Diagnosing interpretability limits of RGB-based explainable deep learning for UAV riparian vegetation reconstruction

Riparian vegetation monitoring using unmanned aerial vehicle (UAV)-based red–green–blue (RGB) imagery has gained traction because it provides centimeter-scale detail at low cost. However, UAV time series are sparse and irregular, making short-horizon vegetation dynamics difficult to model using deep learning methods. We identify compound near-bank–interior attribution, the bimodal concentration of attribution at near-bank and deep-interior zones under peak-vegetation conditions, as the central informational limit of explainable RGB-based UAV deep learning. We evaluate a three-diagnostic interpretability framework for how a convolutional long short-term memory (ConvLSTM) model reconstructs short-horizon structural greenness, represented by a green–red visible-band greenness index (pseudo-SAVI), from UAV RGB imagery. Five UAV RGB acquisitions over a mixed alluvial reach of Yidong Stream (central South Korea) were collected between July and December 2023 at 2 cm ground sampling distance (GSD), aggregated and resampled to a common 256 × 256 analysis grid, yielding five base frames from which training samples were generated through radiometric and geometric augmentation rather than spatial partitioning. Benchmarking against five alternative architectures confirmed competitive spatiotemporal reconstruction. ConvLSTM was retained as the diagnostic substrate for its recurrent spatial-grid preservation rather than as the most accurate predictor. Model behavior was diagnosed through Integrated Gradients (IG), an extreme gradient boosting (XGBoost) surrogate with tree-based SHapley Additive exPlanations (SHAP), and a structured RGB perturbation experiment, complemented by hydrologic sensitivity analysis and Monte Carlo dropout (MCD) uncertainty quantification. ConvLSTM yielded mean squared error (MSE) = 0.0047, structural similarity index measure (SSIM) = 0.90–0.94, temporal consistency index (TCI) = 0.94, and expected calibration error (ECE) = 0.013, with seasonal coefficients of determination (R 2 ) of 0.89, 0.86, and 0.85. Channel-wise sensitivity was 1.6× stronger for the green than red channel, and IG attribution exhibited a bimodal distance-to-edge distribution (top-20% mean edge distance 5.3 versus 4.7 pixels, p < 0.001). Under peak-vegetation conditions, IG attribution concentrated jointly at near-bank (4–7 px) and deep-interior (14–18 px) zones rather than at the immediate channel edge, while channel-wise sensitivity tracked the pseudo-SAVI index formula and hydrologic response surfaces reflected climatic-mean coupling rather than process-level learning. Within this single-reach case study, these patterns are consistent with shortcut-learning concerns identified in natural-image classifiers, with index-formula tracking and climatic-mean coupling as candidate pathways under sparse temporal sampling. Rather than improving prediction, this five-date study quantifies, as a site-level demonstration requiring multi-site replication, when explainable UAV deep learning produces ecologically misleading yet internally consistent information.