DeepRootz: Classifying satellite imagesof the Amazon rainforest

Scott Longwell, Tyler Shimko, and Alex Williams

Dataset analysis

Each chip is labelled with exactly one atmospheric condition labeland one or more ground condition labels. An atmospheric label of"cloudy" negates any possible ground labels, as the ground cannotbe observed from the satellite imagery.

Certain ground labels are highly correlated. For instance, if an areahas agricultural activity, it is also likely to have roadways and othersigns of human settlement.

Model Design Results

40,479 ~1km x 1km "chips"Each chip 256px x 256px4 channelsRedGreenBlueNear-IR

Best mean F2 = 0.88Current leader achievestest set F2 = 0.93Best decision thresholdis rarely near 0.5Mid-frequency labelsare hardest to recall

Try larger models with further regularizationTry residual network models to improve training speed/reliabilityContinue adjusting decision thresholds to improve training accuracy

Fig. 1 - Example "chips" from the KagglePlanet challenge dataset.

Fig. 2 - A) Probabilities of observing any groundlabel conditioned on observing a given atmosphericlabel. B) Probabilities of observing any atmosphericlabel conditioned on observing a given ground label.

Fig. 4 - Current model architectureFig. 3 - A) Total number of images labelled with a given label. B) Co-occurrence matrix for allground labels.

Fig. 8 - A) Total possible improvement if every instance of any particular label was correctlyrecalled. B) Possible improvement as a function of the frequency of any given label.

A

A

B

B

conv2d - 3x3, 32 filters, stride 1, pad 1

input images - 256x256px, 4 channels

= ReLU= batchorm= dropout= sigmoid

conv2d - 3x3, 32 filters, stride 1, pad 1

conv2d - 3x3, 32 filters, stride 1, pad 1

conv2d - 3x3, 32 filters, stride 1, pad 1

conv2d - 3x3, 32 filters, stride 1, pad 1

flatten

output labels

linear - 524288 input, 40 output

linear - 40 input, 17 output

conv2d - 2x2, 32 filters, stride 2

Raw data

Atmospheric/ground label co-occurrence

Ground label co-occurrence

F2-measure

Final entries are judged on meanexample-wise F2-measure uponsubmission.

Binary cross entropy loss was usedfor each possible label. Initialthresholds for label inclusion wereset at 0.5, but were adjustedas training progressed to provide aboost in model performance.

Batch normalization was employedfollowing every layer in the model,with the exception of the final layer.Batchnorm has been shown to bothease the training process as well asserve as a type of regularization.

Dropout with a probability of 0.2was employed on the second to lastset of activations for regularization.

With the exception of the final layerof the model, ReLU units are usedto add non-linearities to the model.This decision was based primarilyon the performance superiority ofReLU units in similar imagelabelling tasks.

Our model is composed of sixconvolutional layers, including onelayer with filter size equal to stride toeffectively downsample the image.Stacked layers with smaller filtersizes serve to increase effectivereceptive field size without the sameparameter scaling as larger sizedfilters.

Loss function

ReLU non-linearities

Batch normalization

Dropout

Convolutional layers

Future Directions

Fig. 7 - F2 score as a function of the decision boundary forevery possible label. In most cases, F2 score can beimproved by adjusting the threshold from the default of 0.5.

A B

Fig. 5 - Training and validation F2 scoresper epoch.

Fig. 6 - Conditional co-occurrence matrix for groundlabels applied by our model.