The global attention to the utilization of unmanned aerial vehicle remote sensing drones in crop disease-wide detection has led to the urgent need to find an adapted model for different environmental conditions. Therefore, the current study has focused on spatiotemporal usage of different multispectral cameras in acquiring spectral reflectance models of in-field rice bacterial blight stresses. Where, long short-term memory (LSTM) model was compared with the other models in transfer learning strategy for assessing the blight stress severity. The results revealed that by extracting 30% of the data from the target domain and transferring it to the source domain, the adaptability of the model across different sites was effectively enhanced. Besides, LSTM showed high tuning transfer efficiency that demonstrated optimal predictive performance and the shortest training time in transfer tasks. Its coefficient of the prediction set was 0.82, and its residual prediction deviation has reached 2.26. In practice, LSTM enabled the acquisition of reliable prediction results at a minimal sample collection cost while circumventing feature reduction resulting from inter-domain data alignment. When the transfer ratio reached 20%, the coefficient of determination of the prediction set reached 0.71, and the residual prediction deviation reached 1.79. The novelty of this study came from the transfer learning efficiency in improving the model’s application capabilities across the different sites, environment, and unmanned aerial vehicle in farmland disease detection.

全球对无人机遥感无人机在作物病害检测中的应用的关注导致迫切需要找到适合不同环境条件的模型。因此，目前的研究集中在不同多光谱相机在获取田间水稻细菌枯萎病胁迫光谱反射率模型中的时空使用。其中，长短期记忆 （LSTM） 模型与迁移学习策略中的其他模型进行比较，以评估枯萎病压力的严重程度。结果表明，通过从目标域提取 30% 的数据并将其传输到源域，模型在不同站点的适应性得到了有效增强。此外，LSTM 表现出高调优传输效率，在传输任务中表现出最佳的预测性能和最短的训练时间。其预测集的系数为 0.82，其残差预测偏差已达到 2.26。在实践中，LSTM 能够以最小的样本采集成本获得可靠的预测结果，同时避免了域间数据对齐导致的特征减少。当迁移率达到 20% 时，预测集的决定系数达到 0.71，残差预测偏差达到 1.79。这项研究的新颖性来自于迁移学习效率，提高了模型在不同地点、环境和无人机在农田病害检测中的应用能力。