Abstract: A novel algorithm to detect the multi-scale edge features on multispectral remotely sensed imagery which is based on the concept of combining vector field model with dyadic wavelet transform was proposed, and two different neighborhood models in the algorithm was introduced to lead to locate edge points more complete. Firstly, multispectral images are defined by using of the vector field model. And then the dyadic wavelet transform is applied to produce the multi-scale edge detail coefficients, and first fundamental form is used for detecting the gradient magnitude and orientation of multispectral images at different levels. Lastly, edge points are located along with the quantified orientation of gradient by using the optimal neighborhood model at different scales. This representation can provide a local measure for the contrast of a high-resolution multispectral image at different scales. A variety of experiments by using QuickBird multispectral images of Nanjing area were presented to demonstrate the representation efficient. It is shown from the results that the edge information of the objects, i.e. factory, paddy, can be detected clearly from coarse to fine at different scale levels. This paper analyzed the relationship of the size of the ground features between the spatial resolution of image and to try to find a suitable level to demonstrate the feature of different objects. And the local maximum of the gradient magnitude provides information of the image edge feature which can be detected from the gradient modulus along with the quantified gradient orientation. And this paper figured out that quantification of the gradient orientation should consider the direction of objects in the image. Using F-measure, the results by the proposed method has higher precision than the traditional edge detectors.