A New Method to Derive Precise Land-use and Land-cover Maps Using Multi-temporal Optical Data

Abstract

Here we propose an accurate and robust method for large-area land-use and land-cover (LULC) mapping using multi-temporal optical data. The conventional method for LULC classification usually uses time-series data at regular intervals to consider the seasonality of LULC. However, high-resolution optical data have considerable seasonal biases, making it difficult to use time-series data. Our basic idea for the accurate classification of LULC using high-resolution optical satellite data is to implement a classification for each scene considering seasonality first, and to then integrate multi-temporal classification results. In the per-scene classification, we accurately estimated the class-conditional spectral-seasonal densities of observation values from training data by conducting a kernel density estimation (KDE), and we used the densities in a Bayesian inference to obtain the class posterior probability. After the multi-temporal per-scene classification, we calculated the classification score by integrating class posterior probabilities in multi-temporal scenes. We conducted an 8-class classification for the entirety of Japan with 10-m spatial resolution using 1,876 scenes from the Advanced Visible and Near Infrared Radiometer type 2 (AVNIR-2) low-cloud-cover data, and we evaluated the accuracy of the classification by conducting a cross validation test and comparing the results to that obtained with existing methods: maximum likelihood classifier (MLC) and support vector machines (SVMs). The evaluation results showed that the overall accuracy of the proposed method is the best of all of the methods examined.