Understanding deep learning in land use classification based on Sentinel-2 time series.

Area of Science:

• Remote Sensing
• Artificial Intelligence
• Environmental Science

Background:

• Deep learning (DL) is increasingly used for remote sensing (RS) data analysis, showing promise in tasks like image classification.
• A significant gap exists in understanding the interpretability of DL model predictions, hindering wider adoption, especially in policy-driven applications.
• Accountability and transparency are crucial for DL models used in managing public funds and ensuring policy compliance.

Purpose of the Study:

• To enhance the interpretability of a recurrent neural network (RNN) used for land use classification.
• To analyze Sentinel-2 time series data within the European Common Agricultural Policy (CAP) framework.
• To identify the most relevant predictors influencing the RNN's classification decisions.

Main Methods:

• Utilized a recurrent neural network (RNN) for land use classification.
• Employed Sentinel-2 satellite time series data.
• Performed an analysis to determine the relevance of different spectral bands and temporal features.

Main Results:

• The study identified that red and near-infrared Sentinel-2 bands are the most informative spectral predictors.
• Temporal analysis revealed that features derived from summer satellite acquisitions hold the most influence.
• The interpretability analysis provided insights into the RNN's decision-making process for land use classification.

Conclusions:

• Understanding DL model interpretability is key for their application in policy-making, such as the European Common Agricultural Policy (CAP).
• The findings support the use of specific Sentinel-2 bands and temporal features for effective land use classification.
• This research contributes to the accountable use of DL in achieving European Green Deal objectives, including climate action and biodiversity protection.