Land Use Land Cover Classification in the Amhara Region, Northwest Ethiopia, Using Convolutional Neural Networks
Abstract
More than 90% of agricultural production in East Africa comes from small-holder farmers, which are solely dependent on traditional rain-fed farming systems. Located on the northwestern highlands of Ethiopia, the Amhara Region occupies one-fifth of the area of the country (>160,000 km2), while supporting more than one-third of the national cereal crop production. The region has complex topography and highly fragmented croplands (averaging half a hectare), where different land use land cover (LULC) classes can occur in very short distances. Mapping such fragmented LULC areas require very high-resolution satellite imageries and robust classification methodologies. To this end, we have used multi-temporal and panchromatic very-high resolution (VHR) WorldView imageries (1 2 meters), via the U.S. government license agreement to access commercial VHR data, to extract information needed to classify LULC of the study area. Training and test LULC class datasets were obtained from the Land Investment for Transformation (LIFT) Ethiopia project that were digitized from centimeter-level special mission aerial photos. In this study, we have used several Convolutional Neural Networks (CNNs) for the task of semantic segmentation and compared their performance to other machine learning algorithms such as decision tree models. Extensive geometric data augmentation and pre-processing techniques were performed to account for the spectral distribution of landscape features in the imagery at different dates. We have classified the region into eight broad land cover classes that comprise cropland, grassland, shrubland, forest, bareland, settlement, water, and wetland. While both LULC classification methods achieve sufficient overall accuracies, CNNs outperform the Random Forest algorithm in classifying the LULC of the study area. With CNNs we were able to classify small features and classes with small occurrence in the training dataset by taking into consideration spatial features and class weighting. These research outputs will be beneficial in making informed decisions by the Amhara Region government sectors, and non-governmental Index-Insurance organizations on food security, market management, and in the crop insurance sector. We will use the cropland maps as a baseline for yield estimation and forecasting in our subsequent research work, which in turn can be used for recurrent seasonal yield estimation and forecasting that are useful for planning purposes to alleviate food insecurity.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFMGC34B..02A