Tweet
All of the bars are designed to flag the possible presence of a detrimental (negative) factor. The longer the red zone, the higher the probability each factor may exist in the recorded colony. In other words, you want to see lots of blue bars.
The AI's look for sound patterns, not simple frequency or amplitude changes. If a sound source has some parts that are similar to one or more colony sounds, it will assign some value to the degree of 'match'.
Each bar is independent of the others. The scale is 0 to 100. The value attempts to assign a probability that the detrimental factor(s) exist.
Let's consider an example. Africanized honey bee (AHB) sounds are one of the factors where the app can score very high on accuracy. For highly Africanized bees, confirmed by wing venation and/or genetic screening, we've seen accuracies as high as 98% (using high-end recorders). Whether a smartphone's audio system is good enough to achieve this level of discrimination is the type of question that app tuning is intended to resolve.
In southern states, like southern FL and TX, I wouldn't be surprised to get recordings, analyses, and inspection reports that indicate some degree of Africanization. At this time, it's too soon to know - we're looking for recordings and analysis that will be useful for calibrating the app on smartphone hardware. So how do we do that?
The ideal Tuning recordings for AHB would be from areas where AHB occurs. One of our go-to groups is bee inspectors. We conducted our original app training for AHB on recordings taken from quarantined AHB colonies in FL. That was before we had the app on a smartphone. Based on colonies that were then screened by an FL lab for a degree of Africanization, we scored as high as 98% accuracy. That's our achievement goal for AHB using a smartphone. However, AHB can vary from mostly European to mostly Africanized, so we need to see how well the app can do at the intermediate hybrid levels.
For most bee colonies in the US, the app tends to report some small probability of AHB. A New York beekeeper reported that on his phone the app always has some low percent probability of AHB for his New York colonies. He went to visit family in the UK. He used the same phone to test some UK colonies, all scored 0 percent probability of AHB!
For Yes/No results like the colony does or does not have a queen, the tuning is fairly straight forward. We want as many verified recordings with visual inspections to verify the presence or absence of a laying queen.
We will want to see a note that indicates how long the colony has been without a queen. Does it have any brood at all? If it has brood, what stages are present? Eggs? Larvae? Pupae? The sound change in queenless colonies tends to decay with time. We need to know when it's been too long to discern a queen loss. Please upload any recordings from queenless colonies with a note that indicates brood presence, absence, and stages. Obviously, if there are eggs, you've likely got a laying queen, even if you don't see her. Few or no eggs, but larvae and pupae, she either stopped laying or was lost with 3-4 days. No eggs, only larvae, could be longer. Only pupae, she's been out of commission for at least 11 days.
For American Foul Brood, the app seems to be working, if not a bit overly sensitive. If you get more than 50% probability of AFB, you should look. We have some preliminary evidence that low levels of AFB that are not visually symptomatic may be detected. Recordings in excess of 50% with a careful visual inspection are useful, but only if you are sure that you can properly identify AFB in early stages (i.e, only a few scattered cells). For AFB and for mites, with recordings and mite wash or sticky trap counts, check the level that seems appropriate, write the actual count number and method posted in a note.
When filling out reports, notice that there is a pencil icon for submitting notes, and a camera option to supplement inspections. Do remember, audio recordings are large files, pictures are large files. Take a lot of pictures and you may spend a lot of time waiting for the files to upload. Reserve pictures for either question - I don't know what this is, or to show something important, like a colony overrun with small hive beetle.
As per Not Normal. That means that the app found some fragment of a sound profile that wasn't normally recorded in a hive of bees. That result could simply be from a poor quality recording, an external sound, or any number of other sources that don't match a healthy colony. To tune for this, we want recordings of hives with bees and stores, and recordings of similar equipment with frames, combs, and honey stores, but without bees. Honey harvest time is perfect. Record your hive or a group of your hives, then after pulling the honey supers, makeup bee less hives by adding a bottom board and a cover and recording. In the notes, tell us the type of bee hive, the number of supers in the stack. We can tune on the boxes with bees versus the boxes without bees.
