Feat: long tracker

notebooks/track-floes/track-floes.ipynb

@@ -27,9 +27,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "using IceFloeTracker: pairfloes, deserialize, PaddedView, float64, mosaicview, Gray\n",
-    "using DataFrames\n",
-    "imshow(x) = Gray.(x);\n"
+    "using IceFloeTracker: deserialize, addfloemasks!, adduuid!, add_passtimes!, addψs!, long_tracker, addlatlon!, imshow"
    ]
   },
   {
@@ -80,26 +78,133 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# Load data\n",
-    "pth = joinpath(HOME, \"test\", \"test_inputs\", \"tracker\")\n",
-    "floedata = deserialize(joinpath(pth,\"tracker_test_data.dat\"))\n",
-    "passtimes = deserialize(joinpath(pth,\"passtimes.dat\"))\n",
+    "begin # Load data\n",
+    "    pth = joinpath(HOME, \"test\", \"test_inputs\", \"tracker\")\n",
+    "    floedata = deserialize(joinpath(pth, \"tracker_test_data.dat\"))\n",
+    "    # test/test_inputs/tracker/tracker_test_data.dat\n",
+    "    passtimes = deserialize(joinpath(pth, \"passtimes.dat\"))\n",
+    "    props, imgs = deepcopy.([floedata.props, floedata.imgs])\n",
     "\n",
-    "latlonimgpth = joinpath(HOME, \"test\", \"test_inputs\", \"NE_Greenland_truecolor.2020162.aqua.250m.tiff\")\n",
+    "    # Add required fields to props\n",
+    "    addfloemasks!(props, imgs)\n",
+    "    add_passtimes!(props, passtimes)\n",
+    "    addψs!(props)\n",
+    "    adduuid!(props)\n",
+    "end\n",
+    "\n",
+    "begin # Filter out floes with area less than `floe_area_threshold` pixels\n",
+    "    floe_area_threshold = 400\n",
+    "    for (i, prop) in enumerate(props)\n",
+    "        props[i] = prop[prop[:, :area].>=floe_area_threshold, :];\n",
+    "        sort!(props[i], :area, rev=true);\n",
+    "    end\n",
+    "end\n",
     "\n",
-    "props, imgs = deepcopy(floedata.props), deepcopy(floedata.imgs);\n",
+    "# Delete some floes\n",
+    "deleteat!(props[1], 1); # delete the first floe in day 1 so it doesn't have a match in day 2\n",
+    "deleteat!(props[2], 5); # delete the fifth floe in day 2 so it doesn't have a match in day 1\n",
     "\n",
-    "# Filter out floes with area less than 350 pixels\n",
-    "for (i, prop) in enumerate(props)\n",
-    "    props[i] = prop[prop[:, :area].>=350, :]\n",
-    "end"
+    "# All floes in days 1 and 2 have a match in day 3\n",
+    "# Expected: 5 trajectories, 3 of length 3 and 2 of length 2\n",
+    "nothing # suppress output -- not particularly informative. See the next block."
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "### 4. Pair and label floes"
+    "### 4. View floe data"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Day 1\n",
+    "props[1][!, [:uuid, :passtime, :area]]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(imgs[1])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(props[1][1, :mask])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Day 2\n",
+    "props[2][!, [:uuid, :passtime, :area]]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(imgs[2]) # slightly rotated version of the image in day 1"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(props[2][1, :mask])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# Day 3\n",
+    "props[3][!, [:uuid, :passtime, :area]]"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(props[3][1, :mask]) # missing in day 1"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "imshow(props[3][5, :mask]) # missing in day 2"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 5. Pair and label floes"
    ]
   },
   {
@@ -109,13 +214,57 @@
    "outputs": [],
    "source": [
     "# Get paired floes with labels\n",
-    "pairs = pairfloes(imgs, props, passtimes, latlonimgpth, condition_thresholds, mc_thresholds)"
+    "trajectories = long_tracker(props, condition_thresholds, mc_thresholds);"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "### 6. View trajectories and _goodness_ of pairings"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "trajectories[!, [:ID, :passtime, :area_mismatch, :corr]]"
+   ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "7. Add latitude/longitude data to trajectories"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "latlonimgpth = joinpath(HOME, \"test\", \"test_inputs\", \"NE_Greenland_truecolor.2020162.aqua.250m.tiff\")\n",
+    "addlatlon!(trajectories, latlonimgpth)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "# View trajectories with lat/lon data\n",
+    "cols = [:ID, :area, :passtime, :latitude, :longitude]\n",
+    "trajectories[!, cols]"
    ]
   }
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "Julia 1.9.0",
+   "display_name": "Julia 1.9.4",
    "language": "julia",
    "name": "julia-1.9"
   },

