Local vs. Global: Local Land-Use and Land-Cover Models Deliver Higher Quality Maps

TL;DR

This paper addresses the poor accuracy and inconsistencies of global LULC maps in Africa by introducing a data-centric, teacher-student framework (DATS) that leverages high-resolution Maxar imagery for a local, high-detail teacher and freely available Sentinel-2 data for a scalable student, with knowledge transfer between the models. Applied to Murang'a county, Kenya, the approach yields higher-quality LULC maps than leading global maps, achieving an external F1 of $0.96$ and IoU of $0.86$ for the teacher, and improvements of $0.14$ in F1 and $0.21$ in IoU for the student relative to the best global baseline. The study also highlights substantial inconsistencies among existing global maps when evaluated locally, emphasizing the need for region-specific modeling to support food-security decisions and policy compliance. The proposed framework demonstrates practical impact through downstream use in crop type mapping and yield estimation, with clear plans to scale to larger regions and incorporate temporal dynamics for longitudinal monitoring.

Abstract

In 2023, 58.0% of the African population experienced moderate to severe food insecurity, with 21.6% facing severe food insecurity. Land-use and land-cover maps provide crucial insights for addressing food insecurity by improving agricultural efforts, including mapping and monitoring crop types and estimating yield. The development of global land-cover maps has been facilitated by the increasing availability of earth observation data and advancements in geospatial machine learning. However, these global maps exhibit lower accuracy and inconsistencies in Africa, partly due to the lack of representative training data. To address this issue, we propose a data-centric framework with a teacher-student model setup, which uses diverse data sources of satellite images and label examples to produce local land-cover maps. Our method trains a high-resolution teacher model on images with a resolution of 0.331 m/pixel and a low-resolution student model on publicly available images with a resolution of 10 m/pixel. The student model also utilizes the teacher model's output as its weak label examples through knowledge transfer. We evaluated our framework using Murang'a county in Kenya, renowned for its agricultural productivity, as a use case. Our local models achieved higher quality maps, with improvements of 0.14 in the F1 score and 0.21 in Intersection-over-Union, compared to the best global model. Our evaluation also revealed inconsistencies in existing global maps, with a maximum agreement rate of 0.30 among themselves. Our work provides valuable guidance to decision-makers for driving informed decisions to enhance food security.

Figures (7)

• Figure 1: Overview of our framework to build local land-use and land-cover (LULC) model that produces high quality map, using Murang'a county in Kenya as our area of interest. We propose a setup of teacher and student models to be trained on high- and low-resolution satellite images, respectively.
• Figure 2: Block diagram of our high-resolution teacher model. Deep learning models are trained recursively using high-resolution Maxar imagery and label examples. We use non-overlapping train and test sets for training and testing the model, respectively.
• Figure 3: High-resolution LULC map generated using the teacher model. Zoomed-in version of the LULC map from the test set, which was not seen during training, shows a high-quality map with clear delineation.
• Figure 4: Confusion matrices of LULC maps from the high-resolution teacher model across (a) Whole, (b) Test and (c) External sets.
• Figure 5: Comparison of LULC maps across Murang'a county. (a) GDW brown2022dynamic, (b) ESA zanaga2022esa, (c) ESRI karra2021global, and (d) DATS. Both the ESRI and DATS maps demonstrate similar patterns, such as a higher observation of croplands. Overall, the DATS map exhibits higher quality compared to the global maps (a) - (c).