Nitrogen is a critical nutrient for tobacco growth, substantially impacting leaf development and photosynthesis. Excessive nitrogen negatively affects tobacco quality, while insufficient nitrogen lowers yield per unit area. Traditional fertilization methods often rely on farmers’ experience, making it difficult to precisely meet the specific nitrogen requirements of tobacco. This study proposes a recommended nitrogen application index (RNAI), derived from multispectral imagery obtained by unmanned aerial vehicles (UAVs), to enable precise nitrogen application and improve tobacco yield. Key agronomic traits—including leaf area index (LAI), leaf biomass (LB), chlorophyll content (LCC), and leaf nitrogen content (LNC)—were used to construct RNAI through the entropy weight method (EWM). The Pearson correlation coefficient and recursive feature elimination (RFE) were employed to identify vegetation indices (VIs) from UAV imagery that significantly correlated with each agronomic trait. Estimation models for LAI, LB, LCC, and LNC were optimized using random forest (RF) and extreme gradient boosting (XGBoost) algorithms, with XGBoost consistently outperforming RF for all indicators. Based on RNAI, a nitrogen application decision-making scheme for tobacco growth was developed. The RNA decision-making model, informed by RNAI fitting curves, guided nitrogen application rates and produced a decision-making map for nitrogen management in tobacco fields. Applying this map reduced pure nitrogen usage by 3.71 kg/ha, increased yield by 54.56 kg/ha, and enhanced economic benefits by 2.58% while improving the uniformity of the tobacco population. The RNA decision-making model established in this study provides a robust scientific framework for precise nitrogen management in tobacco cultivation.

Post-curing nitrogen content (cNg) and nicotine content (cNt) constitute fundamental chemical attributes that govern the quality of flue-cured tobacco leaves. In this study, we investigate the use of unmanned aerial vehicle (UAV)-borne hyperspectral imagery in conjunction with advanced sequence-learning models to obtain canopy-level spectral information and to quantitatively predict cNg and cNt at the field scale. Nevertheless, because field experiments conducted in different growing seasons follow heterogeneous observation schedules, the resulting time series differ in both length and temporal sampling density, which poses a substantial obstacle to the direct deployment of recurrent neural-network models. To mitigate this limitation, we devised three autoencoder-based temporal-alignment schemes to reconcile the temporal dimension of multi-year spectral–phenotypic sequences, namely a fully connected autoencoder (AEF), a one-dimensional convolutional autoencoder (AEC), and a long short-term memory (LSTM) autoencoder (AEL). Within this framework, the autoencoders project the original multi-temporal observations into compact latent representations and subsequently reconstruct them on a standardized temporal grid, thereby preserving salient spectral–phenotypic information while enforcing dimensional consistency across years. For comparative analysis, we additionally considered a straightforward time-step removal (TSR) procedure, which discards non-overlapping observation dates among different years and thus serves as a baseline temporal-harmonization strategy. On the temporally aligned sequences, two representative recurrent-neural architectures—long short-term memory (LSTM) networks and gated recurrent unit (GRU) networks—were trained to establish predictive models for cNg and cNt. Overall, models trained on inputs aligned by the autoencoder-based schemes exhibited markedly higher predictive skill than their TSR-based counterparts, with the convolutional and LSTM autoencoders consistently yielding greater improvements than the fully connected variant. Among all evaluated configurations, the combinations of AEL with LSTM and with GRU delivered the best performance, attaining coefficients of determination of 0.73 for cNg and 0.56 for cNt on the independent test set. Taken together, these results indicate that the proposed autoencoder-based temporal-alignment framework can effectively distil informative features from multi-year UAV hyperspectral and phenotypic observations and constitutes a promising tool for the quantitative prediction of post-curing quality indicators in flue-cured tobacco production.

