基于GF-1与Landsat8 OLI影像的作物种植结构与产量分析

作物种植结构监测和估产是精准农业遥感的重点领域,其研究对于指导作物种植结构和制定农业政策具有重要意义。该文以黑龙江省北安市为研究区,以2015年的Landsat8 OLI和多时相GF-1为遥感数据源,基于物候信息和光谱特征确定的农作物识别关键时期和特征参数,构建面向对象的决策树分类模型,开展作物种植结构监测研究;综合植被光谱指数和地面采样数据,采用逐步回归方法建立产量遥感估算模型。结果表明:多源与多时相的遥感数据可以反映不同农作物的季相特征,应用本文所构建的决策树分类模型,作物分类效果较好,总体精度达87.54%,Kappa系数为0.8115;2015年,北安市的主要作物类型为大豆、玉米、水稻和小麦,面积分别为2204、1955、122和19 km~2,其中大豆的种植面积最大,占作物种植面积的51.24%。基于NDVI、EVI和GNDVI构建的多元回归模型为北安市大豆和玉米产量估算最优模型(R~2=0.823 7,均方根误差135.45 g/m~2,精度80.55%);北安市玉米高产区集中分布在西部,大豆的高产区主要分布在东部;2015年北安市玉米和大豆的单产分别为8 659、2 846 kg/hm~2,总产量分别为16.93×10~8、6.27×10~8 kg。利用作物关键物候期的多源多时相遥感数据能够精确高效地提取作物种植结构,构建的产量估算多元回归模型,为精准农业科学发展提供参考。

Analysis crops planting structure and yield based on GF-1 and Landsat8 OLI images

Abstract: Crop classification and yield estimation are key research in remote sensing-based precision agriculture, which have important significance in making agricultural policies. To improve the accuracy of classification based on single-source and single-season images, multi-temporal, multi-source and high spatial resolution image data were used to extract information of crops. Multi-source remote sensing data can play an important role in the coupling process. Multi-spectral data are used to distinguish between different crop species. Based on Landsat8 OLI (operational land imager) and GF-1 images, crop structure was mapped and yield was estimated for Bei''an County, Heilongjiang Province. According to phonology information and spectral characteristics, the critical period of crop identification and the characteristic parameters were determined, and the model of object-oriented decision-tree classification was built and crop structure was explored. Meanwhile, compositing multi-spectral images of crop maturation period and yield crop data, vegetation indexes were selected. Using correlation analysis, stepwise regression analysis and one-way ANOVA (analysis of variance), the correlation was explored and the model was built between yields of maize and soybeans and vegetation indices, which included NDVI (normalized differential vegetation index), EVI (enhanced vegetation index), GNDVI (green normalized difference vegetation index), OSAVI (optimal soil adjusted vegetation index), RVI (ratio vegetation index), SIPI (structure intensive pigment index), SAVI (soil adjusted vegetation index), NRI (nitrogen reflectance index) and DVI (difference vegetation index). Results show that the multi-source and multi-temporal remote sensing data can be used to show seasonal characteristics of different crops. Characteristic parameters of crops (including NDVI, NDWI, RVI, brightness, rectangular fit and texture) can be used to identify crop characteristics in landsat8 OLI and GF-1 images. After verified by ground investigation, the results of classification were accurate. The overall accuracy and Kappa coefficient were 87.54% and 0.811 5, respectively. The soybean had the largest area (2 204 km2) and the areas of maize, rice and wheat were 1 955, 122 and 19 km2, respectively. The high-yield maize was concentrated in the western area and the high-yield soybean was distributed in the east of study area. Correlation coefficients between crop yields and vegetation indices were more than 0.85 (P<0.001), which indicated that vegetation indices (including NDVI, EVI, GNDVI, OSAV and RVI ) were closely related with the production of maize and soybean. Meanwhile, the sensibility of each vegetation index was different (GNDVI > RVI > NDVI > EVI > OSAVI > SAVI > NRI > SIPI > DVI). After cross validation for the yield-estimation model, the NDVI, EVI and GNDVI model can be used to estimate accurately the yield of maize and soybean, and the yield estimation was significantly correlated to the actual production (R2=0.823 7, RMSE=135.45 g/m2, accuracy was 80.55%) based on regression analysis which indicated these vegetation indices can be used for crop yields estimation with the yield-estimation model. Total yields of maize and soybean were estimated to be 16.93×108 and 6.27×108 kg, with per unit area yields of 8 659 and 2 846 kg/hm2, respectively. Crop planting structure can be mapped accurately and efficiently using crop key phonological phase, multi-source and multi-temporal remote sensing data. The results provide the reference for the study on remote sensing indicators and the scientific and technological support for the development of precision agriculture science.