Predicting soil organic carbon with ensemble learning techniques by using satellite images for precision farming.

Soil plays a major role in the agricultural system. Soil composition detection can help farmers to take appropriate decision leading to proper crop growth. Soil organic carbon is crucial for many soil activities and ecological characteristics, is at the centre of sustainable agriculture. The goal of the research is to create a system for evaluating soil organic carbon based on topographic features and soil properties incorporating machine learning algorithms. A group of covariates has been chosen to function as potential predictor factors for soil properties, including four topographical variables, two soil-related remote sensing indices, and four climate variables which were retrieved from satellite images. Along with predictor variables, soil health card data as dependent variable was used for training the model. It was notified that bagging and boosting showed good results for training than for testing. XGBoost algorithm noted highest R as 0.95 and lowest RMSE as 0.03 with sMAPE as 0.04 while using Random Forest it was identified that R was 0.86, RMSE was 0.06 and sMAPE was 0.08. For testing dataset, RMSE ranged between 0.15 and 0.16 while sMAPE recorded as 0.19-0.20 and R was recorded as 0.12 for Random Forest and 0.03 for XGBoost algorithm. Stacking method proved its significance prominently in overcoming the problem of overfitting as compared to other two methods. For stacking method, R was recorded low having numeric value for training dataset as 0.17 and testing dataset as 0.07 but RMSE for both datasets was nearly same, as 0.16 and sMAPE as 0.18-0.20. This system will assist farmers in making decisions about applying fertilizer precisely which will increase crop yield. Application of ML techniques on remote sensing data can help to build a decision support system in precision farming for improving crop yield.