基于SupReMe影像重建和RF的玉米冠层LAI反演

针对Sentinel-2卫星影像拥有3个对植被生长状况非常敏感、空间分辨率为20m的红边波段（705、740、783nm）,其空间分辨率与可见光和近红外波段10m的空间分辨率不一致,使Sentinel-2影像应用受到限制的问题,基于多光谱多分辨率估计的超分辨率（Super-resolution for multispectral multiresoltion estimation, SupReMe）算法将空间分辨率20m的6个波段重建为10m;以重建后的影像为数据源,耦合PROSAIL辐射传输模型和随机森林模型反演玉米冠层叶面积指数（LAI）,并以野外实测LAI验证其反演精度。结果表明,采用SupReMe算法对Sentinel-2影像进行重建后,在保持光谱特性不变的同时提高了影像的空间细节;基于重建影像和原始影像的LAI反演决定系数R2分别为0.70、0.68,均方根误差RSME分别为0.240、0.262。研究表明,利用SupReMe算法重建后的Sentinel-2卫星影像,能够在提高玉米冠层LAI反演空间分辨率的同时提高反演精度,在挖掘高分辨率农作物生长信息方面具有很大潜力。     

基于视觉里程计的森林样地调查系统研究

以视觉里程计技术恢复连续摄影序列图像位姿,并以恢复位姿的图像为基础构建样地调查系统。该系统通过对图像位姿尺度恢复、定义样地坐标系、标记立木等过程估计样地中立木位置及胸径。用相机对12块半径为7. 5 m的圆形样地进行连续摄影,获取有序图像序列,并使用构建的样地调查系统对图像序列进行处理,以获取样地中立木位置及胸径。实验结果表明,所有样地立木位置估计值x轴与y轴方向的偏差(BIAS)分别为0. 04、-0. 03 m,均方根误差(RMSE)分别为0. 21、0. 17 m;样地中立木胸径估计值的BIAS及RMSE分别为0. 09 cm(0. 51%)和0. 88 cm(5. 03%)。     

乡镇尺度森林景观格局特征与影响因素定量分析

为更客观、精细地刻画森林生态系统格局的完整性、稳定性和森林资源健康状况,为森林资源源头保护和合理经营管理提供科学依据和决策支持,以河北省易县森林资源清查数据、DEM数据和社会经济资料为基础,运用景观指数、结构方程模型和统计学方法,分析乡镇尺度森林景观类型、数量与分布,研究自然地理、社会经济和土地开发利用主导下森林景观格局特征差异及其影响机制。结果表明,西部山区乡镇森林景观资源优势明显,东部平原乡镇森林资源则较小,森林资源主要分布于海拔100～300 m浅山丘陵区和海拔500～800 m中山区。在景观和类型水平上,斑块密度(PD)、凝结度(COHESION)对地理环境支配和人类活动干扰的响应更大,低破碎化、高连通性、多样化森林景观与高海拔、高起伏度的地形条件关系密切,乡镇耕地和居民建设用地面积占比(PCLA、PRCLA)越大,林地景观破碎化越强和连通性程度越低。平均坡度(MS)、土层厚度(SD)和高差(ED)作为基础性的控制因素,直接影响耕地、居民建设用地占比等方面,间接作用于森林景观格局。在一定的地形土壤条件下,降低人口数量、耕地数量和居民建设用地占比,可降低森林景观格局破碎化,提升森林景观连通性。     

基于随机森林回归的玉米单产估测

为了提高玉米单产估测精度,以河北省中部平原为研究区域,以条件植被温度指数(VTCI)和上包络线S-G滤波的叶面积指数(LAI)为特征变量,通过随机森林回归确定玉米主要生育时期VTCI和LAI的权重,构建加权VTCI和LAI与玉米单产的单变量和双变量估产模型。结果表明,基于随机森林回归的双变量估产模型精度最高(R~2=0. 303),达极显著水平(P 0. 001)。将随机森林回归双变量估产模型用于研究区域2012年各县(区)玉米单产估测,结果表明,53个县(区)玉米估测单产与实际单产的平均相对误差为9. 85%,均方根误差为824. 77 kg/hm~2,模型精度较高。基于随机森林回归双变量估产模型逐像素估测研究区域2010—2018年玉米单产,结果表明,玉米单产在空间上的分布特征为西部地区最高、北部和南部次之、东部地区最低,年际间的分布特征为在波动中呈先减少后增加的趋势。     

