Hedge detection with Airbus Pléiades satellite imagery

The challenge lies in the ability to provide up-to-date data that can accurately reflect changes in hedges across territories. There is a need to refine the mapping of hedges and to monitor their evolution. Since hedges are small elements of the landscape, ranging in width from a few meters to 20 meters, it is necessary to use very high-resolution spatial imagery to detect them with precision. The objective of this work is to assess the contribution of spectral and 3D data from a pair of Pléiades Neo stereo images compared to a pair of Pléiades stereo images for the automatic detection of hedges using artificial intelligence.

Our study area is the protected perimeter of the Kerne Huel pond, which is a bocage area of approximately 18.5 km² located in Brittany, France. The Fédération Départementale des Chasseurs des Côtes-d'Armor has provided us with field surveys and data. We compared the performance of hedge detection using deep learning with Pléiades Neo stereo and Pléiades stereo respectively acquired in October and November 2022. We used the spectral bands of the sensors, pan-sharpened to a spatial resolution of 30 cm for Pléiades Neo and 50 cm for Pléiades. We also used height information from the digital surface model (DSM) calculated using the CARS software (CNES) for each pair of stereo. We trained three convolutional neural networks using the following data: Pléiades with 4 bands (blue, green, red, near-infrared) and DSM, Pléiades Neo with 4 bands and DSM, Pléiades Neo with 6 bands (4 bands + deep blue and red edge) and DSM.

The predictions were validated on areas that were not used for model training. Both sensors performed almost equivalently, Pléiades Neo with 6 bands reaching the highest accuracy score.

When considering equivalent spectral bands, Pléiades Neo allows for the detection of more hedges but with more errors compared to Pléiades. When adding the two additional bands from Pléiades Neo, more hedges are detected than when using only four bands. Overall, it can be assumed that the red edge band of Pléiades Neo helps improve hedge detection.

The advantage of Pléiades Neo over Pléiades seems to lie in its additional spectral bands (particularly the red edge band) rather than its higher spatial resolution for hedge detection. Further investigations should be made on Pléiades Neo images acquired over various landscapes.