Assess the value of bidirectional LSTM context under small-sample constraints

Determine whether the bidirectional context benefit of a bidirectional LSTM outweighs its approximately doubled recurrent parameter count under the same small-sample overfitting risk identified for the unidirectional LSTM baseline in the Sentinel-2 wheat classification framework.

Background

The operational classifier was evaluated primarily against unidirectional sequence-model and XGBoost baselines using a curated dataset of 849 samples. The paper identifies small-sample overfitting as a constraint on higher-capacity recurrent architectures and proposes a bidirectional LSTM as a possible extension because bidirectional temporal context may improve crop-classification performance.

The authors explicitly leave unresolved whether the potential benefit of bidirectional temporal context would justify the approximately doubled recurrent parameter count in the same small-sample setting. The problem is therefore an empirical comparison requiring evaluation under the framework’s sample-size and validation constraints.

References

First, a bidirectional LSTM, which has demonstrated accuracies exceeding 94\% in smallholder crop classification by capturing phenological context from both temporal directions \citep{khan2024bilstm}, is a candidate replacement for the unidirectional LSTM baseline evaluated in Section~\ref{sec:baselinecomp} -- though whether its bidirectional context benefit outweighs its approximately doubled recurrent parameter count, under the same small-sample overfitting risk this study identifies for the unidirectional case, is an open empirical question rather than an assumed improvement.

— A Dual-Track Curation-and-Classification Framework for Resolving Ground-Truth Label Noise in Operational Sentinel-2 Wheat Area Estimation  (2609.36975 - Agharia et al., 29 Sep 2026) in Section 6, Limitations and Future Work, subsection “External validity”