WEBVTT

1
00:00:00.000 --> 00:00:05.667
AquaLens summarizes freshwater specimen images by sampling site and year.

2
00:00:05.667 --> 00:00:11.833
It combines the views of each specimen, estimates label-group counts, and exports a report.

3
00:00:11.833 --> 00:00:20.583
For a monitoring analyst, the report brings together the recorded images, model predictions, and uncertainty for each sampled batch.

4
00:00:20.583 --> 00:00:27.417
Environmental monitoring depends on comparing observations across places and sampling periods.

5
00:00:27.417 --> 00:00:36.083
Researchers can use these reports to examine a sample's composition and check uncertain predictions before carrying out further analysis.

6
00:00:36.083 --> 00:00:42.875
AquaLens groups the records by site and year and provides matching tables for communication or reuse.

7
00:00:42.875 --> 00:00:46.958
Our evidence comes from standardized Finnish lake specimens.

8
00:00:46.958 --> 00:00:54.917
Applying this workflow to citizen-collected urban-stream observations will require validation on those images.

9
00:00:54.917 --> 00:00:58.833
Select a site-year batch to open its completed report.

10
00:00:58.833 --> 00:01:03.042
Twenty-nine reports cover over eighteen thousand test specimens.

11
00:01:03.042 --> 00:01:07.875
The image browser shows real specimen views and their stored predictions.

12
00:01:07.875 --> 00:01:11.375
Change views to inspect the image evidence.

13
00:01:11.375 --> 00:01:15.708
Top-label counts assign each specimen to its most likely group.

14
00:01:15.708 --> 00:01:20.208
Probability-weighted counts sum the model's probabilities across specimens.

15
00:01:20.208 --> 00:01:23.125
Both estimates are shown here for comparison.

16
00:01:23.125 --> 00:01:26.292
Low confidence is flagged automatically.

17
00:01:26.292 --> 00:01:32.208
Download matching CSV and JSON records for further analysis.

18
00:01:32.208 --> 00:01:40.250
AquaLens combines existing recognition models with specimen-level prediction alignment, batch aggregation, and a browser interface.

19
00:01:40.250 --> 00:01:45.000
We evaluate both classification scores and errors in the resulting reports.

20
00:01:45.000 --> 00:01:53.583
The L/14 comparison shows why this matters: fine-tuning improved classification and count estimates, while richness error increased.

21
00:01:53.583 --> 00:01:57.000
Report quality therefore needs its own evaluation.

22
00:01:57.000 --> 00:02:02.042
The interface offers both count estimates so users can inspect the difference.

23
00:02:02.042 --> 00:02:10.500
The implementation combines established methods; our contribution is the integrated reporting workflow and its evaluation on the same

24
00:02:10.500 --> 00:02:12.208
specimens.

25
00:02:12.208 --> 00:02:20.542
We compared two biological vision models: BioCLIP 2.5, H/14, and BioCLIP 2, L/14.

26
00:02:20.542 --> 00:02:25.375
Each used a frozen control and six fine-tuning epochs on the same specimen split.

27
00:02:25.375 --> 00:02:30.667
We used mixed precision and gradient checkpointing to fit training within the available memory.

28
00:02:30.667 --> 00:02:35.583
Gradient accumulation kept the effective batch at sixty-four specimens.

29
00:02:35.583 --> 00:02:41.833
We aligned specimen and class identities before selecting the blend on validation data.

30
00:02:41.833 --> 00:02:48.750
Validation selected sixty percent H/14 predictions and forty percent L/14 predictions.

31
00:02:48.750 --> 00:02:52.792
Known-test accuracy reached 92.35%.

32
00:02:52.792 --> 00:03:00.917
Macro F1 was 54.99%: slightly below H/14 alone, despite higher accuracy.

33
00:03:00.917 --> 00:03:06.792
Mixture weights were chosen on validation; test scores were calculated afterwards.

34
00:03:06.792 --> 00:03:13.792
The experiments cover forty-three thousand specimens and three hundred forty-five thousand selected frame identities.

35
00:03:13.792 --> 00:03:19.542
The local dashboard runs with Python and a browser, without a GPU or paid service.

36
00:03:19.542 --> 00:03:22.125
This screen replays completed results.

37
00:03:22.125 --> 00:03:28.958
A separate image-inference command accepts specimen folders or a manifest and writes new predictions and batch reports.

38
00:03:28.958 --> 00:03:37.083
It requires a trained checkpoint. The repository documents the data, configuration files, and report formats.

39
00:03:37.083 --> 00:03:44.708
The prototype includes a specimen browser, batch reports, model comparisons, and an image-inference command.

40
00:03:44.708 --> 00:03:48.167
Testing on urban-stream imagery is the next step.

41
00:03:48.167 --> 00:03:56.125
The current results describe the sampled organisms; they do not establish water safety or human-health outcomes.