Efficient selection of SAR features using ML based algorithms for accurate FVC estimation

Abstract

Fractional vegetation cover (FVC) is a ratio of vertical projection area of green vegetation to the total area under consideration. Crops infested by pests, diseases or nutrient deficiency show their impact on the crop coverage. Therefore, FVC is a good indicator of crop health and arid soil. Recently, various models have been reported for FVC estimation using optical data, but it is still limited to different weather conditions. Therefore, it is not feasible to continue crop monitoring using optical data. On the other hand, synthetic aperture radar (SAR) data is weather independent, and has a good potential for crop monitoring in all-weather conditions. SAR data has been used for many crop parameters estimation, however has not been much explored for FVC estimation. Plenty SAR features are available which are sensitive to vegetation parameters. Some of the features are sensitive during early crop stages (e.g., entropy, DpRVI-dual polarization radar vegetation index), while others are sensitive during different stages of crops (backscattering signal of VH and VV polarization). Therefore, there is a need to critically assess all the features and find the optimum combination that provides exemplary results during the entire crop cycle. For this purpose, sixteen features are considered using the different combinations of Sentinel-1 SLC data and their temporal analysis is observed for their different phenology stages. Four machine learning (ML) based models i.e., LightGBM, Xgboost, K-nearest neighbor (KNN), and Random Forest have been explored on these features for FVC estimation. The performance of each model is assessed with the error metrics. Xgboost emerges as the best model with a minimum RMSE value of 0.159. Xgboost model has the capability to recognize the most important features. Due to the stochastic nature of the algorithm, feature priority sequence may vary, therefore, algorithm runs multiple times and the probability of each feature for every position is calculated and on the basis of the highest probability, feature importance sequences is decided. Xgboost model is developed by increasing the input features in the order of their importance sequence and the RMSE value is calculated for each input combination. It is noted that initially, the RMSE value improved from 0.22 to 0.15 for the top five input features. When additional features were included, no further improvement in the RMSE was observed. Therefore; entropy, alpha, VH, VH/VV, and VV are the top five features which are used in the Xgboost model for FVC estimation instead of all sixteen features, which delivers satisfactory results.