agtuary / machine-learning-challenge

Agtuary Machine Learning Technical Challenge

The purpose of this challenge is to test your ability in working with satellite imagery, imbalanced datasets and machine learning models.

Problem

You must implement a machine learning model to perform crop classification, and determine what crop types are grown where in an image of farm paddocks (also known as one-shot in-season crop classification).

Sentinel-2 is a satellite that captures 12 different wavelengths of light (also known as bands) in an image. These range from visible light (red, green, blue) to infra-red. The values for each band changes depending upon the material/object that is on the ground.

The Task

Note: Use the Agtuary ML Challenge template notebook provided in this repo to start.

1. Download the dataset at: https://agtuary-data-public.s3.ap-southeast-2.amazonaws.com/machine-learning-challenge/agtuary-ml.tar.gz

2. Prepare the data for training and testing. The pixels.csv file has 14 columns, and each row is a different pixel from satellite imagery. The first 12 columns are the different bands B01..B12, then the cloud_prob column which is the probability that the pixel of satellite imagery is cloud (ranges from 0. to 100.), and finally the label column which is the ground-truth crop type.

   The band values go from 0. to 1. Remove any pixels that have a cloud probability of over 2. Split the dataset into a train and test set as you feel is appropriate. Encode the crop type labels into a numeric value. Drop B04 from the dataset and it must not be used in any calculations. You may use Satellite indices to augment the data. Any form of augmentation/calculated feature generation can be used.

3. Create a plot showing the mean band values of each crop type, as well as +- 2 standard deviations.

4. Justify the usage of features that went into training the model backed by statistical evidence.

5. Create/use a machine learning model of your choice to perform multiclass classification, and train it on the any features/bands/indices deemed necessary.

6. Plot a confusion matrix and print a classification report of your model. Plot/print any other interesting metrics you may have (optional).

7. Inside the dataset, there are also satellite imagery band files from a test region in Australia (B01.tif to B12.tif). These image files are in uint16 data type and have values that go from 0 to 10000, which must be rescaled to fit the training data range. There is also a cropland mask file (mask.png), which is int8 data type and the values go from 0 to 255. Use your model to perform inference over this test region, where a value greater than 0 in the mask represents the only pixels where inference should be performed.

8. From your inference results, create an image which contains purple pixels where the result was Other, red pixels where the result was Sorghum and white pixels where the result was Cotton. Save it as a PNG file.

Restrictions

• The usage of B04 is not allowed.

Evaluation

Explain your approach inside your code (write it as a comment block or notebook cell). Some of the following questions are good places to start:

How would you improve the model?

• Are there any specific areas in the satellite imagery where you see the model not performing as well as other areas?
• Are there farm paddocks where there is a lot of noise in the output (e.g. lots of mixed crop type pixels)?
• Are there any other features you could use as inputs?
• Could you use the different band values to engineer new features/inputs (such as the difference between particular bands)?
• What machine learning model or type would best suit this problem?
• How would you ensure the training data is high quality and only includes pixels over valid crop land?

Additionally

• Did you change the model hyperparameters? If so, which ones and why?
• What are the downsides to your approach?
• What other methods could you choose? Why would that be the next best choice?
• Put in in references to scientific papers, websites or blogs that you used to inform your choices.

Submission

Save your notebook/code as well as output images into a .zip file and send it to [e-mail]