基于CNN和LSTM深度学习算法的作物播期差异遥感识别

Remote Sensing Identification of Crop Sowing Date Differences Based on CNN and LSTM Deep Learning Algorithms

摘要:

作物播期差异遥感识别可为精准农业和智慧农业提供空间数据支撑，对推动农业生产从“经验驱动”向“数据驱动”转型具有重要价值。本研究融合双阈值决策与深度学习算法，基于Landsat 8/9 OLI影像数据和野外采样数据，提取作物NDVI时序特征曲线，揭示作物出苗至生长旺盛期（4月25日—7月30日）的动态特征，识别作物早晚播差异的敏感生育阶段。计算敏感生育阶段作物整体生长斜率和相邻时相NDVI差值。生长斜率大于样本平均值且NDVI差值大于样本平均值时判定为早播样本，否则为晚播样本。采用随机森林（RF）、人工神经网络（NNC）机器学习模型和卷积神经网络（CNN）、长短期记忆网络（LSTM）深度学习模型对作物播期差异进行遥感识别。结果表明，作物早晚播敏感生育阶段内早播作物NDVI均值普遍比晚播作物高0.1~0.3，且早播作物生长进程显著提前，玉米快速生长期提前7~13d，葵花提前10~16d，西葫芦生长高峰期提前12~24d，番茄提前5~20d，瓜类提前18~35d。作物分类结果表明RF与CNN模型表现较优，其总体分类精度分别达91.77%和90.66%，Kappa系数分别为0.91和0.90，均能有效区分玉米、葵花、西葫芦、瓜类和番茄这5类作物的早播与晚播情况。经过对各模型分类结果的细节对比，CNN分类图像更连续且破碎度更低，在田埂识别与地块边界区分上优于其他模型，选取CNN模型为最优分类模型。

Abstract:

Remote sensing identification of the difference in crop sowing timing can provide spatial data support for precision agriculture and smart agriculture, and is of great value in promoting the transformation of agricultural production from “experience-driven” to “data-driven”. The double threshold decision-making and deep learning algorithms were integrated, based on Landsat 8/9 OLI image data and field sampling data, the time series characteristics curves of crop NDVI was extracted, revealing the dynamic characteristics of crops from emergence to vigorous growth period (April 25th-July 30th), and identifying the sensitive growth stages of crop sowing timing differences. The growth slope of crops in the sensitive growth stage and the difference in NDVI between adjacent images were calculated. When the growth slope was greater than the sample average and the NDVI difference was greater than the sample average, it was determined as an early sowing sample;otherwise, it was a late sowing sample. Random forest, artificial neural network machine learning models, and convolutional neural network, long short-term memory network deep learning models were used for remote sensing identification of crop sowing timing differences. The results showed that the NDVI mean value of early sowing crops within the sensitive growth stage was generally higher than that of late sowing crops by 0.1~0.3, and the growth process of early sowing crops was significantly earlier. The rapid growth period of early sowing corn was 7 to 13 days earlier than that of late sowing corn, and early sowing sunflower was 10 to 16 days earlier. The growth peak of early sowing zucchini was 12 to 24 days earlier than that of late sowing, early sowing tomato was 5 to 20 days earlier, and early sowing melon was 18 to 35 days earlier. The classification results showed that the RF and CNN models performed better, with overall accuracies reaching 91.77% and 90.66% respectively, and Kappa coefficients of 0.91 and 0.90, which could effectively distinguish the early and late sowing situations of the five types of crops: corn, sunflower, zucchini, melon, and tomato. Through detailed comparison of the classification results of each model, the CNN classification image was more continuous and had lower fragmentation, and the CNN model was selected as the optimal classification model in this study.