Unmanned Aerial Vehicles (UAVs) equipped with RGB cameras are increasingly used as low-cost tools for crop monitoring, offering a range of vegetation indexes in the visible spectral range. These indexes have often been reported to correlate with other multispectral indexes such as the Normalized Difference Vegetation Index (NDVI) during active growth stages. However, still efforts should be done about their performance under conditions of canopy degradation. In this study, UAV flights were conducted over a cereal field immediately after harvest, when the canopy consisted mostly of bare soil and dry residues. RGB-based indexes were calculated from the orthomosaic, normalized to a [0–1] scale, and compared to NDVI derived from a multispectral sensor. Data preprocessing included ground control point (GCP) georeferencing, removal of NoData pixels, and raster alignment. Results revealed very weak correlations between RGB indexes and NDVI (Pearson r < 0.15), with Visible Atmospherically Resistant Index (VARI) showing almost no variability across the field. Although the Leaf Index (GLI), yielded the lowest error values, all RGB indexes failed to reproduce the variability of NDVI under post-harvest conditions. These findings highlight a critical methodological limitation: RGB indexes are unsuitable for vegetation monitoring when canopy cover is severely reduced. While they remain useful during active growth, their reliability diminishes in degraded or post-harvest scenarios, thereby limiting their application in assessing abiotic stress in cereals.
Soft robotics has emerged as a promising direction for enabling safe, adaptive, and energy-efficient interactions with unstructured environments due to its inherent compliance. Recently, Deep Reinforcement Learning (DRL) has become a powerful tool for autonomous behavior generation in soft robots, surpassing limitations of classical model-based control. However, despite rapid growth of publications in this domain, there is still a lack of systematic comparative surveys that clarify how different DRL approaches have been used for soft mobile robots, what types of tasks they address, and what performance evaluation criteria have been used. In this article, we review and classify existing works in DRL-enabled soft robotics, focusing particularly on soft mobile systems, and present a structured synthesis of contributions, algorithms, training strategies, and real-world applications. Unlike previous reviews that discuss soft robotics or DRL separately, this paper explicitly provides cross-comparison across DRL paradigms and soft robot tasks, enabling researchers to identify suitable DRL approaches for different soft mobile robotic behaviors. Finally, major challenges and promising future directions are proposed to advance this interdisciplinary research area.
