability, accuracy, and flexibility, while SVR used TensorFlow and permutation important to calculate features important in each model. The model was validated by dividing the training data into 70% and the testing data into 30%. The performance of the model was then assessed using R2, RMSE, and MAE.

The result demonstrates that slope, vegetation index, rainfall, soil moisture, and land use play a significant role in SOCS prediction. According to the results of machine learning, XGBOOST has the highest SOCS prediction accuracy among the three models, with R2 = 0.803, RMSE = 5.569, and MAE = 3.204, in combination with the top five features: EVI, SLOPE, Rainfall, TVI, and Soil moisture. The significance of features of SOCS prediction is exceptional in three models: SLOP, EVI, Land use, and Soil Moisture variable. However, RI, NDVI, and Temperatures performed effectively in the SVR model, whereas Red-edge2 performed well in the Random Forest model. Finally, the prediction results from XGBOOST model display the highest SOC storage in rubber tree field with average of 14.29 ton C/ha-1, but overall of SOCS in agriculture area are close proximate with approximately 10-12 ton C/ha-1. In which the terrain is low It has the lowest carbon storage capacity among all plants. The results of this study indicate that remote sensing data combined with environmental variables can predict SOCS on large, complex agricultural lands in Thailand's tropical climate zone.

-