May 28, W8

Features scaling before inputting into models? log-scaling on skewed distributions? (i remember area)

May 21, W7

Non-transit/stopped tracks pipeline:

• Use (cleaned) radar detections feed through activity XGboost model and generated inferred activity
• Filter out all tracks predicted as transit & stopped with confidence score >= 0.5
• Concat these tracks w/ labelled tagged detections tracks that are also not transit/stopped
• So final dataset (1880 tracks) are from tagged_detections where manual labels are not transit/stopped and radar_detections where inferred labels are not transit/stopped
• Additionally: if these tagged_detections are labelled as one of the fishing categories, label their type_m2_agg field as fishing_boat
• Finally, trained another XGBoost on type classification, test set results: link

Some Analysis on External AIS data source, looks promising with caveats link

May 8, W6

More on CNN architecture exploration:

• Tried using more than 3 input channels for track preprocess: counts, mean speed, mean turning, std speed, std turning.
• Performance degraded severely (~40% accuracy) as model collapses towards predicting the most common class

Next Steps:

• Main: Rerun previous models on non-transit tracks predicted by activity-xgboost-model, see if anything new comes up
• Check XGBoost confidence score: maybe filter out transit tracks with very high confidence score
• External AIS data source?
• Image augmentation for CNN training: rotation & horizontal/vertical flip
• CNN + LSTM architecture? (probably need more data on this front)

April 22, W4

Preprocessing: Establish preprocessing pipeline on raw radar_detections that:

1. One-to-one matching between ais_tracks and radar_detections (results match with Songyu's)
2. Remove tracks with less than 50 detections counts
3. Remove disrupted tracks if max instant speed >= 150 (~8% removed) Results ~14K radar tracks for later use

Time-series Modelling: Treating each detection track as a (T_max, M) matrix, where T_max is padded to the longest sequence in batch, and M features (speed, heading, lat, lon).Fitted preliminary GRU on time-series, concat final hidden layers with sum_stats. Very poor fitting tho with 53% overall acc on test set compared to 76% on XGboost. More literature research on this aspect

Activity: From Songyu's labelled activity data, XGboost can do rly well (almost 100% sensitiviy) on transit and stopped trajectories. Can use this model to infer on radar_detections activity type, and remove tracks in transit to focus on the rest, more interesting trajectories.

• Good for type-activity seperation: e.g. some AIS-labelled fishing tracks seem to just in transit in radar_detections, which shouldn't be of any concern to authority.
• In the remaining tracks, the boat is likely doing something which is definitely correlated with its type: better use of AIS labels

CNN: Computer-Vision Based Model Experiment

• Rasterized vessel trajectory according to methods outlined in this paper: https://arxiv.org/html/2401.01676v1
  • Red channel: Number of detection points within each pixel (capped at 255)
  • Green channel: Average speed of detection points within each pixel (clipped to range (0, 25.5))
  • Blue channel: Average turning of detection points within each pixel (clipped to range (0, 180) unit: degree)
• ResNet-18 Model vastly overfits to the training data even with moderate dropout (p = 0.2, at the end of each resnet block)
  • Training Accuracy: 0.99, Validation Accuracy: 0.89 (XGBoost has 0.92 validation accuracy)
  • Model is really struggling with Hook and Line: very likely to label it as transit, drifting and lobster.
  • Generalization gap presumably caused by the lack of observations in certain activity classes (eg. purse seine, trawl, lobster, hook and line)
  • Highlights the need to perform some form of data augmentation
  • "fishing" can be misleading to the model -> It contains some tracks that should belong to seine, trawl, lobster, or hook, but specific label could not be determined due to a lack of annotation.
  • ResNet is not as effective at vessel type prediction compared to activity: 90% Train Accuracy, 75% Validation Accuracy. Still overfitting. (Too much variance of track pattern within each vessel type?)
  • Currnetly not considering to try deeper models (eg. ResNet-50 or ViT) since moderately large models are already overfitting.