In China, tobacco production must strictly follow the yield plan set by the higher authorities. In this context, accurate and stable yield estimation is meaningful for effective production management. In this paper, we adopted a multi-source data fusion strategy to develop the yield estimation models for tobacco. The data used include unmanned aerial vehicle (UAV)-borne hyperspectral features (HF), biophysical parameters (BPP) collected in the field, and biochemical parameters (BCP) measured in the laboratory. Since the crop state at different growth stages both affect the final yield, we employed two typical recurrent neural network (RNN) algorithms, long short-term memory (LSTM) and gated recurrent unit (GRU), for modeling. The random forest (RF) algorithm was selected as the baseline scheme. In addition, we designed a one-dimensional convolutional autoencoder (AEC1D) to unify the input dimensions of raw data from different years. It was found that yield estimation performance from multi-source data was more accurate than using any single feature. The GRU model with the HF+BCP+BPP combination achieved the highest estimation accuracy, with an Rv2 of 0.705. The overall performance of LSTM and GRU models was also better than that of RF. We also quantified the contribution of each feature to the model, with HF, BPP, and BCP accounting for approximately 45%, 32%, and 23%, respectively. This study demonstrated the benefits of multi-source data fusion and RNN algorithms in estimating tobacco yields, which can be used to assist in site-specific crop management.

This study pioneers a rapid, non-destructive methodology for detecting crab roe in live Chinese mitten crabs using near-infrared transmittance spectroscopy (NIR-TS), motivated by the high edible and economic value of this component. A customized optical system was used to measure NIR transmission spectra from 62 live crabs, including intact specimens, dissected tissues (roe, shell, and roe-removed), and nine specific body regions, while simultaneously measuring crab central thickness and roe distribution. A classification model was developed based on regional spectra to distinguish roe-containing and roe-free regions, enabling prediction of crab roe distribution within live specimens. A dataset of 558 specta was split into training, validation, and test sets at a ratio of 6:2:2. The results indicated measurable non-zero transmittance in both intact crabs and isolated shells, suggesting the potential ability of short-wave NIR (SW-NIR, 780-1100 nm) to penetrate multi-layered tissues. Regional spectral analysis revealed that crab roe was the primary transmission modulator, surpassing thickness effects, with roe-containing regions consistently exhibiting lower transmittance than roe-free. Modeling result showed that after multiplicative scatter correction (MSC) preprocessed for spetra, Backpropagation Neural Network (BPNN) model combining both LASSO-selected feature and 30 spectral statistic features (SFs30) achieved the best performance with 96.2 % accuracy. Furthermore, roe distribution prediction showed 94.4 % concordance with ground-truth validation. The model's classification capability was attributed to specific NIR wavelengths associated with key biochemical signatures, forming a distinctive fingerprint for crab roe detection. This methodology provides a novel solution for quality grading of crabs, addressing environmentally friendly, and cost-effective requirements for aquatic supply chain.

Dissolved oxygen (DO) is an important indicator of the water health of the freshwater aquaculture pond. However, since DO is a non-photosensitive parameter, it is difficult to directly inverse using UAV imaging technology. We proposed an estimation method of DO based on UAV multispectral data and machine learning algorithms. The method utilizes chlorophyll-a (Chl-a) and spectral indices as input features to accurately estimate DO content in water bodies. UAV images were collected in six periods at two aquaculture ponds. Machine learning algorithms were applied to map Chl-a concentration in each aquaculture pond, and a DO estimation model was developed through the relationship between Chl-a, spectral index and DO. The model was validated using measured samples, and the spatial and temporal variations in DO at the two freshwater aquaculture ponds were analyzed. The findings demonstrated that the model exhibited suboptimal performance when solely utilising spectral index. However, the incorporation of Chl-a as an input feature resulted in a substantial enhancement in model performance, in comparison to the utilisation of only spectral index. The RF model performed well during both training and testing phases at the first freshwater aquaculture pond, achieving R2 = 0.87, RMSE = 1.785 mg/L, and MAE = 1.512 mg/L for the testing set. Concurrently, the validation in the other two periods(GC - August and October 2023 and PK-April and May 2024) further confirmed the model's generalization ability, with R2 = 0.84, RMSE = 2.245 mg/L, and MAE = 1.251 mg/L. Similarly, the model showed robust performance at the second freshwater aquaculture pond, achieving R2 = 0.85, RMSE = 3.743 mg/L, and MAE = 2.730 mg/L. UAV multispectral imaging technology combined with this method can efficiently and accurately capture the spatial and temporal distribution of DO in freshwater aquaculture pond, supporting aquaculture management.

