Samsung Solve for Tomorrow

Earth has lost half its topsoil in the last 150 years. In my hometown of Pleasanton, erosion has caused entire backyards to collapse into creeks. Current detection methods require manually sampling soil over years, which is slow, expensive, and limited to accessible areas. We wanted to build something that could survey large areas quickly and flag problem zones before they caused damage.

I validated the problem by talking with local landowners who had lost their land, city officials, and researchers at Lawrence Livermore National Laboratory. Their insights confirmed that a scalable detection system would be valuable, especially one that could work on both farmland and natural landscapes like the fields and mountain range near my hometown.

The first stage uses a DJI Phantom 2 running DroneLink. The drone captures geo-tagged aerial imagery during automated survey flights, and these images are fed into a Bayesian CNN we trained. The model outputs erosion probabilities rather than binary yes/no classifications, and ambiguous regions get flagged for ground verification instead of forcing a decision on low-confidence predictions.

We trained the model on a dataset of labeled aerial images, split 80/20 for training and validation. Classification accuracy reached 64% on the validation set. The uncertainty quantification reduces false negatives by routing uncertain areas to the rover for physical confirmation.

Flagged coordinates are transmitted to a rover for ground verification. The rover uses GPS to navigate to target locations, then captures close-range imagery and takes physical soil measurements, to be analyzed for nutrient content and moisture levels. Tank-style differential steering handled tight turns and uneven terrain.

Ground-level data confirms or rejects the aerial classification, building a feedback loop that improves the CNN over time. Our two stage approach minimizes both false positives (wasted rover trips) and false negatives (missed erosion). Based on our testing, the system can identify ~83% of eroded land at < a tenth of the cost of traditional surveying.

Results are displayed on a website dashboard as a colormap overlay on the surveyed area. We designed the visualization so that non-technical users, like farmers and city planners, can immediately see which zones need attention without interpreting raw model outputs. The interface highlights high-probability erosion sectors and provides a quick analysis for each flagged region.

The project placed in the California state finals, meaning top 1% of over 30,000 competing teams nationwide. The $5,000 grant funded continued development and additional equipment for our school's engineering program.