Agricultural internet of things (IOT) and sensor technology has been widely used in the informationalized and mechanized orchard. The research aimed at both constructing an automatic assistant diagnosis and a real time alerting for plant disease and insect pest. The purpose also covered to realize an unmanned pest disease monitoring and to release some human interaction in making a diagnosis. A method for pathologic image recognition diagnosis based on deep learning neural network was designed and an innovative method for updating free parameters of the network was proposed on the basis of analyzing the error propagation of the network, so called the gradient descendent with flexible momentum. Then, computer recognizing pathologic images of fruit sphere was researched into systematically, where the apple was selected as a subject. Experiment result revealed the method manifested a recall rate at 98.4%. And in parallel with several well known updating schemes based momentum, the proposal was able to obviously improve the accuracy of learning network with a flatter converging curve, at a cost of short converging time. The test upon the several popular benchmark data sets also demonstrated it could perform an effective recognition on the image pattern.