【作者】 周俊；

【导师】 金秀；

【作者基本信息】 安徽农业大学 ， 计算机科学与技术， 2024， 硕士

【摘要】 土壤养分含量是土壤肥力的重要指标,对作物生长和农业生产具有重要影响。近红外和短波红外(VNIR-SWIR)漫反射光谱技术因其快速、经济、无损等特点,在土壤养分预测分析中的应用越来越受到人们的重视。近年来,深度学习技术在土壤光谱数据分析方面取得了显著成效,为土壤养分含量预测提供了新的思路。然而在使用一维光谱数据建立土壤养分含量预测模型时仍面临不少挑战,这些挑战包括数据维度不匹配、模型训练效率低下、数据分布不均匀以及数据量较少等问题。因此,利用深度学习在图像处理领域的优势,建立简单普适又准确的土壤养分预测模型具有重要意义。本文的工作围绕土壤养分的预测方法展开,主要包括如下三个方面的内容:(1)研究了光谱数据的不同维度转换方法,以及探究不同维度下的光谱数据在卷积神经网络下建立土壤养分预测模型。实验中选择了世界上最大的土壤光谱数据库LUCAS作为研究对象,对原始光谱数据经过预处理,使用了3种不同的维度转换方法(CR、GAF、MTF)将一维光谱数据转换为二维光谱图像。依据数据的不同维度,使用不同卷积神经网络(1D-CNN、2D-CNN)设计了不同的土壤养分预测模型,用于预测八种土壤养分含量(OC、N、P、K、CEC、pH、Sand和Clay)。结果表明:经过SG预处理后使用GAF中的GADF维度转换方法,利用2D-CNN训练模型的方法在测试集上的结果表现最优,其中,预测OC、N、CEC、pH、Sand、Clay的RMSE分别为24.18、1.82、32.40、192.07、8.71、0.60、18.41和7.24,R~2分别为0.93、0.89、0.13、0.25、0.78、0.88、0.63和0.80,比其他两种维度转换方法的模型效果好,并且优于1D-CNN的建模结果。从而证实了维度转换方法的有效性,同时说明了GADF方法在数据转换方面的优势。(2)设计并实现了一种基于Swin Transformer的土壤养分预测模型。对于土壤光谱图像的模型处理,实验对比并选择最优的模型配置。将其他深度学习算法(Res Net50、Mobile Net、Efficient Net和Vision Transformer)应用于建立土壤养分预测模型,结果表明,使用GADF与Swin Transformer建立的土壤养分预测模型的性能最佳,预测OC、N、P、K、CEC、pH、Sand、Clay的RMSE分别为20.25、0.98、25.61、142.32、7.55、0.41、15.33和6.14,R~2分别为0.96、0.94、0.37、0.55、0.81、0.91、0.74和0.84,同时该方法与使用相同数据集的其他现有方法进行了性能比较,同样领先于其他方法,验证了GADF-Swin Transformer方法的优越性。(3)探究了基于深度扩散生成的土壤养分预测模型。针对土壤样本分布不均和光谱数据采集困难的问题,设计了一种基于扩散模型DiT的光谱图像生成模型,对DiT中不同的Transformer模块进行了实验研究,包括In-context conditioning、Cross-attention、adaLN和adaLN-zero,以及DiT中Transformer block大小与patch大小的不同组合。经对比分析得出了最优的土壤光谱图像生成模型配置。经过对DiT模型生成土壤光谱图像测试,结果表明,不仅生成的图像数据集可以有效地应用于现有的土壤养分预测模型中,而且在使用混合数据进一步提升了模型在预测土壤养分含量上的测试精度。综上所述,本文针对土壤养分的精准识别所提出的方法能够显著提高模型在土壤养分含量上的准确性,其研究结果对于土壤养分检测具有重要的参考价值。更多还原

【Abstract】 Soil nutrient content is an important indicator of soil fertility,which has a significant impact on crop growth and agricultural production.The application of near-infrared and short-wave infrared(VNIR-SWIR)diffuse reflectance spectroscopy techniques in soil nutrient prediction and analysis has received more and more attention because of its fast,economical and non-destructive features.In recent years,deep learning techniques have achieved remarkable results in soil spectral data analysis,providing new ideas for soil nutrient content prediction.However,there are still a number of challenges in using one-dimensional spectral data to build soil nutrient content prediction models,which include data dimensionality mismatch,inefficient model training,uneven data distribution,and low data volume.Therefore,it is of great significance to utilize the advantages of deep learning in the field of image processing to establish simple and universal yet accurate soil nutrient prediction models.This paper’s work is centered on the prediction method of soil nutrients,which mainly includes the following three aspects:(1)Research was conducted on different dimensional conversion methods for spectral data,as well as exploring different dimensions of spectral data to build soil nutrient prediction models under convolutional neural networks.LUCAS,the world’s largest soil spectral database,was chosen as the research object in the experiment,and the raw spectral data were preprocessed,and three different dimensional conversion methods(CR,GAF,and MTF)were used to convert one-dimensional spectral data into two-dimensional spectral images.Based on the different dimensions of the data,different soil nutrient prediction models were designed using different convolutional neural networks(1D-CNN,2D-CNN)for the prediction of eight soil nutrient contents(OC,N,P,K,CEC,pH,Sand and Clay).The results showed that the method of training models using 2D-CNN after SG preprocessing using the GADF dimension transformation method in GAF performed optimally on the test set,where the RMSEs for predicting OC,N,CEC,pH,Sand,and Clay were 24.18,1.82,32.40,192.07,8.71,0.60,respectively,18.41,and 7.24,and R~2 of 0.93,0.89,0.13,0.25,0.78,0.88,0.63,and 0.80,respectively,which are better than the models of the other two dimension transformation methods and superior to the modeling results of 1D-CNN.This confirms the effectiveness of the dimension transformation methods and illustrates the advantages of the GADF method in data transformation.(2)Designed and implemented a soil nutrient prediction model based on Swin Transformer.For model processing of soil spectral images,experiments are conducted to compare and select the optimal model configuration.Other deep learning algorithms(Res Net50,Mobile Net,Efficient Net,and Vision Transformer)were applied to build a soil nutrient prediction model,and the results showed that the soil nutrient prediction model built using GADF with Swin Transformer had the best performance,predicting OC,N,P,K,CEC,pH,Sand,and Clay with RMSE of 20.25,0.98,25.61,142.32,7.55,0.41,15.33,and 6.14,respectively,and R~2 of 0.96,0.94,0.37,0.55,0.81,0.91,0.74,and 0.84,respectively,and also that the method was compared with other existing methods,and the performance of the method was compared,which was also ahead of other methods,verifying the superiority of the GADF-Swin Transformer method.(3)A soil nutrient prediction model based on deep diffusion generation was explored.Aiming at the problem of uneven distribution of soil samples and the difficulty of spectral data acquisition,a spectral image generation model based on the diffusion model DiT was designed,and different Transformer modules in DiT were experimentally investigated,including In-context conditioning,Cross-attention,adaLN and adaLN-zero,and different combinations of Transformer block size and patch size in DiT.The optimal model configuration for generating soil spectral images is derived from the comparative analysis.After testing the DiT model for generating soil spectral images,the results show that not only the generated image dataset can be effectively applied in existing soil nutrient prediction models,but also the test accuracy of the model in predicting soil nutrient content is further improved when using mixed data.更多还原