src/IceFloeTracker.jl

@@ -116,6 +116,7 @@ include("segmentation_f.jl")
 include("tracker/tracker-funcs.jl")
 include("tracker/matchcorr.jl")
 include("tracker/tracker.jl")
+include("tracker/long_tracker.jl")
 
 """
     MorphSE

src/tracker/long_tracker.jl

@@ -0,0 +1,161 @@
+"""
+    long_tracker(props, condition_thresholds, mc_thresholds)
+
+Track ice floes over multiple days.
+
+Trajectories are built in two steps:
+0. Get pairs of floes in day 1 and day 2. Any unmatched floes, in both day 1 and day 2, become the "heads" of their respective trajectories.
+1. For each subsequent day, find pairs of floes for the current trajectory heads. Again, any unmatched floe in the new prop table starts a new trajectory.
+
+# Arguments
+- `props::Vector{DataFrame}`: A vector of DataFrames, each containing ice floe properties for a single day. Each DataFrame must have the following columns in the order specified:
+    - "area"
+    - "min_row"
+    - "min_col"
+    - "max_row"
+    - "max_col"
+    - "row_centroid"
+    - "col_centroid"
+    - "convex_area"
+    - "major_axis_length"
+    - "minor_axis_length"
+    - "orientation"
+    - "perimeter"
+    - "mask": 2D array of booleans
+    - "ID": unique identifier
+    - "passtime": A timestamp for the floe
+    - "psi": the psi-s curve for the floe
+- `condition_thresholds`: 3-tuple of thresholds (each a named tuple) for deciding whether to match floe `i` from day `k` to floe j from day `k+1`
+- `mc_thresholds`: thresholds for area mismatch and psi-s shape correlation
+
+# Returns
+A DataFrame with the above columns, plus two extra columns, "area_mismatch" and "corr", which are the area mismatch and correlation between a floe and the one that follows it in the trajectory. Trajectories are identified by a unique identifier, "uuid".
+"""
+function long_tracker(props::Vector{DataFrame}, condition_thresholds, mc_thresholds)
+    begin # 0th iteration: pair floes in day 1 and day 2 and add unmatched floes to _pairs
+        props1, props2 = props[1:2]
+        matched_pairs0 = find_floe_matches(props1, props2, condition_thresholds, mc_thresholds)
+
+        # Get unmatched floes from day 1/2
+        unmatched1 = get_unmatched(props1, matched_pairs0.props1)
+        unmatched2 = get_unmatched(props2, matched_pairs0.props2)
+        unmatched = vcat(unmatched1, unmatched2)
+        consolidated_matched_pairs = consolidate_matched_pairs(matched_pairs0)
+
+        # Get _pairs: preliminary matched and unmatched floes
+        trajectories = vcat(consolidated_matched_pairs, unmatched)
+        trajectories[:, [:uuid, :passtime, :area_mismatch, :corr]]
+    end
+
+    begin # Start 3:end iterations
+        for i in 3:length(props)
+            trajectory_heads = get_trajectory_heads(trajectories)
+            new_pairs = IceFloeTracker.find_floe_matches(trajectory_heads, props[i], condition_thresholds, mc_thresholds)
+            # Get unmatched floes in day 2 (iterations > 2)
+            unmatched2 = get_unmatched(props[i], new_pairs.props2)
+            new_pairs = IceFloeTracker.get_matches(new_pairs)
+
+            # Attach new matches and unmatched floes to trajectories
+            trajectories = vcat(trajectories, new_pairs, unmatched2)
+            DataFrames.sort!(trajectories, [:uuid, :passtime])
+            _swap_last_values!(trajectories)
+        end
+    end
+    IceFloeTracker.reset_id!(trajectories)
+    trajectories.ID = trajectories.uuid
+    # list the uuid in the leftmost column
+    cols = [col for col in names(trajectories) if col ∉ ["ID", "uuid"]]
+    return trajectories[!, ["ID", cols...]]
+end
+
+"""
+    find_floe_matches(
+    tracked,
+    candidate_props,
+    condition_thresholds,
+    mc_thresholds
+)
+
