In the context of the Image Processing for Earth Observation EPFL course (ENV-540)

In the context of climate crisis we are facing, deforestation is an ever increasing problem, destroying precious carbon dioxide intake. In Ivory coast, the deforestation has reached an impressive magnitude mainly for cocoa plantations. As a results only few places of wild forest remains among which the Taï national park, situated South-West of the country close to the border with Liberia. This 3300 km² protected area is a refuge for a multitude of species. Preserving it is thus an important challenge. The present project propose a machine learning approach based on remote sensing images. The goal is to build a classifier able to decide whether a pixel from Sentinel-2 image belongs to a coca plantations.

The classifier is developed on the free Sentinel-2 L1C images dated of the 29 of December 2018 (a day without clouds over the whole park). The Sentinel-2 images have 13 bands (4 at 10 meter resolution, 6 at 20 meters, and 3 at 60 meters resolution) that goes from the visible spectrum to the infrared.

The model is trained on a small area in which there are known plantations (cf Figure Overview). The training area have been completely labeled for the binary problem : plantation vs no plantation using the high resolution images from Google Earth. Other testing areas have been labeled for some known plantation. Those labeled area will be used to test if the classifier generalize over space.

The model development unfolds on four steps summarized on figure [Processing Pipeline](#processing pipeline)

1. Feature Engineering make use of the 10m and 20m bands to build 14 new bands. The classical and green NDVI. Then each of the R, IR and NDVI are transformed using Opening and Closing by reconstruction, Local Binary Pattern and the Local Entropy.

2. Model Performances estimation The 24 bands are reshaped in a 2D array where each sample is a pixel in represented by 24 features. The trained model is composed of a Normalized followed by a PCA and an estimator. Four estimator are tried : KNN, SVM, Random Forest and a MLP. The performances are estimated through a 5-folds Cross-validation (CV). Note that the normalized, PCA, and estimator are fitted after the training set is defined (through the sklearn.pipeline.Pipeline). The CV report the training and testing performance through the accuracy, the f1-score, the recall and the precision.

3. Model Fitting The same architecture are trained on the whole data available (all the pixels in the training area).

4. Prediction on Testing area The four models are used to predict the unseen nearby and distant plantations to assess their space generalization capabilities. Because the whole image have not been labeled, the quality of the predictions is assessed through the detection rate which represent the fraction of the labeled polygon that have been detected. Therefore a model predicting plantations everywhere would have a 100% detection rate. To control for that pitfall, the image are also qualitatively observed.

The figure [Feature engineering](#Feature engineering) presents the additional features obtained from the Red, Near-Infrared and NDVI bands. The NDVI is an index reflecting the of vegetation and can thus help to discriminate plantations from native forest. The goal of the entropy, the LBP, the opening and closing by reconstruction is to add informations about local structures in the pixel as the arrangement of pixel together carries a lot of information.

In the table below is presented the performances of the four models obtained through a 5-folds cross-validation. The accuracy, f1-score, precision and recall are reported for both the train and test fold(s) as the mean and standard deviation over the different folds arrangements.
The best performing model on the training area appears to be the MLP. The RF seems to have overfitted as the train and test scores strongly differ. The rather low detection rate (recall) may be due to a sparse spatial prediction : the overall patch is detected but with a lot non-detected pixel. The morphological post processing aims at correcting this behavior and increase the spatial relevance of the predictions.

The figure [Morphological Post-Processing](#Morphological Post-Processing) presents the effect of the post treatment applied to the predictions reshaped as images. The successive closing and opening removes the salt and pepper noise to yield satisfying prediction patches.

The model is trained only on a single spot around the Taï national park. Therefore, the models built may be too specific to the training area. In order to test how the models generalize over space, they are used to predict two area where some known plantations have been labeled :

• Nearby predictions represent plantations that are close to the training area (blue polygons).
• Distant predictions represent plantations away from the training area, around 60 km north close to a lake.

The results of the testing over space are presented on the figure [Testing predictions](#Testing predictions). All the four models seems to generalize locally as the blue patches are mostly filled. However, over a longer distance, KNN seems to be the best as it detect 20% of the labeled plantations. The location may be hard to predict with the presence of the water body that may induce atmospheric effect different than the training area.

KNN is used to make some prediction on various part of the park to control whether cocoa plantations are detected within the park. The interpretation must be made carefully as the model does not fully generalize over space as seen before. Therefore the prediction closer to the training area are consider more trustable. The predictions are presented on figure predictions.