Abstract
Earth observation data processing requires interpretable deep learning (DL) models that learn physically significant and meaningful features. The current study proposes approaches to make the network to learn meaningful features. In addition, a set of interpretability- and explanation-based evaluation strategies are proposed to evaluate the DL models. Adversarial variational encoding along with constraints to regulate latent representations and embed label information are employed to learn interpretable manifold. The proposed architecture, called interpretable adversarial encoding network (IAENet), significantly improves the results compared to other main existing DL models. The proposed IAENet learns the features which are essential in distinguishing the different classes thereby improving the interpretability of the model. The explanations for the different models are generated through analysis of the concepts learned by each model using activation maximization. Besides, the relevance assigned by the model to input features is also estimated using the layer-wise relevance propagation approach. Experiments on the phenological curve-based crop classification illustrate that IAENet learn relevant features (giving importance to the non-rainy season) to distinguish different irrigation schemes. The performance can be attributed to the learned interpretable manifold, and the refinement of architectural units and convolutions considering the point-nature and irregular sampling of the input data. Experiments on learning crop-specific features from multispectral images for crop-type classification indicate that IAENet learns red and green edge features crucial in distinguishing the studied crops. The improvement in interpretability of the DL models is found to reduce the sensitivity toward network parameters. The proposed evaluation measures facilitate ascertaining the physical significance of the learned manifold.
Original language | English |
---|---|
Pages (from-to) | 10929-10948 |
Number of pages | 20 |
Journal | Neural Computing and Applications |
Volume | 34 |
Issue number | 13 |
DOIs | |
State | Published - 1 Jul 2022 |
Keywords
- Classification
- Crop-specific features
- Deep learning
- Interpretability
- Phenological curves
- VENµS
ASJC Scopus subject areas
- Software
- Artificial Intelligence