Research On Semantic Segmentation Of High-resolution Remote Sensing Field Based On Deep Learning

Farmland remains the most important component of agricultural production,and the use of remote sensing for monitoring large-scale agriculture has become an important foundation for precision agriculture.However,in China,farmland fragmentation is significant,and farmland is relatively small and intertwined with irrigation channels,drainage ditches,villages,and roads.The high revisit period remote sensing satellites commonly used in agricultural monitoring typically only have moderate resolution,making it difficult to obtain accurate monitoring results due to mixed pixels.Therefore,obtaining ultra-high-resolution basic information on farmland boundaries is of great significance for more accurate use of high revisit period satellites for agricultural monitoring.In this study,we first constructed a dataset of 1500 farmland and non-farmland training samples with wide regional and seasonal representativeness based on Google Earth’s ultra-high spatial resolution(0.3m)RGB remote sensing images,and proposed an optimized DeepLabv3+deep learning field block recognition network model with ResNet50 as the backbone network.The optimized model was then compared and evaluated with several other models.Subsequently,the optimized model was applied to three regions,and the main results obtained were as follows:(1)Comparative analysis of seven deep learning field block recognition models(ResN et5 0-DeepLabv3+,MobileNet-DeepLabv3+,Xception-DeepLabv3+,VGG19-SegNet,VGG16-SegNet,VGG16-FCN)and traditional image semantic segmentation support vector machine(SVM)algorithm showed that the optimized DeepLabv3+deep learning field block recognition model proposed in this study with ResNet50 as the backbone network can quickly and accurately extract farmland blocks from high-resolution satellite images.(2)The farmland block recognition network model proposed in this study can accurately extract densely and sparsely distributed farmland blocks with a global accuracy(GA)of up to 97.23%and an average intersection over union(mIoU)of 94.02%.The global accuracy and average intersection over union were improved by 3.50 and 4.70 percentage points,respectively,compared to the Xception-based DeepLabv3+model.(3)The proposed model has better recognition performance for objects with fuzzy edges and is more widely applicable,providing an effective method for extracting farmland blocks from complex background remote sensing images in larger areas with good edge recognition performance.The model achieved high-precision extraction of farmland blocks in some regions of three irrigation areas in Ningxia,Tibet,and Anhui.The farmland block coverage areas were calculated as 42.37 km2,38.42 km2,and 56.76 km2,accounting for 57%,49%,and 75%of the total study area,respectively.Based on these results,farmland blocks in irrigation areas can be further monitored and managed.However,due to the limitations of semantic segmentation methods,the model is difficult to accurately count farmland blocks and extract the area of each individual block,especially for farmland blocks with small spacing.Therefore,further research is needed to address these limitations.