基于多源遥感数据融合和LSTM算法的作物分类研究

准确、及时地获取农作物的空间分布信息,对于指导农业生产、制定农业政策具有重要意义。为了检验长短时记忆网络(long short-term memory,LSTM)算法在基于时序遥感数据进行作物分类中的优势,该文以临汾盆地为研究区域,利用Savitzky-Golay滤波对MODIS NDVI进行平滑处理,并采用ESTARFM(enhanced spatial and temporal adaptive reflectance fusion model)算法对滤波后的MODIS NDVI和Landsat NDVI进行融合,生成空间分辨率为30 m、时间分辨率为8天的时序NDVI。基于Landsat NDVI利用LSTM算法进行作物分类,同时,基于融合NDVI分别利用LSTM算法和神经网络(neuralnetwork,NN)算法进行作物分类,并对比3种方法的分类精度。结果表明,Savitzky-Golay滤波后的时序MODISNDVI能够反映不同作物的物候特征;基于融合NDVI的分类精度明显高于基于LandsatNDVI的分类精度,表明融合后的时序NDVI由于具有更高的时间分辨率,能够更加突出不同作物的物候特征,显著提高作物分类精度;基于融合NDVI和LSTM算法的分类精度高于基于融合NDVI和NN算法的分类精度,前者的冬小麦面积估测精度高于后者的估测精度,表明LSTM算法的分类精度高于NN算法。该文可为基于遥感影像进行不同作物种植区域提取的研究提供重要的方法参考。

Crop classification based on multi-source remote sensing data fusion and LSTM algorithm

Accurate distribution information of crop types is vital for monitoring crop growth, guiding agricultural production, and making effective management measurements. Time series remote sensing data can reflect phenological characteristics of crops, which have more advantages than single temporal data in identifying crop types or planting patterns. MODIS and Landsat data can be fused to obtain time series data with medium spatial resolution and high temporal resolution, which can be used for classifying different crops based on phenology characteristics. In this study, in order to test the accuracy of combining long short-term memory (LSTM) algorithm with time series remote sensing data in crop classification, the Linfen basin was chosen as the study area for obtaining crop distribution map. At first, the Savitzky-Golay filter was used to denoise and reconstruct time series MODIS NDVI data. Then the filtered MODIS NDVI and Landsat NDVI were merged by the enhanced spatial and temporal adaptive reflectance fusion model (ESTARFM) to generate time series NDVI with a spatial resolution of 30 m pixel by pixel. Based on field investigation, GlobeLand30 data, Google Earth images and agro-meteorological stations data, we obtained the coordinate information of several sampling sites representing different land cover and crop types. The phenological characteristics of time series NDVI of the pixels covering the sampling sites were analyzed, and the types of randomly selected pixels were determined based on the phenological characteristics for increasing the number of sampling sites. Three methods were used for crop classification in this study: 1) the Landsat NDVI of training samples were used to train the LSTM model, and the trained LSTM model was adopted to determine the crop type pixel by pixel (called the Landsat NDVI+LSTM method); 2) the fused NDVI of training samples were used to train the LSTM model for crop classification (called the fused NDVI+LSTM method); and 3) the fused NDVI of training samples were used to train the neural network (NN) model for crop classification (called the fused NDVI+NN method). In order to compare the accuracies of the three methods, the classification accuracies were evaluated with the validation samples. The evaluation indexes included overall accuracy (OA) and Kappa coefficient. Also, the planting area of winter wheat for each county of the study area was estimated according to the crop classification map, and the relative error (RE) and root mean square error (RMSE) between estimated and statistical wheat areas were calculated for further validating the accuracies of the three methods. Results showed that the Savitzky-Golay filter can remove the influence of factors such as cloud and atmosphere, thus the reconstructed time series MODIS NDVI curves could reflect the phenological characteristics of crops effectively. Positive correlation between the fused NDVI and the Landsat NDVI indicated the fused NDVI can reflect the information of Landsat NDVI effectively. The classification accuracies based on the fused NDVI, either using the fused NDVI+LSTM (OA=90.00%, Kappa=0.88) or fused NDVI+NN (OA=88.10%, Kappa=0.86) methods, were significantly higher than the accuracy of the Landsat NDVI+LSTM method (OA=82.86%, Kappa=0.80). The RE and RMSE of the formers were lower than those of the latter. These results indicated that the fused time series NDVI could highlight the phenological information of different crop types, thus the classification accuracy can be improved significantly. In addition, the classification accuracy of the fused NDVI+LSTM method was slightly higher than that of the fused NDVI+NN method, and the RE and RMSE of the former were lower than those of the latter. These indicated that the classification accuracy of LSTM algorithm was higher than that of the NN algorithm. This study can provide an important reference for accurately extracting distribution information of different crops in the study area.