Ensemble Deep Learning for the Classification of Strategic Crops Using Sentinel-2 Time Series

Abstract

Reliable and up-to-date information on crop distribution is essential for ensuring food security in the face of rapid population growth and climate change, and with recent advancements in remote sensing and artificial intelligence, large-scale analyses of agricultural landscapes can now be performed more quickly and accurately. The main objective of this study is to improve the classification of strategic crops (wheat, barley, and corn) through an ensemble learning approach that integrates three complementary neural network architectures—TempCNN, BiLSTM, and Transformer-Encoder—using Sentinel-2 time series data. To achieve this objective, several vegetation indices were first computed from the spectral bands of Sentinel-2 time series data. A Random Forest (RF) model was then employed to identify and select the most relevant features. Finally, these features were used to train the individual models, and their prediction probabilities were combined using soft voting to produce the final ensemble predictions. Model performance was evaluated across four spatial cross-validation configurations, in which the data from each of the four departments of Brittany was held out in turn as an independent test region, ensuring a rigorous assessment of the model’s spatial generalization capability. The ensemble model achieves a mean Overall Accuracy (OA) of 98.14% across all configurations, with per-class F1-scores exceeding 0.93 for barley, 0.97 for wheat, and 0.99 for corn. Compared to TempCNN and LSTM baseline models, as well as Recurrent Convolutional Neural Network (R-CNN), RF, and Support Vector Machine (SVM), the proposed ensemble model consistently achieves higher OA and per-class metrics. It also outperforms the same ensemble trained without vegetation indices. These findings highlight the effectiveness of combining complementary deep learning architectures and demonstrate the added value of vegetation indices for crop type classification using satellite time series.