1. Advanced speckle filtering
Our raster processing got a major upgrade that significantly improves the quality of our backscatter and coherence rasters. We analyzed, modified and combined multiple published methods when designing a new filter for KappaOne. Although the filter adds significant load to the computations, the output raster will become much sharper with it as the edge structures are preserved much better.
To demonstrate the impact of a new filter to the backscatter images, we chose an area that has comb-shaped parcel on it (see the bottom right corner of figure 1). The narrow “teeth” of the comb are about 20 meters in width. Although “teeth” were also visible previously, their sharpness on KappaOne output is drastically higher.
The newly created KappaOne filter allows improving the quality of coherence images as well. On figure 2 the same area is used for demonstration as previously. Now the comb-like structure is gone, as for coherence more averaging is needed to produce meaningful output. There are large high coherence areas on the previous coherence image, which are created by objects that create extremely high coherence. One of those areas is marked by an ellipse on figure 2. Our novel filter helps to significantly shrink the area where those coherence values dominate. As a result, we can produce an image with much finer details on it.
Mihkel Veske, software developer
2. Towards production: Synthetic NDVI (SNDVI) rasters
The ultimate test of any machine learning model waits in production. Hence in the last month, we have focused on operationalizing the NDVI model as a layer of a new Analysis-Ready Data (ARD) product – KappaOne. In the ARD Germany demo alone, we prepared 35 rasters (each spanning 2500 sqkm) with a newly implemented semi-automated pipeline, that takes collocated Sentinel-1 and -2 products through preprocessing, inference, and post-processing tasks, to create NDVI or SNDVI rasters. 31 of these rasters in Northern Germany were entirely or partially created by our generative model due to cloudy Sentinel-2 images. For larger demo areas such as these, the production objective is to have a new NDVI or SNDVI image within 24 hours of newly acquired data from Sentinel-2.
However, we are tackling new challenges at this stage, foremost of which is validating the SNDVI rasters. Validation involves evaluating the quality of images generated by our model, and the accuracy of parcel-level SNDVI compared with available NDVI. We have made significant progress in the former as reported in previous issues, however by the next issue on SNDVI we plan to have a comparable SNDVI timeseries (in addition to partial NDVI timeseries) in the demo map to validate newly created rasters. Most importantly, we want to learn more about the model as we scale and validate on new data and unseen areas, to improve the accuracy and reliability of predictions.
Hudson Taylor Lekunze, data analyst
