Research And Application Of Forest Remote Recognition Techniques Based On Deep Learning

Remote sensing imageries are acquired from remote detection of ground objects at the aid of non-contact sensors.Based on inference from low-level spectral features shown on remote sensing imageries,such a procedure is called "remote recognition" or"interpretation" where thematic maps are produced to reflect visually the spatial distribution of detected ground objects.Due to its real-time and low-cost,remote recognition techniques are widely applied in agriculture,forestry,military,and environmental fields."Fores",is a most important category of ground objects to be remotely recognized,but difficulties are still encountered in its recognition:First,inconsistency exists between image features of forests due to the variance of locations and vegetation species.Second,misclassifications often occur due to isolated trees or small gaps among forests.These problems on one hand bring loss of recognition accuracy and on the other hand render the generated thematic maps noisy.To tackle those problems,this article proposes a novel forest recognition approach based on cascaded convolutional neural networks.The main work is as follows:First,an image set is collected illustrating forests of multiple species,covering the most common image features of forests in China.Second,a cascaded model is designed for forest recognition consisting of a res-net and a res-u-net allowing dynamic determination of the threshold used during the thresholding operation on the res-u-net's output for thematic map generation.Furthermore,a "multi-scale connection" structure is introduced to the res-u-net to expand the network's receptive field.Finally,the cascaded model is trained on the collected image set for evaluation of its performance of thematic map generation,measured by accuracy and continuity metrics.As shown by experimental results,the proposed model can be trained for recognition of forests in various provinces and of multiple species.Due to the model's balanced performance over various forest samples and superiority over traditional res-u-net model on accuracy and continuity metrics,the feasibility of the model is proved for forest recognition.With its ability to eliminate isolated map spots and fill tiny gaps on a generated thematic map,the application of the proposed model can be generalized for recognition of ground objects that are spread out over vast areas with certain continuity.