One of the most important nutrients needed for fruit tree growth is nitrogen. For orchards to get targeted, well-informed nitrogen fertilizer, accurate, large-scale, real-time monitoring, and assessment of nitrogen nutrition is essential. This study examines the Leaf Nitrogen Content (LNC) of the custard apple tree, a noteworthy fruit tree that is extensively grown in China’s Yunnan Province. This study uses an ensemble learning technique based on multiple machine learning algorithms to effectively and precisely monitor the leaf nitrogen content in the tree canopy using multispectral canopy footage of custard apple trees taken via Unmanned Aerial Vehicle (UAV) across different growth phases. First, canopy shadows and background noise from the soil are removed from the UAV imagery by using spectral shadow indices across growth phases. The noise-filtered imagery is then used to extract a number of vegetation indices (VIs) and textural features (TFs). Correlation analysis is then used to determine which features are most pertinent for LNC estimation. A two-layer ensemble model is built to quantitatively estimate leaf nitrogen using the stacking ensemble learning (Stacking) principles. Random Forest (RF), Adaptive Boosting (ADA), Gradient Boosting Decision Trees (GBDT), Linear Regression (LR), and Extremely Randomized Trees (ERT) are among the basis estimators that are integrated in the first layer. By detecting and eliminating redundancy among base estimators, the Least Absolute Shrinkage and Selection Operator regression (Lasso)model used in the second layer improves nitrogen estimation. According to the analysis results, Lasso successfully finds redundant base estimators in the suggested ensemble learning approach, which yields the maximum estimation accuracy for the nitrogen content of custard apple trees’ leaves. With a root mean square error (RMSE) of 0.059 and a mean absolute error (MAE) of 0.193, the coefficient of determination (R2) came to 0. 661. The significant potential of UAV-based ensemble learning techniques for tracking nitrogen nutrition in custard apple leaves is highlighted by this work. Additionally, the approaches investigated might offer insightful information and a point of reference for UAV remote sensing applications in nitrogen nutrition monitoring for other crops.

Dissolved organic matter (DOM) is a pivotal indicator for assessing aquatic health and ecological functions. Monitoring DOM in aquaculture ponds using satellite requires validation through field measured samples. However, due to the inherent spatial variability of DOM in aquaculture ponds, individual samples are insufficient to represent the entire pond. Consequently, directly applying field measurements to satellite remote sensing can compromise the accuracy of estimation models. A spatial mapping approach was proposed in the study, which integrated UAV multispectral data with Sentinel-2 images to address scale mismatches between satellite images and ground-based measurements. Then a self-optimizing model was used to estimate and map DOM concentration at county scale. Firstly, high-resolution spatial distribution of DOM in some aquaculture ponds were obtained through field samples and UAV multispectral images. Secondly, a spatial mapping relationship was established between the UAV-derived DOM distribution and the corresponding satellite image pixels, thereby providing high-quality samples for large-scale monitoring of DOM in aquaculture. Results showed that: (1) Among the four models constructed using UAV data, the simulated annealing-optimized random forest (SA-RF) achieved the highest performance, with the R2 of 0.84, RMSE of 2.66mg/L, and MAE of 2.21mg/L. (2) The spatial mapping method improved the accuracy of DOM concentration estimation based on satellite images. Specifically, the accuracy of SA-RF model increased by 10% compared with the model constructed directly using satellites and ground measurements, achieving an R2 of 0.78. This study demonstrates that the spatial mapping method provides a novel method for UAVsatellite collaborative inversion of DOM concentration in aquaculture ponds.

