dl-time-series

Deep-learning applied to time series classification of remote sensing data, according to this workflow:

Workflow Execution (Deforestation toy data)

Download the Deforestation toy data in https://storage.googleapis.com/nextgenmap-dataset/dl-time-series/deforestation_toy.zip and follow the instructions below:

1. Stack all the images inside deforestation_toy/images to produce a image time-series, organized per band.

$ ./generate_img_series.py -i deforestation_toy/images -b 1 2 3 4 -o deforestation_toy/image_series

2. Generate the time-series data (e.i. numpy arrays) considering the samples:

$ ./generate_series.py -i deforestation_toy/image_series/ -s deforestation_toy/samples/toy_samples.shp -o deforestation_toy/data_series/ -n 2

3. Train a LSTM model, for 100 epochs, using default hyperparameter (see usages):

$ ./train_model.py -i deforestation_toy/data_series/ -e 100 -o deforestation_toy/model

• Follow the trainning process using tensorboard:

$ tensorboard --logdir=deforestation_toy/model/log

4. Classify the image time-series using the last epoch model:

$ ../classify_img_series.py -i deforestation_toy/image_series/ -m deforestation_toy/model/last_model.H5 -o deforestation_toy/classification/result.img

• Check the classification result, deforestation_toy/classification/result.img, in QGIS:

Usages

generate_img_series.py

usage: generate_img_series.py [-h] -i INPUT_DIR [-b BANDS [BANDS ...]] -o OUTPUT_DIR

STEP 01/04 - Stack all images of input directory, producing one Virtual
Dataset-VRT per band in output directory

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT_DIR, --input-dir INPUT_DIR
                        <Required> Input image directory.
  -b BANDS [BANDS ...], --bands BANDS [BANDS ...]
                        The bands that should be considered. [DEFAULT=All]
  -o OUTPUT_DIR, --output-dir OUTPUT_DIR
                        <Required> Output VRTs directory

generate_series.py

usage: generate_series.py [-h] -i INPUT_DIR -s SAMPLES -n NUM_CLASSES
                          [-c COLUMN_LABEL] -o SERIES_DIR

STEP 02/04 - Generate the time-series considering the vector file informed as
samples.

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT_DIR, --input-dir INPUT_DIR
                        <Required> Input directory that contains the VRTs
                        images.
  -s SAMPLES, --samples SAMPLES
                        <Required> Vector file with the point geometries and
                        class labels.
  -n NUM_CLASSES, --num-classes NUM_CLASSES
                        <Required> Number of possible class labels.
  -c COLUMN_LABEL, --column-label COLUMN_LABEL
                        Name of column that contains the class label values.
                        [DEFAULT=class]
  -o SERIES_DIR, --series-dir SERIES_DIR
                        <Required> The name of output directory

train_model.py

usage: train_model.py [-h] -i SERIES_DIR [-s SEED] [-n] [-v VALIDATION_SPLIT]
                      [-t TEST_SPLIT] [-e EPOCHS] [-b BATCH_SIZE]
                      [-l LEARNING_RATE] -o OUTPUT_DIR

STEP 03/04 - LSTM training approach using several time-series

optional arguments:
  -h, --help            show this help message and exit
  -i SERIES_DIR, --series-dir SERIES_DIR
                        <Required> Input directory that contains the VRT
                        images.
  -s SEED, --seed SEED  Seed that will be used to split the time-series in
                        train, validation, test groups. [DEFAULT=2]
  -n, --only-evaluate   Execute only the evaluation, using the test group.
                        [DEFAULT=False]
  -v VALIDATION_SPLIT, --validation-split VALIDATION_SPLIT
                        Percentage size of the validation group.
                        [DEFAULT=0.15]
  -t TEST_SPLIT, --test-split TEST_SPLIT
                        Percentage size of the test group. [DEFAULT=0.15]
  -e EPOCHS, --epochs EPOCHS
                        Number of epochs of the training process.
                        [DEFAULT=100]
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        Batch size of training process.
  -l LEARNING_RATE, --learning-rate LEARNING_RATE
                        Learning rate of training process. [DEFAULT=0.00005]
  -o OUTPUT_DIR, --output-dir OUTPUT_DIR
                        <Required> The output directory that will have the
                        trained model and the tensorboard logs

classify_img_series.py

usage: classify_img_series.py [-h] -i INPUT_DIR -m MODEL -o OUTPUT

04/04 - Classify image series using a trained model.

optional arguments:
  -h, --help            show this help message and exit
  -i INPUT_DIR, --input-dir INPUT_DIR
                        <Required> Input directory that contains the VRTs
                        images.
  -m MODEL, --model MODEL
                        <Required> The model filepath that should be used in
                        the classification approach.
  -o OUTPUT, --output OUTPUT
                        <Required> The output filepath. The file will be
                        generated in ERDAS_IMG format.