An inter-comparison of different PSO-optimized artificial intelligence algorithms for thermal-based soil moisture retrieval

Soil moisture is one of the most important variables which affects different aspects of human life such as agriculture, flood, landslide, water resources, etc. There are different methods for modeling soil moisture such as conceptual, empirical and physically based models. Conceptual and physically based models are robust but they need several different data and information. However data driven models can be run using limited number of data. In this study, four data driven models i.e. MLP, ANFIS, SVR and GMDH were used to model soil moisture. Particle Swarm Optimization technique was used to optimize the structure of the four aforementioned models. Several different remote sensing-based indices were calculated using Landsat Imagery e.g. NDVI, TVDI, VTCI and TVX. An extensive field survey was conducted to collect soil moisture data in the region. A 70/30 ration was used to separate train and test data. 30 additional samples were used for a final validation of produced maps. Results showed a relatively poor performance of PSO-MLP model. The performance of PSO-ANFIS and PSO-SVR was moderate with R 2 of 0.74 and 0.84 and RMSE of 3.4% and 3.1%, respectively. PSO-GMDH had a superior performance with R 2 of 0.91 and RMSE of 2.4%. Therefore, PSO-GMDH can be suggested as a powerful modeling approach to produce soil moisture maps.