Counting and locating every tree from laser scans is essential for tracking forest carbon and health, but the data look very different depending on where the laser sits—on the ground, on a backpack/car, or on a drone. We propose TLNet, a neural network that “thinks in layers,” mimicking how forests are vertically organized (understory, mid-story, canopy). It learns which height bands are most useful and merges information from bottom to top and top to bottom. Trained first on realistic simulations and then fine-tuned on limited real data, TLNet detects trees accurately across all three sensor types, keeping errors to just a few centimeters. The internal weights show that the model naturally focuses on the crown–stem junction, a biologically meaningful region, making its decisions easier to interpret for ecologists.