Haiti

Last updated 2 March 2023

Current serverless endpoint: arn:aws:sagemaker:us-east-1:053061259712:endpoint/haiti3

Training summary

This is the first validated Greenstand model. Succinctly, we performed transfer learning using a pre-trained Pl@ntNet Inception v4 backbone that had been trained on the Pl@ntNet300K dataset. The model was further trained on Greenstand seedling data from The Haiti Tree Project (THTP) using 1,945 RGB images resized to 256 x 256 pixels from the 9 most common species. The data was augmented by applying horizontal flips (p=0.5), Gaussian blurs (kernel size of 3), brightness jitter (0.05), and contrast jitter (0.02).

Hyperparameters

• 1240 train images, 312 validation images, 393 test images

• 50 epochs

• 16 batch size

• Adam Optimizer with 0.001 learning rate and 0.0 weight decay

• Focal loss (i.e. class-weighted cross-entropy)

• 0.5 horizontal flip probability for augmentation

• 3 Gaussian blur kernel size for augmentation

• 0.05 brightness jitter for augmentation

• 0.02 contrast jitter for augmentation

Results

Species	True Positive	False Positive	True Negative	False Negative	Frequency	Accuracy	Precision	Recall
ACACAURI	42	2	347	2	44	0.954545	0.954545	0.954545
ANACOCCI	23	1	367	2	25	0.920000	0.958333	0.920000
CATALONG	39	2	351	1	40	0.975000	0.951220	0.975000
CEDRODOR	61	4	325	3	64	0.953125	0.938462	0.953125
DOMBTORR	13	2	378	0	13	1.000000	0.866667	1.000000
GREVROBU	14	1	377	1	15	0.933333	0.933333	0.933333
MANGINDI	145	6	233	9	154	0.941558	0.960265	0.941558
SENNSIAM	10	0	381	2	12	0.833333	1.000000	0.833333
SIMAGLAU	25	1	366	1	26	0.961538	0.961538	0.961538

Overall test set accuracy: 0.947