Address comments

Author: prantogg

Analyzing_Data/Raster_Text_To_Segments.ipynb

@@ -39,9 +39,9 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Load NAIP imagery for Maimi Airport\n",
+    "## Load NAIP imagery for Miami Airport\n",
     "\n",
-    "We will use the native raster reader to load Geotiffs as out-of-database or \"out-db\" rasters and perform dyanamic tiling on read.\n",
+    "We will use the native raster reader to load GeoTIFFs as out-of-database or \"out-db\" rasters and perform dyanamic tiling on read.\n",
     "Spliting the large GeoTIFF into small tiles improves the distribution of workload across the cluster."
    ]
   },
@@ -72,7 +72,7 @@
     "Before running inference, it’s useful to explore the imagery itself.  \n",
     "We’ll visualize the footprints of the raster tiles with SedonaKepler and preview a few raw images using SedonaUtils. This gives us confidence that the data is being read correctly and aligned spatially before applying any models.\n",
     "\n",
-    "Tip: You can also save the Kepler map as an interactive HTML file with `kepler_map.save_to_html()`."
+    "> Tip: You can also save the Kepler map as an interactive HTML file with `kepler_map.save_to_html()`."
    ]
   },
   {
@@ -102,17 +102,19 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## Running Segementation using SAM2 model\n",
+    "## Running segementation using SAM2 model\n",
     "\n",
-    "Now we’ll run inference over the raster tiles using Wherobots’ SQL function `RS_Text_to_Segments()`.\n",
+    "Now we’ll run inference over the raster tiles using Wherobots SQL function `RS_Text_to_Segments()`.\n",
     "\n",
-    "For this example, we specify the follwing parameters -\n",
+    "For this example, we specify the following parameters -\n",
     "- Model: `\"sam2\"`\n",
     "- Text prompt: `\"airplanes\"`\n",
     "- Confidence threshold: `0.5`\n",
     "\n",
-    "The function returns predicted segments for each raster in our region of interest.\n",
+    "The confidence threshold controls which detections are returned (scores range from 0 to 1).  \n",
+    "For this particular model, most positives have confidence scores of at most ~0.7, so we start with `0.5` to favor **recall** on this model/dataset. This helps us understand how the model performs before tightening the threshold later.\n",
     "\n",
+    "The function returns predicted segments for each raster in our region of interest.\n",
     "We’ll cache the results for efficiency and register them as a temporary view so we can explore them further."
    ]
   },
@@ -343,6 +345,14 @@
     "preds_filtered.show()"
    ]
   },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "Each raster tile returns multiple predictions as arrays.  \n",
+    "We use `explode` to flatten these arrays so every detected object becomes its own row for easier mapping and analysis."
+   ]
+  },
   {
    "cell_type": "code",
    "execution_count": null,
@@ -442,7 +452,7 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "## <span style=\"color: #7b73e2;\">Next Steps with Raster Inference</span>\n",
+    "## Next Steps with Raster Inference\n",
     "\n",
     "With access to general-purpose, text-promptable models, what will you predict and georeference next?\n",
     "\n",
@@ -454,6 +464,11 @@
     "\n",
     "We're excited to hear about what you're doing with SAM2 and OWLv2! "
    ]
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": []
   }
  ],
  "metadata": {