XAI for Early Crop Classification

TL;DR

This work tackles early crop classification by identifying the most informative timesteps through Layer-wise Relevance Propagation (LRP) applied to an attention-based Transformer classifier. By computing timestep relevances $R_t$ and selecting the top $n$ timesteps, the authors derive shortest effective timeframes $\Delta t_n$ and demonstrate that a window of $\Delta t_3$ (2019-04-21 to 2019-08-09) achieves only a $0.75\%$ accuracy loss relative to the full timeseries. The approach yields crop-specific dominant peaks in $R_t$ that align with growth milestones and spectral bands (e.g., $B6$, $B8$, $B8A$), enabling accurate early classification with substantially reduced data requirements. These findings suggest practical pathways to near-real-time crop mapping and motivate cross-year transfer to account for interannual variability.

Abstract

We propose an approach for early crop classification through identifying important timesteps with eXplainable AI (XAI) methods. Our approach consists of training a baseline crop classification model to carry out layer-wise relevance propagation (LRP) so that the salient time step can be identified. We chose a selected number of such important time indices to create the bounding region of the shortest possible classification timeframe. We identified the period 21st April 2019 to 9th August 2019 as having the best trade-off in terms of accuracy and earliness. This timeframe only suffers a 0.75% loss in accuracy as compared to using the full timeseries. We observed that the LRP-derived important timesteps also highlight small details in input values that differentiates between different classes and

Figures (2)

• Figure 1: Input spectral reflectances and corresponding $R_t$ for three summer crops. The solid black line represents the median $R_t$ for the crop type while the shaded blue region represents the 25 to 75 percentile. For each timeseries only a few dominant peaks that have significantly larger $R_t$ exist. These peaks tend to be crop-specific and represents the most important timesteps for the classification of that crop. For , this is in late June and early August, Cucurbits have them in mid July to early August and have them in July.
• Figure 2: A sample parcel with the input values at the top, the relevance scores at each timestep $R_t$ in the middle and the corresponding band-specific relevance scores at each timestep $R_{b,t}$ in the bottom. The four dominant peaks indicate the most important timesteps are from mid to late June and early August. The corresponding important bands are B6, B8 and B8A