Earth ecology often studies the interaction of several living beings and environmental factors. This is typically achieved by localizing and quantifying objects in large expanses of land. While studying a particular problem, many ideas come to mind and being able to experiment quickly in large amounts of data saves time and produces results. Here, we'll present one such experiment. 

Object Detection

In this example we are trying to automatically detect livestock enclosures, called Boma, in Serengeti in order to see any livestock influence on wildebeest migration patterns. We start with about 100 GeoTIFF images with Bomas present. ViQi recognizes GeoTIFF metadata and will be able to export automated annotations into geo-spatial formats such as GeoJSON or KML.

image-13732.png

Training

We have started by creating a semantic type "Boma" and clicking on a few points per image creating 210 annotations of Bomas in our dataset in about 20 minutes. Training a model on this binary problem quickly created an automated recognizer of Boma objects. However applying this model to detect Boma objects showed that many uniform grassy areas were also detected as Boma. 

Then, we spent another 30 minutes by clicking on locations containing "Bushes", "Grassland" and "Livestock" with 198, 176, and 108 samples respectively. Since there are so few training samples (700 for 4 classes) the system augmented their number by 10 times by rotating, scaling and otherwise perturbing the samples giving a bit more data for the deep network to chew on.

model_bomas.png

This time the model demonstrated much better accuracy although "Bushes" class has a high detection error. Since our true objective is to only detect Bomas we could simply ignore all the other classes to improve the Boma detection accuracy. Moreover, in order to further increase the Boma detection accuracy we increased sample-level goodness to 99%. This discarded 32% of less certain samples and the final Boma detection accuracy increased to 71% and the error decreased to only 0.3%.

Using region detection mode of the Connoisseur system we can create polygons delineating detected Boma structures in hundreds of satellite images using this model.  

Conclusion

With very little training data and approximately one hour of user interaction the Connoisseur system created a reasonably performant model that could be applied at scale to detect Boma structures. Annotations automatically created by this model can be exported to geo-spatial analysis software thanks to ViQi's flexible metadata model. Interested in trying your own data? Contact us today to get a demo.

Learn more about ViQi Connoisseur for your machine learning needs.