Semantic Segmentation of L&S Band SAR Data after Tuning the Hyper Parameters in Machine Laearning Models

The machine learning (ML) algorithms are being widely used for image segmentation. Land use and land cover (LULC) classification using remote sensing data can be more fine-tuned by various state-of-the-art ML algorithms. In this work, semantic segmentation on L&S band SAR data is implemented which is collected under NASA-ISRO Synthetic Aperture Radar (NISAR) airborne SAR mission. It is a unique combination of dual band. The Random Forest (RF), and Multi-Layer Perceptron (MLP) models have been applied in present work for image segmentation task and the performance of the model is measured in pixel accuracy and mIOU (mean of intersection over union). The important aspect of this work is to fine tune or evolve the hyper parameters of these models so that the classification accuracy can be improved. The improvement in pixel accuracy after fine tuning the RF classifier is 5.1% and for MLP classifier is 4.35% and in terms of mIOU, improvement is 0.062 for RF and 0.054 for MLP model.