To minimize pesticide use it is necessary to detect at the early stage the present of plant disease and perform local treatment instead of global systematic treatment. To achieve this goal, one of the techniques may be used is image processing through the deployment of WMSN in the cultivated field. However, transmitting massive image data wirelessly will increase significantly network traffic and particularly energy consumption. In this paper, we propose a plant disease detection approach which is designed to run on the resource-constrained WMSN nodes. Through the analysis of plant images acquired by the node, this new approach is able to make a preliminary local decision on the health condition of the plant and determine the necessity of sending back images to the control centre for further inspection, thus improving the efficiency of the monitoring network. The complete method includes image segmentation based on both color and shape, and uses 2D histogram as the feature for classification. Experiments on the plant images with nutrient deficiency symptoms show the classification accuracy of the new method reaches 87.5%.