基于CNN-LSTM的苹果树种植区域提取

Apple Planting Area Extraction Based on Improved CNN-LSTM Model

摘要:

苹果树种植区域提取有利于农业资源高效管理。为解决苹果种植区域提取中存在的分类精度不高、时效性滞后等问题，提出一种基于Sentinel-2和MODIS融合影像的卷积神经网络-长短期记忆网络 （CNN?LSTM） 时序分类模型。首先采用ESTARFM 时空融合算法构建融合影像，对不同卫星影像在空间和时间监测能力优势和缺陷进行互补，得到高空间和高时间分辨率共存的影像数据。在特征选择方面，通过随机森林模型进行重要性分析并结合后向特征消除法从25个原始特征中选15个关键特征变量作为最优特征组合。分类模型方面，卷积神经网络（Convolutional neural network, CNN）可以很好地在空间域、光谱域提取有效特征。长短期记忆网络（Long short-term memory, LSTM）作为循环神经网络（Recurrent neural network, RNN）的改进，可以处理不等长的输入序列。二者结合能够提取“时空谱”有效特征，实现更精准的图像分类和遥感数据分析。以烟台市牟平区观水镇为研究区，利用时空融合弥补原始 Sentinel-2的影像缺失，使用 CNN?LSTM模型进行苹果树种植区域提取，并与常用的机器学习分类算法进行对比，进而确定最优分类模型。研究表明在苹果种植区域提取方面 CNN?LSTM 模型总体精度为 97.98%，Kappa 系数为 0.958 6，总体精度对比其他 4 种机器学习算法 CART、SVM、RF、GBDT分别高15.43、5.25、4.00、3.31个百分点，与LSTM模型总体精度和Kappa系数相比分别提高2.11个百分点和0.0148。所提出的苹果树种植区域精准遥感提取方法可为制定科学合理的农业管理措施提供有力支持。

Abstract:

The efficient management of agricultural resources can be significantly improved through the accurate extraction of apple cultivation areas. In order to solve the problems of poor classification accuracy and time lag in apple planting area extraction, a CNN?LSTM temporal classification model was proposed based on Sentinel-2 and MODIS fusion images. The ESTARFM spatio-temporal fusion algorithm was firstly used to construct the fusion image, which complemented the strengths and weaknesses of different satellite images in spatial and temporal monitoring capabilities, and obtained image data with high spatial and temporal resolution. The random forest model was utilized to select the most optimal feature combinations from the initial 25 features, narrowed down to 15 key variables using backward feature elimination. In terms of classification models, convolutional neural networks（CNN）can well extract effective features in the spatial and spectral domains. As an improvement of recurrent neural network, long short-term memory network （LSTM） can handle unequal input sequences. The combination of the two networks proposed can extract effective features in the spatial, temporal and spectral domains to achieve more accurate image classification and remote sensing data analysis. Taking Guanshui Town, Muping District, Yantai City as the study area, the spatio-temporal fusion algorithm was utilized to compensate for the lack of images from a single Sentinel-2, and the CNN?LSTM model was used for apple tree planting area extraction. The CNN?LSTM model achieved an overall accuracy of 97.98% and a Kappa coefficient of 0.9586, outperforming the other four machine learning algorithms by 15.43 percentage points，5.25 percentage points，4.00 percentage points, and 3.31 percentage points，respectively. The overall accuracy and Kappa coefficient of the CNN?LSTM model were improved by 2.11 percentage points and 0.0148, respectively, compared with that of the LSTM model. The precise remote sensing extraction method for apple tree planting areas proposed can provide strong support for the development of scientific and rational agricultural management.