UAV imaging technology has become one of the means to quickly monitor water quality parameters in freshwater aquaculture ponds. The change of sunlight during a long flight affects the quality of UAV images, which will reduce the accuracy of monitoring water quality. This study aims to propose a method to correct spectral variation during UAV imaging and apply it to detect dissolved organic matter (DOM) concentration and dissolved oxygen (DO) content in freshwater aquaculture ponds. Firstly, a spectral correction method was used to transform UAV-based multispectral images. The spectral data before and after correction was extracted. Secondly, 18 spectral indices before and after correction were constructed. The optimal combination of indices was identified using correlation analysis algorithm. The estimation models of water quality parameters were then constructed and compared using the Random Forest (RF), Support Vector Regression (SVR), and BP neural network (BP) methods. The results showed that the accuracy of estimating DOM concentration using corrected spectral indices was significantly improved compared to pre-correction models, with the highest improvement of 38 % (SVR), the lowest of 23 % (BP), and an average improvement of 31 %. The RF model performed best, achieving R-2 = 0.81, RMSE = 3.34 mg/L, and MAE = 2.17 mg/L. For DO content estimation, the accuracy of models using corrected spectral indices was also improved significantly, with the highest improvement of 97 % (RF), the lowest of 39 % (SVR), and an average improvement rate of 67 %. The Random Forest model was again optimal, with R-2 = 0.69, RMSE = 1.97 mg/L, and MAE = 1.47 mg/L. This study indicates that the proposed spectral correction method helps to map the concentration of DOM and DO in freshwater aquaculture ponds with high accuracy using UAVbased multispectral images.

Monitoring water quality is crucial for water exchange, precise feeding, and quality control of water products in freshwater aquaculture. In light of the issue of spatial heterogeneity in freshwater aquaculture pond waters and the constraints of conventional sensor detection techniques and traditional machine learning models. In this study, UAV multispectral images were combined with four machine learning algorithms (Ridge, XGBoost, Cat- Boost, RF) and the Stacking model to model the estimation of Chlorophyll a (Chl-a) and Turbidity and map their spatial distribution. The findings indicate that, in contrast to machine learning models, the Stacking model of water quality parameter performs better with higher accuracy. Meanwhile,for Chl-a and Turbidity the optimal sub-model combination in the Stacking model varies, with the most effective estimation model for Chl-a concentration identified as RF-XGB-Ridge (R-2 = 0.84, RMSE=1.882 =1.882 mu g/L, MAE=3.433 =3.433 mu g/L and Slope = 0.791). As to Turbidity, the RF-CAB-Ridge model demonstrates superior performance, with macro-averaged precision (macro-p) of 93.3 %, macro-averaged recall (macro-R) of 88.8 %, macro-averaged F1-score (macro-F1) of 0.895, and Kappa coefficient of 0.813. Furthermore, the results of the joint analyses, which included measured samples and management measures at the test site, demonstrated that the spatial distribution maps of Chl-a and Turbidity were in alignment with the current status of water quality at the test site. This consistency was observed across both temporal and spatial scales. The results demonstrate that the integration of UAV multispectral images with the Stacking model can enhance the precision of water quality parameter models, facilitates the examination of the spatial and temporal distribution of water quality parameters and the underlying influencing factors, and advances the capability for dynamic monitoring of water quality parameters in freshwater aquaculture regions. Concurrently, it offers fundamental theoretical and methodological assistance for the precise regulation of water quality in freshwater aquaculture ponds and the formulation of optimal production management strategies.