+Find matches for floes in `tracked` from floes in  `candidate_props`.
+
+# Arguments
+- `tracked`: dataframe containing floe trajectories.
+- `candidate_props`: dataframe containing floe candidate properties.
+- `condition_thresholds`: thresholds for deciding whether to match floe `i` from tracked to floe j from `candidate_props`
+- `mc_thresholds`: thresholds for area mismatch and psi-s shape correlation
+"""
+function find_floe_matches(
+    tracked::T,
+    candidate_props::T,
+    condition_thresholds,
+    mc_thresholds
+) where {T<:AbstractDataFrame}
+    props1 = deepcopy(tracked)
+    props2 = deepcopy(candidate_props)
+    match_total = MatchedPairs(props2)
+    while true # there are no more floes to match in props1
+
+        # This routine mutates both props1 and props2.
+        # Get preliminary matches for floe r in props1 from floes in props2
+
+        # Container for props of matched floe pairs and their similarity ratios. Matches will be updated and added to match_total
+        matched_pairs = MatchedPairs(props2)
+        for r in 1:nrow(props1) # TODO: consider using eachrow(props1) to iterate over rows
+            # 1. Collect preliminary matches for floe r in matching_floes
+            matching_floes = makeemptydffrom(props2)
+
+            for s in 1:nrow(props2) # TODO: consider using eachrow(props2) to iterate over rows
+                Δt = get_dt(props1, r, props2, s)
+                @debug "Considering floe 2:$s for floe 1:$r"
+                ratios, conditions, dist = compute_ratios_conditions(
+                    (props1, r), (props2, s), Δt, condition_thresholds
+                )
+
+                if callmatchcorr(conditions)
+                    @debug "Getting mismatch and correlation for floe 1:$r and floe 2:$s"
+                    (area_mismatch, corr) = matchcorr(
+                        props1.mask[r], props2.mask[s], Δt; mc_thresholds.comp...
+                    )
+
+                    if isfloegoodmatch(
+                        conditions, mc_thresholds.goodness, area_mismatch, corr
+                    )
+                        @debug "** Found a good match for floe 1:$r => 2:$s"
+                        appendrows!(
+                            matching_floes,
+                            props2[s, :],
+                            (ratios..., area_mismatch, corr),
+                            s,
+                            dist,
+                        )
+                        @debug "Matching floes" matching_floes
+                    end
+                end
+            end # of s for loop
+
+            # 2. Find the best match for floe r
+            @debug "Finding best match for floe 1:$r"
+            best_match_idx = getidxmostminimumeverything(matching_floes.ratios)
+            @debug "Best match index for floe 1:$r: $best_match_idx"
+            if isnotnan(best_match_idx)
+                bestmatchdata = getbestmatchdata(
+                    best_match_idx, r, props1, matching_floes
+                ) # might be copying data unnecessarily
+                addmatch!(matched_pairs, bestmatchdata)
+                @debug "Matched pairs" matched_pairs
+            end
+        end # of for r = 1:nrow(props1)
+
+        # exit while loop if there are no more floes to match
+        @debug "Matched pairs" matched_pairs
+        isempty(matched_pairs) && break
+
+        #= Resolve collisions:
+        Are there floes in day k+1 paired with more than one
+        floe in day k? If so, keep the best matching pair and remove all others. =#
+        resolvecollisions!(matched_pairs)
+        deletematched!((props1, props2), matched_pairs)
+        update!(match_total, matched_pairs)
+    end # of while loop
+    return match_total
+end