超载过牧是我国草原退化的主要原因之一，而载畜情况是评估草畜平衡的关键。面向牧畜高效、精准监管工作需求，针对牧畜“小（微）目标”监测的痛点难点，基于亚米级卫星遥感数据，综合利用牧畜“点”特征、“群”特征、“移动不固定”特征等多种特征，融合深度学习、面向对象等多种识别技术方法，构建了一种多尺度特征和多模型方法相融合的牧畜高分卫星遥感监测技术方法。该方法通过对牧畜弱信号的有效增强、“牧畜群”和“牧畜斑点”的分阶段检测与相互融合增强，实现了对牧畜群分布、牧畜斑点分布和牧畜群规模的监测提取，推动牧畜卫星遥感监测向“点数”式精细化监测迈进。锡林郭勒草原区域监测实验数据显示，模型检出率约0.802，误检率约0.244，具有较好效果。该方法的应用，可为草原区载畜情况的监测监管提供支撑，也可为其他“小（微）目标”的遥感监测提供借鉴；无论是在技术创新还是业务应用方面都具有十分重要的意义。

China is a large country of grassland and animal husbandry. Overloading and overgrazing is one of the main causes of grassland degradation in China. To protect the grassland, it is necessary to precisely monitoring livestock carrying capacity, which is the key to evaluation and control grass-livestock balance. However, traditional method of livestock carrying capacity such as hierarchical statistics, sampling field survey and online camera monitoring is whether time-consuming, labor-intensive, costly or poor quality. So it is very urgent to find a kind of efficient and precise monitoring method of livestock carrying capacity in grassland. To achieve this goal, this research proposes an efficient and precise monitoring method of the livestock in grassland by using of sub-meter resolution satellite image. The method not only fuses multi-scale features of livestock in sub-meter resolution satellite image such as "blob feature"、"flock feature"、"moving feature" , but also integrates deep learning technology and object oriented recognition technology. Firstly, considering that the livestock in satellite image is a kind of small (tiny) target, it uses kinds of image enhancement method such as bi-lateral filtering and Laplace of Gaussian (LoG) operator to enhance the weak livestock signal successfully. Secondly, in consideration of "flock feature" of livestock flock, it use a kind of "livestock flock detection model" based on deep learning technology to get the rough distribution area of livestock flocks. Thirdly, in consideration of "blob feature" and "moving feature" of livestock, it use a kind of "livestock blob detection method" based on LoG Gradient Difference and object oriented recognition technology to get the possible livestock blobs. Finally, by integrating the detecting result of both livestock flocks and livestock blobs, it uses the livestock blobs result to enhance and verify the livestock flocks result, and the enhanced and verified livestock flocks and the livestock blobs within them is finally get by using some simple manually revising work. Through an experiment in Xilingol grassland, it is found that the approach has good effect on livestock flock detection: with positive detection rate about 0.802 and false detection rate about 0.244, especially as to big livestock flock, the positive detection rate is up to 0.937, the false detection rate is low to 0.072. It is very helpful for the monitoring and supervision of livestock flock in grassland, and can also provide reference for remote sensing monitoring of other "small (tiny) targets". It is of great significance both in terms of technological innovation and business application. It makes a litter effort in promoting the livestock satellite remote sensing monitoring into an intuitive and fine monitoring era - "Number-Counting Era".