Recognition And Classification Of Maize Drought On UAV Images Based On Semantic Segmentation

Maize is one of the most important food crops in the world and the largest food crop in China.Maize in the northern region plays an important role in China's maize industry.Northern China is the main maize producing area.Due to the changes of temperature and precipitation distribution,drought has become an important factor affecting maize yield in this area.With the continuous improvement of agricultural fine management requirements,it is of great significance to accurately and efficiently identify the drought level and estimate the area of maize.Traditional methods for crop drought identification use satellite multispectral remote sensing images to calculate the vegetation index to obtain drought conditions.However,due to various factors such as weather,transit time,and low spatial resolution,satellite imagery is not satisfactory in terms of timeliness and accuracy.With the development of Unmanned Aerial Vehicle(UAV)technology,it is possible to obtain very high resolution remote sensing image easily and flexibly.It has provided new ideas for the solution of many agricultural questions.With the development of deep neural networks,the accuracy of semantic segmentation of remote sensing images has been further improved with the introduction of full convolutional neural networks.In this paper,we designed and implemented a scheme based on deep learning semantic segmentation to identify and classify maize drought disasters,using UAV images.The specific contents of the study are as follows:(1)In this paper,traditional drought identification methods were analyzed and summarized.A maize drought label generation method based on four-band UAV data is proposed for the actual needs of agricultural insurance.UAV could collected RGB+NIR images.Then the Normalized Difference Vegetation Index(NDVI)calculated from the two bands of R and NIR is subjected to a certain round of mean smoothing.The index range corresponding to the disaster level is determined by expert evaluation,so as to generate ground truth labels.(2)The three channels of RGB are used as the input of the network,and the ground truth labels evaluated by experts are used to supervise the training of the neural network model.This verified the feasibility of using deep learning model to classify drought levels on RGB images.(3)In order to improve the accuracy of the segmentation results,a series of optimizations and improvements were carried out based on the U-Net network structure.These include using the SE-ResNeXt-50 as the backbone in the downsampling section,adding Atrous Spatial Pyramid Pooling(ASPP)structure to improve the network's ability to recognize multi-scale features,using Jaccard loss instead of the traditional Categorical Cross Entropy(CCE)as the loss function,and using transposed convolution as the upsampling method.The final trained network can identify and classify the drought of corn on RGB images taken by consumer UAVs,ensuring accuracy and reducing subsequent data acquisition costs.The experimental results showed that our pipeline achieved an F1-score of 0.9034 and a Jaccard index of 0.8287 on the test set.