基于多光谱遥感和CNN的玉米地上生物量估算模型

Estimating Aboveground Biomass of Maize Based on Multispectral Remote Sensing and Convolutional Neural Network

摘要:

目前玉米地上生物量（Aboveground biomass，AGB）的预测方法集中在使用从无人机图像中提取光学植被指数，通过线性模型或机器学习算法与AGB建立关系，原始图像信息损失严重，玉米生长后期的饱和效应会严重降低模型精度。针对此问题，本文收集了玉米拔节期、吐丝期和乳熟期的无人机图像和地面数据。分析了不同生育期玉米干地上生物量、鲜地上生物量与8个植被指数(Vegetation indexes, VIs)之间的相关性。分别以最优植被指数作为输入建立多层感知机(Multilayer perceptron, MLP）模型、以无人机多光谱图像作为输入建立卷积神经网络（Convolutional neural network, CNN）模型来估算玉米干地上生物量、鲜地上生物量。结果表明，基于MLP的玉米干地上生物量估算模型随着玉米生育期推进，模型的精度急剧下降，3个生长期MLP模型验证集R2分别为0.65、0.23、0.32，RMSE分别为0.27、2.15、5.03 t/hm2。CNN模型能够较好地克服光谱饱和问题，具有良好的精度和适用性，3个生育期验证集R2分别提高27.69%、191.30%、171.88%，RMSE分别降低22.22%、38.14%、45.53%。基于MLP的玉米鲜地上生物量估算模型在玉米生长后期模型的精度同样较低，吐丝期、乳熟期验证集的R2分别为0.27、0.37，RMSE分别为11.57、14.98 t/hm2。CNN模型2个生育期验证集的R2分别提高159.26%、129.73%，RMSE分别降低26.62%、54.01%。使用原始多光谱图像作为输入的CNN模型取得了最好的估计结果，可为玉米不同生育期的监测研究、精准管理提供指导。

Abstract:

Rapid and accurate estimation of aboveground biomass (AGB) in maize is crucial for evaluating maize growth and precise field management. Current AGB prediction methods mainly use optical vegetation indexes extracted from UAV images, employing linear models or machine learning algorithms. These methods often result in significant loss of raw image information and face saturation effects during later stages of maize growth, severely degrading model accuracy. UAV images and ground data were collected from maize at the nodulation, silking, and milking stages. The correlations between dry and fresh AGB of maize and eight vegetation indexes at different fertility stages were analyzed. Optimal vegetation indexes were used to build a multilayer perceptron (MLP) model, while UAV multispectral images were used to construct a convolutional neural network (CNN) model to estimate dry and fresh AGB, respectively. The results showed that the accuracy of the MLP-based maize AGB dry weight estimation model was decreased sharply with the advancement of maize fertility, and the R2 values of validation set of MLP model for the three growing seasons were 0.65, 0.23, and 0.32, and the RMSE values were 0.27 t/hm2, 2.15 t/hm2, and 5.03 t/hm2, respectively. The CNN model can better overcome the spectral saturation problem with good accuracy and applicability. The R2 values of the three fertility validation sets were improved by 27.69%, 191.30% and 171.88%, and the RMSE was reduced by 22.22%, 38.14% and 45.53%, respectively. The accuracy of the MLP-based maize AGB fresh weight estimation model was similarly low in the late maize growth stage model, with R2 values of 0.27 and 0.37, and RMSE values of 11.57 t/hm2 and 14.98 t/hm2 for the validation sets of the spatula and milk maturity stages, respectively. The R2 values of the validation set for the two fertility stages of the CNN model was improved by 159.26% and 129.73%, and the RMSE was decreased by 26.62% and 54.01%, respectively. The CNN model using original multispectral images as inputs achieved the best estimation results, providing valuable guidance for monitoring research and precise management of maize at different fertility stages.