Multi Sensor Data Application with Machine Learning for Classification and Soil Moisture Retrieval

Abstract

The spectral reflectance and/or backscattering characteristics of certain land cover classes or targets can be similar, and are termed as mixed classes. Examples of these include agricultural crops with similar canopy characteristics, non-oriented urban structures, tall vegetation classes, crops at early growth stages, in-field crop variations etc. Classifying these complex scenarios using single-sensor data is challenging due to limited information. Multi-sensor data, capturing at different parts of the electromagnetic spectrum, provides unique insights into various complex scenarios. Synthetic aperture radar (SAR) sensors offer ideal capabilities for modelling complex land cover scenarios, surpassing weather-dependent optical sensors due to their all-day, all weather imaging and sensitivity to various structural and electrical characteristics of the target. The sensitivity of SAR backscatter to different structural components of the vegetation relative to the wavelength and polarization of the incoming signal, SAR data obtained at different wavelengths and polarizations provide valuable complementary information for distinguishing mixed classes by capturing differences in scattering mechanisms. While full polarimetric data improves crop classification accuracy, it is often limited and expensive. Dual polarization SAR data, though less detailed, offers advantages like larger swath coverage and more frequent revisit periods. Other complex land cover scenarios, such as intra-field classification and early-stage crop classification are challenging using coarse resolution satellite data. On the other hand, unmanned aerial vehicles (UAVs) or drones provide high-resolution data at the field level. Hence, there is a need to fuse drone and satellite data for precision agriculture information at a large scale. Traditional parametric-based classifiers often struggle with complex scenarios, highlighting the need for advanced machine learning (ML) and deep learning (DL) techniques combined with multi-sensor data for complex land cover classification. Further, integrating land cover classifications with large-scale agricultural drought assessments can enhance monitoring systems, helping the decision makers make informed decisions at both field and regional levels. Soil moisture (SM) retrieval is an important surface parameter, particularly under vegetation where multiple scattering effects complicate accurate estimation. Several theoretical, empirical, and semi-empirical methods exist for SM retrieval using SAR data, but their application is limited by model complexity. ML models offer an alternative, requiring less a priori information but needing extensive training data. Synthetic SAR data generation provides a cost-effective solution for generating diverse training datasets. Microwave Modelling of different crops enhances SM retrieval by accurately characterizing vegetation parameters.