Extraction of farmlands in Djibouti from satellite imagery using deep learning

Abstract

In Djibouti, about 25% of the population engage in livestock production including nomadic herders. Recent frequent occurrence of droughts caused by climate change has resulted in grass shortage, making sustainable nomadic condition difficult. In order to improve this situation, the government has formulated the plan which aims to help nomadic in rural areas strengthen their means of livelihood through agriculture. To establish more farmland areas, it is necessary to develop access points for sustainable water resource. For this purpose, assessment and analysis for location environment. Remote sensing is an effective technology for wide-area analysis. However, most of the papers focus on the development on algorithms of deep learning (DL). In this study, the DL analyses to identify the in-use farmlands were conducted using satellite image covers all over Djibouti.

Firstly, the ground truth data of the farmland polygons were extracted to create polygon datasets from both using ALOS pan-sharpen images (2006-2011) and Google Earth images (2018-2020). Environment of farmland location was analyzed by overlaying with the geology map, DEM, and wadi flow paths derived from the DEM. Training image datasets were set from ALOS 3-layer composite image of RGB. We used Mask R-CNN, the DL toolsets of ArcGIS Pro, to identify the agricultural land in use by location and area. The accuracy of the model was evaluated by the indicator of precision, recall, target rate, and hit rate.

As results, from the Pan-sharpen imagery of ALOS (2006-2011) and Google Earth image (2018-2020), 337 and 670 farmland parcels with 687 and 1090 ha in total area, respectively, were discovered to be exist throughout Djibouti. The total area of the farmlands increased by about 60%. Many farmlands (44%) were discovered to be locate on Fluvial alluvium and sand geological features, and 36% of the farmlands locate at elevations of less than 100 m. It was also discovered that 87% of the farmlands located within 100 m from the center line of wadi flow paths. The DL model of automatic farmland extraction for ALOS pan-sharpen images showed a low precision of 1.1% with recall rate of 34.8%. The target rate and hit rate were 27.7% and 21.1%, respectively. The densely distributed farmland regions were discovered to locate in the eastern and the south-western part of Djibouti, where the accuracy of the prediction was very low.