Determination of agricultural land use: incidence of atmospheric corrections and the implementation in multi-sensor and multi-temporal images

Abstract:

Soil coverage and its modifications are critical variables in human-environmental sciences. Changes in land use are causes and consequences of climate change. This situation makes that detailed and updated information is needed for many applications. Remote sensing provides data of large areas periodically, so it becomes a useful input to soil classification. The objectives of this work were to determinate algorithm and images combination that produces the best results to classify agricultural land and, simultaneously, evaluate the need of making atmospheric corrections over them, when classifying multi-temporal/multi-sensor series. The two supervised classification algorithms used were neural networks and maximum likelihood. In the study area, agriculture is the main land use, predominantly summer crops, and the area sown with soybean, corn and sorghum account for over 90% of the total. Time series of Landsat 8 and SPOT 5 images were used, and 164 plots were registered to train and validate the models as ground truth. Maximum likelihood and neural networks models produce very good results when multi-temporal/multi-sensor series are used, with global accuracy between 79.17% to 90.14% and Kappa index between 60% to 82%. The radiometric correction at surface level did not improve the results when the reflectance was corrected at the top of the atmosphere. The time series that use images taken in more advanced phenological stages of crops produce better coverage classifications than time series that use images from early stages.

Show more Show less