The normalized difference vegetation index (NDVI) is a key parameter in precision agriculture used globally since the 1970s. The NDVI is sensitive to the biochemical and physiological properties of the crop and is based on the Red (~650 nm) and NIR (~850 nm) spectral bands. It is used as a proxy to monitor crop growth, correlates to the crop coefficient (Kc), leaf area index (LAI), crop cover, and more.

Yet, it is susceptible to clouds and other atmospheric conditions which might alter the real NDVI value of the crop. Synthetic Aperture Radar (SAR), on the other hand, can penetrate clouds, hardly affected by atmospheric conditions, but is sensitive to the physical structure of the crop, therefore, does not give a direct indication of NDVI.  

Several SAR vegetation indices have been suggested to estimate NDVI via SAR, however, they tend to work for specific spatial and temporal settings and do not work well globally. As the effort to estimate NDVI from SAR continues, we understand that the solution should be local, and even hyper-local. Meaning, not only that each field has a different NDVI-SAR relationship, but this relationship might also change during the growing season.  

In this study, we present a hyperlocal machine learning approach to estimate NDVI from SAR images for agriculture fields. Each time a SAR image is available over an agricultural field, a machine learning model will learn the relationship between past NDVI and SAR values, for that specific field, and consequently will make an estimation of the crop NDVI value from the current SAR image. Then, when the next SAR image is available, the model will re-learn the relationship (based on past values) which might have changed, thus, the model is kept Up-To-Date. Results from various crops around the world show improvement in NDVI estimation compared to published SAR indices.  

The outcome of this study is ensuring a constant stream of NDVI values, regardless of atmospheric conditions, which is crucial in cloudy areas and at specific times during the growing season such as when crops start their development stage. This has the potential to detect and attend to agronomic events such as irrigation malfunction or crop diseases seamlessly during the growing season.