A carregar...
Publicação
Machine Learning Approach for Estimating Soil Organic Carbon with Earth Observation Data
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 3.17 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Soil Organic Carbon (SOC) is an important measure of soil fertility and accurate information on SOC is crucial for agriculture and biodiversity in a changing climate. However, the conventional techniques to measure SOC are time-consuming and expensive which produce a coarse-resolution and limited information by excluding climate and topographic characteristics of the soil. Therefore, the integration of Earth Observation (EO) data with Machine Learning (ML) algorithms has the potential to predict SOC. This study estimates SOC in the District Beja, Portugal by preparing the four model datasets (Model-I incorporates Sentinel data (S1, S2, and S3) with other environmental variables; Model-II expands upon Model-I by adding S1 time series data; Model-III simplifies Model-II by Principal Component Analysis (PCA); Model-IV reduces Model-II by Pearson's Correlation). Based on in-situ SOC samples (139), modeling was performed in Google Earth Engine (GEE) on four model datasets using three ML algorithms: Random Forest (RF), Support Vector Machine (SVM), and Gradient Tree Boost (GTB). The results showed that the RF algorithm outperformed achieving the highest R2 value of 0.85 and Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) values of 82.2 g/kg and 59.9 g/kg respectively. Model-II dataset was more predictive in the RF algorithm as compared to other model datasets. Furthermore, to improve the ML algorithms, two band reduction methods (Model-III and Model-IV) were implemented, from which Model-IV performed well in the RF algorithm by obtaining the R2, RMSE, and MAE values of 0.84, 69.0 g/kg, and 53.0 g/kg respectively. Model-IV was unable to achieve the highest R2 value, however, a significant reduction of RMSE and MAE was recorded as compared to the aforementioned Model-II results. The outcomes of these results further prove the effectiveness of the RF algorithm in predicting SOC. Also, this research will provide beneficial information on land with agricultural potential in district Beja contributing to sustainable food production, which is needed to achieve Sustainable Development Goal (SDG) 2 and SDG 13 (Zero Hunger and Climate Change respectively).
Descrição
Dissertation submitted in partial fulfilment of the requirements for the Degree of Master of Science in Geospatial Technologies
Palavras-chave
Soil Organic Carbon estimation Machine Learning Google Earth Engine Sentinel Series Feature Selection Digital soil mapping SDG 2 - Zero hunger SDG 13 - Climate action