Chlorophyll, a key pigment in leaf photosynthesis, is crucial for monitoring tobacco growth, evaluating quality, and determining optimal harvest timing. Hyperspectral remote sensing (HRS) via unmanned aerial vehicles (UAVs) provides a viable method to assess tobacco leaf chlorophyll content (LCC) due to its real-time and high-throughput capabilities. However, the existing spectral feature extraction methods often suffer from variability due to external factors or internal parameters, resulting in unstable results. To address this, we proposed a novel feature construction method, termed segmented fitting for integral (SFI). This method divides the entire spectral curve into five areas based on the spectral response properties of chlorophyll and nitrogen and then selects a suitable fitting function for each area to calculate the integrals. This idea not only fully utilizes the spectral reflectance and curve shape characteristics but also remains largely unaffected by external factors and internal parameters. To further verify the stability and predictive capability of the SFI method, we compared it against three different feature extraction methods, including the successful projections algorithm (SPA), the recursive feature elimination (RFE), and the principal component analysis (PCA), and also with two ensemble learning-based modeling approaches, namely, random forest (RF) and adaptive boosting (AdaBoost). The results demonstrated that the SFI method can effectively reduce data dimensionality and enhance model performance. Finally, we applied the best-performing SFI-AdaBoost model in field prediction to generate chlorophyll distribution and error maps, which closely aligned with actual chlorophyll measurements and demonstrated its potential for practical evaluation of tobacco LCC.

Fast, accurate, and real-time detection of nitrogen content in tobacco leaves is of great significance for monitoring the quality of tobacco leaves. Hyperspectral remote sensing (HRS) coupled with unmanned aerial vehicle (UAV) platform can provide unprecedented spectral information of field plants on a large scale. And with the support of various machine learning algorithms, a series of efficient models for leaf nitrogen content (LNC) assessment can be developed. This study aimed to develop a high-performance model to estimate the LNC of tobacco using UAV-borne HRS image data. Meanwhile, to cope with the heterogeneous performance problem of the individual model, ensemble learning strategies were applied to assemble multiple estimators, including multiple linear regression (MLR), decision tree regression (DTR), random forest (RF), adaptive boosting (Adaboost), and stacking to mine more valid data features. To accurately assess the performance of the established models, the coefficient of determination (R2), root mean square error (RMSE), and mean absolute percentage error (MAPE) were introduced as the evaluation indicators, and partial least squares regression (PLSR) was selected as the baseline model. Results on the test set showed that all ensemble learning methods outperformed PLSR (R2=0.680, RMSE=5.402 mg/g, 19.72%). Specifically, the stacking-based models achieved the highest accuracy as well as relatively high stability (R2=0.745, RMSE=4.825 mg/g, 17.98%). This study provides a reference for efficient and non-destructive detection of LNC or other vegetation phenotypic traits using UAV-borne HRS technology.

Tobacco is an important economic crop and the main raw material of cigarette products. Nowadays, with the increasing consumer demand for high-quality cigarettes, the requirements for their main raw materials are also varying. In general, tobacco quality is primarily determined by the exterior quality, inherent quality, chemical compositions, and physical properties. All these aspects are formed during the growing season and are vulnerable to many environmental factors, such as climate, geography, irrigation, fertilization, diseases and pests, etc. Therefore, there is a great demand for tobacco growth monitoring and near real-time quality evaluation. Herein, hyperspectral remote sensing (HRS) is increasingly being considered as a cost-effective alternative to traditional destructive field sampling methods and laboratory trials to determine various agronomic parameters of tobacco with the assistance of diverse hyperspectral vegetation indices and machine learning algorithms. In light of this, we conduct a comprehensive review of the HRS applications in tobacco production management. In this review, we briefly sketch the principles of HRS and commonly used data acquisition system platforms. We detail the specific applications and methodologies for tobacco quality estimation, yield prediction, and stress detection. Finally, we discuss the major challenges and future opportunities for potential application prospects. We hope that this review could provide interested researchers, practitioners, or readers with a basic understanding of current HRS applications in tobacco production management, and give some guidelines for practical works.

