TensorFlow implementation of a Multi-Input ConvLSTM, as presented at the Machine Learning Advances Envrionmental Sciences workshop at ICPR 2020. Repo previously titled malawi-flood-prediction
Flooding is among the most destructive natural disasters in the world. As a result of this, there is a great need to develop accurate flood prediction models to prevent flooding ahead of time. However, due to the many factors that are related to predicting flood extent, floods are highly complex to model. Machine learning techniques have shown promise in this space but the current state-of-the-art techniques fail to generalise to other flood events. This study shows that a novel deep learning technique, namely the Multi-Input ConvLSTM, is effective at predicting the level of flood extent and generalising to other types of flood events.
Link to full paper.
Example (see experiment.py)
def run_lstm_exp(model_list, folds=3, seed=SEED):
tf.random.set_seed(seed)
exp_dict = defaultdict(list)
for model in model_list:
if model == lmi:
for fold in range(1, folds+1):
print(f"*Training on fold {fold}")
model.fit(X_train_temp, X_train_const, y_train, X_val_temp, X_val_const, y_val)
preds = model.predict(X_test_temp, X_test_const).reshape(len(X_test,))
preds[preds < 0.0] = 0.0
preds[preds > 1.0] = 1.0
y = y_test.flatten()
res = rmse(preds, y)
exp_dict[model].append(np.round(res, 6))
else:
for fold in range(1, folds+1):
print(f"*Training on fold {fold}")
model.fit(X_train, y_train, X_val, y_val)
preds = model.predict(X_test).reshape(len(X_test,))
y = y_test.flatten()
preds[preds < 0.0] = 0.0
preds[preds > 1.0] = 1.0
res = rmse(preds, y)
exp_dict[model].append(np.round(res, 6))
return exp_dict