Wireless Sensor Networks are being used in multiple applications and they are becoming popular particularly in precision-agriculture and environmental monitoring. Their low-cost enables to build distributed deployments with large spatial density of nodes. They have been traditionally used to build maps describing scalar fields varying in time and space. However, in the recent years, image capturing capable nodes have appeared allowing to measure more complex data but imposing new challenges for the processor and memory constrained nodes. Transmission of large images over a Wireless Sensor Network is a costly operation since most of the power consumption at the node is due to the operation of its radio. Hence, it is desirable to process and extract interesting features from the images at the node in order to transmit the important information and not all the images. However, image processing is also complicated by low processor and memory resources at the node. An image is usually delivered in JPEG format by the node’s camera and stored in flash memory but, with current typical node configurations, memory resources are insufficient to open the image file and perform the image processing algorithms on the pixels of the image. To overcome this limitation, image processing can be done in the compressed domain parsing the JPEG file and working directly on the Discrete Cosine Transform coefficients of the compressed image blocks as soon as they are decoded. In this article, we present an agricultural Wireless Sensor Network application that implements block based classification in the compressed domain. In this application, image-sensor nodes are placed on insect pest traps to quantify pest population in fruit trees.