Next Steps:

• Incorporate XGBoost activity prediction model to assist the labeling of more activity data points (especially for fishing activities)
• Merge observations in the "fishing" class to one of the four predefined fishing activities
• Selectively perform data augmentation (eg. only perform data augmentation for the fishing activity classes)

April 15, W3

Insights from Summary Statistics Calc

• Some tracks, with various issues, disrupt summary statistics calculations (ref PPT)

Activity Data Progress:

• Preliminary screening: Removed all invalid tracks as well as tracks with fewer than 50 detections. Currently have 4366 Tracks.
• Tentatively defined the following labels: Transit, Low Speed, Stopped / Anchored, Work, Other / Unspecified, Fishing - Hook and Line, Fishing - Seine, Fishing - Trawl, and Fishing - Others
• For vessels traveling in a straight line (previosuly all manually tagged as transit), categoriezed avg speed >= 5kts as transit, and <5kts as low speed
• Categorized all kayaking activities as low speed
• Categorized fishing activities based on user submitted notes.
• Categorized vessles with an average speed of <= 1km/hr (0.54 kts) as stopped. Manually removed some mistakes the heuristics made.data
• May consider separating lobster fishing into its own category.

Questions and Notes:

• radar_detection file has attribute like speed, course etc. How does radar pick up these attributes? They seem to be really robust to coordinate disruptions
• Radar data seems to be quite noisy on some tracks, denoising techniques are applied to the radar detection time series data (basically a convolution with a gaussian kernel).

Meeting Notes

• Sailboats: two distinctive behaviors: sails down using motors vs sails up (quite some in the dataset), useful to think about for activity prediction
• Site 45: bay area where pleasure crafts (like sailboats putting down anchor), hence lot of detection points w/ few movements
• activity tag: achor/stop (might be different from loitering)
• transit vs. travelling in a straight line w/o changing speed

Objectives

• Think about sensible activity categories
• Look into sum stats calculation file

April 8, W2

Discovery

• Dataset understanding: current working on cleaning and labelling (vessel type, activity) rows with notes the labelled_tracks to produce a better-quality detection datasets (~4000 tracks). Combine w/ radar_detections (~15000 tracks that have info in AIS) to generate around 20K tracks as complete set.
• Radar tracks correspond to 16591 ais tracks, but only 15345 are present in ais_tracks/radar_tracks provided. Examples of some missing ['id_track', 'assoc_id']: (36979855 , 36979840), (32063462 , 32065809), (31251315, 31251147). A total of 1246 missing
• Size EDAs: less informative than EDAs on vessel type, less clear on which direction to focus on. Also size is very dependent on type of vessel as well. One idea will be to come up with a few different size-bucket (small, medium, large) but this is just kinda like categorizing vessels by type
• Follow-up on summary statisitcs calculation functions: tried self-implementation but does have (slight) mismatches w. provided radar_tracks
• Current vessel type aggregation scheme:

Original Category	Aggregated Category
tanker_ship	cargo/tanker
cargo_ship	cargo/tanker
tug	tug/tow
towing_ship	tug/tow
fishing_boat	fishing_boat
commercial_fishing_boat	fishing_boat
military_ship	military_ship
class_b	class_b
passenger_ship	passenger_ship
pleasure_craft	pleasure_craft
sailboat	pleasure_craft
search_and_rescue_boat	other
pilot_boat	other
high_speed_craft	other
law_enforcement_boat	other
other	other
unknown	other

Data Cleaning

• For tagged data: Removed most unwanted tracks (eg. tracks with notes identifying as false targets, disrupted tracks, and tracks I have significant evidence for being false targets, such as those going inland / thru land.)
• Tagged most observations with a straight trajectory as transit. (Added about 1k transit tags during data cleaning.) Training a model for transit tags was not necessary since I removed about 4k unwanted observations during preliminary filtering.
• Tagged many noted tracks with vessel type information to be incorporated into Peter's dataset
• Tagged some tracks with patterns that I believe could be fishing as the unused tag "repairs." ~100 of such tags.

