Palm tree detection with CNN

  The university project aimed to develop a convolutional neural network (CNN) capable of detecting palm trees and palm plantations using remotely sensed data. This is because illegal palm plantations and poorly managed plantations in tropical areas pose a serious threat to the environment and local wildlife. By combining remotely sensed data with neural networks, it is possible to effectively detect these hard-to-reach and hard-to-identify areas and determine the type of vegetation planted and quantify them.

  CNN was developed in Cognition Network Language (CNL) using eCognition software. The study area was an area of about 300 hectares on the island of Tongatapu in Polynesia. For this local scale study, we used three bands of very high-resolution aerial imagery. The training samples were based on GIS vector data, which included the marked palm trees. Using this input data, we created 40 000 small subsets of the original image. Each of these subsets represented a single plant for training the model.

  The model was a 2-layer convolutional neural network using a Rectified Linear Units (ReLU) activation function. The developed model successfully identified almost all palm trees in the area (where it was optically possible based on aerial photography). Misclassification was not significant. The model, – used in the current form, – identified was able to identify a plant from the samples multiple times, so there are multiple points around a single palm in the final result map.
This university project did not aim to improve the algorithm and eliminate duplications (as shown in the accompanying figure).