地形异质下光谱组合与机器学习融合的水稻种植结构提取研究

Extraction of Rice-planting Structures under Terrain Heterogeneity by Fusing Multispectral Indices with Machine-learning Algorithms

摘要:

精确掌握区域尺度作物种植结构分布特征，对精细化农业生产管理、优化水资源配置和保障粮食安全具有重要意义。本研究聚焦江西省典型地貌单元，选取平原地貌的信丰灌区和丘陵地貌的大山灌区为研究区，基于Sentinel-2遥感数据，构建包含归一化植被指数（NDVI）、增强型植被指数（EVI）、改进型归一化水体指数（MNDWI）及其多元组合的遥感特征矩阵，耦合朴素贝叶斯（NB）、支持向量机（SVM）、多层感知机（MLP）、随机森林（RF）及极端梯度提升（XGBoost）5种机器学习算法，引入交叉验证，网格搜索进行模型参数优化，研究不同光谱指数与算法组合提取水稻种植结构的最优融合模型，最后引入平移、噪声扰动等数据增强方法，模拟年际气候波动和作物生长时序变化对分类性能的影响，并对两个灌区水稻种植结构（双季稻、一季晚稻、中稻）进行遥感分类制图。结果表明：在信丰灌区和大山灌区中，NDVI+MNDWI-XGBoost均展现出最优分类性能，大山灌区总体精度高达97.30%，Kappa系数达0.958；信丰灌区总体精度高达95.40%，Kappa系数达0.915，并且两个灌区的平均生产者精度和用户精度均超过95%。地形特异性、田块状况和种植复杂度是决定最优光谱指数组合和机器学习算法耦合提取特定灌区水稻种植结构的关键要素，制图结果显示信丰灌区不同田块之间的边界较为明显，并且田块形状规则统一，大山灌区田块呈现不规则和破碎特征，并且田块间的边界并不清晰。信丰灌区以种植双季稻为主，占比80.50%，大山灌区以种植单季稻为主，占比78.60%。本研究明确了地形异质条件下遥感分类的最优算法与光谱指数组合方案，为灌区尺度水稻种植结构提取和农业精准管理提供了有效技术支撑和方法参考。

Abstract:

Accurately capturing crop planting structures at regional scales is crucial for precise agricultural management, optimal water resource allocation, and food security assurance. Focusing on typical geomorphological units in Jiangxi Province, selecting the Xinfeng irrigation area with plain terrain and the Dashan irrigation area characterized by hilly terrain as case study areas. Utilizing Sentinel-2 remote sensing data, a remote sensing feature matrix incorporating the normalized difference vegetation index (NDVI), enhanced vegetation index (EVI), modified normalized difference water index (MNDWI), and their various combinations was constructed. Five machine learning algorithms, Naive Bayes (NB), support vector machine (SVM), multi-layer perceptron (MLP), random forest (RF), and extreme gradient boosting (XGBoost), were coupled with this matrix. Cross-validation and grid search were implemented for optimizing model parameters, aiming to identify the best-performing combination of spectral indices and algorithms for extracting rice-planting structures. Additionally, data augmentation techniques such as translation and noise perturbation were introduced to simulate the impacts of inter-annual climate variability and crop growth temporal dynamics on classification accuracy. Rice-planting structures, including double-cropping rice, single-season late rice, and medium rice, were mapped for both irrigation areas. Results indicated that the combination of NDVI+MNDWI with the XGBoost algorithm achieved the highest classification performance in both irrigation areas. Specifically, Dashan exhibited an overall accuracy of 97.30% and a Kappa coefficient of 0.958, whereas Xinfeng achieved an overall accuracy of 9540% and a Kappa coefficient of 0.915. Both areas demonstrated average producer and user accuracies exceeding 95%. Terrain specificity, field conditions, and cropping complexity emerged as critical factors influencing the optimal spectral index and machine learning algorithm combination. Mapping results revealed clearly defined field boundaries and uniform shapes in the Xinfeng irrigation area, whereas Dashan featured irregular and fragmented fields with indistinct boundaries. Double-cropping rice dominated in Xinfeng, accounting for 80.50% of the total area, while single-season rice predominated in Dashan, covering 78.60%. This research established optimal remote sensing classification methods and spectral index combinations under varied terrain conditions, providing robust technical support and methodological guidance for extracting rice-planting structures at the irrigation-area scale and enhancing precision agricultural management.