Mapping the extent and land use intensity of shifting cultivation with Planet Scope imagery and deep learning in the Democratic Republic of Congo 

Abstract

Shifting cultivation, in which primary or secondary forest plots are converted into agriculture for one to two years and then left fallow, is often deemed responsible for tropical deforestation. However, the general attribution of deforestation to areas under shifting cultivation is debatable if considering also the component of forest regrowth during the fallow phase, which is an essential part of a mature shifting cultivation system. Yet, little is known about the extent of small-size cropped fields and fallow stages, which are needed to derive information about the temporal development between small-size cropped fields and fallow in shifting cultivation landscapes. The primary objective of our study is to develop a deep learning-based framework to quantify land use intensity in tropical forest nations such as the Democratic Republic of Congo (DRC) using 4.7-m multi-temporal Planet Basemaps from 2015 to 2023. By employing a convolutional neural network image classification model, we first identified the shifting cultivation landscapes. Secondly, utilizing two-phase imagery, we delve into the temporal development of shifting cultivation, determining whether the landscape continues to be characterized by this practice. Thirdly, the shifting cultivation landscapes were segmented into cropped fields, young fallow, old fallow and old-growth forest/primary forest. Lastly, we used a deep learning regression model to quantify the intensity of shifting cultivation within identified areas. This last step adds depth to our analysis, by offering nuanced insights into the varying practices associated with shifting cultivation practices. Our study in DRC offers a detailed spatio-temporal dataset of the dynamics of shifting cultivation serving as a stepping stone to better understand its impacts on forest loss.