基于GEE的山东省近30年农业大棚时空动态变化研究

针对精确获取大尺度空间范围内农业大棚的分布情况并进行长时间的序列动态监测存在数据量大、计算效率低、精度不高等问题,利用Google Earth Engine(GEE)云平台能够实现快速存取、实时处理海量卫星数据,基于多时相Landsat影像进行农业大棚时序光谱特征和纹理特征的自动提取,采用随机森林算法实现山东省农业大棚的遥感分类,从而生成了山东省近30年农业大棚的空间分布和时空动态变化图。结果表明,本文分类流程具有较高的分类精度,其平均总体精度达到91.63%,Kappa系数均值为0.8642。经分析,山东省农业大棚从1990年的6.67 km^2增加到2018年的9919.40 km^2,增长速度为354.03 km^2/a。     

基于无人机LiDAR的天然林与人工林林隙提取

为研究主动遥感进行森林地物分类和林隙提取的效果,分别在天然林和人工林中比较了无人机激光雷达(Li DAR)数据的阈值法、逐像元法和面向对象法3种方法的分类精度和适用性。选取天然林(黑龙江省哈尔滨市帽儿山林场)和人工林(内蒙古自治区赤峰市旺业甸林场)两处试验区,应用阈值法、逐像元法和面向对象法3种方法,对两个试验区采集的无人机Li DAR数据进行林隙、非林隙、其他类型划分。研究结果表明,面向对象法在天然林和人工林试验区中的分类精度和Kappa系数均最高,天然林为82. 43%、0. 73,人工林为91. 74%、0. 88;逐像元法次之,天然林为76. 62%、0. 64,人工林为78. 68%、0. 68;阈值法的分类精度和Kappa系数差异较大,在天然林中的精度极低,为50. 54%、0. 27,人工林的精度较高,为79. 12%、0. 69。面向对象法和逐像元法在天然林和人工林普遍适用,均可以达到理想的分类精度和Kappa系数。阈值法在天然林的精度较低,更适合于人工林的分类,即林分高度趋于一致,且建筑、道路等其他类型干扰较少的区域。天然林的最佳分类方法为面向对象法,人工林的最佳分类方法为阈值法。     

基于随机森林的高寒湿地地区土地覆盖遥感分类方法

高寒湿地是青藏高原典型独特的生态系统,是全球气候变化的敏感地带和预警区。利用遥感技术快速、准确地分类提取高寒湿地的土地覆盖信息,对当地生态安全监测和保护具有重要意义。本文以若尔盖湿地国家级自然保护区为研究区,首先,以高分一号(GF-1)遥感影像为数据源,融合光谱特征、水体指数、地形特征、植被指数和纹理信息等26个变量进行随机森林(Random forest,RF)分类实验;然后,根据袋外数据(Out of bag,OOB)的特征变量重要性得分和精度评价结果,选出高寒湿地地区土地覆盖类型的最优分类方案和特征;最后,对特征变量进行降维,并基于相同的变量,采用极大似然法(Maximum likelihood classification,MLC)、支持向量机(Support vector machine,SVM)、人工神经网络(Artificial neural network,ANN)和RF等方法进行分类,比较不同方法的优适性。结果表明:结合GF-1影像光谱、水体、植被、纹理特征和地形信息,使用26个变量的RF模型的分类精度最高,总体精度(Overall accuracy,OA)为90.07%,Kappa系数为0.86;通过RF模型的变量重要性分析可以有效选出重要的特征信息,在降低特征变量维度的同时,还能保证较高的分类精度; 4种分类方法中,RF算法是高寒湿地地区较合适的分类方法,OA比MLC基准方法高17.63个百分点,比SVM和ANN等机器学习算法分别高6.98、6.56个百分点。     