For alternative hives, like top bar, I find it harder to think about an easy way to simulate bee less hives. That seems to be a good discussion point. Please post your suggestions.
The AI's look for sound patterns, not simple frequency or amplitude changes. If a sound source has some parts that are similar to one or more colony sounds, it will assign some value to the degree of 'match'.
Each bar is independent of the others. The scale is 0 to 100. The value attempts to assign a probability that the detrimental factor(s) exist.
Let's consider an example. Africanized honey bee (AHB) sounds are one of the factors where the app can score very high on accuracy. For highly Africanized bees, confirmed by wing venation and/or genetic screening, we've seen accuracies as high as 98% (using high-end recorders). Whether a smartphone's audio system is good enough to achieve this level of discrimination is the type of question that app tuning is intended to resolve.
In southern states, like southern FL and TX, I wouldn't be surprised to get recordings, analyses, and inspection reports that indicate some degree of Africanization. At this time, it's too soon to know - we're looking for recordings and analysis that will be useful for calibrating the app on smartphone hardware. So how do we do that?
The ideal Tuning recordings for AHB would be from areas where AHB occurs. One of our go-to groups is bee inspectors. We conducted our original app training for AHB on recordings taken from quarantined AHB colonies in FL. That was before we had the app on a smartphone. Based on colonies that were then screened by an FL lab for a degree of Africanization, we scored as high as 98% accuracy. That's our achievement goal for AHB using a smartphone. However, AHB can vary from mostly European to mostly Africanized, so we need to see how well the app can do at the intermediate hybrid levels.
For most bee colonies in the US, the app tends to report some small probability of AHB. A New York beekeeper reported that on his phone the app always has some low percent probability of AHB for his New York colonies. He went to visit family in the UK. He used the same phone to test some UK colonies, all scored 0 percent probability of AHB!
For Yes/No results like the colony does or does not have a queen, the tuning is fairly straight forward. We want as many verified recordings with visual inspections to verify the presence or absence of a laying queen.
We will want to see a note that indicates how long the colony has been without a queen. Does it have any brood at all? If it has brood, what stages are present? Eggs? Larvae? Pupae? The sound change in queenless colonies tends to decay with time. We need to know when it's been too long to discern a queen loss. Please upload any recordings from queenless colonies with a note that indicates brood presence, absence, and stages. Obviously, if there are eggs, you've likely got a laying queen, even if you don't see her. Few or no eggs, but larvae and pupae, she either stopped laying or was lost with 3-4 days. No eggs, only larvae, could be longer. Only pupae, she's been out of commission for at least 11 days.
For American Foul Brood, the app seems to be working, if not a bit overly sensitive. If you get more than 50% probability of AFB, you should look. We have some preliminary evidence that low levels of AFB that are not visually symptomatic may be detected. Recordings in excess of 50% with a careful visual inspection are useful, but only if you are sure that you can properly identify AFB in early stages (i.e, only a few scattered cells). For AFB and for mites, with recordings and mite wash or sticky trap counts, check the level that seems appropriate, write the actual count number and method posted in a note.
When filling out reports, notice that there is a pencil icon for submitting notes, and a camera option to supplement inspections. Do remember, audio recordings are large files, pictures are large files. Take a lot of pictures and you may spend a lot of time waiting for the files to upload. Reserve pictures for either question - I don't know what this is, or to show something important, like a colony overrun with small hive beetle.
As per Not Normal. That means that the app found some fragment of a sound profile that wasn't normally recorded in a hive of bees. That result could simply be from a poor quality recording, an external sound, or any number of other sources that don't match a healthy colony. To tune for this, we want recordings of hives with bees and stores, and recordings of similar equipment with frames, combs, and honey stores, but without bees. Honey harvest time is perfect. Record your hive or a group of your hives, then after pulling the honey supers, makeup bee less hives by adding a bottom board and a cover and recording. In the notes, tell us the type of bee hive, the number of supers in the stack. We can tune on the boxes with bees versus the boxes without bees.
For alternative hives, like top bar, I find it harder to think about an easy way to simulate bee less hives. That seems to be a good discussion point. Please post your suggestions.
Comment