MapMaintainer is a system for updating street maps with newly constructed roads by analyzing the progression of satellite imagery over time. In contrast to map extraction methods that process a single satellite image, MapMaintaner compares old and new images of the same location to detect new roads with high precision.
This repository contains the MapMaintainer code for "Updating Street Maps using Changes Detected in Satellite Imagery" (SIGSPATIAL 2021). For more details about the project, see our project website at https://favyen.com/mapupdate/ where our paper and talk are available.
tensorflow-gpu 1.15 with Python 3 is needed to run the components.
pip install tensorflow-gpu==1.15 rtree pillow scipy scikit-image
Go 1.11 is also needed; for newer Go versions, you may need to disable module support:
export GO111MODULE=off
First, download the Massachusetts dataset. This dataset includes 5,000 square kilometers of satellite imagery from each of two years in Massachusetts.
mkdir /data/
cd /data/
wget https://favyen.com/files/mass.zip
unzip mass.zip
The dataset includes satellite imagery from 2013 in sat-2013/, imagery from
2018 in sat-2018/, and road network graphs in graphs/.
Two road networks are included, taken from 2018:
graphs/mass.graphexcludes parking and service roads and should be used for training the tracing model, because we want the model to focus on tracing major roads.graphs/mass-withall.graphincludes all roads and should be used as the base map during inference, because during inference we want to make sure we are only looking for new roads.
The dataset format is detailed at https://github.com/mitroadmaps/roadtracer-beta. Here is a quick summary:
mass_X_Y_sat.jpgis the satellite image spanning pixel (4096X, 4096Y) to (4096*(X+1), 4096*(Y+1)).- The road network graph files are in a text format, and each vertex is labeled
with a pixel coordinate (i, j). For example, a vertex at (4200, 1000)
corresponds to pixel (104, 1000) in
mass_1_0_sat.jpg.
In the commands below, we will assume that the dataset has been extracted in
/data/.
MapMaintainer runs in three steps:
- Change-seeking iterative tracing: trace newly constructed roads that are visible in the new satellite image, but neither appear in the current map dataset nor are visible in the old satellite image.
- Selective change detection: filter the detected roads from tracing to prune false positive detections arising from changes in lighting, off-nadir angle, etc.
- Post-processing: re-construct a new, updated street map by combining the existing map with the new road detections.
Change-seeking iterative tracing produces an initial set of road detections. We train a model that extracts a road network from a single satellite image. Then, we compare model's outputs when independently applied on old and new imagery to robustly identify newly constructed roads.
Process the road network graphs to create directional labels for training the iterative tracing model:
cd changeseeking_tracing
mkdir /data/angles/
go run angle_tiles.go mass -20 -10 5 10 /data/graphs/ /data/angles/
This will create labels for tiles spanning from mass_-20_-10_sat.jpg to
mass_5_10_sat.jpg, which is the training dataset.
Now, we can train the model. It will run for 200 iterations.
mkdir ./model/
mkdir ./model/{model_latest,model_best}
python run_m6.py ./model/ /data/sat-2013/ /data/sat-2018/ /data/graphs/ /data/angles/
Before running inference, create crops of the road network graph corresponding to each image tile. These will be used as base maps that will be extended through tracing.
mkdir /data/tile-graphs-remaining/
python tile_graphs.py /data/graphs/mass-remaining.graph /data/sat-2018/ /data/tile-graphs-remaining/
Now we can apply the model on the test set, which spans from
mass_-13_-22_sat.jpg to mass_12_-10_sat.jpg.
mkdir ./outputs/
python infer_change.py ./model/model_latest/model /data/sat-2013/ /data/sat-2018/ /data/graphs/ /data/tile-graphs-remaining/ ./outputs/
We can visualize the inferred graph using vis.go. For example, to visualize
roads detected in mass_0_0_sat.jpg:
go run vis.go /data/graphs/mass-withall.graph ./outputs/0_0.graph /data/sat-2018/ 0 0
Finally, concatenate all of the output graphs into one road network graph. Here, we use several confidence thresholds so we can get a precision-recall tradeoff.
python graph_cat.py ./outputs/ 30 out_30.graph
python graph_cat.py ./outputs/ 35 out_35.graph
python graph_cat.py ./outputs/ 40 out_40.graph
python graph_cat.py ./outputs/ 45 out_45.graph
python graph_cat.py ./outputs/ 55 out_55.graph
python graph_cat.py ./outputs/ 60 out_60.graph
python graph_cat.py ./outputs/ 65 out_65.graph
python graph_cat.py ./outputs/ 70 out_70.graph
Selective change detection filters the new roads detected through change-seeking iterative tracing to prune false positives that arose due to differences in lightning, off-nadir angle, and other factors between the old and new satellite image.
Prepare per-tile graphs for training. This time, we use the road network with parking and service roads excluded:
cd ../change_model/
mkdir /data/tile-graphs/
python prepare_graphs.py /data/graphs/mass.graph /data/sat-2013/ /data/sat-2018/ /data/tile-graphs/
Next, we train the model. This uses a self-supervisory learning signal so that no labels are required.
mkdir ./model/
mkdir ./model/{model_latest,model_best}
python run.py ./model/ /data/sat-2013/ /data/sat-2018/ /data/tile-graphs/
We can now apply the model on the .graph files containing new roads detected
through change-seeking iterative tracing. This will loop through each connected
component, and score that for matching/mismatched classes.
python apply.py ./model/model_latest/model ../changeseeking_tracing/out_30.graph out_30.graph /data/sat-2013/ /data/sat-2018/
Finally, we can evaluate the output from the second stage:
cd ..
go run postprocess/eval.go /data/graphs/mass-pruned.graph change_model/out_30.graph /data/unmatched/