Logging pattern and landscape changes over the last century at the boreal and deciduous forest transition in Eastern Canada

Forestry practices associated with the industrial era (since ~1900) have altered the natural disturbance regimes and greatly impacted the world’s forests. We quantified twentieth century logging patterns and regional scale consequences in three sub-boreal forest landscapes of Eastern Canada (117,000, 49,400 and 92,300 ha), comparing forestry maps depicting age and forest cover types for early industrial (1930) and present-day (2000) conditions. Results were similar for the three landscapes, indicating large-scale forest change during the twentieth century. In 1930, previous logging activities had been concentrated in the lowlands and along the main hydrographical network, as compared to a more even distribution over the landscapes in 2000, reflecting a decreasing influence of the environmental constraints on forest harvesting. In 1930, old-aged forests (>100 years) accounted for more than 75% of the unlogged areas of the three landscapes, as compared to less than 15% for the present-day conditions. Logging practices have thus inverted the stand age distribution of the landscapes that are currently dominated by young and regenerating stands. The 1930 forest cover types showed a clear relationship with elevation, with conifers located in the lowlands and mixed and deciduous stands restricted to the upper slopes. Between 1930 and 2000, 58–64% of the conifer areas transformed to mixed and deciduous forests, such that no clear altitudinal relationships remained in 2000. We conclude that twentieth century logging practices have strongly altered the preindustrial vegetation patterns in our study area, to the point that ecosystem-based management strategies should be developed to restore conifer dominance, altitudinal gradients, as well as the irregular structure inspired from old forest stands.     

视频和GIS技术在森林防火中的应用

针对浙江省森林覆盖率高,并且发生森林火灾的频率也高,单纯的视频监控在夜晚无法识别等问题,提出了视频技术与地理信息系统技术相结合来监控森林火灾的方案。该方案是在森林的某个高程点上架设视频摄像设备,通过远程传输到达视频监控系统的后台,显示在森林地理信息系统上,然后根据显示的区域来监控森林火情。该方案通过同步跟踪算法实现视频技术与地理信息系统的融合,在森林防火地理信息系统上实时同步跟踪,能够提高森林防火的效率。     

Tracking change in a human‐dominated landscape: developing conservation guidelines using freshwater fish

• 1. To conserve biodiversity in a human‐dominated landscape, a science‐based inventory and monitoring plan is needed that quantifies existing resources, isolates drivers that maintain natural communities, determines harmful stressors, and links ecological drivers and human stressors. A tactical approach is proposed for conservation planning using freshwater fish at the Cape Cod National Seashore.
• 2. Freshwater fish are well studied and occur across environmental gradients. The lentic systems at the national park are relatively pristine yet are enveloped by a region of high population density. Using fish community data, three steps were taken for tracking anthropogenic impacts in a human dominated landscape. First, fish and potential drivers were sampled intensively along a gradient to determine which fish metrics reflect natural communities and which abiotic and biotic factors structure them. Second, emerging and existing regional human threats were identified. Third, these human threats were linked to the potential drivers that maintain natural communities to identify the most informative metrics to monitor and track change.
• 3. Fish communities, water quality, habitat, and food resources were sampled concurrently in 18 ponds in 1999 and 2000. Three common fish species explained 98% of variation in numbers across systems. Based on ecological relationships, pH, depth, vegetation, prey, and community complexity were determined to maintain biodiversity of freshwater fish communities.
• 4. The primary human threats here included: development‐related, land‐use changes; non‐point source pollution; eutrophication from septic systems; and introduced species that are a byproduct of high human visitation. These are common threats in many rapidly urbanized areas and are likely to have relevance to many sites.
• 5. To track the impact of emerging threats to freshwater ponds related to increased human population, monitoring changes in water quality, vegetated habitat, fish diversity, and trophic interactions are recommended.

Copyright © 2008 John Wiley & Sons, Ltd.