Estimate Soil Organic Matter from Remote Sensing Data by Using Statistical Predictive Models

Soil organic matter (SOM) is essential in terms of soil fertility and conservation, and it has an impact on soil physical and biological activity. It is necessary to track the status of SOM and its dynamics frequently. This task is expensive and time-consuming if only a chemical analysis is used. This study was carried out in Sidi Bennour, situated in Morocco's Doukkala irrigated area. Satellite data may offer an alternative and provide low-cost filling for this void. A special value was shown by predicting the SOM using statistical models and satellite images. This study aims to compare SOM predictions using four different models and RapidEye images. The selected models are multiple linear regression (MLR), k-nearest neighbors (k-NN), decision trees (DT), and artificial neural networks (ANN). The obtained results indicate that the ANN provides the best fitted model compared to other models with R-2 = 0.482 in training and R-2 = 0.472 in the test. The ANN model provided the lowest RMSE (0.251) and MAE (0.185). The MLR reveals nearly the same results except that it gives a larger difference between training and test than ANN, with R-2 = 0.455 in training and 0.481 in the test. MLR provided the same RMSE (0.251) as the ANN and an approximate MAE of 0.185.