stelar_3dunet

Bias mitigation strategies for agricultural crop type prediction using 3D U-net for spatio-temporal semantic segmentation

Code for 3D unet is adopted from Bhattiprolu, S. (2023). python_for_microscopists. GitHub. https://github.com/bnsreenu/python_for_microscopists/blob/master

Below is the dictionaly of crop types to predict with indices :

{ 0:'NA' , 1: 'ALFALFA', 2: 'BEET', 3: 'CLOVER', 4: 'FLAX', 5: 'FLOWERING_LEGUMES', 6: 'FLOWERS', 7: 'FOREST', 8: 'GRAIN_MAIZE', 9: 'GRASSLAND', 10: 'HOPS', 11: 'LEGUMES', 12: 'VISTA_NA', 13: 'PERMANENT_PLANTATIONS', 14: 'PLASTIC', 15: 'POTATO', 16: 'PUMPKIN', 17: 'RICE', 18: 'SILAGE_MAIZE', 19: 'SOY', 20: 'SPRING_BARLEY', 21: 'SPRING_OAT', 22: 'SPRING_OTHER_CEREALS', 23: 'SPRING_RAPESEED', 24: 'SPRING_RYE', 25: 'SPRING_SORGHUM', 26: 'SPRING_SPELT', 27: 'SPRING_TRITICALE', 28: 'SPRING_WHEAT', 29: 'SUGARBEET', 30: 'SUNFLOWER', 31: 'SWEET_POTATOES', 32: 'TEMPORARY_GRASSLAND', 33: 'WINTER_BARLEY', 34: 'WINTER_OAT', 35: 'WINTER_OTHER_CEREALS', 36: 'WINTER_RAPESEED', 37: 'WINTER_RYE', 38: 'WINTER_SORGHUM', 39: 'WINTER_SPELT', 40: 'WINTER_TRITICALE', 41: 'WINTER_WHEAT'}

Data preparation:

1. make directory mkdir dataset/france2/lai_ras
2. move LAI.zip to ./dataset/france2/lai_ras/ and unzip it.
3. make directory mkdir stelar_3dunet/dataset/france2/processed_lai_npy
4. Run python data_saving_france_RAS_npy.py to save time series .npy files in the directory ./dataset/france2/processed_lai_npy/

Scarcity-Aware Data Slicing and Sampling:

We perform spatio-temporal data slicing and sampling as the dataset is as large as (10002x10002x256). In order to ensure representation of all necessary crop types in the training subsets to train the ensamble of models, we perform data slicing and sampling conditioned with positional information of the crop types .

Run the below code :

export CUDA_VISIBLE_DEVICES=1

cd stelar_3dunet/

conda activate inn

python traiing_data_per_crop.py --chosen_crop_types 27

Crop subgrouping

Each of the models in the ensemble is trained with the below subgroups of crop types Subgroup_indices = [ 1, 2, 3],[4, 5, 7],[ 8, 9, 10],[ 11, 12, 13],[14, 15, 16],[18, 19, 20],[21, 23, 27],[28, 30, 32],[33, 34, 35],[36, 37, 40, 41]

Training:

We train each of the models in the ensamble separately

export CUDA_VISIBLE_DEVICES=0

conda deactivate

conda deactivate

cd stelar_3dunet/

source spt19/bin/activate

python3 3D_unet_data_generator_check_pt.py --crop_1 28 --crop_2 30 --crop_3 32

How to save the test sets of all the subsets of data with which the models are trained

export CUDA_VISIBLE_DEVICES=1

conda deactivate

conda deactivate

cd stelar_3dunet/

source spt19/bin/activate

python3 test_set_saving_10000_spt_points.py --crop_1 21 --crop_2 23 --crop_3 27

python3 test_set_saving_10000_spt_points.py --crop_1 28 --crop_2 30 --crop_3 32

python3 test_set_saving_10000_spt_points.py --crop_1 33 --crop_2 34 --crop_3 35

python3 test_set_saving_10000_spt_points.py --crop_1 36 --crop_2 37 --crop_3 40

python3 test_set_saving_10000_spt_points.py --crop_1 37 --crop_2 40 --crop_3 41

To save the ground truths and predictions for a given experiment and prepare for evaluation

cd stelar_3dunet/

python outputs_and_ground_truth_saving.py

To replicate the quantitative results

Check : ensamble_quantitative analysis_of_results.ipynb

Ensamble of models

Experiment 0

This is a small scale example with lower coverage of landscape and less number of crop types. In this experiment we just considered the 10 percent of overall landscape. We consider corpp types ALFALFA, FOREST, GRAIN_MAIZE, GRASSLAND, LEGUMES, PERMANENT_PLANTATIONS and TEMPORARY_GRASSLAND and train three models to predict these crop types:

Below are the results:

Considered crop type distribution

Intersection over union for crop types

Qualitative results :

1. Sample 1

1. Sample 2

1. Sample 3

Experiment 1

We train 11 different models such that each model predicts three crop types. We then create an ensamble of models and post process a cosolidated crop type prediction by the ensamble of models on a given spatio-temporal patch.

Each of the 11 models trained as follows :

model 1 : types : NA, ALFALFA, BEET, CLOVER samples : 850 each

model 2 : types : NA, FLAX, FLOWERING_LEGUMES, FOREST

.
.
.

model 11: ...

Crop type representation in training data

Intersection over union for crop types

TOo plot ther IOU for experiment 1, refer to the jupyter notebook ensamble_quantitative analysis_of_results_exp1.ipynb

Qualitative results :

1. Sample 1

2. Sample 2

3. Sample 3

4. Sample 4

Experiment 2: Bias mitigation measures

We incorporate measures to mitigate bias through scarcity-aware sampling and class weights in the loss function. We used higher amount of training data in this experiment. The training set consisted of 9000 spatio-temporal training samples

In-order to replicate the results, run the jupyter notebook direct_testing_with saved_test_set.ipynb and follow the instructions in the notebook.

In order to plot the inputs and outputs of this experiment with and without missing values(cloud) interpolation at the inference stage and also to save the ground truths and predictions to further calculate IOU and F1 score, run :

export CUDA_VISIBLE_DEVICES=1 conda deactivate conda deactivate cd stelar_3dunet/ source spt19/bin/activate python3 direct_testing_with_saved_test_set.py

Intersection over union for crop types

F1 score

For quantitive plots of this experimet check the jupyter notebook ensamble_quantitative analysis_of_results_exp2.ipynb.

Qualitative results :

1. Sample 1

2. Sample 2

3. Sample 3

4. Sample 4

Interpolation of missing LAI values during inference.

We performed imputation of missing values through time series pixel-wise interpolation. Below are the examples.

1. Sample 1

2. Sample 2

3. Sample 3

4. Sample 4

We expected that this imputation right before inference would improve the results. But we didn't see significant improvement in results.