Land use/land cover recognition in arid zone using A multi-dimensional multi-grained residual Forest

Monitoring arid areas could effectively improve economic, ecological and humanity benefits. It is an effective monitoring approach to recognize the land cover or land use of arid areas through machine learning methods using satellite images. However, there is no public classified dataset for arid areas currently, and hence remote sensing image monitoring in desert areas is restricted. Existing classification methods are not able to fully utilize effective features of satellite images and multi-spectral optical parameters. In this paper, our contributions are as follows: Firstly, we presented a new satellite dataset named the ARID-5 for arid area land cover/land use (LULC) classification, the LULC in arid areas included desert, oasis, Gobi, and water system. Second, we proposed a machine learning algorithm named the multi-dimensional multi-grained residual forest algorithm for LULC recognition on arid areas. In this algorithm, the multi-dimensional multi-grained structure was able to effectively extract image features and spectral information. The residual forest structure mapped probability feature vectors to higher levels for prediction, which effectively improved the reflection of the forest structure on the sample. At the same time, the base estimator was transmitted in cascade layers, and thus the diversity and the accuracy were improved. Experimental results proved that the multi-dimensional multi-grained residual forest showed good classification abilities. Last, we also tested our algorithm on SAT-4 and SAT-6 datasets, which proved the generalization performance of our algorithm.