It says you should use 2015 data (target_2015 and its features) to train your model to predict flood extent (a.k.a, target_2015). Make your sure your model is good. Then use the model and the 2019 features to predict target_2019 (a.ka. flood extent in 2019 which we do not know). After you submit your work, Zindi will use your predicted target_2019 and the actual target_2019 (not the one in samplesubmission.csv) to measure RMSE.
I think square = sqaure = rectangle in this project.