Overview and objectives:
- The purpose of creating the model is to predict the size of cassava roots based on scan results using Ground Penetrating Radar without having to damage the plants.
- The data provided consists of several varieties, and around 81 layers of radar scan results in the form of images consisting of 7 plants.
Architecture diagram:
From the Yolo mask results, by calculate all the segment yolo predictions, I made an estimate of the boundaries between plants 1 to 7. I used these points to cut the image based on the plant number for the model to use to produce image for training and predictions.
Yolo segment boundary lines based on plant number
To speed up the process of loading masked images, I put the masking results of all images into a dictionary and save to single pickle file.
For training and prediction, from each radar plant image, I stacked images from “Start” to “Start” + 40 vertically.
Then, I will insert the mask value from yolo into each image by taking the maximum value of each pixel between the original image and the image mask.
I also added a picture of the opposing plant number to the right of the plant picture. The mapping of opposing plant numbers is based on the following pairs:
- 1L and 7R
- 2L and 6R
- 3L and 5R
- 4L and 4R
- 5L and 3R
- 6L and 2R
- 7L and 1R
Images used as input for the pytorch-timm model
I also combined the tabular values to combine with the image embed values from timm,
-
difference between End and Start,
-
Genotype
-
Stage
For genotype and stage data, I coded them into numbers.
Data modeling:
I tried several models to get a lot of insight into which model is the best. I used SVR, LGBM, CatBoost, and NN.
For fairly good results I tend to SVR and LGBM, I also tried using CNN but the results were not so good.
For post processing, I calculate the average of each related plant pair and assign it to both plants.
Model validation uses StratifiedGroupKFold by RootVolume with grouped by FolderName with 5 folds, and CV score is calculated using RMSE.
Training takes about 10 minutes, and inference takes a total of about 5 minutes.
The metric I use is RMSE, with 5 StratifiedGroupKFold CV. From the training results, the OOF RMSE I got was 1.45120 with a public score of 1.025004044 and private score of 1.2613859