Figure 1 shows the total atmospheric absorption of the Standard US Atmosphere with humidity levels of 0, 30, 70, and 100% within the oxygen absorption complex between 50 and 70 GHz and atmospheric windows on either side of it. This oxygen absorption complex is a unique natural resource for temperature sounding, since the oxygen absorption, used for atmospheric temperature profiling, is about two orders of magnitude stronger than water vapor absorption. There is no other oxygen absorption line that has the same properties.

The green areas show the frequency spectrum that is reserved for passive use, such as for passive microwave remote sensing. Black sections delineate the channels of the Advanced Technology Microwave Sounder, ATMS. Red areas below 52 GHz are Starlink frequencies, and the red between 57 and 71 GHz is unlicensed spectrum.

Despite the importance of this portion of the spectrum for passive microwave remote sensing, only a very narrow bandwidth is reserved for passive applications. The clear overlap between the bands showcases just how vulnerable these channels are to interference.

BEST has developed a hardware solution and an algorithm to minimize the impact of Radio Frequency Interference (RFI) both within and near the oxygen complex. Such an approach can be generalized to other frequency ranges as well.

Our algorithm is a convolutional neural net RFI detector, meant to effectively filter any RFI-contaminated observations, while retaining as much uncontaminated data as allows. It has been developed at BEST with data sourced from the ECO1280 nature run dataset, in conjunction with our proprietary radiative transfer simulation. These compose the foundation of our algorithm’s training and validation data, with uses synthetic RFI generation to simulate contamination on various channels.  

Our solution also utilizes simple, low power consumption analog multiplexers designed for low channel NEDT, which can significantly improve microwave sensors observations in this and other frequency ranges. Most RFI detection algorithms proposed to date, utilize power-hungry digital backend processing.

Figure 2 is an example which shows the effectiveness of our algorithm for RFI detection, in physical scenarios in tropical regions, with precipitation, over the sea. Each entry shows model recall, which is defined as:

Where the TP is true positives, e.g., properly detecting the RFI when it is present. FN is false negatives, the failure to recognize the presence of RFI on a contaminated channel. In other words, when RFI was present on a channel, how often did the model recognize it?

The vertical axis of the plot corresponds to the number of channels that were randomly contaminated for a given observation during testing, and the level of interference (in Kelvin) added to these contaminated channels is on the horizontal axis. The development and testing of the algorithm considers the sensors’ internal noise, which is about 0.32 K on average for all channels. As can be seen, the detection the RFI is very likely, above 80% likelihood, even when five channels are contaminated at just 1.28 K. The algorithm can detect a very low level, e.g., 1.6 K of the RFI in just one channel or even when up to 15 channels are contaminated. All recall scores also correspond to a less than 2% False Positive Rate, meaning low loss in usable data. The sensitivity level of the algorithm for RFI detection can also be adjusted. For example, if the goal of detection is identifying any possible RFI, the detection rate can be raised at the expense of increased false negatives.