Datasets and Pretrained Models

Datasets

We use a DJI Mavic Pro 2 to collect many aerial images of waterfowl in Missouri at different altitudes under different weather conditions. The table below provide examples of waterfowl of many colors, shapes, and sizes across many different habitat types in different weather conditions, all of which may affect the performance of deep learning models due to differences in the appearance of the waterfowl. Our datasets available here.

Datasets

Datasets	Location	Number of images	sizes	Number of Birds	Altitude(m)	Season	Bird/Decoy
Bird-A	Mizzou	162	3584*5120	9010	30-120	Spring	Decoys
Bird-B	Eagle Bluff	36	3648*5472	2175	30-90	Fall	Decoys
Bird-C	A Pond	104	3648*5472	964	30-90	Spring	Decoys
Bird-D	Illinois River	339	3840*5760	28837	90	Winter	Birds
Bird-E	Grand Pass	694	512*512	79447	90	Fall/Winter	Birds
Bird-F	Pershing park	26	3648*5472	1936	30-60	Winter/Spring	Birds

Example images of six datasets

Pretrained models

We also provide pretrained model for our six datasets from A to F. Also, we provide general models which are trained by part or all of six datasets. These Pretrained models available here . Download and unzip it into folder models

Pretrained models

Model name	Training dataset	Inference dataset	Precision	Recall	F1-Score
Fasterrcnn-A	Bird-A	Bird-A	0.823	0.893	0.857
Fasterrcnn-B	Bird-B	Bird-B	0.914	0.840	0.876
Fasterrcnn-C	Bird-C	Bird-C	0.540	0.896	0.674
Fasterrcnn-D	Bird-D	Bird-D	0.659	0.804	0.724
Fasterrcnn-E	Bird-E	Bird-E	0.998	0.818	0.899
Fasterrcnn-F	Bird-F	Bird-F	0.916	0.965	0.940
Fasterrcnn-high_density	Bird-D,E
Fasterrcnn-Bird	Bird-D to F
Fasterrcnn-Decoy	Bird-A to C
General_Model	Bird-A to F