Soil organic matter (SOM) is a critical indicator of soil nutrient levels, and the precise mapping of its spatial distribution through remote sensing is essential for soil regulation, precise fertilization, and scientific management and protection. This information can offer decision support to agricultural management departments and various agricultural producers. In this paper, two new soil indices, NLIrededge2 and GDVI(rededge2), were proposed based on the sensitive spectral response characteristics of SOM in Northeast China. Nine parameters suitable for SOM mapping and modeling were determined using the competitive adaptive reweighted sampling (CARS) method, combined with spectrum reflectance, mathematical transformations of reflectance, vegetation indices, and so on. Then, utilizing unmanned aerial vehicle (UAV)-based multispectral images with centimeter-level resolution, a random forest machine learning algorithm was used to construct the inversion model of SOM and mapping SOM in the study area. The results showed that the random forest algorithm performed best for estimating SOM (R-2 = 0.91, RMSE = 0.95, MBE = 0.49, and RPIQ = 3.25) when compared with other machine learning algorithms such as support vector regression (SVR), elastic net, Bayesian ridge, and linear regression. The findings indicated a negative correlation between SOM content and altitude. The study concluded that the SOM modeling and mapping results could meet the needs of farmers to obtain basic information and provide a reference for UAVs to monitor SOM.

智慧烟草农业是现代烟草农业矩阵构建的重要组成,是解决烟叶生产手段落后、生产资源供给不足的有效途径.该研究面向云南12个植烟地州、5个烟草农业科学院研究团队、1个烟草质量监督检测站开展了关于烟草农业智慧化发展现状调查.基于调查数据展开对云南烟草农业智慧化发展现状和问题的分析,并从夯实发展基础、加快重大技术和装备研究、开展典型应用示范等方面提出了烟草农业智慧化发展的对策建议,为云南省智慧烟草农业建设的科学推进和战略目标的制定提供参考.

害虫检测是害虫测报的关键步骤,对于害虫防治具有重要意义,也是保证农作物产量和品质的前提.近年来,随着卷积神经网络的迅速发展,害虫检测技术进入智能化时代,使用深度学习相关技术实现精确的害虫检测已成为研究人员重点关注的课题.为了促进深度学习害虫检测技术的发展,对检测算法和现有数据集进行综述.总结了当前面临的数据匮乏、小目标检测、多尺度检测和密集与遮挡检测等四大难点问题,并分析了其主要成因.重点针对以上难点问题,总结归纳了近年来提出的深度学习害虫检测算法的改进策略和技术细节,以及面向实际场景的应用算法,对比分析了各类算法的性能表现、改进策略的适用场景及其优缺点.从面向复杂检测场景、解决数据匮乏问题、模型增量更新和应用落地等方面分析并展望了未来的研究趋势.

At present, tobacco leaf grading relies mainly on manual classification, which is highly intensive with respect to labor, materials, and cost; in addition, the performance of manual grading is poor. The realization of automatic grading is an urgent requirement in the tobacco industry. To address this need, we developed assembly line equipment and an RGB image classification method for tobacco grading. There is little difference in appearance between different grades of tobacco leaves, but there are some differences in high-level semantic features; therefore, tobacco grading is fundamentally a fine-grained visual categorization task. It is difficult to classify tobacco using hand-crafted image features. Therefore, in our method, we use a pyramid structure, attention mechanism, and angle decision loss function to improve the bilinear convolutional neural network (a fine-grained visual categorization framework) for tobacco grading. The feature extraction and classification model was trained using 66,966 images, to effectively extract high-level semantic features from a global tobacco image and multi-scale features from a local tobacco image. In an online test on assembly line equipment, the proposed model was able to classify six main grades of tobacco with an accuracy of 80.65%, which is higher than that of the state-of-the-art model. Additionally, the time required to classify each tobacco leaf was 42.1 ms. This work is of great significance for the industrial application of tobacco grading models and grading equipment, and it provides a theoretical reference for the quality grading of other agricultural products.

为探索深度学习技术在烟叶图像上的特征提取效果,提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)模型的烟叶等级分类方法,并对模型关注的烟叶特征进行了可视化分析.通过图像预处理得到高分辨率的局部烟叶图像,以弥补全局烟叶图像缩放后导致烟叶细节信息丢失;利用改进的CNN模型VGG-16和ResNet-50分别提取烟叶全局和局部图像特征;构建分类器对烟叶全局和局部图像的特征向量进行分类和结果融合;采用类别激活图(Class Activation Map,CAM)技术绘制模型关注烟叶特征的热力图.结果表明:提出的方法对6个等级的烟叶分级准确率达到84.71%,单张烟叶图像测试时间为17.87 ms;特征热力图显示ResNet-50模型对烟叶病斑、皱褶、主脉和纹理走势等局部特征较为敏感.该方法可为实现烟叶快速、准确分级提供支持.

