Cropland: Refugee LCLUC North Uganda 2022

[cnakalembe]

• Labeling projects created and links added below
• Set 1 Labeling: Gedeon, Manthan, Bhanu, Taryn, Isha, Ivan, Abena, Adebowale
• Set 2 Labeling: Sheila, Snehal, Aninda, Diana, Aditya, Mirali, Alana
• Data added to repository
• Intercomparison analysis
• Model trained
• Map made
• Expert check

[hannah-rae]

@MsPixels will confirm details with @cnakalembe and break out into individual issues

[cnakalembe]

We have to do annual-- something worth exploring later which is outside the current scope is looking at rapid changes (month to month) when new settlements are established

[MsPixels]

New CEO Institution created for this project @cnakalembe

[MsPixels]

CEO Labeling update: 53.6% analyzed for Set 1 and 50.70% completed for Set 2

[MsPixels]

@hannah-rae, CEO Labeling Update: 61.10% for Set A and 51.70% for Set B

[MsPixels]

@hannah-rae, CEO Labeling Update: 902 (90.20%) for Set A and 673 (67.30%) for Set B

[MsPixels]

Intercomparison analysis for Uganda North @cnakalembe

[MsPixels]

Uganda North Trained model @cnakalembe :

"Uganda_North_2022_V1": {
"params": "https://wandb.ai/nasa-harvest/crop-mask/runs/v5f8m489",
"test_metrics": {
"accuracy": 0.8018,
"f1_score": 0.6457,
"precision_score": 0.5775,
"recall_score": 0.7321,
"roc_auc_score": 0.8606
},
"val_metrics": {
"accuracy": 0.7982,
"f1_score": 0.6667,
"precision_score": 0.6571,
"recall_score": 0.6765,
"roc_auc_score": 0.8647
}

Random forest cropland Map V1

[MsPixels]

After adding corrective labels, here is the metrics @cnakalembe:

"Uganda_North_2022_V2": {
"params": "https://wandb.ai/nasa-harvest/crop-mask/runs/0ngyleht",
"test_metrics": {
"accuracy": 0.6388,
"f1_score": 0.5341,
"precision_score": 0.3917,
"recall_score": 0.8393,
"roc_auc_score": 0.8004
},
"val_metrics": {
"accuracy": 0.6886,
"f1_score": 0.6321,
"precision_score": 0.488,
"recall_score": 0.8971,
"roc_auc_score": 0.8259
}
},

[MsPixels]

Corrective Labeling App

[MsPixels]

Generated a Random Forest approach in GEE to train the machine learning model and compare the results with the LSTM approach. 2975 crop and non crop labels were collected as training data and the CEO labeled dataset was used as the validation set. Code. The F1-score is 0.719

[MsPixels]

The 2975 labels were imported into the LSTM data pipeline as training data and used to retrain the model. The metrics of the model is stated below:

Uganda_North_2022_V3": {
"params": "https://wandb.ai/nasa-harvest/crop-mask/runs/9plwjt8m",
"test_metrics": {
"accuracy": 0.8106,
"f1_score": 0.6446,
"precision_score": 0.6,
"recall_score": 0.6964,
"roc_auc_score": 0.8742
},
"val_metrics": {
"accuracy": 0.8465,
"f1_score": 0.7445,
"precision_score": 0.7391,
"recall_score": 0.75,
"roc_auc_score": 0.9045
}

[MsPixels]

Link @cnakalembe

[MsPixels]

LSTM vrs RF 2022 @cnakalembe