基于Sentinel-1/2时序数据与历史信息融合的甘蔗分布提取方法

Sugarcane Distribution Extraction Method Based on Sentinel-1/2 Time-series Data and Historical Information Fusion

摘要:

精准提取甘蔗种植分布对我国糖料作物监测和食糖供应安全评估具有重要意义。针对现有方法以单年遥感观测为主、未能充分利用甘蔗多年连续种植先验信息的问题,本研究以广西崇左市为研究区,基于Sentinel-2 L2A与Sentinel-1 SAR时序数据,构建了融合多年历史信息的长短期记忆网络(LSTM)多分支分类框架,系统对比了历史数据(前n年原始遥感特征)与历史概率(前n 年基线模型输出的连续甘蔗归属概率)两种先验信息引入策略在2022—2025 年4 个目标年份中的分类效果。结果表明,引入3 年历史信息后,各数据源方案的分类精度均显著提升。其中,“Sentinel-1/2融合+3年历史数据”方案表现最优,其在4年平均条件下,F1值达到90.6%、Kappa系数为0.854;相较于同源单年基线模型,该方案的F1值提升约8.3个百分点;在3年回溯期条件下,历史数据与历史概率两种策略的F1 值均稳定在90% 以上,表现出较高的一致性;随着回溯年限由1 年延长至3 年,模型分类精度呈整体上升趋势,该最佳回溯年限与广西甘蔗“一年新植、两年宿根”的3 年农艺种植周期高度吻合。此外,Sentinel-1与Sentinel-2的融合在基线场景下未能展现显著协同效应,但在引入3年历史信息后,融合方案精度明显优于仅光学方案,表明SAR数据的结构敏感性在时序先验信息的支持下能够充分发挥作用。基于最优方案,本研究生成了崇左市2022—2025年10m分辨率甘蔗分布结果,验证了该方法在多年连续制图中的可行性。

Abstract:

Accurately mapping sugarcane planting distribution is of great significance for monitoring the sugar crop industry and assessing sugar supply security in China. To address the limitation of existing methods that primarily relied on single-year remote sensing observations and failed to fully utilize multi-year continuous planting prior information, the Chongzuo City in Guangxi Zhuang Autonomous Region was selected as the study area. A multi-branch long short-term memory (LSTM) classification framework integrating multi-year historical information was constructed based on time-series data from Sentinel-2 L2A and Sentinel-1 SAR imagery. Two strategies for incorporating prior information were systematically compared: the historical data strategy (using raw remote sensing features from the previous n years) and the historical probability strategy (using continuous sugarcane probability maps output by a baseline model for the previous n years). The classification performance was evaluated for four target years from 2022 to 2025. The results demonstrated that introducing three years of historical information significantly improved classification accuracy across all data source schemes. Among all configurations, the fusion of Sentinel-1 and Sentinel-2 combined with three years of historical data achieved the best performance, with an average F1-score of 90.6%, a Kappa coefficient of 0.854, representing an improvement of approximately 8.3 percentage points in F1-score compared with the single-year baseline model. Under the three-year retrospective condition, both the historical data and historical probability strategies yielded F1-scores consistently above 90%, showing high consistency between the two approaches. As the retrospective period increased from one to three years, the classification accuracy exhibited an overall upward trend, and this optimal retrospective period of three years coincided well with the typical three-year agronomic cycle of sugarcane cultivation in Guangxi, characterized as “one year of newly planted cane followed by two years of ratoon cane”. Furthermore, the fusion of Sentinel-1 and Sentinel-2 data did not demonstrate a significant synergistic effect in the baseline scenario without historical information; however, after incorporating three years of historical information, the fusion scheme significantly outperformed the optical-only scheme, indicating that the structural sensitivity of SAR data can be fully leveraged with the support of temporal prior information. Based on the optimal scheme, it generated 10-meter resolution sugarcane distribution maps for Chongzuo City from 2022 to 2025, confirming the feasibility and effectiveness of the proposed method for multi-year continuous sugarcane mapping.