Deep learning based high-throughput phenotyping of chalkiness in rice exposed to high night temperature

Table 6 Chalkiness Segmentation: results of the weakly supervised Grad-CAM approach with the best performing classification models as backbone

Model	GT-known Loc. Acc. (%)	Loc. Acc. (%)	Avg. IoU (%)	Layer	T (%)
Grad-CAM (DenseNet-121)	51.20 = 085/166	51.20 = 085/166	47.44	Features_ denseblock2_ denselayer7_ conv2	60
Grad-CAM (ResNet-101)	84.34 = 140/166	83.13 = 138/166	68.11	Layer2_0_ conv2	60
Grad-CAM (SqueezeNet-1.0)	15.06 = 025/166	0 = 00/166	31.01	Features_12 _expand1x1	60
Grad-CAM (VGG-16)	7.23 = 012/166	7.23 = 012/166	24.92	Features_ module_5	60
Grad-CAM (EfficientNetB4)	28.92 = 048/166	28.92 = 048/166	35.40	Stem _conv	50
Mask R-CNN (ResNet-101)	18.67 = 031/166	N/A	29.63	N/A	N/A

The results of Mask R-CNN with ResNet-101 as backbone are also shown. Only the 166 chalky seed images in the test set were used for chalkiness segmentation evaluation. Performance is reported using the following metrics (as applicable): Ground-Truth Localization Accuracy (GT-known Loc. Acc.), which represents the fraction of ground-truth chalky seed images with \(\text{ IoU } \ge 0.5\); Localization Accuracy (Loc. Acc.), which represents the fraction of ground-truth chalky images, with \(\text{ IoU } \ge 0.5\), correctly predicted by the model; Average IoU (Avg. IoU), which represents the average IoU for the set of chalky seed images. To calculate the IoU, the mask of the predicted chalkiness is obtained using a threshold \(T=60\%\) of the maximum pixel intensity. The last two columns show the layer that was used for generating the heatmap and the threshold used to binarize the heatmap when calculating IoU, respectively

ISSN: 1746-4811