Qs

• More on background algorithm: my understanding is, radar detecting a signal, keeps a log of its tracks, backend algorithm calculates the track summary statistics and feed it to XGBoost model to decide whether the signal is true/false positive (we will need to implement new-feature-functions, need to think about how can it integrate w/ the existing system)
• There are a lot of trajectories in site 45 (I believe thousands) where I see a tracked object lingering / not moving for a very long period of time. (Sometimes lasting for almost a day.) Should I tag these activities as loitering or ignore these observations all together? (Since I cant verify if these are false targets or not.)
• Site 45 also has many kayaking / sailing activities going on. I think it is most appropirate to tag kayaking activities as loitering.
• Songyu: The tagged data has a lot of observations tagged as "kayaks" or "sailboat". Maybe incorporate these observations to Peter's vessel type dataset? For now, I have tagged kayaks as sailboats, but let me know a more approporate type is required.

Next Steps

• For tagged vessel data: I plan to remove all tracks with less than 50 detection points. Such tracks are not very informative on activities and some are likely to be false targets. (I have seen many of such tracks going through land, and yes I removed all the AVES)
• For longer tracks, it may be helpful to break it down into mutliple segments. Most statsitical / ML models expects inputs of fixed dimensions, and the great variation we have in track lengths may need to be considered when we eventually start modeling on the detection data. Splitting detections into segments could also help with potential class imbalance issues later down the line.
• Possibly investigate the transit tags a bit more using the summary statistics. My current vessel tracker app does not display the summary statistics: I could have falsely tagged some fishing / loitering as transit just because they appear to be a straight line on the map. Asking Sam about the typical vessel transit speed could help as well.

...

Objectives

1. Perform some EDAs on vessel size
2. Data Cleaning: Dig deeper into dataset relationship, and try using Sam's method on dataset cleaning: i.e. only find ais-radar track pairs with the most number of detection points
3. Look into vessel activity (distinguish from type), currently three main types of interests: in transit, fishing, leisure
4. Work on the tagger application; perhaps establish a model to predict transit vs not transit? Consider some extent of hand-labelling thru the tagger?
5. (If we get our hands on sum_stats functions): Look into it and try it on tracks

April 5, W1

Trajectory Viewer: different vessel type track viz

Baseline Vessel Type: idea is to build a baseline type classifier (XGBoost) using raw track-level features (0.84 overall accuracy) to which future models can be compared. Currently have 9 types: ['class_b' 'passenger_ship' 'other' 'fishing_boat' 'tug/tow' 'military_ship' 'search_and_rescue_boat' 'pleasure_craft' 'cargo/tanker']

Questions

• Prioritize certain vessel types?
• Birds/Aves in labels ?
• How are summary statistics of tracks calculated? provide functions ?
• Original binary classifier: is it dynamic or static (after gathering enough datapoints of a tracks to obtain summary statistics). Can we look at the code? Should our type prediction follow its paradigm?

Features

Current features

• Min, max, avg of speed over ground
• Curviness ??
• Mean, sd of heading (degrees)
• Mean, sd of turning (the absolute change in heading between consec point)
• Total Distance Travelled
• Max Distance target travelled from origin
• Duration in designated area of interest

New features that might be promising:

• Max lat/lon span (coverage)

Next Steps

• Data requires cleaning ... need to talk abt methods
• Extract more info from notes?

Labeling Application

• I wrote an application (via Python's Tkinter GUI) for visualizing vessel trajectories and modifying the tags in our tagged data.
• Upon investigating the tags, some activity tags appear to be incorrect. (A fishing boat in transit is tagged fishing_c rather than transit.)
• Labels that reflects vessel activities may be more informative for avtivity type classification.
• This application can be later expanded for labeling vessel types, providing additional data for training / validating the vessel type classifier.