ABSTRACT

NDVI (Normalized difference vegetation index) is a critical variable for monitoring climate change, studying ecological balance, and exploring the pattern of regional phenology. Traditional neural network models only consider image features in time series prediction, while historical data and its changes play an important role in time series forecasting. For this study, we proposed convolutional neural networks (CNN) combined feature engineering forecasting model (SF-CNN), which integrated both the advantages of image characteristics learned from CNN and statistic characteristics calculated by historical data related to the forecast period to improve the accuracy of NDVI predictions in the next 3 months with 30-day interval at multiple complex areas. To intuitively show the performance of SF-CNN, it was compared with CNN using the same parameters. Results mainly showed that (1) in terms of visual analysis, the texture, pattern, and structure of predicted NDVI using SF-CNN are similar to the observed NDVI, and SF-CNN exhibits strong generalization ability; (2) in terms of quantitative assessment, SF-CNN generally outperforms CNN, and it can improve the reliability and robustness for predicting NDVI through simple statistical characteristics while reducing the uncertainties; (3) SF-CNN can learn seasonal and sudden changes in four different and complex study areas with considerable accuracy and without extra data.

NDVI（归一化差异植被指数）是监测气候变化，研究生态平衡以及探索区域物候模式的关键变量。传统的神经网络模型在时间序列预测中只考虑图像特征，而历史数据及其变化在时间序列预测中起着重要作用。在这项研究中，我们提出了卷积神经网络（CNN）组合特征工程预测模型（SF-CNN），该模型结合了从CNN中学习到的图像特征的优势和由与预测周期相关的历史数据计算出的统计特征的优势，以提高准确性未来3个月在多个复杂区域的NDVI预测，间隔30天。为了直观地显示SF-CNN的性能，使用相同的参数将其与CNN进行了比较。结果主要表明：（1）在视觉分析方面，使用SF-CNN预测的NDVI的纹理，图案和结构与观察到的NDVI相似，并且SF-CNN具有很强的泛化能力；（2）在定量评估方面，SF-CNN通常优于CNN，它可以通过简单的统计特征提高NDVI预测的可靠性和鲁棒性，同时减少不确定性；（3）SF-CNN可以在四个不同且复杂的研究区域中以相当高的准确性学习季节性和突然变化，而无需额外的数据。SF-CNN通常优于CNN，它可以通过简单的统计特征提高NDVI预测的可靠性和鲁棒性，同时减少不确定性。（3）SF-CNN可以在四个不同而复杂的研究区域中以相当高的准确性学习季节性和突然的变化，而无需额外的数据。SF-CNN通常优于CNN，它可以通过简单的统计特征提高NDVI预测的可靠性和鲁棒性，同时减少不确定性。（3）SF-CNN可以在四个不同而复杂的研究区域中以相当高的准确性学习季节性和突然的变化，而无需额外的数据。