Exploring Self-Attention for Crop-type Classification Explainability

• DOI: 10.48550/arxiv.2210.13167
• Publication Date: 2022-01-01
• Authors: Ivica Obadić; Ribana Roscher; Dário Augusto Bor...; Xiao Xiang Zhu
• Keywords: 2. Zero hunger; FOS: Computer and information sciences; Computer Vision and Pattern Recognition (cs.CV); Computer Science - Computer Vision and Pattern Recognition; 0202 electrical engineering, electronic engineering, information engineering; 02 engineering and technology; 15. Life on land

Automated crop-type classification using Sentinel-2 satellite time series is essential to support agriculture monitoring. Recently, deep learning models based on transformer encoders became a promising approach for classification. Using explainable machine reveal the inner workings of these an important step towards improving stakeholders' trust and efficient In this paper, we introduce novel explainability framework that aims shed light crop disambiguation patterns learned by state-of-the-art encoder model. More specifically, process attention weights trained critical dates use domain knowledge uncover phenological events model performance. We also present sensitivity analysis understand better capability revealing crop-specific events. report compelling results showing strongly relate key dates, consequently, These findings might be relevant stakeholder optimizing monitoring processes. Additionally, our demonstrates limitation identifying in phenology as empirically show unveiled depend other crops data considered during training.