Lodging depresses the grain yield and quality of maize crop. Previous machine learning methods are used to classify crop lodging extents through visual interpretation and sensitive features extraction manually, which are cost-intensive, subjective and inefficient. The analysis on the accuracy of subdivision categories is insufficient for multi-grade crop lodging. In this study, a classification method of maize lodging extents was proposed based on deep learning algorithms and unmanned aerial vehicle (UAV) RGB and multispectral images. The characteristic variation of three lodging extents in RGB and multispectral images were analyzed. The VGG-16, Inception-V3 and ResNet-50 algorithms were trained and compared depending on classification accuracy and Kappa coefficient. The results showed that the more severe the lodging, the higher the intensity value and spectral reflectance of RGB and multispectral image. The reflectance variation in red edge band were more evident than that in visible band with different lodging extents. The classification performance using multispectral images was better than that of RGB images in various lodging extents. The test accuracies of three deep learning algorithms in non-lodging based on RGB images were high, i.e., over 90%, but the classification performance between moderate lodging and severe lodging needed to be improved. The test accuracy of ResNet-50 was 96.32% with Kappa coefficients of 0.9551 by using multispectral images, which was superior to VGG-16 and Inception-V3, and the accuracies of ResNet-50 on each lodging subdivision category all reached 96%. The ResNet-50 algorithm of deep learning combined with multispectral images can realize accurate lodging classification to promote post-stress field management and production assessment.

为实现烤烟等级的快速准确识别,降低人工分级中主观因素对分级结果的影响,提高烟叶分级的准确性和一致性,提出一种基于烤烟RGB图像和深度学习的多尺度特征融合的烟叶图像等级分类方法,采用ResNet50提取烟叶图像特征,并引入基于注意力机制的SE模块（压缩激发模块）,增强不同通道特征的重要程度;同时,采用FPN（特征金字塔网络）对提取的由浅及深不同层级的烟叶特征进行融合,以实现烟叶多尺度特征的表达。采集皖南地区6 068个烤烟的正面和背面图像用于建模和分析。结果表明,提出的烟叶分级方法的分级正确率比经典CNN（卷积神经网络）高出5.21%,分级模型在新批次7个等级烟叶上的分级正确率为80.14%,相邻等级的分级正确率为91.50%。因此,采用RGB图像结合深度学习技术可实现烤烟烟叶等级的良好识别,可为烤烟烟叶收购等级评价提供一种新方法。

Aiming at the time series characteristics of soil moisture changes in the core planting area of tobacco, a soil moisture prediction model based on the combination of principal component analysis (PCA) and long-term short-term memory neural network (LSTM) was established and evaluated for tobacco precise production management and tobacco field water-saving irrigation applications. This study proposed a method using PCA to optimize the model of LSTM to predict soil moisture in tobacco-growing areas at different time intervals. The PCA method is mainly used to reduce the dimension of the original input layer meteorological data. The data are processed by dimensionality reduction to solve the input variables, and the neural network has a larger network size, resulting in reduced efficiency. LSTM network is used to solve the defects of traditional neural network model in time series data analysis. Experimental analysis shows that the LSTM prediction model of tobacco field soil moisture using PCA method can provide good short-term and medium-term prediction. The average absolute percentage error (MAPE) of the predicted soil moisture in tobacco fields is 0.27%, which is better than the full factorial radial basis function neural network (RBF NN), recurrent neural network (RNN) and traditional LSTM. The PCA and LSTM model meets the precise management requirements of actual tobacco cultivation and provides a new method for automated water-saving irrigation control of tobacco cultivation.