Comparisions In Tranditional Stomatal Models And Machine Learning Algorithms For Prediction Of Stomatal Conductance In Tropical Rainforest Trees

Stomata is a common channel for water vapor and carbon dioxide in and out of plants,which plays an important role in the balance regulation of photosynthesis and transpiration.At present,the plant model usually adopts unified stomatal optimization（USO）and the Ball-Berry-Leuning model（BBL）in combination with mesophyll conductance implicit photosynthetic model（g_m-implicit）to predict the response of stomatal conductance(g_sw)and its characteristics.However,the prediction accuracy is not high,especially in tropical rainforest tree species.In order to improve the prediction accuracy of g_sw,this study modified the algorithm structure of the photosynthesis-stomatal coupling model,introduced g_m explicit photosynthetic model（g_m-explicit）,and used three commonly used machine learning algorithms Random Forest（RF）,Deep neural networks（DNN）and Partial Least Squares regression（PLSR）replaced USO or BBL stomatal models to develop a variety of photosynthesis-machine learning algorithm fusion models,providing new ideas and methods for predicting stomatal conductance.In this study,we collected the photosynthetically intercellular CO₂concentration response curve（A_n/C_i curve）,gas exchange diurnal variation and environmental factors of 20 tropical species in Xishuangbanna and Panama rainforests of China,and conducted the algorithm evaluation on 20 species of tropical rain forest trees in Xishuangbanna and Panama rainforests.The main results of this study are as follows:1.The prediction accuracy of leaf assimilation rate（A_n）by g_m explicit photosynthesis model combined with BBL and USO stomatal models was better than that by g_m implicit photosynthetic model.The accuracy of A_n and g_sw predicted by g_mexplicit photosynthetic model coupled with BBL stomatal model(g_m-expP-BBL)was the highest(A_n:R²=0.83,MRE=29.3%;g_sw:R²=0.55,MRE=57.5%),but the prediction accuracy of g_sw was low.2.Three machine learning algorithms,RF,DNN and PLSR,combined with the photosynthesis-stomatal coupling model,significantly improved the prediction accuracy of g_sw,and all of them improved with the increase of training set species type and sample size,among which RF showed the best performance.In the data set divided by species type,when the training set accounted for 80%,the predicted results of RF algorithm were R²=0.73 and MRE=31.31%.When the data set was divided by sample size,when the training set accounted for 80%,the predicted result of RF algorithm was R²=0.91 and MRE=24.8%.In addition,RF algorithm still has good generalization ability under the classification of training data set by species type,and has little loss of model accuracy on test set.And this study also found that when the gradient of the training set was the same,the model prediction effect of dividing data sets by sample size was better than that of dividing data sets by species type.At the same time,RF algorithm was used to rank the importance of input factors.The results showed that the accurate prediction of A_n plays a crucial role in improving the prediction effect of g_sw.3.In order to further verify the prediction of diurnal and seasonal dynamics of tropical rainforest species by RF algorithm,a new photosynthesis-RF deep coupling model(g_m-expP-RF)was proposed by directly coupling RF algorithm with photosynthetic model,that is,replacing USO or BBL stomatal models in the photosynthesis-stomatal coupling model by RF algorithm.The results showed that g_m-expP-RF could better understand the diurnal and seasonal dynamics of 20 species in Xishuangbanna（XTBG）and Panama rainforest（PNM）,and further improved the prediction accuracy of g_sw.The g_sw prediction R²=0.91,RMSE=10.03 mmol m^–2 s^–1;g_sw prediction R²=0.84,RMSE=41.16 mmol m^–2 s^–1 for PNM species.Compared with the traditional photosynthesis-stomatal coupling model,the photosynthesis-RF coupling model can better predict the complex VPD response relationship between stomatal conductance and water vapor pressure difference.4.The photosynthesis-RF coupling model was tested on the independent data set of 15 species of Panamanian rainforest trees,and the results showed that it has good generalization ability in predicting g_sw.When the seasonal variation of 0-100cm soil water content was included into the training data set,the prediction accuracy of g_swseasonal variation by the photosynthesis-RF coupling model was significantly improved,and the RMSE decreased from 66.94 mmol m^–2 s^–1 to 61.92 mmol m^–2 s^–1,with a decrease of 7.5%.