In the heart of Nigeria, a technological revolution is taking root, quite literally. Researchers, led by Joaquin Gajardo from the Swiss Federal Laboratories for Material Science and Technology (Empa) and ETH Zurich, are leveraging the power of machine learning to transform the way we map and monitor croplands in data-scarce regions. Their work, published in the ISPRS Open Journal of Photogrammetry and Remote Sensing (a journal focused on photogrammetry, remote sensing, and spatial information sciences), could have significant implications for the energy sector, particularly in bioenergy and land-use planning.
The challenge is clear: cropland maps are vital for agricultural monitoring, but collecting ground-truth data is time-consuming and expensive. Enter machine learning, a tool that can enable large-scale mapping but relies heavily on the quality and proximity of training data to the target region. Gajardo and his team set out to evaluate the impact of combining global and local datasets for cropland mapping in Nigeria at an impressive 10-meter resolution.
They manually labeled 1,827 data points evenly distributed across Nigeria and leveraged the crowd-sourced Geowiki dataset, evaluating three subsets: Nigeria-only, Nigeria plus neighboring countries, and worldwide. Using Google Earth Engine, they extracted multi-source time series data from Sentinel-1, Sentinel-2, ERA5 climate, and a digital elevation model (DEM). They then compared Random Forest (RF) classifiers with Long Short-Term Memory (LSTM) networks, including a lightweight multi-task learning variant.
The results were eye-opening. “Local training data consistently improved performance, with accuracy gains up to 0.246 for RF and 0.178 for LSTM,” Gajardo explained. Models trained on Nigeria-only or regional datasets outperformed those trained on global data, except for the multi-headed LSTM, which uniquely benefited from global samples when local data was unavailable.
The sensitivity analysis revealed that Sentinel-1, climate, and topographic data were particularly important. Their removal reduced accuracy by up to 0.154 and F1-score by 0.593. Handling class imbalance was also critical, with weighted loss functions improving accuracy by up to 0.071 for the single-headed LSTM.
The best-performing model, a single-headed LSTM trained on Nigeria-only data, achieved an F1-score of 0.814 and accuracy of 0.842. This model performed competitively with the best global land cover product and showed strong recall performance, a metric highly relevant for food security applications.
So, what does this mean for the energy sector? Accurate cropland mapping is crucial for bioenergy planning, land-use management, and sustainable agriculture. As the world shifts towards renewable energy sources, the demand for bioenergy is expected to grow. This research could help energy companies identify suitable locations for bioenergy crops, optimize land use, and ensure sustainable practices.
Moreover, the integration of multi-modal feature data and the handling of class imbalance could pave the way for more accurate and reliable large-scale mapping in other data-scarce regions. As Gajardo puts it, “Our findings underscore the value of regionally focused training data, proper class imbalance handling, and multi-modal feature integration for improving cropland mapping in data-scarce regions.”
The team has released their data, source code, output maps, and an interactive Google Earth Engine web application to facilitate further research. This open-access approach could spur innovation and collaboration, driving the field forward and shaping the future of agricultural monitoring and land-use planning.
In the words of Gajardo, “We hope our work will inspire further research and practical applications, ultimately contributing to food security and sustainable development.” And with the energy sector’s growing interest in bioenergy, this research could be a game-changer, helping to light the way towards a more sustainable and energy-secure future.